A method for constructing a fusion neural network model for negative ion source performance prediction

Abstract

The present invention provides a method for constructing a fusion neural network model for predicting the performance of a negative ion source, comprising: constructing a data-driven module of a fusion neural network model; designing a fusion neural network model: the fusion neural network model comprises a primary fusion neural network and a secondary fusion neural network, the primary fusion neural network is used to solve the problem of spatiotemporal data misalignment caused by timing problems during the operation of a radio frequency negative ion source; the secondary fusion neural network is used to implicitly improve the input dimension of experimental data of a radio frequency negative ion source; in the working process of the fusion neural network, the input and output structure of the two-level neural network is differentiated without changing the original input data dimension to improve the overall prediction accuracy; and training the fusion neural network model. The present invention can assist experimenters in the management of experiments on long pulse neutral beam injection devices, and provide a basis for matching assumed parameters with actual conditions while protecting the ion source device.

Description

Technical Field

[0001] The present invention relates to the research field of performance simulation and prediction of auxiliary heating physical equipment of large scientific facilities, and specifically to a method for constructing a fusion neural network model for predicting the performance of a negative ion source. Background Art

[0002] Neutral beam injection is an effective method for heating and current driving fusion plasma. It is recognized by the fusion community as an essential means to achieve ignition and burning plasma control of fusion reactors. With the continuous increase in the size of fusion devices and the continuous improvement of fusion plasma parameters, the beam energy of the neutral beam injection system is required to increase from tens of keV to the MeV level. Considering the efficiency of the ion neutralization process of the neutral beam injection system and the economy of the entire system, when the required beam energy exceeds 50keV / amu, the neutral beam injection system based on a negative ion source (Negative ion based Neutral Beam Injection system, referred to as NNBI) should generally be considered. Although there are successfully operated NNBI devices in the world, their beam energy is low and the pulse width is short. Domestic research in the field of NNBI is in its initial stage, so it is of great significance to develop NNBI devices that meet the needs of fusion reactors and provide practical and effective support for fusion reactors.

[0003] For NNBI, the beam source system is its core. The beam source system mainly includes two parts: a plasma generator and a negative ion accelerator. The plasma generator is mainly used to generate a large area of ​​uniform density plasma, and the accelerator is used to extract ions from the plasma and accelerate them to the required energy to form an ion beam with certain optical properties. Therefore, the design, implementation and testing of the RF-driven negative ion source are more important for the experimental results. Usually, to obtain an ion beam with certain optical properties, the experimenter needs to set the various parameters of the ion source operation through the control system. This requires the experimenter to have certain operating experience. However, in the development of the negative ion source neutral beam system, the research on the ion source is a process from scratch, which makes it difficult to obtain suitable parameters. At the same time, the RF negative ion source is expensive, and it is easy to be damaged by inappropriate parameters. Therefore, if there is a method to predict the operating results of the RF negative ion source under the set parameters, it can not only determine whether the beam current that meets the requirements can be produced under the parameters, but also protect the ion source device to a certain extent. This is of great significance for both practical applications and theoretical research. Summary of the invention

[0004] In order to solve the above technical problems, the present invention provides a method for constructing a fusion neural network model for predicting the performance of a negative ion source, comprising the following steps:

[0005] Step 1, data set analysis: used to construct a data-driven module for the fusion neural network model. The data-driven module refers to the data set and its loading process required for training the fusion neural network model. The construction process includes: obtaining the data generated by the negative ion source neutral beam injection experiment from the historical database, and then dividing it into strongly correlated data, strongly correlated physical data and weakly correlated physical data based on the relationship. On this basis, all the experimental data obtained are further divided into training data sets, verification data sets and test data sets;

[0006] Step 2, fusion neural network model design: The fusion neural network model includes a first-level fusion neural network and a second-level fusion neural network. The first-level fusion neural network is used to solve the problem of spatiotemporal data misalignment caused by timing problems during the operation of the radio frequency negative ion source. It is built based on an artificial neural network and integrates a residual network during the construction process; the second-level fusion neural network is used to implicitly increase the input dimension of the radio frequency negative ion source experimental data and improve the model prediction accuracy. The second-level fusion neural network is also built based on an artificial neural network and integrates a residual network during the construction process;

[0007] Step 3: Implicit dimension increase algorithm design: used in the working process of the fusion neural network, without changing the original input data dimension, to improve the overall prediction accuracy by differentiating the input and output structure of the two-level neural network;

[0008] Step 4: Train the fusion neural network model and evaluate the learning status of the fusion neural network model through the loss changes of the training set and the validation set. Then run the model on the test set to evaluate the prediction rate and generalization ability of the fusion neural network model.

[0009] Beneficial effects:

[0010] The present invention is designed and implemented by fusing artificial neural networks and residual neural networks, and the prediction results are processed collaboratively by multi-level neural networks. The present invention makes full use of the data collected during the negative ion source experiment and efficiently analyzes its dependencies. While solving the problem of spatiotemporal dislocation, it can predict the performance of the ion source under set parameters. The present invention can assist experimenters in the management of the long pulse neutral beam injection device experiment, and provide a basis for matching the assumed parameters with the actual situation while protecting the ion source device. The present invention is guided by the actual problems in the current NNBI research and development progress, and aims to predict the operating effect of the long pulse experimental negative ion source device under set parameters. A fusion neural network model construction method for negative ion source performance prediction is proposed through a two-level fusion network and an implicit dimension increase method. The present invention uses the experimental data in the NNBI experimental database as the data source, and divides the data set into strongly correlated data, strongly correlated physical data, and weakly correlated physical data to construct the training, verification and test data sets required for this model. The present invention forms a large model with a two-level structure by fusing artificial neural networks and residual neural networks. The final model is determined by adjusting hyperparameters such as learning rate, data sample size, disorder, number of training iterations, hidden layer, optimizer, and data dimension. The present invention enriches the features of the predicted object by implicitly improving the input dimension in the algorithm, strengthens its feature engineering, and thus enhances the overall prediction capability of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 A fusion network model and an implicit dimension-increasing algorithm design diagram of a fusion neural network model construction method for predicting negative ion source performance provided by an embodiment of the present invention;

[0012] Figure 2 A design diagram of a residual neural network fusion method for a fusion neural network model construction method for predicting negative ion source performance provided by an embodiment of the present invention;

[0013] Figure 3 The error variation curve of the fusion neural network model construction method for predicting the performance of a negative ion source provided by an embodiment of the present invention;

[0014] Figure 4 (a) is a schematic diagram showing the comparison between the predicted value and the actual value of co-extracted electrons (EG) by the method for constructing a fusion neural network model for predicting the performance of a negative ion source provided in an embodiment of the present invention;

[0015] Figure 4 (b) is a schematic diagram showing the comparison between the predicted value and the actual value of the hydrogen negative ion flow (H-) by the fusion neural network model construction method for predicting the performance of the negative ion source provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0016] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0017] The embodiment of the present invention provides a method for constructing a fusion neural network model for predicting the performance of a negative ion source, including data set analysis, fusion neural network model design, implicit dimension increase algorithm design, iterative training and model evaluation. The following describes each step in detail:

[0018] Step 1, data set analysis: The data-driven module used to build the fusion neural network model refers to the data set and its loading process required for training the fusion neural network model. The construction process includes: obtaining the data generated by the negative ion source neutral beam injection experiment from the historical database, and then dividing it into strongly correlated data, strongly correlated physical data, and weakly correlated physical data based on the relationship. On this basis, all the experimental data obtained are further divided into training data sets, verification data sets, and test data sets to provide data support for the fusion neural network model. The negative ion source neutral beam injection experiment refers to setting the parameters necessary for the operation of the RF driven negative ion source through the neutral beam injection control system, and generating a beam current that meets the required beam energy. The relationship it is based on refers to the relationship between the working state of the RF driven negative ion source and the experimental data. Strong correlation refers to a strong relationship between the working state of the RF driven negative ion source and the selected experimental data, and weak correlation refers to a weak relationship between the working state of the RF driven negative ion source and the selected experimental data. Strongly correlated data include RF power RF, magnetic field current MAG_I, acceleration voltage ACC_V, extraction voltage EXT_V, bias voltage BIAS_V, bias 2 voltage, gas intake GAS intake, strongly correlated physical data include hydrogen negative ion current I_H- and electron current I_EG, and weakly correlated physical data include incident power P_IN, reflected power PREF, bias 1 current I_B, bias 2 current I_B2, extraction current I_EXT, and ground electrode current I_GG. Strongly correlated data is defined by human analysis, and as input data, it is the dependency of model calculation; strongly correlated physical data is the final prediction data required, and is used as the target value to calculate the difference between it and the prediction value during the model training process; weakly correlated physical data is used to optimize the model input structure, and as the filling part of the input data of the secondary fusion neural network, it implicitly increases the dimension of the input data. Among them, the training data set is used to train the model; the validation data set is used to periodically evaluate the learning status and generalization ability of the model; and the test data set is used to finally evaluate the status of the model and the generalization ability of the model.

[0019] Step 2, fusion neural network model design: The fusion neural network model includes a first-level fusion neural network and a second-level fusion neural network. The first-level fusion neural network is used to solve the problem of spatiotemporal data misalignment caused by timing problems during the operation of the radio frequency negative ion source. It is built based on an artificial neural network, and the residual network is integrated during the construction process; the second-level fusion neural network is used to implicitly increase the input dimension of the radio frequency negative ion source experimental data and improve the model prediction accuracy. The second-level fusion neural network is also built based on an artificial neural network, and the residual network is also integrated during the construction process.

[0020] As mentioned above, in view of the experimental data characteristics of the radio frequency negative ion source, this fusion neural network model adopts a two-level neural network superposition structure. Each level of fusion neural network is designed by integrating artificial neural network and residual neural network. Therefore, this negative ion source performance prediction fusion neural network model consists of two fusion neural networks with similar structures, which are named as the first-level fusion neural network and the second-level fusion neural network as mentioned above.

[0021] In one embodiment, the first-level fusion neural network has a total of four network hidden layers. Hidden layer one is used to store the characteristics of the radio frequency negative ion source experimental data, and at the same time construct a residual copy of the radio frequency negative ion source experimental data, which will be used to form a residual network in hidden layer two; hidden layer two is used to fuse the artificial neural network of the first-level fusion neural network with the residual network in hidden layer two; hidden layer three is used to further extract the characteristics of the radio frequency negative ion source experimental data in the first-level fusion neural network; hidden layer four is used for filtering and normalizing noise in the regression process of the first-level fusion neural network.

[0022] The structure of the secondary fusion neural network is similar to that of the primary fusion neural network, and it contains four hidden layers. Hidden layer 1 is used to implicitly increase the dimension of the input radio frequency negative ion source experimental data. This implicit improvement is achieved by appending the output of the primary fusion neural network to the initial input data to form higher-dimensional input data. After the calculation of the primary fusion neural network is completed, the output data is weakly correlated physical data, which will be passed to the input layer of the secondary fusion neural network. The residual copy of the optimized radio frequency negative ion source experimental data is obtained and further constructed by the hidden layer 1. The residual copy will be used to form a residual network in the hidden layer 2; the hidden layer 2 is used to fuse the artificial neural network of the combined secondary fusion neural network and the residual network in the hidden layer 2; the hidden layer 3 is used to further extract the features of the optimized radio frequency negative ion source experimental data in the secondary fusion neural network; the hidden layer 4 is used to filter the noise in the regression process of the secondary fusion neural network and output the predicted values ​​of hydrogen negative ion current and co-extraction current. The optimized radio frequency negative ion source experimental data refers to the superposition of the output of the primary fusion neural network with the initial input data.

[0023] As can be seen from the above, the negative ion source performance prediction fusion neural network model of the present invention is a superposition neural network model that integrates artificial neural networks and residual neural networks. In this model, each level of fusion neural network integrates residual networks in hidden layers. In other neural network model designs, the design of residual networks includes two ways, one is to directly build, that is, the whole model is a residual network, and the second is as a part, that is, as an independent part of the whole model, the independent calculation result is calculated, and then this calculation result is weighted to the final calculation result. In the present application, the residual network is directly fused into the artificial neural network through the hidden layer, so that the residual network is fused with the overall artificial neural network as a part of the radio frequency negative ion source prediction network model and participates in the calculation process together, and then a first-level fusion neural network and a second-level fusion neural network are formed by such operation, and then the first-level fusion neural network and the second-level fusion neural network are further superimposed to obtain the final negative ion source performance prediction fusion neural network model. Therefore, the calculation of the fusion model is integrated, and its appearance features include input layer, hidden layer and output layer, and the inner table features are the artificial neural network model design that integrates the residual network.

[0024] The fusion neural network model must meet the characteristics, state transitions, practical requirements and accuracy requirements that the radio frequency negative ion source should reflect under the set parameters. Meeting the above-mentioned characteristics, state transitions, practical requirements and accuracy requirements is achieved in the following ways. For the characteristics, it extracts information through feature engineering. For the state transitions, it predicts the values ​​of the hydrogen negative ion flow and the co-extracted electron flow of the radio frequency negative ion source in the initial state and the terminal state through the fusion neural network. The practical requirements are reflected in actual engineering, which includes the prediction of the performance of the radio frequency negative ion source under set parameters. The accuracy requirement is the prediction accuracy of the fusion neural network model for the hydrogen negative ion flow and the co-extracted electron flow in the beam generated by the radio frequency negative ion source.

[0025] Step three, implicit dimension increase algorithm design: Implicit dimension increase processing refers to the process of improving the overall prediction accuracy by differentiating the input and output structure of the two-level neural network without explicitly changing the original input data dimension in the working process of the fusion neural network. The input and output structure of the differentiated two-level neural network refers to the input of the first-level fusion neural network in the main body of the radio frequency negative ion source performance prediction model formed by encapsulating the two-level neural network structure. The input data is gradually calculated to the set output data dimension during the forward propagation of the network, and the calculated data is output to the input layer of the second-level fusion neural network. Then the second-level fusion neural network implicitly expands the data dimension by superimposing the input of the first-level fusion neural network with the above output data, so as to enrich the characteristics of the target algorithm. From the outside, the value and dimension of the input data have not been changed, but through the differentiation algorithm, the model can significantly improve the overall prediction accuracy of the model by implicitly increasing the input dimension of the second-level fusion neural network. The design includes a first-level fusion neural network, a second-level fusion neural network and its input and output neurons. It is used to enrich the characteristics of the predicted object and strengthen its feature engineering to enhance the overall prediction ability of the fusion neural network model. Feature engineering refers to the process of converting raw data into features that can better represent business logic and potential problems, with the goal of improving the performance and accuracy of artificial intelligence models.

[0026] Step 4: Train the fusion neural network model, evaluate the learning status of the fusion neural network model through the loss changes of the training set and the validation set, and evaluate the prediction rate and generalization ability of the fusion neural network model.

[0027] In the training process, the adaptive moment estimation algorithm based on infinite norm is used as the optimizer, and the mean square error is selected as the loss function. The weights and biases of the model neurons are continuously updated in the effective training process of the fusion neural network model for radio frequency negative ion source performance prediction through forward iteration and back propagation. The performance of this model needs to be evaluated during and after the training process of this model.

[0028] Evaluating the learning status of the fusion neural network model through the loss change of the training set and the validation set includes: determining whether the training of the fusion neural network model is normal by comparing the mean square error changes of the training data set and the validation data set during the training process of the radio frequency negative ion source performance prediction fusion neural network model, specifically including: when the training of the radio frequency negative ion source performance prediction fusion neural network model tends to be stable, if the training set mean square error change curve and the validation set mean square error change curve do not continue to decline or have an inflection point, the training of the model is determined to be normal; when the training set mean square error change curve or the validation set mean square error change curve continues to decline or has an upward inflection point, the training of the model is determined to be abnormal.

[0029] Evaluation of the prediction rate and generalization ability of the fusion neural network model includes: further evaluating the prediction rate and generalization ability of the fusion neural network model on the test set. The test set contains a randomly selected historical data set of negative ion source neutral beam injection experiments. First, the trained negative ion source performance prediction model is saved, and then the data in the test set is input in random order to obtain the hydrogen negative ion current and co-extraction electron current values ​​predicted by the model under the set parameters. The values ​​are then compared with the actual values, and the errors less than 0.3 are considered to be valid predictions. The higher the prediction accuracy, the stronger the generalization ability of the negative ion source performance prediction model. The prediction accuracy refers to the prediction accuracy of the fusion neural network model for hydrogen negative ion current and co-extraction electron current, and the generalization ability refers to the actual prediction ability of the fusion neural network model for radio frequency negative ion source performance prediction in actual work.

[0030] Combine the following Figure 1 Describe an example of the design of the above fusion neural network model and implicit dimension increase algorithm. Figure 1 In the example, the input tensor is a tensor type data obtained after the experimental data is processed, and its source is the experimental data stored in MDSplus. The fusion neural network model described in this method is constructed by artificial neural network (ANN) and residual network. ANN performs dimension increase and dimension reduction processing on the input tensor. Then the residual network is used to perform residual processing on the input tensor. The fusion neural network model contains a two-level structure. The first-level fusion neural network is used to solve the problem of spatiotemporal data misalignment caused by timing problems during the operation of the radio frequency negative ion source. The second-level fusion neural network is used to implicitly increase the input dimension of the radio frequency negative ion source experimental data and improve the model prediction accuracy. The fusion neural network model can predict the values ​​of the co-extracted electron flow and hydrogen negative ion flow. The implicit dimensionality increase algorithm is embodied in Figure 1 The first-level fusion neural network, the second-level fusion neural network and their differentiated input and output processing processes. Among them, the input of the first-level fusion neural network is the original input, and its output is the data item required for implicit dimension increase. The input of the second-level fusion neural network is the output of the first-level fusion neural network and the original input, which is used to enrich the features of the predicted object and strengthen its feature engineering to enhance the overall prediction ability of the model.

[0031] Combine the following Figure 2 Describes an example of a fusion method that combines a residual neural network with an artificial neural network. Figure 1 and Figure 2The fusion neural network model shown contains two fusion processes. Taking one fusion process as an example, its architecture contains four artificial neural networks (ANNs). The core function of the four ANNs is to fit linear transformations, and the activation function is selected as the rectified linear unit. The input data first needs to be operated by ANN1 to extract the features of the input tensor preliminarily and create a residual copy. At this time, the output tensor is saved and recorded as x0. Then, while keeping the dimension of the x0 tensor unchanged, all the original information is obtained through ANN2, and the output at this time is recorded as x1. Next, x0 and x1 are added through the weighting function to construct a residual network. Next, ANN3 is used to further extract the features of the input tensor in a higher dimension after optimization. Finally, ANN4 is used to filter the noise in the optimized tensor and reduce it to the dimension of the output tensor to obtain the final predicted value.

[0032] The mean square error change curve of the fusion neural network model construction method for negative ion source performance prediction is shown in Figure 3 As shown. The training loss of the fusion network model is close to flat at the end, so the learning process of the fusion neural network model is sufficient. The loss change of the fusion neural network model in the training set and the validation set is obvious, the model learning is effective, and it shows a convergence trend. At the same time, there is no rising edge in the loss change of the fusion neural network model in the training set and the validation set, so the fusion neural network model is not overfitting. In summary, the fusion neural network model is trained normally, learns normally, has no overfitting and underfitting effects, and the fusion neural network model is effective.

[0033] The prediction results of the fusion neural network model construction method for negative ion source performance prediction are as follows Figure 4 (a) and Figure 4 As shown in (b). Figure 4 (a) is a schematic diagram comparing the predicted value and actual value of co-extracted electrons (EG). Figure 4 (b) is a schematic diagram comparing the predicted value and the actual value of the hydrogen anion current (H-). Figure 4 (a) and Figure 4 As can be seen in (b), the fusion neural network model is effective in learning. After being saved, it is generalized and tested on the test set. Its predicted value can accurately attach to the actual value, and the overall prediction performance is excellent.

[0034] The data set classification table of the fusion neural network model construction method for predicting the performance of the negative ion source is shown in Table 1. The control parameters of the control system for the RF negative ion source are selected as strongly correlated data, including P_Set (RF power), I_MAG (magnetic field current), ACC_V (acceleration voltage), EXT_V (extraction voltage), V_b (bias voltage), V_b2 (bias 2 voltage), GAS (gas intake). These data items are used as the input of the fusion neural network model. From the measured data of the generated beam, the physical data that can be used to evaluate the performance of the RF negative ion source under the set parameters are selected as strongly correlated physical data, including I_H- (hydrogen negative ion current) and I_EG (electron current). These parameters can be used to evaluate the performance of the RF negative ion source under the set parameters, so they are used as the output of the fusion neural network model. After selecting the set parameters, other physical data generated by the RF negative ion source during operation that can be used to describe the status of its own components are used as weakly correlated physical data, including incident power P_in, reflected power P_ref, bias 1 current I_b, bias 2 current I_b2, extraction current I_ext, and ground electrode current I_GG.

[0035] Table 1 Fusion network model dataset

[0036]

[0037] According to the method for constructing a fusion neural network model for predicting the performance of a negative ion source according to an embodiment of the present invention, a multi-level network structure is constructed by fusing an artificial neural network and a residual neural network to collaboratively process the prediction results, thereby realizing the operation simulation and performance prediction of an actual ion source under set parameters.

[0038] The above-mentioned embodiments only express several implementation methods of the present invention, and the description thereof is relatively specific and detailed, but it cannot be understood as limiting the scope of the patent of the present invention. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, which all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be subject to the attached claims.

[0039] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above and that the present invention can be implemented in other specific forms without departing from the spirit or essential features of the present invention. Therefore, the embodiments should be considered exemplary and non-restrictive in all respects, and the scope of the present invention is defined by the appended claims rather than the above description, and it is intended that all changes falling within the meaning and scope of the equivalent elements of the claims be included in the present invention.

[0040] In addition, it should be understood that although the present specification is described according to implementation modes, not every implementation mode contains only one independent technical solution. This description of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment may also be appropriately combined to form other implementation modes that can be understood by those skilled in the art.

Claims

1. A method for constructing a fusion neural network model for predicting negative ion source performance, characterized in that: The following steps are involved: Step 1, data set analysis: the data-driven module used to construct the fusion neural network model. The data-driven module refers to the data set and its loading process required for training the fusion neural network model. The construction process includes: obtaining the radio frequency negative ion source experimental data from the historical database, and then dividing it into strongly correlated data, strongly correlated physical data and weakly correlated physical data based on the relationship. On this basis, all the acquired radio frequency negative ion source experimental data are further divided into training data sets, verification data sets and test data sets; among them, the strongly correlated data are the input data of the fusion neural network model; the strongly correlated physical data are the final prediction data of the fusion neural network model; the weakly correlated physical data are the filling part of the secondary fusion neural network input data, which is used for implicit extraction. The dimension of high input data, where strong correlation refers to the strong relationship between the working state of the RF-driven negative ion source and the selected experimental data, and weak correlation refers to the weak relationship between the working state of the RF-driven negative ion source and the selected experimental data. Strongly correlated data include RF power RF, magnetic field current MAG_I, acceleration voltage ACC_V, extraction voltage EXT_V, bias voltage BIAS_V, bias 2 voltage, gas intake volume GAS, strongly correlated physical data include hydrogen anion current I_H- and electron current I_EG, weakly correlated physical data include incident power P_IN, reflected power PREF, bias 1 current I_B, bias 2 current I_B2, extraction current I_EXT, ground electrode current I_GG; Step 2, fusion neural network model design: the fusion neural network model includes a first-level fusion neural network and a second-level fusion neural network. The first-level fusion neural network is used to solve the problem of spatiotemporal data misalignment caused by timing problems during the operation of the radio frequency negative ion source. It is built based on an artificial neural network, and a residual network is integrated during the construction process; the second-level fusion neural network is used to implicitly increase the input dimension of the radio frequency negative ion source experimental data and improve the model prediction accuracy; the second-level fusion neural network is also built based on an artificial neural network, and a residual network is integrated during the construction process. The first-level fusion neural network has a total of four network hidden layers, wherein hidden layer one is used to save the characteristics of the radio frequency negative ion source experimental data, and at the same time construct a residual copy of the radio frequency negative ion source experimental data, which will be used in hidden layer two to form a residual network; hidden layer two is used to fuse the artificial neural network of the first-level fusion neural network with the residual network of the hidden layer two; hidden layer three is used to further extract the characteristics of the radio frequency negative ion source experimental data in the first-level fusion neural network; hidden layer four is used for noise filtering and normalization processing during the regression process of the first-level fusion neural network; The secondary fusion neural network is used to implicitly improve the input dimension of the radio frequency negative ion source experimental data and improve the prediction accuracy of the model. It includes four network hidden layers, wherein the hidden layer one is used to implicitly improve the dimension of the input radio frequency negative ion source experimental data. This implicit improvement is achieved by appending the output of the primary fusion neural network, i.e., weakly correlated physical data, to the initial input data, and at the same time constructing a residual copy of the optimized radio frequency negative ion source experimental data, which will be used to form a residual network in the hidden layer two; the hidden layer two is used to fuse the artificial neural network of the secondary fusion neural network with the residual network in the hidden layer two; the hidden layer three is used to further extract the features of the optimized radio frequency negative ion source experimental data in the secondary fusion neural network; the hidden layer four is used to filter the noise in the regression process of the secondary fusion neural network and output the predicted values ​​of hydrogen negative ion current and co-extraction current; Step 3: Implicit dimension increase algorithm design: used in the working process of the fusion neural network, without changing the original input data dimension, to improve the overall prediction accuracy by differentiating the input and output structure of the two-level neural network; Step 4: Train the fusion neural network model and evaluate the learning status of the fusion neural network model through the loss changes of the training set and the validation set. Then run the model on the test set to evaluate the prediction rate and generalization ability of the fusion neural network model.

2. The method for constructing a fusion neural network model for predicting negative ion source performance according to claim 1, characterized in that: In the training process, the adaptive moment estimation algorithm based on infinite norm is adopted as the optimizer, and the mean square error is selected as the loss function. The weights and biases of the model neurons are continuously updated in the effective training process of the fusion neural network model through forward iteration and back propagation.

3. The method for constructing a fusion neural network model for predicting negative ion source performance according to claim 1, characterized in that: The training process also includes: judging whether the training of the fusion neural network model is normal by comparing the mean square error change of the fusion neural network model: When the fusion neural network model tends to be stable as a whole, and its training set mean square error change curve and the validation set mean square error change curve do not continue to decline or have an inflection point, the training of the fusion neural network model is judged to be normal; when the training set mean square error change curve or the validation set mean square error change curve continues to decline or has an upward inflection point, the training of the fusion neural network model is judged to be abnormal.

Images