Miniature reactor system real-time filtering method based on hybrid expert system

Abstract

The invention discloses a reactor system real-time filtering method based on a hybrid expert system, and belongs to the technical field of neural networks. The method comprises the following specific steps: 1, determining the size of a time window of an input signal sequence, obtaining noise data and clean data, and constructing a database; step 2, an Encoder-Decoder neural network is constructed, and the Encoder-Decoder neural network is constructed; 3, classifying the data set, and training a plurality of neural network sub-models; 4, testing the neural network, judging whether the maximum error and the average error meet requirements or not, if not, entering the step 3, re-dividing the data set, and if yes, entering the step 5; step 5, constructing a hybrid expert system; and step 6, inputting noise data according to the size of the time window of the input signal sequence, and outputting noise-reduced data by the hybrid expert system. The method has the advantages that a plurality of noise signals can be processed in real time, complex and nonlinear data can be processed, the result precision is high, the constraint of a mathematical physical model is not needed, the distortion problem after noise reduction is avoided, and compared with a single network model, the adopted hybrid expert system has higher accuracy and generalization.

Description

Technical Field

[0001] The invention belongs to the technical field of neural network and micro reactor control, and specifically discloses a real-time filtering method for a reactor system based on a hybrid expert system. Background Art

[0002] Reactor safety is the lifeline of nuclear energy development. Reactor system status monitoring provides operators with key information on the operating status of the reactor, and is an important guarantee and basis for ensuring the safe and healthy operation of the reactor. However, since the reactor is in a harsh environment of high temperature, high pressure, and high radiation, its measurement signal error will gradually increase over time. Generally, nuclear power plants use calibration methods to correct the error. However, for micro reactors, its calibration work is difficult to carry out, considering that micro reactors generally work in unmanned environments, such as deep sea, space, remote areas, islands, etc. Therefore, it is necessary to use a filtering method to filter its measurement noise signal to obtain the true operating status of the reactor.

[0003] Traditional filtering methods such as moving average and Savitsky-Golay filtering have univariate properties, and the redundancy and correlation of multiple variables cannot be removed. Kalman filtering is a commonly used filtering method, but it requires the selection of a suitable mathematical model and a priori estimated parameters as the covariance matrix. Algorithms such as principal component analysis and wavelet transform are mainly aimed at static data and cannot achieve real-time data filtering. In recent years, denoising autoencoders have achieved remarkable results in the field of image denoising and have received widespread attention in other fields. However, in order to improve the accuracy and generalization of denoising encoders, massive amounts of noise data are required for training. The acquisition of a large amount of noise data and model training have become a further problem restricting the development of denoising encoders, making it difficult for them to have further room for improvement.

[0004] Patent document CN116595884A discloses a method for implementing ultra-real-time computing of a micro-nuclear reactor system, including constructing a three-level database database, constructing a neural network including two regression networks and one classification network, and constructing a space thermal ion reactor computing model under the GPU and multi-core CPU computing framework. Although this method overcomes the defects of traditional numerical calculation and realizes high-dimensional ultra-real-time computing, it is still a proxy model in essence, and its fundamental problem has not been effectively solved, that is, the generalization and accuracy of its algorithm cannot be effectively guaranteed. In fact, even if a high-precision traditional calculation method is adopted, the result still has deviations. This error is inevitable due to boundary conditions, model coefficients, environmental factors, etc. In order to compensate for the error between the numerical model and the physical system, it is necessary to carry out online correction. The performance of the online correction method depends largely on the real-time measurement results. At present, most online correction methods believe that their correction data is high-fidelity data. However, high-fidelity data is generally difficult to obtain, especially for the aerospace field. The data generally contains a large amount of noise and outliers, which need to be processed to a certain extent. Therefore, the filtering method proposed in the present invention can effectively solve this problem. Summary of the invention

[0005] In order to compensate for the real error between the numerical model and the physical system, it is necessary to carry out online correction, and the performance of the online correction method depends largely on the real-time measurement results. However, the measurement data generally contains a large amount of noise and outliers, which need to be processed to a certain extent. The existing filtering methods cannot achieve real-time, high-precision filtering. The purpose of the present invention is to provide a real-time filtering method for a reactor system based on a hybrid expert system to overcome the above difficulties and make up for the defects of the above method.

[0006] The technical solution of the present invention is as follows:

[0007] The real-time filtering method of a reactor system based on a hybrid expert system specifically comprises the following steps:

[0008] Step 1. Determine the time window size of the input signal sequence. The data format of the input signal sequence time window is N×M, where N is the number of variables and M is the time step of the signal. The output signal is clean data, and its data format is consistent with the input signal format. The input data needs to be normalized, and the formula is as follows: (x n -x n,min ) / (x n,max -x n,min ), where x n is the value of the clean data of the nth variable, x n,mib is the minimum value of the clean data of the nth variable, x n,max is the maximum value of the clean data of the nth variable.

[0009] The database data contains three levels, and the database includes a test set and a training set. Each data set contains multiple transient conditions, and each transient condition contains multiple input signal sequences and output signal sequences, as well as the maximum and minimum values ​​of the input signal of the transient condition before normalization.

[0010] Step 2. Construct an Encoder-Decoder neural network; the network structure includes a convolutional layer, a channel attention mechanism, a fully connected layer, and a deconvolutional layer. The network uses a mean square error loss function. The Encoder-Decoder neural network can receive an input signal sequence window and output clean data, including:

[0011] S201.Encoder uses a one-dimensional convolutional neural network to encode the input signal sequence and obtain the initial feature vector;

[0012] S202. The channel attention mechanism conducts attention mechanism learning on the initial feature vector and improves the network's fitting ability through the fully connected layer, and finally obtains the representation feature vector input into the Decoder;

[0013] S203.Decoder uses an inverse convolutional neural network to decode the representation feature vector and finally obtain the output signal;

[0014] Step 3. Classify the training data in the dataset according to the requirements, taking transient conditions as units, and train a single network using one or more transient conditions. A single transient condition contains input signal sequences with different noises and clean output signal sequences. The loss function of the neural network uses the root mean square error function: in, is the value of the nth variable output by the neural network.

[0015] Step 4. Use the test set data to test the neural network and determine whether the maximum error and average error meet the requirements. If not, proceed to step 3, reduce the amount of data learned by a single neural network, increase the number of neural networks, and retrain; if satisfied, proceed to step 5;

[0016] Step 5. Construct a hybrid expert system; the hybrid expert system is composed of multiple expert models, and a single expert model is composed of two neural network models and a moving average filter window. Based on the transient working conditions learned by the neural network, the neural network model is selected to form an expert system. When the hybrid expert system obtains the input signal sequence, the processing steps include:

[0017] S501. The expert system receives the noise signal sequence at time t and the expected steady-state fuel average temperature under the transient condition, and selects the expert model according to the thermal power and the training data condition of the expert model. The selection should meet the principle that the transient condition to which the input signal sequence belongs is as close as possible to the transient condition included in the expert model.

[0018] S502. According to the selected expert system, the input signals are normalized accordingly and respectively input into the neural network model included in the expert model.

[0019] S503. The output data of the neural network model included in the expert model is processed by moving average filtering, and then the output data after moving average filtering of each network is weighted averaged. Weighted formula: In the formula It represents the result of moving average filtering of the output data of the first and second neural network models in a single expert model. a1 and a2 are weighted coefficients. The formula is: a1 = (P s -P2) / (P1-P2), a2=1-a1. s It is a parameter when the transient condition of the input signal sequence is stable, and P1 and P2 are the parameters corresponding to the transient condition of the training data of the neural network model.

[0020] Step 6. According to the time window size of the input signal sequence, the noise data at time t and the thermal power from time t-1 to time t are input to the hybrid expert system, and the output of the hybrid expert system is the clean data at time t.

[0021] Compared with the prior art, the present invention has the following beneficial effects:

[0022] The present invention provides a real-time filtering method for a reactor system based on a hybrid expert system, which is suitable for signal measurement filtering of a reactor system and can also be expanded to other fields. The present invention can process multivariable dynamic data, eliminate redundancy and correlation between multiple variables, and output filtered data in real time. The hybrid expert system framework is adopted to reduce the data capacity of a single denoising autoencoder learning, and greatly reduce the difficulty of model training. Data processing by multiple networks can reduce the uncertainty of a single model. Compared with a single model, the hybrid expert system has stronger generalization and accuracy. The neural network and the moving average method are combined in a single model, which fully utilizes the advantages of the two methods, does not require the constraints of complex physical and mathematical models, effectively avoids distortion problems, and can achieve more reliable signal filtering. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] The accompanying drawings are used to further explain the present invention, but do not constitute a limitation to the present invention.

[0024] Figure 1 A flow chart of a real-time filtering method for a reactor system based on a hybrid expert system provided by an example of the present invention;

[0025] Figure 2 System structure of the space thermionic reactor TOPAZ-II provided by the present invention

[0026] Figure 3 A three-dimensional view of the space thermionic reactor TOPAZ-II provided for the example of the present invention;

[0027] Figure 4 A schematic diagram of the denoising autoencoder Encoder-Decoder network structure provided by an example of the present invention;

[0028] Figure 5 The hybrid expert network structure diagram provided for the example of the present invention has three levels, including an Encoder-Decoder network framework, a single expert network framework, and a hybrid expert network structure diagram.

[0029] Figure 6 A schematic diagram for comparing the hybrid expert network, single neural network model filtering, noisy data, and clean data provided for the example of the present invention, wherein a is the noise data, b is the sub-model 1 data, c is the expert network filtered data, d is the sub-model 2 data, and e is the real clean data. DETAILED DESCRIPTION

[0030] In order to more clearly explain the design of the present invention, the present invention is further described below in conjunction with the accompanying drawings and specific examples. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0031] Figure 1 A flow chart of the real-time filtering method for a reactor system based on a hybrid expert system is given. The specific steps are as follows:

[0032] Step 1. Select the heat pipe improved TOPAZ-II reactor system as the physical object. Its system structure and 3D view are as follows: Figure 2 and Figure 3 As shown in the figure, thermal power, electric power, average coolant inlet temperature, average coolant outlet temperature, mass flow rate, and fin temperature are selected as observed variables, and the time step is 5, that is, the input signal sequence format is 6×5, and each row represents the value of a variable from time t-5 to time t.

[0033] Noise data generation method: The high-precision system program of the heat pipe improved TOPAZ-II reactor is used to simulate different transient conditions to obtain the time series of the above six variables under multiple transient conditions. The time series is normalized by the maximum and minimum values, and the processing formula is as follows: (x n -x n,min ) / (x n,max -x n,min ), where x n is the value of the clean data of the nth variable, x n,min is the minimum value of the clean data of the nth variable, x n,max is the maximum value of the clean data of the nth variable.

[0034] Noise data can be generated by adding different degrees of Gaussian noise and outlier data to the normalized time series. Gaussian noise formula: Where σ is the variance, which obeys the uniform distribution of (0.02, 0.2); μ is the mean value, which obeys the uniform distribution of (-0.1, 0.1). Generation of outlier data: After adding Gaussian noise to the data, add outlier data. First, select the time point when the outlier is generated, and its possible generation time is included in the entire transient operating time range, and its generation probability is less than 10%. If outlier data is generated, the outlier noise obeys the exponential distribution, and its generation process is: 1.f(x)=1-e -0.5x ; 2. The data generated that obeys the exponential distribution changes to the range of (0.6,1).

[0035] The database data contains three levels, including test sets and training sets. Each data set contains multiple transient conditions, each of which contains multiple input signal sequences and output signal sequences, as well as the maximum and minimum values ​​of the input signal of the transient condition before normalization. The input data contains a 6×5 matrix composed of the above six variables from t-5 to t, as well as the expected steady-state fuel average temperature of the transient condition where the signal sequence is located. The output data is the clean values ​​of the six variables at time t.

[0036] Step 2. Construct an Encoder-Decoder neural network; the network structure includes a convolutional layer, a channel attention mechanism, a fully connected layer, and a deconvolutional layer, such as Figure 4 As shown. The network adopts a mean square error loss function. The Encoder-Decoder neural network can receive an input signal sequence window and output clean data, including:

[0037] S201.Encoder uses a one-dimensional convolutional neural network to encode the input signal sequence and obtain the initial feature vector;

[0038] S202. The channel attention mechanism conducts attention mechanism learning on the initial feature vector and improves the network's fitting ability through the fully connected layer, and finally obtains the representation feature vector input into the Decoder;

[0039] S203.Decoder uses an inverse convolutional neural network to decode the representation feature vector and finally obtain the output signal;

[0040] Step 3. Classify the training data in the dataset according to the requirements, taking transient conditions as units, and train a single network using one or more transient conditions. A single transient condition contains input signal sequences with different noises and clean output signal sequences. The loss function of the neural network uses the root mean square error function: in, is the value of the nth variable output by the neural network. The training data is generated online in real time until the error tends to be stable. The transient working conditions of the training data include 6 kinds of conditions, and the expected average fuel temperatures in the steady state are 1875K, 1850K, 1829K, 1808K, 1789K, and 1771K, respectively. A total of 6 neural networks are trained. The training parameters of each network are shown in the following table:

[0041] name parameter Epochs 120000 Training Method GPU-accelerated training Learning Rate Starting value: 8e-3 multiplied by 0.85 every 5000 times Training set / test set 0.8 / 0.2 Optimizer Adam Loss Function Root Mean Square Error Function Model selection Select the model that performs best on the test set

[0042] Step 4. Use the validation set data to test 6 neural networks. The validation set data includes 6 transient conditions. The expected average fuel temperatures in steady state are 1875K, 1850K, 1829K, 1808K, 1789K, and 1771K, respectively. 200 types of Gaussian noise and abnormal noise are added. Use the denoising autoencoder Encoder-Decoder network to denoise and analyze whether the average maximum error is less than 5% and the maximum average error is less than 8%. If not, go to step 3 to reduce the amount of data learned by a single neural network, increase the number of neural networks, and retrain; if satisfied, go to step 5. The calculation results are shown in the table below. Analysis shows that the 6 neural networks meet the given requirements and go to step 5.

[0043]

[0044] Step 5. Construct a hybrid expert system; the hybrid expert system consists of multiple expert models, and a single expert model consists of two neural network models and a moving average filter window. Its general structure is as follows: Figure 5 The hybrid expert system contains a total of 5 expert models, as shown in the following table:

[0045] name Submodel 1 Submodel 2 Submodel 3 Expert Model 1 Neural Network 1 Neural Network 2 Moving average filter Expert Model 2 Neural Network 2 Neural Network 3 Moving average filter Expert Model 3 Neural Network 3 Neural Network 4 Moving average filter Expert Model 4 Neural Network 4 Neural Network 5 Moving average filter Expert Model 5 Neural Network 5 Neural Network 6 Moving average filter

[0046] Sliding average filter formula: x n,i,tis the filtered data of the nth variable output by the ith network in the expert model at time t. When the hybrid expert system obtains the input signal sequence, the processing steps include:

[0047] S501. The expert system receives a noise signal sequence at time t. The expert model is selected based on the expected average fuel temperature of the transient condition where the noise signal sequence is located and the training data condition of the expert model. For example, if the noise data comes from a transient condition, the average fuel temperature in the steady state is 1780K. At this time, expert model 4 should be selected, because neural networks 4 and 5 in expert model 4 contain transient conditions. The selection should meet the principle that the transient condition to which the input signal sequence belongs is as close as possible to the transient condition contained in the expert model.

[0048] S502. According to the selected expert system, the input signals are normalized accordingly and respectively input into the neural network model included in the expert model.

[0049] S503. The output data of the neural network model included in the expert model is processed by moving average filtering, and then the output data after moving average filtering of each network is weighted averaged. Weighted formula: In the formula It represents the result of moving average filtering of the output data of the first and second neural network models in a single expert model. a1 and a2 are weighted coefficients. The formula is: a1 = (P s -P2) / (P1-P2), a2=1-a1. s It is a parameter when the transient condition of the input signal sequence is stable, and P1 and P2 are the parameters corresponding to the transient condition of the training data of the neural network model. Here, P1 and P2 select the average fuel temperature of the transient condition in the steady state. For example, when the noise data source is the transient condition and the expert model 4 is selected, a1 = (1780-1808) / (1789-1808), a2 = 1-a1.

[0050] Step 6. According to the time window size of the input signal sequence, input the noise data at time t and the thermal power from time t-1 to time t to the hybrid expert system, and the hybrid expert system outputs the clean data at time t. The test results of various working conditions are shown in the following table:

[0051]

[0052] It can be seen from the above table that compared with a single model, the hybrid expert model has higher accuracy and generalization ability. Under the 1100 medium Gaussian abnormal noise test, including training conditions and non-training conditions, its maximum error is less than 4.5%, the average maximum error is less than 1.5%, and the maximum average error is less than 0.5%. Figure 6A schematic diagram of the comparison of hybrid expert network filtering, single neural network model filtering, noisy data, and clean data for transient conditions with an expected average fuel temperature of 1798 K is given. As can be seen from the figure, the hybrid expert is basically consistent with the clean data and has a stronger data filtering capability than a single model.

[0053] The above contents are examples of the present invention, and the specific implementation methods of the present invention cannot be considered to be limited to these. For ordinary technicians in the technical field to which the present invention belongs, several simple deductions or substitutions can be made without departing from the concept of the present invention, which should be considered to be within the scope of patent protection determined by the claims submitted by the present invention.

Claims

1. A real-time filtering method for a reactor system based on a hybrid expert system, characterized in that: The steps include: Step 1. Collect transient conditions, determine the time window size of the input signal sequence, obtain noise data and clean data, and build a database; Step 2. Construct and train an Encoder-Decoder neural network to receive the input signal sequence window and output clean data; Step 3. Classify the data set and train multiple neural network sub-models; Step 4. Test the neural network to determine whether the maximum error and average error meet the requirements. If not, proceed to step 3 and re-divide the data set. If yes, proceed to step 5. Step 5. construct a hybrid expert system, insert the hybrid expert system according to the neural network, and generate a hybrid expert neural network; Step 6. According to the time window size of the input signal sequence, the noise data is input to the hybrid expert neural network, and the output is the denoised data.

2. The real-time filtering method for a reactor system based on a hybrid expert system according to claim 1, characterized in that: The step 1. collecting transient working conditions, determining the time window size of the input signal sequence, obtaining noise data and clean data, and building a database specifically includes: Determine the time window size of the input signal sequence, the data format of the time window of the input signal sequence is N×M, where N is the number of variables and M is the time step of the signal; The input signal sequence is normalized, and the formula is as follows: (x n -x n,min ) / (x n,max -x n,min ), where x n is the value of the clean data of the nth variable, x n,min is the minimum value of the clean data of the nth variable, x n,max is the maximum value of the clean data of the nth variable; The construction includes three levels of databases, in which each data set contains multiple transient conditions, and each transient condition contains multiple input signal sequences, output signal sequences, the expected steady-state fuel average temperature of the transient condition where the signal sequence is located, and the maximum and minimum values ​​of the input signal of the transient condition before normalization.

3. The real-time filtering method for a reactor system based on a hybrid expert system according to claim 1, characterized in that: The step 2. constructing and training an Encoder-Decoder neural network for receiving the input signal sequence window and outputting clean data specifically includes: S201.Encoder uses a one-dimensional convolutional neural network to encode the input signal sequence and obtain the initial feature vector; S202. The channel attention mechanism conducts attention mechanism learning on the initial feature vector, and improves the fitting ability of the network through the fully connected layer, and finally obtains the representation feature vector input into the Decoder; S203.Decoder uses an inverse convolutional neural network to decode the representation feature vector and finally obtain the output signal.

4. The real-time filtering method for a reactor system based on a hybrid expert system according to claim 1, characterized in that: The step 3. classifying the data set and training multiple neural network sub-models specifically includes: According to the type of transient operating condition of the reactor system, the collected transient operating condition data are classified, and a single Encoder-Decoder network is trained through one or more transient operating conditions; under specific transient operating conditions, the training data includes high-fidelity data and noise data obtained at different times; the high-fidelity data comes from the numerical simulation of high-precision system programs or the measurement of high-precision measuring instruments; the noise data comes from the measurement of low-precision measuring instruments, or the numerical simulation of high-precision system programs with noise added; when training the neural network, the input is a signal parameter sequence containing different noise levels, and under the same transient operating condition and at the same time, the output can obtain a high-fidelity signal parameter sequence; The loss function of the neural network uses the root mean square error function: in, is the value of the nth variable output by the neural network.

5. The real-time filtering method for a reactor system based on a hybrid expert system according to claim 1, characterized in that: The step 5. constructs a hybrid expert system, wherein the hybrid expert system is composed of multiple expert models, and a single expert model is composed of two neural network models and a moving average filter window. When the hybrid expert system obtains an input signal sequence, the processing steps include: S501. The hybrid expert system receives the noise signal sequence at time t and the expected steady-state average fuel temperature under the transient condition, and selects an expert model according to the thermal power and the training data condition of the expert model; S502. According to the selected expert model, the input signal sequence is normalized accordingly and input into the neural network model included in the expert model respectively; S503. The output data of the neural network model included in the expert model is processed by moving average filtering, and then the output data of each network after moving average filtering is weighted averaged. The weighted formula is: In the formula It represents the result of moving average filtering of the output data of the first and second neural network models in a single expert model. a1 and a2 are weighted coefficients. The formula is: a1 = (P s -P2) / (P1-P2), a2=1-a1, where P s It is a parameter when the transient condition of the input signal sequence is stable, and P1 and P2 are the parameters corresponding to the transient condition of the training data of the neural network model.

Images