A neural network-based rapid phase modulation method and device for a pulse tube refrigerator

Abstract

The application discloses a kind of neural network-based pulse tube refrigerator fast phase modulation method, comprising: the present application is based on neural network model to predict valve opening degree, the input of the neural network model is valve opening degree and valve opening degree error, output is the refrigeration performance of pulse tube refrigeration unit, neural network model is trained based on back propagation algorithm, establish the nonlinear relationship between valve opening degree and refrigeration performance, and calculate the appropriate valve opening degree through given refrigeration performance, and then output control signal to valve adjusting device.The present application accurately predicts the nonlinear relationship between the valve opening degree and the refrigeration performance of the pulse tube refrigerator through the neural network model, thereby achieving fast adjustment of the optimal valve opening degree, significantly improving the debugging efficiency.

Description

TECHNICAL FIELD

[0001] The present application relates to a pulse tube cryocooler, and particularly to a neural network-based pulse tube cryocooler rapid phase modulation method and device. BACKGROUND

[0002] The pulse tube cryocooler has been widely used in quantum physics, particle physics and other high-tech fields due to its simple structure, high reliability, small mechanical vibration and resistance to electromagnetic interference. Although the pulse tube cryocooler has shown excellent performance in various low-temperature applications, the optimization of its refrigeration efficiency and performance is highly dependent on the precise control of the phase modulation valve, which is mainly used to adjust the phase relationship between the pressure wave and the mass flow rate inside the pulse tube cryocooler. Currently, the pulse tube cryocooler mostly uses a bidirectional inlet phase modulation method, which includes a bidirectional inlet valve and a small orifice valve. However, the bidirectional inlet structure can cause a direct current inside the pulse tube cryocooler, which in turn can cause temperature peaks in the pulse tube and reduce the refrigeration performance. In order to weaken the influence of the direct current caused by the bidirectional inlet phase modulation method, an additional valve is needed to control the direct current more accurately.

[0003] However, with the increase in the number of phase modulation valves, the complexity of the valve combination also increases, and the influence of the valve opening degree change on the refrigeration performance shows a nonlinear relationship. This makes it extremely complex to accurately describe the interaction model between different valves. Currently, there is a lack of effective models that can clearly describe the relationship between the valve opening degree and the refrigeration performance, resulting in a dependence on a large amount of experimental data during the commissioning process of the cryocooler, and the commissioning process is usually time-consuming and costly.

[0004] As a highly adaptive machine learning model capable of handling complex nonlinear relationships, neural networks have strong function approximation capabilities. In the valve adjustment process of the pulse tube cryocooler, neural networks can accurately establish the nonlinear relationship between the valve opening degree and the refrigeration performance through training on a large amount of experimental data. Compared with traditional statistical models, neural networks show higher prediction accuracy and generalization ability in handling complex nonlinear function fitting tasks. The neural network-based phase modulation method provides a new technical approach for the performance optimization of the pulse tube cryocooler, which can significantly improve the commissioning efficiency, reduce the experimental cost, and optimize the operating performance of the cryocooler. SUMMARY

[0005] The present application provides a neural network-based pulse tube cryocooler rapid phase modulation method and device, aiming to improve the commissioning efficiency of the pulse tube cryocooler, reduce the experimental cost, and optimize the operating performance of the cryocooler. This method establishes the nonlinear relationship between the valve opening degree and the refrigeration performance of the pulse tube cryocooler by using neural networks, solving the high cost and low efficiency problems existing in the current commissioning method.

[0006] A neural network-based pulse tube refrigerator fast phase modulation method, comprising the following steps:

[0007] (1) Collecting actual opening degree combination data of each valve in the phase modulation mechanism of the pulse tube refrigerator to be phase modulated and corresponding refrigeration performance data, and obtaining an initial training set;

[0008] (2) Training a neural network using the current training set data to generate a valve prediction model based on the neural network;

[0009] (3) Inputting different valve opening degree combination data into the obtained valve prediction model to obtain the corresponding refrigeration performance of the different valve opening degree combination data, and obtaining corresponding target valve opening degree combination data according to the target refrigeration performance;

[0010] (3) Inputting the current obtained target valve opening degree combination data into the valve adjusting device to adjust the opening degree of each valve in the phase modulation mechanism;

[0011] (4) Detecting the refrigeration performance data corresponding to the current valve opening degree, if the refrigeration performance meets the target, the phase modulation is completed; if the refrigeration performance does not meet the requirement, the obtained current valve opening degree combination data and the corresponding refrigeration performance data are stored in the training set, the training set is updated, and the neural network is retrained in step (2).

[0012] Considering the influence of the operating frequency on the performance of the refrigerator, in order to further improve the phase modulation quality, in step (1), the operating frequency of the pulse tube refrigerator is also collected, and the operating frequency and the valve opening degree combination data are simultaneously inputted, and the refrigeration performance is taken as the output, which is used for the neural network training in step (2).

[0013] In the present application, the refrigeration performance mainly considers the refrigeration temperature and the refrigeration capacity, and the pressure data at one or more positions can also be considered according to actual needs. The refrigeration temperature mainly refers to the temperature of each stage cold head or other temperatures (such as the temperature of the regenerator). As a preferred embodiment, the temperature signals measured by the refrigeration unit performance test system can also include the temperatures of the cold end heat exchangers of the pulse tube refrigerator, the temperatures of the regenerators at each stage, and the temperatures at the intermediate positions of the pulse tubes. As a preferred embodiment, the pressure fluctuations at the inlet of the regenerator of the pulse tube refrigeration unit and before and after the phase modulation valve can be used as input parameters.

[0014] Further, the neural network is a feedforward neural network model. The input of the feedforward neural network model is the valve opening degree and the enhanced data considering the valve opening degree error, and the output is the refrigeration performance of the pulse tube refrigeration unit. The feedforward neural network model is trained based on the back propagation algorithm, a nonlinear relationship between the valve opening degree and the refrigeration performance is established, and the appropriate valve opening degree is calculated through the given refrigeration performance, and then the control signal is output to the valve adjusting device.

[0015] Further, the neural network is trained by using a back propagation algorithm, and the weights of the network are adjusted by propagating errors and using a gradient descent optimization method.

[0016] Further, when training the neural network, the activation function uses a ReLU function or a Leaky ReLU function.

[0017] Since the phase modulation mechanism of the pulse tube refrigerator usually adopts a needle valve with a mechanical structure, when rotating the valve core, due to the influence of the thread gap and the mechanical transmission error, there will be a certain deviation between the actual opening degree of the valve and the input opening degree. As an optimization, the input layer of the feedforward neural network needs to include the error of the valve opening degree, that is, the valve opening degree is regarded as a random variable, and is input to the mean (expected value) of the valve opening degree variable, and the error caused by the thread gap and the mechanical transmission device can be regarded as the variance of this random variable. As an optimization, based on the above error analysis, the training set data is enhanced: for each actual opening degree combination and refrigeration performance data pair actually obtained in the training set data, a plurality of training samples are generated, the opening degree combination in the plurality of training samples changes, and the corresponding refrigeration performance data remains unchanged. The number of generated training samples is generally determined according to the error size.

[0018] Further, the enhanced processing of the training set data is specifically: adding random noise from the estimated error distribution to each valve opening degree data point to generate a plurality of possible training samples R i , as an optimization, wherein R i should satisfy the normal distribution:

[0019]

[0020] In the formula, D i represents the expected value of the valve opening degree, and it should be noted that the training sample R i considering the error is the same as D i .

[0021] The training set of the feedforward neural network not only contains experimental data, but also generates corresponding enhanced data set according to the measured valve opening degree and valve opening degree error, and adjusts the weights according to the error between the generated data and the original data during back propagation learning, so as to reduce the interference of samples with large error on the feedforward neural network model.

[0022] Further, when training the neural network, the mean square error is selected to calculate the loss function of the model. Specifically, the expression is as follows:

[0023]

[0024] In the formula, y is the true target value, is the prediction value of the feedforward neural network, and p is the number of samples.

[0025] Further, when training the neural network, a weighted mean square error is selected to calculate the loss function of the model, and the weight can be calculated according to the deviation and variance of the valve opening degree.

[0026] Specifically, the loss function expression is as follows:

[0027]

[0028] wherein r i is the weight of the sample, which can be calculated according to the deviation (i.e. the difference from the expected number of turns) and variance of the valve opening degree. As preferred, for the ith sample, the corresponding weight r i is calculated as follows:

[0029]

[0030] wherein and are adjustment parameters for controlling the influence degree of the distance and variance on the weight, p is the number of samples, D i is the expected value of each valve opening degree, and σ i 2 is the variance of each valve opening degree, and R i is to add random noise from the estimated error distribution to each valve opening degree data point to generate a plurality of possible training samples, satisfying R i ~ N(D i , σ i 2 ).

[0031] Further, the parameters should be adjusted according to the influence degree of different valves on the refrigeration performance of the pulse tube refrigeration unit. During the training process, the feedforward neural network will gradually learn the influence of the valve opening degree error on the refrigeration performance and adjust the weight of the network by optimizing the loss function. Since the variance information of the error is introduced, the feedforward neural network can better adapt to the valve opening degree error in actual operation and improve the robustness and accuracy of its prediction.

[0032] In the training process of the feedforward neural network, two stages are included: in the initial stage, when the training set lacks experimental data, the opening degrees of the valves of the phase modulation mechanism of the pulse tube refrigerator are changed and the corresponding refrigeration performance data are recorded, which are used as the initial data set to train the feedforward neural network model, so as to obtain the initially trained feedforward neural network model. In the second stage, different valve opening degree combinations are input to the initially trained feedforward neural network model, the feedforward neural network model outputs the refrigeration performance corresponding to the different valve opening degree combinations, and the corresponding valve opening degree combination is obtained according to the input target refrigeration performance, and a valve adjustment signal is output to the valve adjustment device, so as to adjust the phase modulation mechanism of the pulse tube refrigerator. After actual phase modulation, when the actual refrigeration performance does not meet the target refrigeration performance, the data set is input to the training set, and the complete training set is updated.

[0033] A pulse tube refrigerator rapid phase modulation device based on a feedforward neural network, comprising:

[0034] A data acquisition and analysis module acquires actual valve opening degree combination data of a phase modulation mechanism of a pulse tube refrigerator to be phase modulated and corresponding refrigeration performance data, obtains an initial training set, trains a valve prediction model based on a neural network according to the phase modulation method of steps (2)-(4) of any of the above technical solutions, obtains different valve opening degree combination data corresponding to a target refrigeration performance, and transmits the data to a valve adjustment device.

[0035] A valve adjustment device receives different valve opening degree combination data obtained by the data analysis module, and adjusts the opening degrees of the valves in the phase modulation mechanism according to the data.

[0036] The actual valve opening degree combination data and the corresponding refrigeration performance data can be obtained from a data monitoring system (or element) of the original refrigerator, or an actual opening degree combination data detection system (or element) and a refrigeration performance test system (or element) can be separately arranged, and existing structures or elements can be used.

[0037] More specifically, a pulse tube refrigerator rapid phase modulation device based on a feedforward neural network, comprising at least one refrigeration unit to be phase modulated, a refrigeration unit performance test system, and a valve control system based on a feedforward neural network:

[0038] The refrigeration unit comprises a plurality of compression devices, a gas distribution system, a regenerator, a pulse tube, a heat exchanger, and a phase modulation valve (or a phase modulation mechanism). The phase modulation valve is used to adjust the flow and pressure inside the pulse tube refrigerator;

[0039] The refrigeration unit performance test system comprises a temperature sensor or / and a pressure sensor or / and a data acquisition device, etc., for monitoring the working state of the refrigeration unit in real time. Further, the refrigeration unit performance test system is used for monitoring the performance data at the cold end heat exchanger of each stage of the pulse tube refrigerator in real time, and inputting the data into the valve control system based on the feedforward neural network for analysis.

[0040] The valve control system based on the feedforward neural network comprises a data acquisition and analysis module (containing a valve opening degree adjustment program) and a valve adjustment device. The valve opening degree adjustment program analyzes the output signal of the refrigeration unit performance test system, predicts the valve opening degree based on the feedforward neural network model, and outputs a control signal to the valve adjustment device, thereby controlling the phase modulation valve in the pulse tube refrigerator.

[0041] The valve adjustment system automatically adjusts the valve opening degree to realize rapid optimization of the performance of the refrigerator and improve the production efficiency.

[0042] The pulse tube refrigerator comprises a GM type pulse tube refrigerator or a Stirling type pulse tube refrigerator, and the structure of the refrigerator comprises a single-stage or multi-stage structure. The phase modulation mechanism comprises a plurality of valves for adjusting the flow and pressure inside the pulse tube refrigerator, so as to optimize the refrigeration performance of the pulse tube refrigerator.

[0043] Compared with the prior art, the beneficial effects of the present application are embodied in:

[0044] The present application accurately predicts the nonlinear relationship between the valve opening degree of the pulse tube refrigerator and the refrigeration performance through the feedforward neural network model, so as to realize rapid adjustment of the optimal valve opening degree. By testing the refrigeration performance according to the optimal opening degree predicted by the model, not only the debugging efficiency is significantly improved and the experimental cost is reduced, but also the interaction mechanism between the valves is effectively revealed, further improving the efficiency of valve adjustment. In addition, the method provides key technical support for the analysis of the thermodynamic process in the pulse tube refrigerator and the optimization of the refrigeration performance, and significantly improves the overall operation efficiency of the pulse tube refrigerator. BRIEF DESCRIPTION OF DRAWINGS

[0045] Figure 1 The flowchart of the valve control system based on the feedforward neural network and the training of the feedforward neural network.

[0046] Figure 2 The schematic diagram of the training of the feedforward neural network of the valve opening degree and the refrigeration performance.

[0047] Figure 3 The structure schematic diagram of one embodiment of the rapid phase modulation system of the pulse tube refrigerator based on the feedforward neural network of the present application.

[0048] Figure 4This section presents a comparison between the no-load cooling temperature predicted by the feedforward neural network model and the measured temperature in the experimental example. Detailed Implementation

[0049] The present invention will be further described below with reference to the accompanying drawings. Any component models, material names, connection structures, control methods, algorithms, etc., not explicitly stated in this technical solution are considered common technical features disclosed in the prior art.

[0050] like Figure 1 As shown, a rapid phase adjustment method for a pulse tube refrigerator based on a neural network includes the following steps:

[0051] (1) Collect the actual opening combination data of each valve in the phase adjustment mechanism of the pulse tube refrigerator to be adjusted and the corresponding refrigeration performance data to obtain the initial training set;

[0052] (2) Train a feedforward neural network using the current training set data to generate a valve prediction model based on the neural network;

[0053] (3) Input different valve opening combination data into the obtained valve prediction model to obtain the refrigeration performance data corresponding to different valve opening combination data, and obtain the corresponding target valve opening combination data according to the target refrigeration performance;

[0054] (3) Input the currently obtained target valve opening combination data into the valve regulating device and adjust the opening of each valve in the phase adjustment mechanism;

[0055] (4) Record the refrigeration performance data corresponding to the current valve opening. If the refrigeration performance meets the target, the phase adjustment is completed. If the refrigeration performance does not meet the requirements, the obtained current valve opening combination and the corresponding refrigeration performance data are stored in the training set, the training set is updated, and the neural network is retrained in step (2).

[0056] The following is a further explanation of the feedforward neural network for training valve opening and refrigeration performance mentioned in this invention:

[0057] Figure 2 This demonstrates a feedforward neural network model used to train valve opening and refrigeration performance, where the output of each neuron in the layer can be represented as:

[0058] (1)

[0059] In the formula, It is the first The output of layer neurons, It is the first Layer to The weight matrix of the layer, It is the first Layer bias, is an activation function. In addition, as shown in Figure 2 , represents an input layer, represents an output layer.

[0060] In the model training stage, the back propagation algorithm is used to train the feedforward neural network, and the network weights are adjusted by propagating the error and using gradient descent and other optimization methods. The training network aims to minimize the error between the predicted value and the actual value. We can choose the mean square error to calculate the loss function of the model:

[0061] (2)

[0062] In the formula, is the true target value, is the predicted value of the feedforward neural network, and p is the number of samples.

[0063] According to the deviation between the prediction result of the feedforward neural network model and the actual result, the parameters related to the prediction result of the model can be adjusted, including but not limited to the following ways: first, the proportion of the training set, the validation set and the test set in the data set is optimized and adjusted to improve the stability and generalization ability of the model training. For example, the data set division ratio can be set to 7:2:1 or 6:2:2; second, the selection of the activation function in the feedforward neural network is optimized. The ReLU function or the Leaky ReLU function can be used in the model training stage to enhance the fitting ability of the feedforward neural network to nonlinear features, improve the training speed and convergence performance of the model; third, according to the number of input and output parameters and the complexity of the task, the number of hidden layers and the number of neurons in each layer of the feedforward neural network can be appropriately set, preferably 1 to 3 hidden layers, and the optimal structure is selected through experiments to avoid overfitting or underfitting, thereby further improving the prediction accuracy of the model and the system control performance.

[0064] In order to consider the influence of valve opening error in the feedforward neural network model, data enhancement processing needs to be performed on the valve opening, and the specific operation method is as follows:

[0065] (1). Determine the error of the valve opening: As mentioned earlier, the error of the phase modulation mechanism valve opening of the pulse tube refrigeration unit mainly comes from the error of the thread gap of the phase modulation valve and the error brought by the mechanical transmission components. Therefore, first of all, the thread gap and the mechanical transmission components need to be analyzed. Let the expected value of each valve opening be D i , and the variance of each valve opening be σ i 2 .

[0066] (2). Input layer data enhancement processing: add random noise from the estimated error distribution to each valve opening data point to generate multiple possible training samples Ri As preferred, R i The normal distribution should be satisfied:

[0067] (3)

[0068] It should be noted that the training samples R i The corresponding refrigeration performance and D i are the same.

[0069] (3). Weighted loss function: In order to make the feedforward neural network better handle errors, a weighting mechanism can be introduced in the loss function to ensure that more attention is paid to samples with smaller errors during training. The specific approach is to calculate the weighted loss of each sample, and the weight is related to the error (or variance) of each sample. Samples closer to the expected number of laps (or set value) should be given greater weight, while samples with larger errors should be given smaller weight. As preferred, the loss function can be expressed as:

[0070] (4)

[0071] where r i is the weight of the sample, which can be calculated according to the deviation of the valve opening (i.e. the difference from the expected number of laps) and the variance. As preferred, the weight calculation formula is:

[0072] (5)

[0073] where and are adjustment parameters used to control the degree of influence of distance and variance on weight. As preferred, the parameters should be adjusted according to the degree of influence of different valves on the refrigeration performance of the pulse tube refrigeration unit. During the training process, the feedforward neural network will gradually learn the influence of valve opening error on refrigeration performance and adjust the weights of the network by optimizing the loss function. Since the variance information of the error is introduced, the feedforward neural network can better adapt to the valve opening error in actual operation, improving its prediction robustness and accuracy.

[0074] As Figure 3 shown, the present embodiment aims to quickly obtain the lowest no-load refrigeration temperature of a single-stage GM pulse tube refrigerator. The single-stage GM pulse tube refrigerator mainly includes a pulse tube refrigeration unit A, a refrigeration unit performance test system D, a data acquisition and analysis module C, and a valve adjustment device B. The data acquisition and analysis module C and the valve adjustment device B constitute the valve control system based on the feedforward neural network of the present invention.

[0075] The pulse tube refrigeration unit comprises a compressor 10, a rotary valve 9, a hot end heat exchanger 4, a regenerator 1, a cold end heat exchanger 3, a pulse tube 2, a small hole valve 7, a gas reservoir 8, a two-way inlet valve 5 and a two-way inlet valve 6 connected in sequence, the small hole valve 7 and the two-way inlet valves 5 and 6 being a phase modulation valve group E (i.e. a phase modulation mechanism) of the pulse tube refrigeration unit, and the opening degree of each valve being controlled and adjusted by a valve adjusting device.

[0076] The working process of the embodiment is as follows:

[0077] The pulse tube refrigeration unit performance test system D is started, and Table 1 shows the sensors used in the pulse tube refrigeration unit performance test system in the embodiment.

[0078] Table 1. Details of the pulse tube refrigeration unit performance test system

[0079]

[0080] The refrigerator is started, and the temperature of the refrigerator is recorded by the refrigeration unit performance test system D, and the temperature of the refrigerator is waited to reach stability.

[0081] The training set is established. After the temperature of the refrigerator reaches stability, the valve control system outputs control signals to the phase modulation valve group E of the pulse tube refrigeration unit to control each valve in the phase modulation valve group E. In order to ensure the breadth of the data set, the valve opening degree is set at intervals of 1.0 turns and covers the full range of the valve. After adjusting the valve each time, the temperature of the refrigerator needs to be waited to reach stability, and the opening degree of each valve in the phase modulation valve group E and the no-load refrigeration temperature, pressure and other performance parameters of the refrigerator using the refrigeration unit performance test system D are recorded.

[0082] The training of the feedforward neural network model. In the embodiment, the data acquisition and analysis module C can use a computer combined with a specific computer program to complete the calculation of data and the training of the model. After the valve opening degree and the performance parameters of the refrigerator are preliminarily obtained, the training set data is first enhanced according to the error of each valve opening degree, the weight corresponding to each data is calculated by formula (5), and the coefficients of the feedforward neural network model are trained and adjusted according to the loss function shown in formula (4).

[0083] After the model is trained, in order to verify the accuracy of the valve opening degree and the no-load refrigeration temperature function model constructed, ten sets of opening degree combinations are selected from the non-training data set (previously collected) by using the random sampling method for no-load refrigeration temperature prediction and experimental verification. The comparison results of the ten sets of experimental temperature data and the predicted temperature data are as follows: Figure 4The results show that the average relative error is 5.1%, and the deviation between the overall model predicted temperature and the experimental temperature is small, indicating that the model has good prediction performance. Based on the above experimental results, the experimental data is included in the training model data set constructed in the early stage to further optimize the model performance.

[0084] Optimal valve opening prediction. A set of arrays with 0.1 circle intervals covering the entire range of the valve is generated by the computer program, which is input into the preliminarily trained feedforward neural network model to calculate the corresponding refrigeration performance of the refrigeration unit at each valve opening. The valve opening corresponding to the working condition under the condition of no refrigeration capacity, cold end heat exchanger temperature between 14 K and 15 K, regenerator inlet pressure wave high pressure greater than 2.5 MPa and low pressure less than 0.9 MPa is found, and the control signal is output from the valve control system B to the phase modulation valve group E. After the temperature stabilizes, the performance data corresponding to the valve opening is obtained, and if the performance data meets the requirements, the phase modulation operation is completed. If it does not meet the requirements, the experimental results can be added to the training set to retrain the model, and the related parameters of the feedforward neural network model are adjusted according to the deviation between the experimental value and the predicted result.

[0085] Based on the above operation content, the lowest no-load refrigeration temperature predicted by the feedforward neural network model, the corresponding valve opening and the corresponding measured no-load refrigeration temperature are obtained, and the specific results are shown in Table 2. The opening of the bidirectional inlet valve 5 is denoted as D1, the opening of the bidirectional inlet valve 6 is denoted as D2, and the opening of the small hole valve 7 is denoted as O.

[0086] Table 2. The lowest no-load refrigeration temperature, valve opening and measured results predicted by the feedforward neural network model

[0087]

[0088] As shown in Table 2, for three different combinations of valve openings, the feedforward neural network model predicts temperatures of 14.0 K, 14.5 K and 14.7 K, respectively, while the measured temperatures are all 15.2 K. The three opening combinations can all make the refrigerator reach the lowest no-load refrigeration temperature, which not only shows that there are multiple corresponding relationships between the bidirectional inlet valve and the small hole valve opening when the pulse tube refrigerator reaches the lowest no-load refrigeration temperature, but also verifies the excellent performance of the feedforward neural network model in predicting the optimal valve opening of the pulse tube refrigerator, proving the effectiveness and reliability of the model, which is of great significance for optimizing the performance of the refrigerator and improving the efficiency of valve regulation.

[0089] The foregoing description of the embodiments has been presented for the purpose of illustration and description. It is not intended to be exhaustive or to limit the application to the precise form disclosed. Modifications and variations are possible in light of the above teachings or can be acquired from practice of the application. As well, the description is presented in the context of the preferred embodiments as a number of alternatives. It is not intended to limit the application to the precise form described.

Claims

1. A neural network-based fast phase modulation method for a pulse tube refrigerator, characterized by, The method comprises the following steps: (1) collecting actual valve opening combination data and corresponding refrigeration performance data in a phase modulation mechanism of a pulse tube cryocooler to be phase modulated, and obtaining an initial training set; (2) training a neural network by using current training set data to generate a valve prediction model based on the neural network; (3) inputting different valve opening combination data into the obtained valve prediction model to obtain refrigeration performance data corresponding to the different valve opening combination data, and obtaining corresponding target valve opening combination data according to a target refrigeration performance; (3) inputting the obtained target valve opening combination data into a valve adjusting device to adjust the opening of each valve in the phase modulation mechanism; (4) detecting the refrigeration performance data corresponding to the current valve opening, and if the refrigeration performance meets the target, the phase modulation is completed; If the refrigeration performance does not meet the requirement, the obtained current target valve opening combination and corresponding refrigeration performance data are stored in the training set, the training set is updated, and the neural network is retrained in step (2).

2. The neural network-based rapid phase modulation method of a pulse tube refrigerator according to claim 1, characterized by, In step (1), the operating frequency of the pulse tube cryocooler is collected at the same time, and the operating frequency and the valve opening combination data are used as inputs, and the refrigeration performance is used as output, which is used for neural network training in step (2).

3. The neural network-based quick phase modulation method of a pulse tube refrigerator according to claim 1 or 2, characterized by, The neural network is a feedforward neural network model; the neural network is trained by using a back propagation algorithm, the error is propagated, and the weights of the network are adjusted by using a gradient descent optimization method.

4. The neural network-based quick phase modulation method of a pulse tube refrigerator according to claim 1 or 2, characterized by, When training the neural network, the activation function uses a ReLU function or a Leaky ReLU function.

5. The neural network-based rapid phase tuning method for a pulse tube refrigerator according to claim 1, wherein The input layer of the neural network includes enhanced data considering error information of valve opening.

6. The neural network-based quick phase modulation method of a pulse tube refrigerator according to claim 5, wherein The method for enhancing the training set data is: for each valve opening expected value D i plus random noise from the estimated error distribution, generating a plurality of error-considered training samples R i ; error-considered training samples R i The corresponding refrigeration performance and D i are the same.

7. The neural network-based quick phase modulation method of a pulse tube refrigerator according to claim 1 or 2, characterized by, When training the neural network, the mean square error is selected to calculate the loss function of the model.

8. The neural network-based quick phase modulation method of a pulse tube refrigerator according to claim 1 or 2, characterized by, When training the neural network, the weighted mean square error is selected to calculate the loss function of the model, and the weight can be calculated according to the deviation and variance of the valve opening.

9. The neural network-based quick phase modulation method of a pulse tube refrigerator according to claim 8, wherein For the i-th sample, its corresponding weight r i The calculation formula is: ； where and are tuning parameters that control the degree to which distance and variance affect the weight, p is the number of samples, D i is the desired value for each valve opening, σ i 2 is the variance for each valve opening, R i is to add random noise from the estimated error distribution to each valve opening data point to generate multiple possible training samples that satisfy R i ~ N(D i , σ i 2 ).

10. A neural network-based pulse tube cryocooler rapid phase modulation device, characterized in that, a data collection and analysis module collects actual valve opening combination data and corresponding refrigeration performance data in a phase modulation mechanism of a pulse tube cryocooler to be phase modulated, and obtains an initial training set; the phase modulation method according to steps (2)-(4) of claim 1 trains a neural network-based valve prediction model, obtains different valve target opening combination data corresponding to a target refrigeration performance, and transmits the data to a valve adjusting device; a valve adjusting device receives different valve opening combination data obtained by the data analysis module, and adjusts the opening of each valve in the phase modulation mechanism according to the data.

Images