Method and system for improving performance of CPT atomic clock based on machine learning

Abstract

The invention discloses a method and system for improving the performance of a CPT atomic clock based on machine learning, and relates to the technical field of atomic clocks, and the technical key points are that the method comprises the following steps: S1, building a BP neural network model suitable for the temperature control of a laser tube of the CPT atomic clock; s2, taking the output of the model as the input of a PID (Proportion Integration Differentiation) controller, and establishing a BP-PID control model; s3, acquiring training parameters of the BP neural network model, and performing optimization training on the BP-PID control model; and S4, performing real-time control on the temperature of the CPT atomic clock laser tube by using the trained BP-PID control model. The method is used for performing real-time temperature control on the CPT atomic clock laser tube, adopts a BP neural network PID algorithm to replace a traditional PID control algorithm, has the excellent characteristics of high response speed, small overshoot, short stabilization time, good robustness and the like, and can realize self-tuning of PID parameters in the temperature control process, so that the overall performance of an atomic clock can be greatly improved.

Description

Technical Field

[0001] The present invention relates to the technical field of atomic clocks, and particularly relates to a method and system for improving the performance of a CPT atomic clock based on machine learning. Background Art

[0002] Entering the atomic era, the emergence of atomic clocks has brought a revolutionary change to the definition of the second. An atomic clock is a high-precision timekeeping device that locks the frequency of a local oscillator to the transition frequency between the internal energy levels of an atom and outputs a standard time-frequency signal, greatly improving the accuracy and stability of timekeeping. After more than half a century of development, high-precision atomic clocks, also known as atomic frequency standards, are not only widely used in fields such as timekeeping and time dissemination, navigation and positioning, aerospace, and national defense security, but also play an important role in scientific research such as precision measurement and exploration of basic physical mechanisms.

[0003] Currently, the research on atomic frequency standards mainly has two directions. One is to explore and study atomic clocks with higher accuracy and stability, such as cold atom fountain clocks or optical clocks. The other is to develop miniaturized and low-power small atomic clocks using the physical principle of Coherent Population Trapping (CPT). Because CPT atomic clocks no longer require a microwave resonator and are not restricted by the wavelength of the radiation field in terms of size, they have the most ideal potential for miniaturization and integration at present. The development of traditional atomic clocks is limited by challenges in aspects such as stability and automatic control. Summary of the Invention

[0004] The purpose of the present invention is to solve the above problems and provide a method and system for improving the performance of a CPT atomic clock based on machine learning.

[0005] To achieve the above purpose, the technical solution of the present invention is as follows:

[0006] The present invention provides a method for improving the performance of a CPT atomic clock based on machine learning, including the following steps:

[0007] Step S1: Establish a BP neural network model applicable to the temperature control of the laser tube of a CPT atomic clock;

[0008] Step S2: Use the output of this model as the input of a PID controller to establish a BP-PID control model;

[0009] Step S3: Obtain the training parameters of the BP neural network model and optimize and train the BP-PID control model;

[0010] Step S4: Use the trained BP-PID control model to perform real-time control on the temperature of the laser tube of the CPT atomic clock.

[0011] The present invention is further configured such that the sub-steps of the step S1 are as follows:

[0012] Step S11: Establish a BP neural network model including an input layer, a hidden layer, and an output layer;

[0013] Step S12: Determine the number of nodes in the input layer, hidden layer, and output layer. The input layer has 3 neurons X i (i = 1, 2, 3), which are respectively the change amount of error Δe(k), error e(k), and the change trend of error Δe(k) - Δe(k - 1); the output layer has 3 neurons Y l (l = 1, 2, 3), which respectively correspond to the proportionality coefficient K p , integral coefficient K i and differential coefficient K d ; Calculate the number of neurons in the hidden layer according to the formula (m + n) 1 / 2 + a, where m represents the number of neurons in the input layer, here m = 3; n represents the number of neurons in the output layer, here n = 3; a is a constant between 1 and 10, and here the number of neurons in the hidden layer is set to 8; net1j is the parameter of the hidden layer neuron Hj, and net2j is the parameter of the output layer neuron Yl;

[0014] Step S13: Give the initial values w 1ij and w 2jl of the weight coefficients of each layer, select the learning rate η, inertia coefficient γ, and calculate the neurons in the hidden layer and output layer;

[0015] The calculation formula of the input layer neuron X i (i = 1, 2, 3) is as follows:

[0016] X1 = Δe(k)

[0017] X2 = e(k)

[0018] X3 = Δe(k) - Δe(k - 1)

[0019] The calculation formula of the hidden layer neuron H j (j = 1, 2, …8) is as follows:

[0020]

[0021] where

[0022]

[0023] w 1ij is the connection weight between the input layer and the hidden layer;

[0024] The output layer neuron Y l(l = 1, 2, 3) The calculation formula is as follows:

[0025]

[0026] Among them

[0027]

[0028] w 2jl is the connection weight between the hidden layer and the output layer.

[0029] The present invention is further configured to: use the output value Y l (l = 1, 2, 3) of this model as the input of the PID controller parameters and act on the controlled object.

[0030] K p = Y1

[0031] K i = Y2

[0032] K d = Y3

[0033] Among them, the proportional coefficient K p , the integral coefficient K i , and the differential coefficient K d are the control parameters of the PID controller. The increment generated by the PID operation is:

[0034] u(k) = u(k - 1)+Δu(k - 1)

[0035] Δu(k) = K p [e(k)-e(k - 1)]+K i e(k)+K d [Δe(k)-Δe(k - 1)]

[0036] e(k) = r(k)-y(k)

[0037] Among them, r(k) represents the target value, y(k) represents the output value, u(k) represents the system control quantity acting on the temperature control object, the difference between the target value and the output value gives the error value e(k), the change in the error e(k)-e(k - 1), and the change trend of the error Δe(k)-Δe(k - 1).

[0038] The present invention is further configured to: set the minimum value of the expected error of the BP neural network. When the expected target value cannot be obtained in the output layer, it turns to backpropagation; design the back-calculation process of the neural network, and take the error function E as the following formula, 1 / 2 is the proportional coefficient, and r and y are the input quantity and the output quantity respectively:

[0039]

[0040] Calculation of the adjustment value of each connection weight between the hidden layer and the output layer:

[0041]

[0042] Calculation of the adjustment value of each connection weight between the input layer and the hidden layer:

[0043]

[0044]

[0045] Wherein is unknown. Take the partial derivative of the output quantity with respect to the change of the increment. It should be a positive and negative change trend, which is approximately replaced by the sign function;

[0046] In the reverse calculation process, the error signal is transmitted backward along the network. The BP neural network corrects the weight coefficients on the network nodes according to the gradient descent method, and then outputs forward. After repeating many times of training, when the error value obtained by subtracting the target value from the output value meets the set requirements, the training of the BP-PID control model is completed.

[0047] The present invention also provides a temperature control system for improving the performance of a CPT atomic clock based on machine learning, including: taking a digital signal processor chip as the microcontroller core, combining the BP-PID control principle, calculating the required temperature control quantity in real time through the temperature error value, and transmitting it to the thermoelectric cooler TEC in the form of a pulse width modulation signal, so that the TEC heats or cools the laser tube to realize the temperature control of the laser tube of the CPT atomic clock.

[0048] The present invention is further set as: including a PID temperature control object module VCSEL laser tube, a temperature acquisition module thermistor, a microcontroller module single-chip microcomputer, and a thermoelectric cooler TEC:

[0049] The PID temperature control object module refers to the VCSEL laser tube that needs to be temperature-controlled. The temperature acquisition module and the temperature control module perform real-time temperature acquisition and control on it to realize the stable output of the laser wavelength and power;

[0050] The temperature acquisition module selects a 10K thermistor integrated inside the laser tube;

[0051] The temperature control module is composed of a driver chip and a thermoelectric cooler TEC. The driver chip selects DRV594 produced by Texas Instruments (TI), amplifies the power of the PWM signal output by the MCU module to meet the power requirement for driving the thermoelectric cooler TEC, so as to realize the temperature control of the VCSEL laser tube;

[0052] The microprocessor module is the core module of the temperature control system. The digital signal processor TMS320F28004 produced by Texas Instruments (TI) is selected as the microcontroller for this optimized design.

[0053] Compared with the prior art, the beneficial effects of this solution are as follows: The BP neural network has strong self-learning ability, adaptability, robustness, and fault tolerance. These advantages enable it to perform well in the temperature control process of the CPT atomic clock laser tube. The present invention optimizes the temperature control system of the CPT atomic clock laser tube: The traditional PID control algorithm in the system is replaced by the BP neural network PID algorithm. The introduced BP neural network algorithm can calculate the current required PID parameters in real time, and can learn and train more appropriate PID parameters under different experimental environments, realizing the self-tuning of PID parameters during the temperature control process. At the same time, the obtained PID parameters are used in PID control to dynamically change the control quantity and its change speed, so that the real-time temperature control effect of the CPT atomic clock laser tube reaches the best. Description of the Drawings

[0054] Figure 1 is the flowchart in the embodiment of the present invention;

[0055] Figure 2 is the structure diagram of BP-PID temperature control in the embodiment of the present invention;

[0056] Figure 3 is the framework diagram of the temperature control system in the embodiment of the present invention;

[0057] Figure 4 is the actual online control result diagram of the temperature control system in the embodiment of the present invention, where Figure 4 (a) is the temperature change curve before stabilization under the two temperature control methods of traditional PID and BP-PID in the embodiment of the present invention, Figure 4 (b) is the temperature change curve after stabilization under the two temperature control methods of traditional PID and BP-PID in the embodiment of the present invention. Detailed Embodiment

[0058] In order to enable those skilled in the art to better understand the solution of the present invention, the technical solution of the present invention will be further described in detail below in conjunction with the embodiments and drawings of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative work shall fall within the protection scope of the present invention.

[0059] It should be noted that, without conflict, the embodiments in the present invention and the features in the embodiments can be combined with each other. The present invention will be described in detail below with reference to the embodiments.

[0060] Embodiment:

[0061] A method for improving the performance of a CPT atomic clock based on machine learning, as Figure 3 shown, includes the following steps:

[0062] Establish a BP neural network model applicable to the temperature control of the laser tube of a CPT atomic clock;

[0063] Take the output of this model as the input of a PID controller to establish a BP-PID control model;

[0064] Obtain the training parameters of the BP neural network model and optimize and train the BP-PID control model;

[0065] Use the trained BP-PID control model to perform real-time control on the temperature of the laser tube of the CPT atomic clock.

[0066] Referring to Figure 1 , specifically, establishing a BP neural network model applicable to the temperature control of the laser tube of a CPT atomic clock is as follows:

[0067] During the calculation process, the BP neural network consists of a forward propagation process and a backward calculation process. In the forward propagation process, calculations are stacked layer by layer from the input layer to the output layer. If the expected result is not obtained in the output layer, it will turn to the backward propagation to adjust all the weights between each node in the network. The method of weight adjustment uses the gradient descent method to minimize the error amount.

[0068] First, establish a BP neural network model including an input layer, a hidden layer, and an output layer, and determine the number of nodes in the input layer, hidden layer, and output layer. The input layer has 3 neurons X i (i = 1, 2, 3), which are respectively the change amount of error Δe(k), error e(k), and the change trend of error Δe(k) - Δe(k - 1); the output layer has 3 neurons Y l (l = 1, 2, 3), which respectively correspond to the proportional coefficient K p , integral coefficient K i and differential coefficient K d ; calculate the number of hidden layer neurons according to the formula (m + n) 1 / 2 + a, where m represents the number of neurons in the input layer, here m = 3; n represents the number of neurons in the output layer, here n = 3; a is a constant between 1 and 10, and here the number of hidden layer neurons is set to 8; net1j is the parameter of the hidden layer neuron Hj, and net2j is the parameter of the output layer neuron Yl;

[0069] And give the initial value \(w\) of the weight coefficients of each layer 1ij and \(w\) 2jl 、Select the learning rate \(\eta\) and the inertia coefficient \(\gamma\), and calculate the neurons of the hidden layer and the output layer;

[0070] Input layer neurons \(X\) i (\(i = 1, 2, 3\)) are calculated as follows:

[0071] \(X1=\Delta e(k)\)

[0072] \(X2 = e(k)\)

[0073] \(X3=\Delta e(k)-\Delta e(k - 1)\)

[0074] Hidden layer neurons \(H\) j (\(j = 1, 2, \cdots 8\)) are calculated as follows:

[0075]

[0076] Where

[0077]

[0078] \(w\) 1ij is the connection weight between the input layer and the hidden layer;

[0079] Output layer neurons \(Y\) l (\(l = 1, 2, 3\)) are calculated as follows:

[0080]

[0081] Where

[0082]

[0083] \(w\) 2jl is the connection weight between the hidden layer and the output layer.

[0084] Refer to Figure 2 , and use the output of this model as the input of the PID controller to establish a BP - PID control model, specifically including:

[0085] Use the output value \(Y\) l (\(l = 1, 2, 3\)) of this model as the input of the PID controller parameters and act on the controlled object,

[0086] \(K\) p \(= Y1\)

[0087] \(K\) i \(= Y2\)

[0088] \(K\) d \(= Y3\)

[0089] Among them, the proportionality coefficient K p , the integral coefficient K i , and the differential coefficient K d are the control parameters of the PID controller. The increment generated by the PID operation is:

[0090] u(k) = u(k - 1) + Δu(k - 1)

[0091] Δu(k) = K p [e(k) - e(k - 1)] + K i e(k) + K d [Δe(k) - Δe(k - 1)]

[0092] e(k) = r(k) - y(k)

[0093] Among them, r(k) represents the target value, y(k) represents the output value, u(k) represents the system control quantity acting on the temperature control object. The difference between the target value and the output value gives the error value e(k), the change in error e(k) - e(k - 1), and the change trend of the error Δe(k) - Δe(k - 1).

[0094] Refer to Figure 1 to obtain the training parameters of the BP neural network model and optimize the training of the BP - PID control model, specifically including:

[0095] Set the minimum value of the expected error of the BP neural network. If the expected target value cannot be obtained in the output layer, then transfer to backpropagation. Design the back - calculation process of the neural network. Take the error function E as the following formula, where 1 / 2 is the proportionality coefficient, and r and y are the input quantity and output quantity respectively:

[0096]

[0097] Calculation of the adjustment value of each connection weight between the hidden layer and the output layer:

[0098]

[0099] Calculation of the adjustment value of each connection weight between the input layer and the hidden layer:

[0100]

[0101] Among them is unknown. It is the partial derivative of the change of the output quantity with respect to the increment. The obtained value should be a positive - negative change trend, and the sign function is approximately used to replace it;

[0102] During the reverse calculation process, the error signal is propagated backward along the network. The BP neural network corrects the weight coefficients on the network nodes according to the gradient descent method, and then outputs forward. After repeating this process multiple times of training, when the error value obtained by subtracting the target value from the output value meets the set requirements, the training of the BP-PID control model is completed.

[0103] Refer to Figure 3 , the trained BP-PID control model is used to control the temperature of the CPT atomic clock laser tube in real time, specifically including:

[0104] The entire temperature control system framework includes the PID temperature control object module VCSEL laser tube, the temperature acquisition module thermistor, the microcontroller module single-chip microcomputer, and the thermoelectric cooler TEC.

[0105] The PID temperature control object module refers to the VCSEL laser tube that needs to be temperature-controlled. The temperature acquisition module and the temperature control module perform real-time temperature acquisition and control on it to achieve stable output of the laser wavelength and power.

[0106] For the temperature acquisition module, a 10K thermistor integrated inside the laser tube is selected.

[0107] The temperature control module consists of a driver chip and a thermoelectric cooler TEC. The driver chip selects DRV594 produced by Texas Instruments (TI), which amplifies the power of the PWM signal output by the MCU module to meet the power requirement for driving the thermoelectric cooler TEC, so as to achieve temperature control of the VCSEL laser tube.

[0108] The microprocessor module is the core module of the temperature control system. The digital signal processor TMS320F28004 produced by Texas Instruments (TI) is selected as the microcontroller for this optimized design.

[0109] This temperature control system takes the digital signal processor chip as the core of the microcontroller, combines the BP-PID control principle, calculates the required temperature control quantity in real time through the temperature error value, and transmits it to the thermoelectric cooler TEC in the form of a pulse width modulation signal, so that the TEC performs heating or cooling operations on the laser tube, realizing temperature control of the CPT atomic clock laser tube.

[0110] Refer to Figure 4 , and the temperature control effects of using the traditional PID and BP-PID methods for the VCSEL laser tube are compared.

[0111] The temperature change curves from the start to the stabilization of the internal temperature control are measured respectively using the traditional PID and BP-PID, as Figure 4(As shown in (a)), the blue line represents the change in the temperature of the laser tube when using traditional PID temperature control, and the pink line represents the change in the temperature of the laser tube when using BP-PID temperature control. It can be seen from the figure that the overshoot of BP-PID temperature control is 554.06 times smaller than that of traditional PID, and the time to reach stability is 13.5 s less.

[0112] After reaching stability, the temperature change curves of the laser tube controlled by traditional PID and BP-PID were respectively recorded. As Figure 4 (shown in (b)), the black line represents the change in the temperature of the laser tube when using traditional PID temperature control, and the red line represents the change in the temperature of the laser tube when using BP-PID temperature control. It can be seen from the figure that within 900 seconds, the temperature jitter range when using ordinary PID temperature control is 0.0008 °C, and the temperature jitter range when using BP-PID temperature control is 0.0003 °C, which is reduced by 2.66 times.

[0113] The above specific embodiments are only explanations of the present invention, and they are not limitations of the present invention. Those skilled in the art can make modifications to this embodiment without creative contributions according to needs after reading this specification, but as long as they are within the scope of the claims of the present invention, they are protected by the patent law.

Claims

1. A method for improving the performance of CPT atomic clocks based on machine learning, characterized in that: The following steps are involved: Step S1, establishing a BP neural network model suitable for temperature control of CPT atomic clock laser tube; Step S2, using the output of the model as the input of the PID controller to establish a BP-PID control model; Step S3, obtaining the training parameters of the BP neural network model, and performing optimization training on the BP-PID control model; Step S4: Use the trained BP-PID control model to control the temperature of the CPT atomic clock laser tube in real time.

2. The method for improving the performance of a CPT atomic clock based on machine learning as claimed in claim 1, characterized in that: The sub-steps of step S1 are: Step S11, establishing a BP neural network model including an input layer, a hidden layer, and an output layer; Step S12: Determine the number of nodes in the input layer, hidden layer, and output layer. The input layer has 3 neurons X. i (i=1,2,3), which are the error change Δe(k), error e(k) and error change trend Δe(k)-Δe(k-1); the output layer has 3 neurons Y l (l=1,2,3), corresponding to the proportional coefficient K respectively p , integral coefficient K i and the differential coefficient K d ; According to the formula (m+n) 1 / 2 +a calculates the number of neurons in the hidden layer, where m represents the number of neurons in the input layer, where m = 3; n represents the number of neurons in the output layer, where n=3; a is a constant between 1 and 10, where the number of neurons in the hidden layer is set to 8; net1j is the parameter of the hidden layer neuron Hj, and net2j is the parameter of the output layer neuron Yl; Step S13: Give the initial value w of each layer weight coefficient 1ij and w 2jl , select the learning rate η, the inertia coefficient γ, and calculate the neurons of the hidden layer and the output layer; The input layer neuron X i The calculation formula for (i=1,2,3) is as follows: X1=Δe(k) X2=e(k) X3=Δe(k)-Δe(k-1) Hidden layer neurons H j The calculation formula for (j=1,2,…8) is as follows: in w 1ij is the connection weight between the input layer and the hidden layer; Output layer neuron Y l The calculation formula for (l=1,2,3) is as follows: in w 2jl is the connection weight between the hidden layer and the output layer.

3. The method for improving the performance of a CPT atomic clock based on machine learning as claimed in claim 1, wherein the output of the model is used as the input of a PID controller in step S2 to establish a BP-PID control model, wherein: The output value Y of the model l (l=1,2,3) is used as the input of PID controller parameters and acts on the controlled object. K p =Y1 K i =Y2 K d =Y3 The proportionality coefficient K p , integral coefficient K i , differential coefficient K d is the control parameter of the PID controller. The increment generated by the PID operation is: u(k)=u(k-1)+Δu(k-1) Δu(k)=K p [e(k)-e(k-1)]+K i e(k)+K d [Δe(k)-Δe(k-1)] e(k)=r(k)-y(k) Among them, r(k) represents the target value, y(k) represents the output value, u(k) represents the system control quantity acting on the temperature control object, the difference between the target value and the output value is the error value e(k), the error change is e(k)-e(k-1), and the error change trend is Δe(k)-Δe(k-1). At this point, the BP-PID control model is established.

4. The method for improving the performance of a CPT atomic clock based on machine learning as claimed in claim 3, wherein the training parameters of a BP neural network model are obtained, and the BP-PID control model is optimized and trained, wherein: Set the minimum expected error value of the BP neural network. When the output layer cannot obtain the expected target value, turn to back propagation. Design the reverse calculation process of the neural network, take the error function E as the following formula, 1 / 2 as the proportional coefficient, r and y as the input and output respectively: The adjustment value of each connection weight between the hidden layer and the output layer is calculated as: Calculation of the adjustment value of each connection weight between the input layer and the hidden layer: in If it is unknown, find the partial derivative of the output with respect to the increment, and the result should be a positive or negative trend, so use the sign function to approximate it. During the reverse calculation process, the error signal is transmitted backward along the network. The BP neural network corrects the weight coefficients on the network nodes according to the gradient descent method, and then outputs forward. After multiple trainings, when the error value obtained by subtracting the target value from the output value meets the set requirements, the BP-PID control model training is completed.

5. A temperature control system based on machine learning to improve the performance of CPT atomic clocks, using a trained BP-PID control model to control the temperature of the CPT atomic clock laser tube in real time, characterized by: With the digital signal processor chip as the core of the microcontroller and combined with the BP-PID control principle, the required temperature control amount is calculated in real time through the temperature error value, and it is transmitted to the semiconductor cooler TEC in the form of a pulse width modulation signal, so that TEC heats or cools the laser tube, thereby realizing the temperature control of the CPT atomic clock laser tube.

6. A temperature control system for improving the performance of a CPT atomic clock based on machine learning as claimed in claim 5, characterized in that: Including PID temperature control object module VCSEL laser tube, temperature acquisition module thermistor, microcontroller module single chip microcomputer, semiconductor cooler TEC: The PID temperature control object module refers to the VCSEL laser tube that needs to be temperature controlled. The temperature acquisition module and the temperature control module perform real-time temperature acquisition and control on it to achieve stable output of laser wavelength and power. The temperature acquisition module uses the 10K thermistor integrated inside the laser tube; The temperature control module consists of a driver chip and a semiconductor cooler TEC. The driver chip is DRV594 produced by Texas Instruments. It amplifies the PWM signal power output by the MCU module to meet the power requirements of driving the semiconductor cooler TEC, so as to achieve temperature control of the VCSEL laser tube. The microprocessor module is the core module of the temperature control system. The digital signal processor TMS320F28004 produced by Texas Instruments is selected as the microcontroller for this optimized design.

Images