DC-DC converter PWM / PFM self-detection switching method based on neural network and control circuit

Abstract

The invention discloses a PWM / PFM (Pulse Width Modulation / Pulse Frequency Modulation) self-detection switching method of a DC-DC converter based on a neural network and a control circuit, the working state quantity of the DC-DC converter under PWM and PFM control is collected, a neural network is utilized to train and fit a plurality of working conditions to generate a neural network controller, the working state of the converter is monitored in real time during actual control, and the PWM / PFM self-detection switching method of the DC-DC converter is realized. Therefore, the PWM duty ratio and the PFM frequency can be automatically switched and adjusted.

Description

Technical Field

[0001] The invention relates to the technical field of power electronics, and in particular to a DC-DC converter PWM / PFM self-detection switching method and a control circuit based on a neural network. Background Art

[0002] Since DC grid connection does not require consideration of reactive power, voltage phase, etc., research on DC grids has attracted the attention of many scholars.

[0003] Since the voltage level of renewable energy is usually low, it is necessary to increase its voltage before connecting to the grid, and DC-DC converters have received extensive attention and research. PWM and PFM are commonly used control methods for DC converters. Under heavy load conditions, PWM control has the characteristics of high voltage accuracy and good stability. Under light load conditions, PFM control does not have too large voltage fluctuations, and can optimize efficiency under light load by changing the frequency. Therefore, it is urgent to propose a PWM / PFM automatic switching method for DC-DC converters to achieve self-detection switching between heavy load and light load, and improve overall efficiency. Summary of the invention

[0004] In view of the above technical problems, the present invention discloses a DC-DC converter PWM / PFM self-detection switching method and control circuit based on neural network. By collecting the working state quantity of the DC-DC converter under PWM and PFM control, a neural network controller is generated by using neural network training to fit various working conditions. During actual control, the working state of the converter is monitored in real time, thereby realizing automatic switching and adjustment of PWM duty cycle and PFM frequency.

[0005] To achieve the above object, the present invention provides a neural network-based DC-DC converter PWM / PFM self-detection switching control circuit in a first aspect, comprising: a sampling unit, a driving unit and a neural network controller;

[0006] Preferably, the sampling unit includes: front-end sampling and back-end sampling;

[0007] Preferably, the front-end sampling unit includes: a first Hall voltage sensor, a first Hall current sensor and a resistor R 1 ;

[0008] Preferably, the back-end sampling unit includes: a second Hall voltage sensor, a second Hall current sensor and a resistor R 2 ;

[0009] Preferably, the control quantity of the driving unit includes: duty cycle d and frequency f;

[0010] Preferably, the neural network controller comprises: an input layer, a hidden layer and an output layer, wherein the input layer input is [V in 、i in 、V out 、i out ], the hidden layer can be composed of one or more layers. Here, there are 5 hidden layers and the output layer is [d, f].

[0011] Preferably, the input terminal V in + is connected to the first end of the first Hall current sensor, the second end of the first Hall current sensor is connected to the positive electrode of the DC-DC converter input end, and the voltage sampling end of the first Hall voltage sensor is connected between the second end of the first Hall current sensor and the positive electrode of the DC-DC converter input end;

[0012] Preferably, the input terminal V in - connected to the negative pole of the DC-DC converter input terminal;

[0013] Preferably, the positive electrode of the output end of the DC-DC converter is connected to the first end of the second Hall current sensor, and the second end of the second Hall current sensor is connected to the output end V out + connection, the second Hall voltage sensor is connected to the second end of the second Hall current sensor and the output terminal V out +Between;

[0014] Preferably, the output terminal negative electrode of the DC-DC converter is connected to the output terminal V out -connect;

[0015] Preferably, the input voltage V in 、Input current i in , output voltage V out And the output current i out After sampling, it is used as the input layer of the neural network controller;

[0016] Preferably, the output layer of the neural network controller is connected to the input end of the driving unit, and outputs a duty cycle d and a frequency f;

[0017] Preferably, the output end of the driving unit is connected to the driving signal pole of the switch tube of the DC-DC converter.

[0018] Preferably, in the DC-DC converter PWM / PFM self-detection switching circuit based on a neural network, the sampling unit is a symmetrical structure, sampling the front-end and rear-end voltages and currents.

[0019] Preferably, the constraints of the control parameters of the neural network include:

[0020] 0≤d(g)≤1

[0021] 0≤f(g)≤f max

[0022] 0≤i in (g)≤I inmax

[0023] 0≤V in (g)≤V inmax

[0024] 0≤i out (g)≤I outmax

[0025] 0≤V out (g)≤V outmax

[0026] Among them, d(g), f(g), i in (g), V in (g) i out (g) and V out (g) are duty cycle, frequency, input current, input voltage, output current, and output voltage, respectively.

[0027] Preferably, the DC-DC converter is characterized in that the DC-DC converter is a series / parallel structure of a single switch tube topology or a synchronous dual switch tube topology, and its drive signal input end is connected to the output end of the drive unit.

[0028] In a second aspect, the present invention provides a DC-DC converter PWM / PFM self-detection switching method based on a neural network, the self-detection switching method comprising:

[0029] S1, by collecting the PWM and PFM closed-loop control data of the converter from startup to stable state [V in 、i in 、V out 、i out ], construct a training set as the input sample of the neural network, and input the PWM duty cycle and PFM switching frequency [d, f] of the corresponding cycle as labels into the neural network for fitting training. For example, N cycles of [V in 、i in 、V out 、i out ] Input neural network;

[0030] S2, data preprocessing, N cycles of [V in 、i in 、V out 、i out ]Before sending it into the neural network, the data is normalized and expressed as:

[0031]

[0032] Among them, X max is the maximum value in the input data, X min is the minimum value in the input data;

[0033] S3. Set the neural network control structure and activation function. The input layer consists of 4 neurons, and the real-time state of the DC-DC converter is input. The hidden layer can be composed of multiple neurons, and the nonlinear fitting ability of the neural network is enhanced by the Swish activation function. The output layer consists of two neurons. The activation functions of the hidden layer and the output layer are both Swish functions, which have smoother gradient flow and better training effect. The expression of the Swish activation function is:

[0034]

[0035] Among them, x is the result of weighted summation of the hidden layer or output layer;

[0036] S4. In each training iteration, the input data is calculated through each layer of the forward propagation neural network to obtain its predicted value. The error function can be defined as:

[0037]

[0038] Among them, d(g+1 / g) is the optimal output duty cycle, f(g+1 / g) is the optimal output frequency, F(g+1 / g) and D(g+1 / g) are the output values ​​of the neural network, and the weights and biases of the network are continuously adjusted through back propagation through error calculation;

[0039] The network weights and biases are corrected by gradient descent and chain rule. When the error function value is less than the preset threshold or the error change tends to be stable, the training is completed.

[0040] The gradient descent method can be expressed as:

[0041]

[0042] Among them, λ is the learning rate which affects the step size of weight update, w(i+1) and w(i) are the weights at time i+1 and time i, and λ is multiplied by the partial derivative of the error function with respect to the weight;

[0043] The chain rule can be expressed as:

[0044]

[0045] S5. Transplanting the trained model to the FPGA real-time controller and combining their real-time computing and parallel processing capabilities can significantly improve the response speed of the system and realize automatic switching of control modes under different load conditions, thereby improving the overall conversion efficiency and system stability of the converter.

[0046] Technical effect of the invention: The present invention discloses a DC-DC converter PWM / PFM self-detection switching method and control circuit based on a neural network. The corresponding models are generated after training under different working conditions and transplanted into a real-time controller to achieve self-detection switching of PWM and PFM of the DC-DC converter, and the overshoot is small in the steady-state time period. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] In order to more clearly illustrate the technical solution of the present invention, some of the drawings of this application are used to provide a further understanding of the present application and are not used to limit the present application. In the drawings:

[0048] Figure 1 It is a structural schematic diagram of a PWM / PFM automatic switching method of a DC-DC converter based on a neural network according to the present invention;

[0049] Figure 2 is a neural network propagation flow chart of the present invention;

[0050] Figure 3 This is a schematic diagram of the PWM control waveform;

[0051] Figure 4 It is a schematic diagram of the PFM control waveform;

[0052] Figure 5 A control flow chart of a PWM / PFM automatic switching method of a DC-DC converter based on a neural network according to the present invention;

[0053] Figure 6 is the control result of the Buck circuit in the embodiment of the present invention, where V gs It is the driving signal of the switch tube; DETAILED DESCRIPTION

[0054] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. 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.

[0055] like Figure 1As shown, a DC-DC converter PWM / PFM self-detection switching control circuit based on a neural network is characterized in that it includes: a sampling unit, a driving unit and a neural network controller;

[0056] The sampling unit includes: front-end sampling and back-end sampling;

[0057] The front-end sampling unit includes: a first Hall voltage sensor, a first Hall current sensor and a resistor R 1 ;

[0058] The back-end sampling unit includes: a second Hall voltage sensor, a second Hall current sensor and a resistor R 2 ;

[0059] The control quantity of the driving unit includes: duty cycle d and frequency f;

[0060] The neural network controller comprises: an input layer, a hidden layer and an output layer, wherein the input layer input is [V in 、i in 、V out 、i out ], the hidden layer can be composed of one or more layers, and the output layer is [d, f], where V in 、i in 、V out 、i out They are input current, input voltage, output current, and output voltage respectively;

[0061] The input terminal V in + is connected to the first end of the first Hall current sensor, the second end of the first Hall current sensor is connected to the positive electrode of the DC-DC converter input end, and the voltage sampling end of the first Hall voltage sensor is connected between the second end of the first Hall current sensor and the positive electrode of the DC-DC converter input end;

[0062] The input terminal V in - connected to the negative pole of the DC-DC converter input terminal;

[0063] The positive electrode of the output end of the DC-DC converter is connected to the first end of the second Hall current sensor, and the second end of the second Hall current sensor is connected to the output end V out + connection, the second Hall voltage sensor is connected to the second end of the second Hall current sensor and the output terminal V out +Between;

[0064] The negative output terminal of the DC-DC converter is connected to the output terminal V out -connect;

[0065] The input voltage V in 、Input current iin , output voltage V out And the output current i out After sampling, it is used as the input layer of the neural network controller;

[0066] The output layer of the neural network controller is connected to the input end of the driving unit, and outputs a duty cycle d and a frequency f;

[0067] The output end of the driving unit is connected to the driving signal pole of the switch tube of the DC-DC converter.

[0068] Preferably, the control circuit is characterized in that the sampling unit is a symmetrical structure, sampling the front-end and rear-end voltages and currents.

[0069] Preferably, the control circuit is characterized in that the constraints of the control parameters of the neural network include:

[0070] 0≤d(g)≤1

[0071] 0≤f(g)≤f max

[0072] 0≤i in (g)≤I inmax

[0073] 0≤V in (g)≤V inmax

[0074] 0≤i out (g)≤I outmax

[0075] 0≤V out (g)≤V outmax

[0076] Among them, d(g), f(g), i in (g), V in (g) i out (g) and V out (g) are duty cycle, frequency, input current, input voltage, output current, and output voltage, respectively.

[0077] Preferably, the DC-DC converter is characterized in that the DC-DC converter is a series / parallel structure of a single switch tube topology or a synchronous dual switch tube topology, and its drive signal input end is connected to the output end of the drive unit.

[0078] like Figure 2 As shown, a neural network propagation flow chart of a DC-DC converter PWM / PFM self-detection switching method based on a neural network, the self-detection switching method comprises:

[0079] S1, by collecting the PWM and PFM closed-loop control data of the converter from startup to stable state [V in 、i in 、V out 、i out ], construct a training set as the input sample of the neural network, and input the PWM duty cycle and PFM switching frequency [d, f] of the corresponding cycle as labels into the neural network for fitting training. For example, N cycles of [V in 、i in 、V out 、i out ] Input neural network;

[0080] S2, data preprocessing, N cycles of [V in 、i in 、V out 、i out ]Before sending it into the neural network, the data is normalized and expressed as:

[0081]

[0082] Among them, X max is the maximum value in the input data, X min is the minimum value in the input data;

[0083] S3. Set the neural network control structure and activation function. The input layer consists of 4 neurons, and the real-time state of the DC-DC converter is input. The hidden layer can be composed of multiple neurons, and the nonlinear fitting ability of the neural network is enhanced by the Swish activation function. The output layer consists of two neurons. The activation functions of the hidden layer and the output layer are both Swish functions, which have smoother gradient flow and better training effect. The expression of the Swish activation function is:

[0084]

[0085] Among them, x is the result of weighted summation of the hidden layer or output layer;

[0086] S4. In each training iteration, the input data is calculated through each layer of the forward propagation neural network to obtain its predicted value. The error function can be defined as:

[0087]

[0088] Among them, d(g+1 / g) is the optimal output duty cycle, f(g+1 / g) is the optimal output frequency, F(g+1 / g) and D(g+1 / g) are the output values ​​of the neural network, and the weights and biases of the network are continuously adjusted through back propagation through error calculation;

[0089] The network weights and biases are corrected by gradient descent and chain rule. When the error function value is less than the preset threshold or the error change tends to be stable, the training is completed.

[0090] The gradient descent method can be expressed as:

[0091]

[0092] Among them, λ is the learning rate which affects the step size of weight update, w(i+1) and w(i) are the weights at time i+1 and time i, and λ is multiplied by the partial derivative of the error function with respect to the weight;

[0093] The chain rule can be expressed as:

[0094]

[0095] S5. Transplanting the trained model to the FPGA real-time controller and combining their real-time computing and parallel processing capabilities can significantly improve the response speed of the system and realize automatic switching of control modes under different load conditions, thereby improving the overall conversion efficiency and system stability of the converter.

[0096] like Figure 3 As shown, Figure 3 is the output voltage ripple V under PWM control ripple Schematic diagram.

[0097] like Figure 4 As shown, Figure 4 is the output voltage ripple V under PFM control ripple Schematic diagram.

[0098] pass Figure 3 and Figure 4 By comparing the switching frequency, we can see that the on-time of the switch tube is adjusted by changing the switching frequency. Figure 4 As shown in the figure, the change of switching frequency will affect the fluctuation of output voltage. PFM control can reduce the switching loss of the switch tube by reducing the switching frequency, while PWM control has better stability. Figure 3 As shown in the figure, the output voltage has high accuracy and small ripple. The fixed switching frequency causes the switching device to generate large power loss, which reduces efficiency. Therefore, the advantages of both control can be combined. PWM control can be used to meet the demand for output voltage ripple under heavy load, and PFM control can be used to meet the demand for efficiency under light load. Therefore, in order to achieve higher stability and conversion efficiency under different load conditions, a PWM / PFM self-detection switching method is required.

[0099] like Figure 5 As shown, Figure 5The present invention is a control flow chart of a PWM / PFM automatic switching method of a DC-DC converter based on a neural network, comprising:

[0100] S1, collecting different working state quantities of PWM when the DC-DC converter is overloaded and PFM when it is lightly loaded, collecting and saving the working state quantities of the converter's PWM and PFM dynamic processes into a data area to form a training sample, as shown in Table 1;

[0101] Table 1 Neural network training samples

[0102]

[0103] S2, input state quantity [V in ,i in ,V out ,i out ], the output label [d, f] is normalized and expressed as:

[0104]

[0105] S3, the input layer consists of 4 neurons, which input the real-time state of the DC-DC converter. The hidden layer can be composed of multiple neurons, and the nonlinear fitting ability of the neural network is enhanced by the Swish activation function. The output layer consists of two neurons. The activation functions of the hidden layer and the output layer are both Swish functions, which have smoother gradient flow and better training effect. The output layer is [d, f];

[0106] The Swish activation function expression is:

[0107]

[0108] Among them, x is the result of weighted summation of the hidden layer or output layer;

[0109] S4. In each training iteration, the input data is calculated through each layer of the forward propagation neural network to obtain its predicted value. The error function can be defined as:

[0110]

[0111] Among them, d(g+1 / g) is the optimal output duty cycle, f(g+1 / g) is the optimal output frequency, F(g+1 / g) and D(g+1 / g) are the output values ​​of the neural network, and the weights and biases of the network are continuously adjusted through back propagation through error calculation;

[0112] The network weights and biases are corrected by gradient descent and chain rule. When the error function value is less than the preset threshold or the error change tends to be stable, the training is completed.

[0113] The gradient descent method can be expressed as:

[0114]

[0115] Among them, λ is the learning rate which affects the step size of weight update, w(i+1) and w(i) are the weights at time i+1 and time i, and λ is multiplied by the partial derivative of the error function with respect to the weight;

[0116] The chain rule can be expressed as:

[0117]

[0118] S5. Transplant the trained model to the real-time controller for real-time sampling control.

[0119] like Figure 6 As shown, Figure 6 The circuit diagram of Buck in the embodiment of the present invention is shown in FIG. 1 . This embodiment adopts Buck as the DC-DC converter topology. The V in =48V, V out =24V, inductance L = 330μH, capacitance C = 10μF, P o =48W; collect the working state quantity of the converter under PWM and PFM control, input the neural network for training, and transplant it into the real-time controller to perform real-time control under actual variable load conditions. When the load current is detected to be 0.5A, the PFM control mode is adopted; when the load current is 2A, it switches to the PWM control mode. It can be seen that after the load current changes, the output voltage ripple and control frequency change, and the sudden fluctuation of the output voltage and output current when changing is small.

[0120] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art to which this application belongs. The terms used herein in the specification of this application are only for the purpose of describing specific embodiments and are not intended to limit this application.

[0121] It can be understood that the terms "first", "second", etc. used in the present application can be used in this article to describe various elements, but these elements are not limited by these terms. These terms are only used to distinguish the first element from another element.

[0122] It should be noted that when an element is considered to be "connected" to another element, it can be directly connected to the other element or connected to the other element through an intermediate element. In addition, the "connection" in the following embodiments should be understood as "electrical connection", "communication connection", etc. if there is transmission of electrical signals or data between the connected objects.

Claims

1. A DC-DC converter PWM / PFM self-detection switching control circuit based on neural network, characterized in that: include: Sampling unit, driving unit and neural network controller; The sampling unit includes: front-end sampling and back-end sampling; The front-end sampling unit includes: a first Hall voltage sensor, a first Hall current sensor and a resistor R1; The back-end sampling unit includes: a second Hall voltage sensor, a second Hall current sensor and a resistor R2; The control quantity of the driving unit includes: duty cycle d and frequency f; The neural network controller comprises: an input layer, a hidden layer and an output layer, wherein the input layer input is [V in 、i in 、V out 、i out ], the hidden layer can be composed of one or more layers, and the output layer is [d, f], where V in 、i in 、V out 、i out They are input current, input voltage, output current, and output voltage respectively; The input terminal V in + is connected to the first end of the first Hall current sensor, the second end of the first Hall current sensor is connected to the positive electrode of the DC-DC converter input end, and the voltage sampling end of the first Hall voltage sensor is connected between the second end of the first Hall current sensor and the positive electrode of the DC-DC converter input end; The input terminal V in - connected to the negative pole of the DC-DC converter input terminal; The positive electrode of the output end of the DC-DC converter is connected to the first end of the second Hall current sensor, and the second end of the second Hall current sensor is connected to the output end V out + connection, the second Hall voltage sensor is connected to the second end of the second Hall current sensor and the output terminal V out +Between; The negative output terminal of the DC-DC converter is connected to the output terminal V out -connect; The input voltage V in 、Input current i in , output voltage V out And the output current i out After sampling, it is used as the input layer of the neural network controller; The output layer of the neural network controller is connected to the input end of the driving unit, and outputs a duty cycle d and a frequency f; The output end of the driving unit is connected to the driving signal pole of the switch tube of the DC-DC converter.

2. The control circuit according to claim 1, characterized in that: The sampling unit has a symmetrical structure and samples the front-end and rear-end voltages and currents.

3. The control circuit according to claim 1, characterized in that: The constraints on the control parameters of the neural network include: 0≤d(g)≤1 0≤f(g)≤f max 0≤i in (g)≤I inmax 0≤V in (g)≤V inmax 0≤i out (g)≤I outmax 0≤V out (g)≤V outmax Among them, d(g), f(g), i in (g), V in (g) i out (g) and V out (g) are duty cycle, frequency, input current, input voltage, output current, output voltage, respectively, f max ,I inmax 、V inmax ,I outmax 、V outmax As the maximum value of its constraint condition, it is set according to the rated operating conditions.

4. The control circuit according to claim 1, characterized in that: The DC-DC converter is a series / parallel structure of a single switch tube topology or a synchronous dual switch tube topology, and its driving signal input end is connected to the output end of the driving unit.

5. A DC-DC converter PWM / PFM self-detection switching method based on a neural network, characterized in that: include: S1, by collecting the PWM and PFM closed-loop control data of the converter from startup to stable state [V in 、i in 、V out 、i out ], construct a training set as the input sample of the neural network, and input the PWM duty cycle and PFM switching frequency [d, f] of the corresponding cycle as labels into the neural network for fitting training. For example, N cycles of [V in 、i in 、V out 、i out ] Input neural network; S2, data preprocessing, N cycles of [V in 、i in 、V out 、i out ]Before sending it into the neural network, the data is normalized and expressed as: Among them, X max is the maximum value in the input data, X min is the minimum value in the input data; S3. Set the neural network control structure and activation function. The input layer consists of 4 neurons, and the real-time state of the DC-DC converter is input. The hidden layer can be composed of multiple neurons, and the nonlinear fitting ability of the neural network is enhanced by the Swish activation function. The output layer consists of two neurons. The activation functions of the hidden layer and the output layer are both Swish functions, which have smoother gradient flow and better training effect. The Swish activation function expression is: Among them, x is the result of weighted summation of the hidden layer or output layer; S4. In each training iteration, the input data is calculated through each layer of the forward propagation neural network to obtain its predicted value. The error function can be defined as: Among them, d(g+1 / g) is the optimal output duty cycle, f(g+1 / g) is the optimal output frequency, F(g+1 / g) and D(g+1 / g) are the output values ​​of the neural network, and the weights and biases of the network are continuously adjusted through back propagation through error calculation; The network weights and biases are corrected by gradient descent and chain rule. When the error function value is less than the preset threshold or the error change tends to be stable, the training is completed. The gradient descent method can be expressed as: Among them, λ is the learning rate which affects the step size of weight update, w(i+1) and w(i) are the weights at time i+1 and time i, and λ is multiplied by the partial derivative of the error function with respect to the weight; The chain rule can be expressed as: S5. Transplanting the trained model to the FPGA real-time controller and combining their real-time computing and parallel processing capabilities can significantly improve the response speed of the system and realize automatic switching of control modes under different load conditions, thereby improving the overall conversion efficiency and system stability of the converter.

Images