CN114936632A - Hardware resource dynamic multiplexing neural network controller for digital power supply - Google Patents

Abstract

The invention discloses a hardware resource dynamic multiplexing neural network controller for a digital power supply, which belongs to the field of power electronics and comprises a neural network controller and a control module, wherein the neural network controller is used for generating a digital duty ratio; the neural network controller simultaneously realizes a coarse tuning controller and a fine tuning controller, and the coarse tuning controller is used for adjusting the output voltage to a reference voltage as soon as possible after the digital power supply is disturbed so as to improve the transient response performance; the fine tuning controller is mainly used for eliminating steady-state errors and stabilizing the output voltage at a reference voltage. The invention can reduce the structure and realize the same control performance, and save the hardware resource needed by the hidden layer node.

Description

A Neural Network Controller for Dynamic Multiplexing of Hardware Resources for Digital Power

technical field

The invention relates to the technical field of power electronics, in particular to a hardware resource dynamic multiplexing neural network controller for digital power supply.

Background technique

The structure of the digitally controlled DC-DC switching converter (hereinafter referred to as digital power supply) is shown in FIG. 1 . The analog output voltage V _out (t) at the load end is converted into a digital output quantity V _out [k] by the ADC, and then the error signal e[k] between V _out [k] and the reference voltage V _ref is sent to the digital voltage compensator . In a digital voltage compensator (such as a neural network controller), a specific control algorithm is used to calculate the digital duty cycle signal d[k], and then the digital duty cycle signal d[k] is calculated by DPWM (Digital Pulse Width Modulation). The duty cycle signal d[k] is converted into an analog duty cycle signal d(t), and finally the _gatedriver drives the power stage switches SP and _SN to be turned on or off to adjust the output voltage V _out (t) to make it stable at reference voltage V _ref .

For digital power supplies, neural network controllers have the advantages of high adaptability, strong robustness, and good transient performance, but neural network controllers usually consume a lot of hardware resources. The traditional method of neural network controller to reduce hardware resource consumption is to "prune" the neural network, that is, to zero the weight of the node with low sensitivity (less influence on the output result) or delete the node directly. Although "pruning" can effectively reduce the scale of the network, the amount of calculation and the consumption of hardware resources, the problem of reducing the control accuracy of digital power supplies will increase the steady-state error, and the selection of "pruning" nodes requires After repeated iterative verification, the workload is large, and the "pruned" network will reduce its adaptability.

The document "Dynamic Channel Pruning by Conditional Accuracy Change for Deep Neural Networks", IEEE Transactions on Neural Network and LearningSystems, Vol.32, No.2, February 2021, pp:799-813 proposes a dynamic network channel pruning method, which is Build dynamic evaluation criteria to gradually remove insensitive network channels during training iterations. The method proposed in this paper effectively reduces the size of the network and the amount of computation while achieving higher accuracy. However, the reduction of hardware resources and the acceleration of computing by this method are still based on the principle of network "pruning", a dynamically pruned network structure. Due to missing some nodes or channels, it still causes some loss of accuracy, and more importantly, leads to reduced adaptiveness.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide a neural network controller for dynamic multiplexing of hardware resources of a digital power supply, so as to solve the problem of large consumption of hardware resources of the current neural network controller.

In order to solve the above technical problems, the present invention provides a hardware resource dynamic multiplexing neural network controller for digital power supply, including a neural network controller and a control module, the neural network controller is used to generate a digital duty cycle, wherein The control module includes a look-up table and a dynamic multiplexing controller, the look-up table is used to store the weights and biases of the neural network controller, and the dynamic multiplexing controller dynamically controls the weights of the neural network controller and bias switching, and disable some hidden nodes;

The neural network controller implements a coarse adjustment controller and a fine adjustment controller at the same time, and the coarse adjustment controller is used to adjust the output voltage to the reference voltage as soon as possible after the digital power supply is disturbed, so as to improve the transient response performance; the fine adjustment The controller is mainly used to eliminate steady-state errors and stabilize the output voltage at the reference voltage.

Optionally, the neural network controller is obtained through offline data training, including the following steps:

An error band is established on both sides of the reference voltage, and the training data of the neural network controller is grouped according to the size of the error value. The data with the error within the error band is the training data of the fine adjustment controller, otherwise, it is the coarse adjustment control. training data of the machine;

The two groups of training data are trained by the neural network training method to train two neural networks with the same input and output but different structures, respectively, to obtain the weights and biases of the fine-tuned controller and the weights and biases of the coarse-tuned controller.

Optionally, in the training data of the fine adjustment controller, the output voltage is close to an ideal value; in the training data of the coarse adjustment controller, there are various differences between the output voltage and the ideal value, including changes in the output voltage under different disturbances. .

Optionally, the range of the error band is within ±10% of the reference voltage.

Optionally, the number of hidden layers or hidden layer nodes of the coarse adjustment controller is more, the number of hidden layers or nodes of the fine adjustment controller is relatively small, and the hidden layer of the coarse adjustment If The coarse tuning controller and the fine tuning controller are neural networks with the same number of hidden layers. The coarse tuning controller can get the same structure as the fine tuning controller by disabling some hidden layer nodes. By switching to fine-tune the weights and biases of the controller, the coarse-tune controller can be used to implement the function of the fine-tune controller.

Optionally, based on the principle of time division multiplexing, the structure of the neural network controller is selected to be consistent with the coarse adjustment controller, and some hidden layer nodes are dynamically disabled according to the position where the error is located in the error band, and the weights and biases are switched. , realize the functions of the two controllers, and obtain a neural network controller with dynamic multiplexing of hardware resources to save the hardware resources required by some hidden layer nodes; when the neural network controller works as a fine-tuning controller, some coarse-tuning controllers are disabled. Adjust the hidden layer node used by the controller alone, and switch to use the weights and biases of the fine-tuned controller. The structure of the neural network controller is equivalent to the structure of the fine-tuned controller; When the controller is working, all hidden layer nodes are activated, and the weights and biases of the coarse adjustment controller are switched to be used. At this time, the structure of the neural network controller is the structure of the coarse adjustment controller.

Optionally, the neural network controller is usually divided into two or more sub-controllers, and then based on the principle of time division multiplexing, one or more neural networks are used to switch weights and biases and dynamically activate and disable partial implicits. The layer-containing node implements the functions of two or more sub-controllers, so as to accelerate the neural network controller or reduce hardware resource consumption.

In the hardware resource dynamic multiplexing neural network controller for digital power supply provided by the present invention, the complete neural network controller is divided into two sub-controllers: a fine adjustment controller and a coarse adjustment controller, and then based on the time division multiplexing Using the idea of using a small-scale neural network to achieve the functions of two sub-controllers by dynamically switching weights and biases, and disabling part of the hidden layer nodes. Therefore, by using the hardware resource dynamic multiplexing neural network controller for digital power supply proposed by the present invention, the trained hidden layer nodes will not be deleted, so that the control performance will not be reduced, and some hidden layer nodes can be saved at the same time. Hardware resources required by a layered node.

Description of drawings

Fig. 1 is a schematic diagram of the structure of a digitally controlled DC-DC switching converter;

2 is a schematic structural diagram of a hardware resource dynamic multiplexing neural network controller for digital power supply provided by the present invention;

3 is a schematic diagram of the working principle of the neural network controller provided by the present invention;

4 is a schematic diagram of the switching sequence of the coarse adjustment controller and the fine adjustment controller;

Figure 5 (a) is a schematic diagram comparing the output voltage response curve of the digital power supply in a steady state using the technology of the present invention and a traditional neural network controller;

Fig. 5 (b) is the output voltage response curve comparison schematic diagram of the digital power supply when the load current jumps ±1A using the technology of the present invention and the traditional neural network controller;

FIG. 5( c ) is a schematic diagram comparing the output voltage response curve of the digital power supply when the input voltage jumps ±1V using the technology of the present invention and the traditional neural network controller.

Detailed ways

The following is a further detailed description of a hardware resource dynamic multiplexing neural network controller for a digital power supply proposed by the present invention with reference to the accompanying drawings and specific embodiments. The advantages and features of the present invention will become apparent from the following description and claims. It should be noted that, the accompanying drawings are all in a very simplified form and in inaccurate scales, and are only used to facilitate and clearly assist the purpose of explaining the embodiments of the present invention.

The present invention provides a neural network controller for dynamic multiplexing of hardware resources for digital power supply. As shown in FIG. 2, it is mainly composed of a neural network controller and a control module in structure, wherein the control module is composed of a look-up table (storage details It consists of the weights and biases of the tuning and coarse tuning controllers) and a dynamic multiplexing controller. The neural network controller is used to calculate the digital duty cycle, and the control module is used to control the dynamic switching of weights and biases, and to dynamically disable or activate the nodes of some hidden layers. The neural network controller functions to realize the coarse adjustment controller and the fine adjustment controller. As shown in Figure 3, the coarse adjustment controller is used to pull the output voltage back to the reference voltage as soon as possible after the digital power supply is disturbed, so as to reduce the stabilization time and overrun. The fine-tuning controller is mainly used to eliminate the steady-state error, maintain the steady state, and make the output voltage stable at the reference voltage. The input signal of the neural network controller is {e _o [k], e _oc [k], I _ind [k], V _in [k]}, where e _o [k] is the voltage error of the current cycle, I _ind [k] is the inductor current in the current cycle, V _in [k] is the input voltage in the current cycle, and e _oc [k]=e[k]-e[k-1] is the voltage error rate of change in the current cycle. The output signal of the neural network controller is the digital duty cycle d _nn [k], and its specific design process is as follows:

(1) Establish error bands and classify the data. First, an error band E _b is established on both sides of the reference voltage, usually within ±10% of the reference voltage is selected as the error band, as shown in Figure 4 . Then, the training data of the neural network controller is classified according to the size of the error value, that is, the data with the error within the error band is the training data of the fine-tuning controller, and vice versa, it is the training data of the coarse-tuning controller. In the training data of the fine-tuning controller, the output voltage is close to the ideal value; in the training data of the coarse-tuning controller, there are various differences between the output voltage and the ideal value, including the output voltage changes under different disturbances.

(2) Obtain coarse adjustment and fine adjustment controllers. The classified data is trained by a specific training method (such as gradient descent method) to train two neural networks with the same input and output but different structures, respectively, to obtain the weights and biases of the fine-tuning controller and the coarse-tuning controller, respectively. And the number of hidden layer nodes of the coarse-tuning controller and the fine-tuning controller are N _c and N _f respectively, as shown in Figure 3, where the traditional neural network controller is obtained by directly training the neural network with training data, and its hidden If the number of nodes with layers is N _n , there must be N _n ≤N _c +N _f and N _f ≤N _c <N _n .

(3) The realization of hardware resource dynamic multiplexing neural network controller. The coarse-tuned controller disables _Nc - _Nf hidden layer nodes, and the fine-tuned controller can be implemented by using the weights and biases of the fine-tuned controller. According to the principle of classification of training data, the fine-tuning controller and the coarse-tuning controller cannot work synchronously, that is, when the fine-tuning controller is working, the coarse-tuning controller is in an idle state; The controller is working. Therefore, based on the idea of time division multiplexing, a neural network (whose structure is consistent with that of the coarse controller) can be implemented by dynamically switching weights and biases and dynamically disabling or activating N _c -N _f hidden layer nodes The functions of the two sub-controllers, specifically, when the controller works as a fine-tuning controller, disable N _c -N _f hidden layer nodes, and the structure of the neural network is equivalent to the structure of the fine-tuning controller at this time, And switch to fine-tune the weights and biases of the controller; and when the controller works as a coarse-tuned controller, activate all hidden layer nodes, at this time the structure of the neural network is the structure of the coarse-tuned controller, and switch are the weights and biases of the coarse adjustment controller.

(4) The switching sequence of the fine-tuning controller and the coarse-tuning controller is shown in Figure 4. When the output voltage is within the steady-state error band in three consecutive cycles, and at the same time in the second cycle of the three consecutive cycles The error rate of change e _c [k] = 0 is detected in , which is expressed as follows:

{(e _oc [k]≠0)&(|e _o [k]|≤|E _b |)} ^(k=1,3) &{(e _oc [k]=0)&(|e _o [ k]|≤|E _b |)} ^(k=2) (1)

Then the controller is switched from a coarse adjustment controller to a fine adjustment controller. At this time, some (N _c -N _f ) hidden layer nodes are disabled, and at the same time, the weights and biases of the fine adjustment controller are switched.

When the output voltage is outside the steady-state error band for three consecutive cycles, it is expressed as follows:

{|e _o [k]|>|E _b |} ^(k=1,2,3) (2)

Then the controller is converted from a fine-tuned controller to a coarse-tuned controller, at which time all hidden layer nodes are activated and switched to the weights and biases of the coarse-tuned controller.

For Buck-type digital power supply, Figure 5(a), Figure 5(b) and Figure 5(c) are respectively using the technology of the present invention and the traditional neural network controller in steady state, load current jump ±1A and input voltage jump By comparing the output voltage response curve of the digital power supply when it changes to ±1V, it can be seen that the digital power supply using the hardware resource dynamic multiplexing neural network controller proposed by the present invention is different from the output voltage response curve of the digital power supply using the traditional neural network controller. Almost coincident, indicating that the performance of the two is similar, and the technology of the present invention has no obvious performance reduction.

Compared with the traditional neural network controller, the structure of the neural network controller for dynamic multiplexing of hardware resources of the digital power supply proposed by the present invention is reduced and almost the same control performance is achieved, which can save N _n -N _c hidden functions. The hardware resources required by the tiered node.

The above description is only a description of the preferred embodiments of the present invention, and is not intended to limit the scope of the present invention. Any changes and modifications made by those of ordinary skill in the field of the present invention based on the above disclosure all belong to the protection scope of the claims.

Claims (7)

1. A hardware resource dynamic multiplexing neural network controller for a digital power supply is characterized by comprising a neural network controller and a control module, wherein the neural network controller is used for generating a digital duty ratio, the control module comprises a lookup table and a dynamic multiplexing controller, the lookup table is used for storing a weight and an offset of the neural network controller, and the dynamic multiplexing controller dynamically controls the switching of the weight and the offset of the neural network controller and disables a part of hidden nodes;

the neural network controller simultaneously realizes a coarse tuning controller and a fine tuning controller, and the coarse tuning controller is used for adjusting the output voltage to a reference voltage as soon as possible after the digital power supply is disturbed so as to improve the transient response performance; the fine tuning controller is mainly used for eliminating steady-state errors and stabilizing the output voltage at a reference voltage.

2. The hardware resource dynamic multiplexing neural network controller for a digital power supply of claim 1, wherein the neural network controller is obtained by offline training of data, comprising the steps of:

establishing error bands on two sides of the reference voltage, grouping training data of the neural network controller according to the error values, wherein the data with the error within the error bands are the training data of the fine tuning controller, and otherwise, the data are the training data of the coarse tuning controller;

and (3) respectively training two groups of training data to two neural networks with the same input and output and different structures by adopting a neural network training method to respectively obtain the weight and the offset of the fine tuning controller and the weight and the offset of the coarse tuning controller.

3. The hardware resource dynamic multiplexing neural network controller for a digital power supply of claim 2, wherein in training data of the fine tuning controller, an output voltage is close to an ideal value; in the training data of the coarse tuning controller, various differences exist between the output voltage and an ideal value, including output voltage changes under different disturbances.

4. The hardware resource dynamic multiplexing neural network controller for a digital power supply of claim 3, wherein the error band ranges within ± 10% of a reference voltage.

5. The hardware resource dynamic multiplexing neural network controller for digital power supply of claim 4, wherein the number of implicit layers or implicit layer nodes of the coarse tuning controller is more, the number of implicit layers or implicit layer nodes of the fine tuning controller is relatively less, and the number of implicit layers or implicit layer nodes of the coarse tuning controller and the number of implicit layers or implicit layer nodes of the fine tuning controller are all less than the number of implicit layers or implicit layer nodes of the neural network controller directly trained by data; if the coarse tuning controller and the fine tuning controller adopt the neural network with the same hidden layer number, the coarse tuning controller can obtain the same structure with the fine tuning controller by forbidding part of hidden layer nodes, and if the weight and the bias are switched into the weight and the bias of the fine tuning controller at the same time, the coarse tuning controller can be used for realizing the function of the fine tuning controller.

6. The hardware resource dynamic multiplexing neural network controller for digital power supply of claim 5, wherein based on the time division multiplexing principle, the structure of the neural network controller is selected to be the same as that of the coarse tuning controller, and according to the position of the error in the error band, part of hidden layer nodes are disabled dynamically and the weight and the bias are switched, the functions of the two controllers are realized, and the hardware resource dynamic multiplexing neural network controller is obtained, so as to save the hardware resources required by part of hidden layer nodes; when the neural network controller works as a fine tuning controller, forbidding part of hidden layer nodes used by the coarse tuning controller independently, and switching the weight and the bias of the fine tuning controller, wherein the structure of the neural network controller is equivalent to that of the fine tuning controller; when the neural network controller works as a coarse tuning controller, all hidden layer nodes are activated, and the weight and the bias of the coarse tuning controller are switched to use, so that the structure of the neural network controller is the structure of the coarse tuning controller.

7. The hardware resource dynamic multiplexing neural network controller for digital power supply of claim 6, wherein all the neural network controller is divided into two or more sub-controllers, and then based on the time division multiplexing principle, one or more neural networks are adopted to realize the functions of two or more sub-controllers by switching weights and biases and dynamically activating and deactivating part of hidden layer nodes, thereby realizing the acceleration of the neural network controller or the reduction of the hardware resource consumption.

Images