A neural network controller for dynamic reuse of hardware resources for digital power supply

Abstract

The present invention discloses a hardware resource dynamic multiplexing neural network controller for digital power supply, which belongs to the field of power electronics and includes 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 lookup table and a dynamic multiplexing controller, the lookup table is used to store the weights and biases of the neural network controller, and the dynamic multiplexing controller dynamically controls the switching of the weights and biases of the neural network controller, and disables some hidden nodes; the neural network controller simultaneously implements a coarse adjustment controller and a fine adjustment controller, the coarse adjustment controller is used to adjust 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 adjustment controller is mainly used to eliminate steady-state errors and stabilize the output voltage to the reference voltage. The present invention reduces the structure while achieving almost the same control performance, and can save the hardware resources required for the hidden layer nodes.

Description

Technical Field

[0001] The invention relates to the technical field of power electronics, and in particular to a hardware resource dynamic multiplexing neural network controller for digital power supply. Background Art

[0002] The structure of the digital controlled DC-DC switching converter (hereinafter referred to as digital power supply) is shown in the attached figure. Figure 1 The analog output voltage V at the load end is out (t) is converted into digital output by ADC V out [k], then V out [k] and reference voltage V ref The error signal e[k] between the two is sent to the digital voltage compensator. In the 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 converted into an analog duty cycle signal d(t) through DPWM (Digital Pulse Width Modulation), and finally driven by the gate driver. P and S N The on or off of the out (t) Make it stable at the reference voltage V ref .

[0003] For digital power supply, neural network controller has the advantages of high adaptability, strong robustness, good transient performance, etc., but neural network controller usually consumes a lot of hardware resources. The traditional method of reducing hardware resource consumption of neural network controller is to "prune" the neural network, that is, to set the weights of nodes with low sensitivity (little impact on the output result) to zero or directly delete the node. Although "pruning" can effectively reduce the scale of the network, the amount of calculation and the consumption of hardware resources, it will cause the problem of reduced control accuracy for digital power supply, thereby increasing the steady-state error. At the same time, the selection of "pruning" nodes needs to be repeatedly iterated and verified, which is a large workload, and the "pruned" network will reduce its adaptability.

[0004] The paper "Dynamical Channel Pruning by Conditional Accuracy Change for Deep Neural Networks", IEEE Transactions on Neural Network and Learning Systems, Vol. 32, No. 2, February 2021, pp: 799-813, proposes a dynamic network channel pruning method, which gradually removes insensitive network channels during training iterations by constructing dynamic evaluation criteria. The method proposed in the paper effectively reduces the size and amount of computation of the network while achieving higher accuracy. However, the reduction of hardware resources and computational acceleration achieved by this method are still based on the principle of network "pruning". The dynamically pruned network structure will still cause partial loss of accuracy due to the loss of some nodes or channels, and more importantly, it will lead to reduced adaptability. Summary of the invention

[0005] The purpose of the present invention is to provide a neural network controller for dynamically multiplexing 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.

[0006] To solve the above technical problems, the present invention provides a hardware resource dynamic multiplexing neural network controller for digital power supply, comprising a neural network controller and a control module, wherein the neural network controller is used to generate a digital duty cycle, wherein the control module comprises a lookup table and a dynamic multiplexing controller, wherein the lookup table is used to store weights and biases of the neural network controller, and the dynamic multiplexing controller dynamically controls the switching of weights and biases of the neural network controller, and disables some implicit nodes;

[0007] The neural network controller simultaneously implements a coarse adjustment controller and a fine adjustment controller. The coarse adjustment controller is used to adjust the output voltage to the reference voltage as quickly as possible after the digital power supply is disturbed to improve the transient response performance; the fine adjustment controller is mainly used to eliminate steady-state errors and stabilize the output voltage to the reference voltage.

[0008] Optionally, the neural network controller is obtained through offline data training, comprising the following steps:

[0009] 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-tuning controller, and vice versa, it is the training data of the coarse-tuning controller;

[0010] The two sets of training data are used to train two neural networks with the same input and output but different structures using a neural network training method, and the weights and bias of a fine-tuning controller and the weights and bias of a coarse-tuning controller are obtained respectively.

[0011] Optionally, 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 output voltage changes under different disturbances.

[0012] Optionally, the error band ranges from ±10% of the reference voltage.

[0013] Optionally, the coarse tuning controller has more hidden layers or hidden layer nodes, and the fine tuning controller has relatively fewer hidden layers or hidden layer nodes, and both the hidden layers or hidden layer nodes of the coarse tuning and the fine tuning controller are smaller than the hidden layers or hidden layer nodes of the neural network controller directly trained by data; 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 obtain the same structure as the fine tuning controller by disabling some hidden layer nodes, and if the weights and biases are switched to those of the fine tuning controller at the same time, the coarse tuning controller can implement the functions of the fine tuning controller.

[0014] Optionally, based on the principle of time division multiplexing, the structure of the neural network controller is selected to be consistent with that of the coarse adjustment controller. According to the position of the error in the error band, some hidden layer nodes are dynamically disabled and the weights and biases are switched to realize the functions of the two controllers, so as to obtain a neural network controller with dynamic reuse of hardware resources to save the hardware resources required for some hidden layer nodes. When the neural network controller works as a fine adjustment controller, some hidden layer nodes used solely by the coarse adjustment controller are disabled, and the weights and biases of the fine adjustment controller are switched to use. The structure of the neural network controller is equivalent to that of the fine adjustment controller. When the neural network controller works as a coarse adjustment controller, all hidden layer nodes are activated, and the weights and biases of the coarse adjustment controller are switched to use. At this time, the structure of the neural network controller is the structure of the coarse adjustment controller.

[0015] Optionally, the neural network controller is 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 implement the functions of the two or more sub-controllers by switching weights and biases and dynamically activating and disabling some hidden layer nodes, thereby accelerating the neural network controller or reducing hardware resource consumption.

[0016] In the hardware resource dynamic multiplexing neural network controller for digital power provided by the present invention, the complete neural network controller is divided into two sub-controllers: a fine-tuning controller and a coarse-tuning controller, and then based on the idea of ​​time division multiplexing, a small-scale neural network is used to realize the functions of the two sub-controllers by dynamically switching weights and biases and disabling some hidden layer nodes. Therefore, the hardware resource dynamic multiplexing neural network controller for digital power provided by the present invention will not delete the trained hidden layer nodes, thereby not causing a reduction in control performance, and can save the hardware resources required for some hidden layer nodes. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is a structural schematic diagram of a digitally controlled DC-DC switching converter;

[0018] Figure 2 It is a schematic diagram of the structure of a hardware resource dynamic multiplexing neural network controller for digital power supply provided by the present invention;

[0019] Figure 3 It is a schematic diagram of the working principle of the neural network controller provided by the present invention;

[0020] Figure 4 It is a schematic diagram of the switching timing between the coarse control controller and the fine control controller;

[0021] FIG5(a) is a schematic diagram showing a comparison of the output voltage response curves of a digital power supply using the technology of the present invention and a traditional neural network controller in a stable state;

[0022] FIG5( b ) is a schematic diagram showing a comparison of the output voltage response curves of a digital power supply using the technology of the present invention and a traditional neural network controller when the load current jumps by ±1A;

[0023] FIG5(c) is a schematic diagram showing a comparison of the output voltage response curves of a digital power supply using the technology of the present invention and a traditional neural network controller when the input voltage jumps by ±1V. DETAILED DESCRIPTION

[0024] The following is a further detailed description of a hardware resource dynamic multiplexing neural network controller for digital power supply proposed by the present invention in combination with the accompanying drawings and specific embodiments. The advantages and features of the present invention will become clearer according to the following description and claims. It should be noted that the accompanying drawings are all in a very simplified form and are not in precise proportions, and are only used to conveniently and clearly assist in explaining the purpose of the embodiments of the present invention.

[0025] The present invention provides a hardware resource dynamic multiplexing neural network controller for digital power supply, such as Figure 2As shown in FIG. 1 , the structure is mainly composed of a neural network controller and a control module, wherein the control module is composed of a lookup table (storing the weights and biases of the fine-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 the weights and biases, and dynamically disable or activate some nodes of the hidden layer. The neural network controller implements the coarse-tuning controller and the fine-tuning controller in terms of function, such as Figure 3 As shown, 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, reduce the stabilization time and overshoot, and improve the transient performance. The fine adjustment controller is mainly used to eliminate the steady-state error, maintain the stable state, and stabilize the output voltage to 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 of the current cycle, V in [k] is the input voltage of the current cycle, e oc [k] = e[k] - e[k-1] is the voltage error change rate of the current cycle. The output signal of the neural network controller is the digital duty cycle d nn [k], the specific design process is as follows:

[0026] (1) Establish error bands and classify data. First, establish error bands E on both sides of the reference voltage. b , usually the reference voltage is selected within ±10% as the error band, such as Figure 4 As shown. Then, the training data of the neural network controller is classified according to the size of the error value, that is, the data with errors 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.

[0027] (2) Obtaining coarse-tuning and fine-tuning controllers. The classified data is trained using a specific training method (e.g., gradient descent method) to train two neural networks with the same input and output but different structures, and the weights and biases of the fine-tuning controller and the coarse-tuning controller are obtained respectively. The number of hidden layer nodes of the coarse-tuning controller and the fine-tuning controller is N, respectively. c and N f ,like Figure 3 As shown, the traditional neural network controller is obtained by directly training the neural network using training data, and the number of hidden layer nodes is N n , there must be Nn ≤N c +N f And N f ≤N c <N n .

[0028] (3) Implementation of hardware resource dynamic multiplexing neural network controller. c -N f The function of the fine-tuning controller can be realized by using the weights and biases of the fine-tuning controller. According to the classification principle 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; conversely, when the fine-tuning controller is in an idle state, the coarse-tuning controller is in a working state. Therefore, based on the idea of ​​time division multiplexing, a neural network (whose structure is consistent with that of the coarse-tuning controller) can be used to dynamically switch weights and biases and dynamically disable or activate N c -N f The hidden layer nodes realize the functions of two sub-controllers. Specifically, when the controller works as a fine-tuning controller, N c -N f When the controller works as a coarse-tuning controller, all hidden layer nodes are activated, and the structure of the neural network is equivalent to that of the fine-tuning controller, and is switched to the weights and bias of the coarse-tuning controller. When the controller works as a coarse-tuning controller, all hidden layer nodes are activated, and the structure of the neural network is equivalent to that of the coarse-tuning controller, and is switched to the weights and bias of the coarse-tuning controller.

[0029] (4) The switching sequence between the fine-tuning controller and the coarse-tuning controller is as follows: Figure 4 As shown in Figure 1, when the output voltage is within the steady-state error band for three consecutive cycles, and the error change rate e is detected in the second cycle of the three consecutive cycles, c [k] = 0, which means as follows:

[0030] {(e oc [k]≠0)&(|e o [k]|≤|E b |)} (k＝1,3) &{(e oc [k]=0)&(|e o [k]|≤|E b |)} (k＝2) (1)

[0031] The controller switches from a coarse control to a fine control, and the disabled part (N c -N f ) nodes of the hidden layer and simultaneously switch to the weights and bias of the fine-tuning controller.

[0032] When the output voltage is outside the steady-state error band for three consecutive cycles, it is expressed as follows:

[0033] {|e o [k]|>|E b |} (k＝1,2,3) (2)

[0034] The controller is converted from a fine-tuning controller to a coarse-tuning controller. At this time, all hidden layer nodes are activated and switched to the weights and bias of the coarse-tuning controller.

[0035] For the Buck type digital power supply, Figure 5(a), Figure 5(b) and Figure 5(c) are comparisons of the output voltage response curves of the digital power supply using the technology of the present invention and the traditional neural network controller in a stable state, when the load current jumps to ±1A and the input voltage jumps to ±1V. It can be seen that the output voltage response curve of the digital power supply using the hardware resource dynamic multiplexing neural network controller proposed by the present invention is almost the same as that of the digital power supply using the traditional neural network controller, indicating that the performance of the two is similar and there is no obvious performance degradation of the technology of the present invention.

[0036] Compared with the traditional neural network controller, the hardware resource dynamic multiplexing neural network controller for digital power supply proposed in the present invention has a smaller structure while achieving almost the same control performance, which can save N n -N c The hardware resources required for each hidden layer node.

[0037] 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 or modifications made by a person skilled in the art in the field of the present invention based on the above disclosure shall fall within the scope of protection of the claims.

Claims

1. A hardware resource dynamic multiplexing neural network controller for digital power supply, characterized in that: It includes a neural network controller and a control module, wherein the neural network controller is used to generate a digital duty cycle, wherein the control module includes a lookup table and a dynamic multiplexing controller, wherein the lookup table is used to store weights and biases of the neural network controller, and the dynamic multiplexing controller dynamically controls the switching of weights and biases of the neural network controller, and disables some hidden nodes; The neural network controller implements a coarse adjustment controller and a fine adjustment controller at the same time. The coarse adjustment controller is used to adjust the output voltage to the reference voltage as quickly as possible after the digital power supply is disturbed, so as to improve the transient response performance; the fine adjustment controller is mainly used to eliminate steady-state errors and stabilize the output voltage to the reference voltage. 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-tuning controller, and vice versa, it is the training data of the coarse-tuning controller; The two sets of training data are used to train two neural networks with the same input and output but different structures by using a neural network training method, and the weights and biases of a fine-tuning controller and the weights and biases of a coarse-tuning controller are obtained respectively; 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, the output voltage has various differences from the ideal value, including output voltage changes under different disturbances; The error band is within ±10% of the reference voltage; The coarse tuning controller has more hidden layers or hidden layer nodes, and the fine tuning controller has relatively fewer hidden layers or hidden layer nodes, and the hidden layers or hidden layer nodes of the coarse tuning and the fine tuning controller are both smaller than the hidden layers or hidden layer nodes of the neural network controller directly trained by data; 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 obtain the same structure as the fine tuning controller by disabling some hidden layer nodes, and if the weights and biases are switched to the weights and biases of the fine tuning controller at the same time, the coarse tuning controller can realize the function of the fine tuning controller.

2. The hardware resource dynamic multiplexing neural network controller for digital power supply according to claim 1, characterized in that: Based on the principle of time division multiplexing, the structure of the neural network controller is selected to be consistent with that of the coarse adjustment controller. According to the position of the error in the error band, some hidden layer nodes are dynamically disabled and the weights and biases are switched to realize the functions of the two controllers, and a neural network controller with dynamic reuse of hardware resources is obtained to save the hardware resources required for some hidden layer nodes; when the neural network controller works as a fine adjustment controller, some hidden layer nodes used alone by the coarse adjustment controller are disabled, and the weights and biases of the fine adjustment controller are switched to use, and the structure of the neural network controller is equivalent to that of the fine adjustment controller; and when the neural network controller works as a coarse adjustment controller, all hidden layer nodes are activated, and the weights and biases of the coarse adjustment controller are switched to use, and the structure of the neural network controller is now the structure of the coarse adjustment controller.

3. The hardware resource dynamic multiplexing neural network controller for digital power supply according to claim 2, characterized in that: Whenever the neural network controller is divided into two or more sub-controllers, based on the principle of time division multiplexing, one or more neural networks are used to realize the functions of the two or more sub-controllers by switching weights and biases and dynamically activating and disabling some hidden layer nodes, thereby accelerating the neural network controller or reducing hardware resource consumption.

Images