A neural network controller based on data mining

Abstract

The application discloses a neural network controller based on data mining, and belongs to the field of power electronics. First, a data subset under different characteristic states is obtained from a data set under different working points of multiple different control algorithms through a data mining method; then, a new data subset is formed by combining data subsets with optimal performance under different characteristic states through performance evaluation, and then the data subsets under different working points are recombined into an optimized complete data set; finally, the input-output relationship of the optimized complete data set is fitted in a supervised learning mode to obtain the neural network controller based on data mining. The neural network is trained by the optimized data in the supervised learning mode, and a neural network controller with better performance can be fitted, so that the performance of the neural network controller is further improved.

Description

Technical Field

[0001] This invention relates to the field of power electronics technology, and in particular to a neural network controller based on data mining. Background Technology

[0002] The structural principle of a neural network-controlled DC-DC switching converter is as follows: Figure 1 As shown, the analog output at the load side of the power stage includes the output voltage V. out (t), input voltage V in (t), inductor current I ind (t), which are converted into digital quantities V by the ADC respectively. out [k]、V in [k] and I ind [k], then V in [k],I ind [k] and V out [k] and reference voltage V ref Error signal e between o [k] and its rate of change e oc [k] and others are fed into a neural network controller that has been trained through supervised learning. The neural network controller then processes the input information {V}. in [k],I ind [k],e o [k],e oc [k]} calculates the digital duty cycle signal d[k], then converts the digital duty cycle signal d[k] into an analog duty cycle signal d(t) through DPWM (Digital Pulse Width Modulation), and finally drives the switch S through the gate driver. p and S n The switching on or off of the circuit regulates the output voltage V. out (t) stabilize it at the reference voltage V ref .

[0003] Traditional neural network controllers are typically trained directly from acquired control data in a supervised learning model, and the training data usually comes from a single control algorithm. Since the directly acquired raw control data is not optimized, the training data may contain missing data or outliers, affecting the fitting performance of the neural network and thus limiting the performance improvement of the neural network controller. The paper "Model Predictive Control Using Artificial Neural Network for Power Converters" (IEEE Transactions on Industrial Electrics, Vol. 69, No. 4, Apr. 2022, pp: 3689-3699) proposes a model predictive controller based on artificial neural networks for power converters to improve converter control performance while reducing computational burden and required hardware resources; however, the performance improvement of the power converter is limited. This is because the training data for the neural network comes only from traditional model predictive control algorithms and is not optimized, thus limiting the performance improvement of the neural network-based controller. Summary of the Invention

[0004] The purpose of this invention is to provide a data mining-based neural network controller to solve the problem that the performance improvement of the switching converter controlled by the traditional neural network controller is limited due to the unoptimized training data.

[0005] To address the aforementioned technical problems, this invention provides a data mining-based neural network controller, comprising:

[0006] Data mining methods were used to obtain subsets of data under different characteristic states from datasets of various control algorithms at different operating points;

[0007] Then, through performance evaluation, the data subsets with the best performance under different feature states are selected and combined into new subsets, and then the subsets under different working points are recombined into the optimized complete dataset.

[0008] Finally, the neural network controller fits the optimized input-output relationship of the complete dataset under supervised learning mode to obtain a data mining-based neural network controller.

[0009] In one embodiment, the data mining method includes the following steps:

[0010] Step S1: Collect control data of different control algorithms at different operating points to establish a raw dataset for data mining;

[0011] Step S2: Divide the control data of different control algorithms at each operating point into multiple data subsets according to different characteristic states;

[0012] Step S3: Design a performance evaluation function to evaluate the performance of each data subset under different control algorithms, and select the data subset with the best performance of different control algorithms under the same characteristic state at the same operating point;

[0013] Step S4: Combine the data subsets with the best performance under different feature states at the same selected working point into a new subset, and then recombine the subsets at different working points into a complete optimized dataset.

[0014] In one implementation, the different operating points include: different input voltages and the same load current, the same input voltage and different load currents, and different input voltages and different load currents.

[0015] In one implementation, the data subset includes a steady-state data subset and a transient data subset; wherein the characteristic states of the transient data subset include startup, input voltage mutation, and load current mutation.

[0016] In one implementation, the raw dataset includes: the voltage error e between the switching power supply output voltage and the reference voltage. o [k] and its error rate of change e oc [k] Input voltage of switching power supply V in [k], Switching power supply inductor current I ind [k] and the digital duty cycle d[k].

[0017] In one embodiment, the different characteristic states in step S2 are: power supply startup, power supply in steady-state output, power supply input voltage suddenly increases, power supply input voltage suddenly decreases, load current suddenly increases, and load current suddenly decreases.

[0018] In one implementation, the function model for performance evaluation of the transient data subset is shown in Equation (1):

[0019] F1 = t s ×(|V max -V ref |+|V ref -V min |)(1)

[0020] Where t s V represents the steady-state time. ref Indicates the reference voltage, V max and V minV represents the maximum and minimum output voltage of the control algorithm under this characteristic state, respectively. max -V ref and V ref -V min Let F1 and F2 represent the maximum and minimum overshoot, respectively. F1 represents the product of the overshoot and the settling time. The smaller the function value F, the better the performance of the data subset. The performance evaluation function of the steady-state data subset is the cumulative value of the output voltage error, as shown in equation (2).

[0021]

[0022] Where p represents the number of steady-state cycles evaluated, the smaller F2 is, the more stable the output voltage and the smaller the steady-state error, and the better the performance.

[0023] In the data mining-based neural network controller provided by this invention, the training data can be optimized. The training data no longer comes from a single control algorithm, but consists of subsets of data representing different feature states extracted from different control algorithms. Each subset corresponds to the feature state with the best performance among the different control algorithms, selected through a performance evaluation function. Other non-optimal subsets are discarded, resulting in optimized training data. Because the dataset obtained through data mining filters out inferior data, the impact of training data on neural network performance can be reduced. Therefore, training the optimized data in supervised learning mode can fit a better-performing neural network controller, thereby further improving the performance of the neural network controller. Attached Figure Description

[0024] Figure 1 This is a block diagram of a neural network-controlled DC-DC switching converter.

[0025] Figure 2 This is a schematic diagram of a neural network controller based on data mining proposed in this invention;

[0026] Figure 3 This is a schematic diagram of the data mining process proposed in this invention;

[0027] Figure 4 This is a schematic diagram of six different characteristic states of control data;

[0028] Figures 5(a) to 5(c) This is a comparison chart of the output voltage waveforms of a digital power supply using a data mining-based neural network controller and a traditional neural network controller in steady state and transient state. Detailed Implementation

[0029] The following detailed description, in conjunction with the accompanying drawings and specific embodiments, provides a further detailed explanation of the data mining-based neural network controller proposed in this invention. The advantages and features of this invention will become clearer from the following description. It should be noted that the accompanying drawings are all in a very simplified form and use non-precise proportions, and are only used to facilitate and clarify the illustration of the embodiments of this invention.

[0030] This invention provides a neural network controller based on data mining, the structure of which is as follows: Figure 2 As shown, the neural network controller of this invention obtains data subsets (data subset 1, data subset 2, ..., data subset 5, data subset 6) under different feature states from datasets of different control algorithms (algorithm 1, algorithm 2, ..., algorithm N-1, algorithm N) through data mining. Then, through performance evaluation, the data subsets with the best performance under different feature states are selected and combined into new subsets. Furthermore, the subsets under different operating points are recombine to form an optimized complete dataset. Finally, the neural network fits the input-output relationship of the optimized complete dataset in a supervised learning mode, thereby obtaining a data mining-based neural network controller, including the following steps:

[0031] Step S1: First, collect control data of different control algorithms at different operating points (i.e., different input voltages and / or different load currents) to establish a raw dataset for data mining;

[0032] Step S2: Divide the control data of different control algorithms at each operating point into multiple data subsets according to different characteristic states, including: steady-state data subset and transient data subset; wherein the characteristic states of the transient data subset include startup, input voltage mutation (including power supply input voltage mutation increase and power supply input voltage mutation decrease) and load current mutation (including load current mutation increase and load current mutation decrease);

[0033] Step S3: Design a performance evaluation function to evaluate the performance of each data subset under different control algorithms, and select the data subset with the best performance of different control algorithms under the same characteristic state at the same operating point;

[0034] Step S4: Combine the best-performing subsets of data under different feature states at the same working point into a new subset, and then recombine the subsets at different working points into a complete optimized dataset.

[0035] The specific process of data mining is as follows: Figure 3 As shown, firstly, an original dataset is established using control data from different control algorithms at different operating points, including: voltage error e o [k] and its rate of change e oc [k], Input voltage V in[k], Inductor current I ind [k] and the digital duty cycle d[k]. In this embodiment, control data from traditional PID control, fuzzy PID (FLC-PID) control, and sliding mode (SM) control algorithms are selected to establish the original dataset. Then, the control data of different control algorithms at different operating points are divided into 6 data subsets according to different characteristic states, including: startup, steady state, sudden increase in input voltage, sudden decrease in input voltage, sudden increase in load current, sudden decrease in load current, etc. Figure 4 As shown.

[0036] Design a performance evaluation function and use it to select the best-performing subset of data for different control algorithms with the same characteristic states at a specific operating point. The function model for evaluating the performance of the transient data subset (including: startup, sudden increase in input voltage, sudden decrease in input voltage, sudden increase in load current, and sudden decrease in load current) is shown in Equation (1):

[0037] F1 = t s ×(|V max -V ref |+|V ref -V min |)(1)

[0038] Where t s V represents the steady-state time. ref Indicates the reference voltage, V max and V min These represent the maximum and minimum output voltages within this subset of data, i.e., V. max -V ref and V min -V ref These represent the maximum and minimum overshoot, respectively. F1 represents the performance evaluation value; the smaller the value, the better the transient performance of the data subset. The performance evaluation function for the data subset in steady state is the cumulative value of the output voltage error, as shown in equation (2):

[0039]

[0040] Where p represents the number of steady-state cycles evaluated, the smaller F2 is, the more stable the output voltage and the smaller the steady-state error, and the better the performance.

[0041] The optimal data subsets under each feature state at a specific operating point are recombined to form the complete subset for that specific operating point. Finally, the subsets from different operating points are reassembled into the complete optimized dataset. Finally, the neural network controller fits the input-output relationship of the optimized dataset under supervised learning, thus obtaining a data mining-based neural network controller.

[0042] A buck-type DC-DC switching converter controlled by a data mining-based neural network controller and a traditional neural network controller directly trained from training data were built on the Matlab / Simulink platform for simulation verification. The waveforms of the output voltage of the digital power supply controlled by the two controllers in steady state and transient state were compared. Figure 5(a) shows the start-up and steady-state output voltage when the load current is 1.0A; Figure 5(b) shows the transient response when the load current decreases by 1.0A; and Figure 5(c) shows the transient response when the load current increases by 1.0A. The simulation results show that the steady-state error of the digital power supply using the present invention is significantly reduced, and the output voltage settling time and overshoot are significantly shortened when the load current is disturbed, effectively improving the control performance of the digital power supply.

[0043] The above description is merely a description of preferred embodiments of the present invention and is not intended to limit the scope of the present invention in any way. Any changes or modifications made by those skilled in the art based on the above disclosure shall fall within the protection scope of the claims.

Claims

1. A neural network controller based on data mining, characterized in that, include: Data mining methods were used to obtain subsets of data under different characteristic states from datasets of various control algorithms at different operating points; Then, through performance evaluation, the data subsets with the best performance under different feature states are selected and combined into new subsets, and then the subsets under different working points are recombined into the optimized complete dataset. Finally, the neural network controller fits the input-output relationship of the optimized complete dataset under supervised learning mode to obtain a data mining-based neural network controller. The data mining method includes the following steps: Step S1: Collect control data of different control algorithms at different operating points to establish a raw dataset for data mining; Step S2: Divide the control data of different control algorithms at each operating point into multiple data subsets according to different characteristic states; Step S3: Design a performance evaluation function to evaluate the performance of each data subset under different control algorithms, and select the data subset with the best performance of different control algorithms under the same characteristic state at the same operating point; Step S4: Combine the data subsets with the best performance under different feature states at the same selected working point into a new subset, and then recombine the subsets at different working points into a complete optimized dataset.

2. The neural network controller based on data mining as described in claim 1, characterized in that, The different operating point scenarios include: different input voltages and the same load current, the same input voltage and different load currents, and different input voltages and different load currents.

3. The neural network controller based on data mining as described in claim 1, characterized in that, The data subset includes a steady-state data subset and a transient data subset; wherein, the characteristic states of the transient data subset include startup, input voltage mutation, and load current mutation.

4. The neural network controller based on data mining as described in claim 1, characterized in that, The original dataset includes: the voltage error between the output voltage of the switching power supply and the reference voltage. e o [ k ] and its error change rate e oc [ k Input voltage of switching power supply V in [ k Switching power supply inductor current I ind [ k and digital duty cycle d [ k ].

5. The neural network controller based on data mining as described in claim 1, characterized in that, The different characteristic states in step S2 are: power supply startup, power supply in steady state output, power supply input voltage suddenly increases, power supply input voltage suddenly decreases, load current suddenly increases, and load current suddenly decreases.

6. The neural network controller based on data mining as described in claim 3, characterized in that, The function model for performance evaluation of the transient data subset is shown in Equation (1): (1) in t s Indicates the steady-state time. V ref Indicates the reference voltage. V max and V min These represent the maximum and minimum output voltages of the control algorithm under this characteristic state, respectively. V max - V ref and V ref - V min These represent the maximum and minimum overshoot, respectively. F 1 The function value represents the product of overshoot and settling time. F The smaller the value, the better the performance of the data subset; the performance evaluation function of the steady-state data subset is the cumulative value of the output voltage error, as shown in equation (2): (2) in p Indicates the number of steady-state cycles evaluated. F 2 The smaller the value, the more stable the output voltage and the smaller the steady-state error, indicating better performance.

Images