CN117997183A

CN117997183A - Neural network training method for motor control, motor control method and motor control device

Info

Publication number: CN117997183A
Application number: CN202211342082.9A
Authority: CN
Inventors: 陈宇; 童思雨; 臧晓云
Original assignee: Robert Bosch GmbH
Current assignee: Robert Bosch GmbH
Priority date: 2022-10-28
Filing date: 2022-10-28
Publication date: 2024-05-07

Abstract

The invention relates to a neural network training method and a neural network training device for motor control. The method comprises the following steps: constructing a neural network; constructing a first objective function by adopting a first input sample, training a neural network by utilizing the first input sample and the first objective function to obtain a first neural network, and obtaining a first standard input sample and a first standard output sample of the first neural network; constructing a second objective function by adopting a second input sample, training the neural network by utilizing the second input sample and the second objective function to obtain a second neural network, and obtaining a second standard input sample and a second standard output sample of the second neural network; and taking the first standard input sample and the second standard input sample as mixed input samples, taking the first standard output sample and the second standard output sample as mixed output samples, constructing a third objective function, and training the neural network to obtain the fused neural network.

Description

Neural network training method for motor control, motor control method and motor control device

Technical Field

The invention relates to the technical field of motor control, in particular to a neural network training method for motor control, a motor control method based on a neural network and a device thereof.

Background

The speed-free sensor algorithm is used for replacing a mechanical encoder, so that the system cost can be reduced, the mechanical failure probability can be reduced, and the system reliability can be improved. In a permanent magnet synchronous motor (PERMANENT MAGNET Synchronous Machine, PMSM) sensorless control application, the following problems are encountered: when the load change is large (for example, the load increases by about 100%), the controller cannot control the system at a constant speed, current overshoot occurs, and even the system is slightly poor.

A conventional solution to this problem is to manually fine tune the PI (proportional integral ) parameter of the speed controller, but this approach has the disadvantage of requiring repeated manual adjustments under different operating conditions, which is time consuming and laborious.

As other methods of solving this problem, it is also attempted to replace the PI controller with an advanced controller. Advanced controllers, however, tend to be very complex, requiring a high level of hardware computing power and expertise by the designer.

Disclosure of Invention

In view of the above problems, the present invention aims to provide a neural network training method for motor control, a motor control method based on a neural network, and a device thereof, which do not require manual fine tuning of PI parameters.

According to an aspect of the present invention, there is provided a neural network training method for motor control, the method comprising:

Constructing a neural network;

A first training step of adopting a dynamic response parameter of a motor controller under a first working condition as a first input sample, constructing a first objective function for training purpose by improving the first dynamic response performance of the motor controller under the first working condition, training the neural network by using the first input sample and the first objective function until a trained first neural network is obtained, and obtaining a first standard input sample and a first standard output sample of the trained first neural network;

A second training step of adopting a dynamic response parameter of the motor controller under a second working condition as a second input sample, constructing a second objective function for training purpose by improving the second dynamic response performance of the motor controller under the second working condition, training the neural network by using the second input sample and the second objective function until a trained second neural network is obtained, and obtaining a second standard input sample and a second standard output sample of the trained second neural network; and

And in the fusion training step, the first standard input sample and the second standard input sample are used as mixed input samples, the first standard output sample and the second standard output sample are used as mixed output samples, so that the first dynamic response performance of the motor controller under the first working condition and the second dynamic response performance of the motor controller under the second working condition are improved, a third objective function is built for training purposes, and the neural network is trained until a trained fusion neural network is obtained.

According to still another aspect of the present invention, there is provided a motor control method based on a neural network, including:

An acquisition step of acquiring dynamic response parameters of a motor controller for controlling the motor; and

And a control step of inputting the dynamic response parameters into the fusion neural network obtained by the neural network training method for motor control, obtaining the output of the fusion neural network, correcting the reference value i x _q of the q-axis current output by the motor controller by adopting the output of the fusion neural network, and controlling the motor by adopting the corrected reference value i x _q of the q-axis current.

According to still another aspect of the present invention, there is provided a computer readable medium having stored thereon a computer program which, when executed by a processor, implements the neural network training method for motor control or implements the neural network-based motor control method.

According to a further aspect of the invention, a computer device is provided, comprising a memory module, a processor and a computer program stored on the memory module and executable on the processor, the processor implementing the neural network training method for motor control or the neural network-based motor control method when executing the computer program.

According to still another aspect of the present invention, there is provided a motor control system based on a neural network, including: the computing equipment is used for realizing the neural network training method for controlling the motor and obtaining the trained fusion neural network; and a motor controller, configured to obtain a dynamic response parameter, input the dynamic response parameter to the trained fusion neural network, obtain an output of the fusion neural network, correct a reference value i× _q of a q-axis current output by the motor controller using the output of the fusion neural network, and control the motor with the corrected reference value i× _q of the q-axis current.

According to yet another aspect of the present invention, there is provided a training system for a motor-controlled neural network, comprising:

A motor including a rotation speed sensor for measuring a rotation speed of the motor, the rotation speed sensor outputting the measured rotation speed of the motor;

The load is driven by a driving system formed by the motor and the motor controller, and the driving system and the load are suitable for working under a first working condition or a second working condition, the load change in a time period under the first working condition is located outside a first preset range, and the load change in a time period under the second working condition is located in the first preset range;

the computer equipment is used for realizing the neural network training method for controlling the motor and obtaining the trained fusion neural network; and

And the motor controller is used for controlling the motor, acquiring the dynamic response parameters, inputting the dynamic response parameters into the fusion neural network to obtain the output of the fusion neural network, correcting the reference value i x _q of the q-axis current output by the motor controller by adopting the output of the fusion neural network, and controlling the motor by using the corrected reference value i x _q of the q-axis current.

Drawings

Fig. 1 is a schematic diagram of the dynamic response of the drive system (motor and motor controller) with normal load changes.

Fig. 2 is a schematic diagram of the dynamic response of the drive system (motor and motor controller) during large load changes.

Fig. 3 shows a schematic diagram of a motor control method based on a neural network according to the present invention.

FIG. 4 is a schematic diagram of a control system with a dynamically compensated neural network.

Fig. 5 shows a schematic diagram of simplifying dynamic data by averaging.

Fig. 6 is a schematic diagram showing an iterative process.

FIG. 7 is a schematic diagram of a data collection and training process for neural network training under abrupt load conditions.

FIG. 8 is a schematic diagram of a data collection and training process for neural network training under normal conditions.

FIG. 9 is a schematic diagram showing neural network training process under two conditions, a normal condition and a load ramp condition.

FIG. 10 is a schematic diagram showing the performance of the neural network of the present invention fused to two conditions.

Fig. 11 is a block diagram of a motor control device based on a neural network according to the present invention.

Fig. 12 is a block diagram of the motor control system based on the neural network of the present invention.

Fig. 13 is a block diagram of the training system for motor control neural network of the present invention.

Detailed Description

The following presents a simplified summary of the invention in order to provide a basic understanding of the invention. It is not intended to identify key or critical elements of the invention or to delineate the scope of the invention.

For the purposes of brevity and explanation, the principles of the present invention are described herein primarily with reference to exemplary embodiments thereof. Those skilled in the art will readily recognize that the same principles are equally applicable to all types of neural network-based motor control methods and neural network-based motor control systems, and that these same principles may be implemented therein, and that any such variations do not depart from the true spirit and scope of the present patent application.

The invention provides a motor control method based on a neural network and a motor control system based on the neural network, which are used for solving the problems in the prior art, do not need to finely adjust PI parameters, can ensure the system safety in the dynamic process under the load abrupt change working condition and can improve the dynamic response under the normal working condition.

The inventors of the present invention found that: the controller cannot control the system at a constant speed, and the main reasons for the problem of current overshoot mainly include slow dynamic response of the speed controller and incorrect motor parameters, which lead to rotor position estimation errors. On the other hand, if the PI parameters are fine-tuned, motor control systems with conventional PI controllers can typically react appropriately under most operating conditions, however the tuning process is very time consuming and the performance improvement is very limited, furthermore there is always a trade-off between dynamic response and static stability, in industrial applications the controller is always tuned to give higher priority to static stability, dynamic response is more challenging without sensor control, because the accuracy of estimating rotor position has a greater impact on the system dynamic response, since the controller may not respond well to extreme load changes, a design "feed forward controller" is proposed in the present invention to counteract load changes, however, performance improvement is also limited by the trade-off between dynamic response and static stability.

FIG. 1 is a schematic representation of the dynamic response of a drive system (including a motor and motor controller) when normal load changes occur (i.e., under normal operating conditions). The motor controller includes an inverter, and the motor controller may be integrated.

In fig. 1, the rotation speed is 1600r/min, the curve denoted by i _q in fig. 1 is the curve of the q-axis current i _q, the curve denoted by n is the curve of the rotation speed n, and imax is the maximum current. As can be seen from fig. 1, the q-axis current i _q does not exceed the current limit and the rotational speed n can be maintained at its desired level after a short period of dynamics.

FIG. 2 is a schematic representation of the dynamic response of a drive system (including motor and motor controller) during large load changes (i.e., abrupt load conditions). In the event of an overcurrent event in fig. 2 (circled in fig. 2), the system must be shut down to protect the motor controller from overheating, and a drop in dynamic response is required to prevent the system from tripping and overheating.

From the situation shown in fig. 1 and 2, the inventors of the present invention have found that the dynamic response of the current motor control system is not yet able to meet the desired requirements, for example, there is a problem of excessive overshoot or a problem of excessively long required settling time, which occurs mainly because the PI controller of the control system can meet the general operation requirements, but may not be able to cope with abrupt large load changes.

In order to solve the problems, a motor control method based on a neural network of the present invention is proposed, and the motor control method based on the neural network of the present invention can be improved in the following two aspects: the dynamic response of the controller is improved during a first operating condition (including abrupt load conditions) and the dynamic response of the controller is improved during a second operating condition (including normal operating conditions).

Next, a specific embodiment of the motor control method based on the neural network of the present invention will be described.

In step 1, the inventive neural network is initially constructed and the insertion position of the inventive neural network in the existing motor control process is determined.

The motor control method based on the neural network is suitable for vector control of the permanent magnet synchronous motor, in the motor control method of the invention, i _d、i_q is respectively the d-axis current and the q-axis current of the stator winding of the permanent magnet synchronous motor, and the vector control of i _d =0 is adopted for the whole system according to the general design in the field.

In a vector control system employing "i _d =0", the q-axis current i _q represents the output electromagnetic torque of the system, and may represent the q-axis current of the motor, or may represent the q-axis current generated by the motor controller for controlling the motor. Therefore, the q-axis current i _q can be controlled to achieve the purposes of adjusting the dynamic performance of the motor system and ensuring the safety of the motor controller. The reference value i x _q of the given q-axis current is used as a tracking reference for the q-axis current i _q, and the dynamic response performance of the motor controller can be more effectively adjusted by adjusting the reference value i x _q of the q-axis current. Thus, in the present invention, a neural network is applied to the reference value of the q-axis current, which operates as a feedforward controller, and is capable of compensating for the reference value of the q-axis current when the PI controller fails to respond properly.

In fig. 3, a box with a plurality of small circles inside located above fig. 3 represents a neural network, and two small circles and an input arrow above the box represent an input of the neural network. One PI on the left side in fig. 3 represents a speed loop controller, and two PI on the right side represent two current loop controllers of i _d and i _q, respectively. ωli _q and ωli _d are cross compensation terms for implementing the decoupling function of the two current controllers i _d and i _q. dq/αβ is the coordinate transformation, u× _q and u× _d are the voltages supplied to the motor by the two current loop controllers i _d and i _q, respectively.

Specifically, the inputs to the neural network in fig. 3 are two dynamic response parameters: q-axis current i _q and reference value of q-axis current i _q. The output of this neural network is a compensation value for the reference value of the q-axis current, denoted by "nn_out" in fig. 3. In the invention, after the compensation value NN_out of the reference value of the q-axis current output by the neural network is used for correcting the reference value i x _q (namely, the reference value PI-ref) of the q-axis current output by one PI controller at the left side of the neural network, the reference value i x _q of the q-axis current of the motor controller after correction is obtained, and the reference value i x _q of the q-axis current of the motor controller after correction is finally provided for the motor controller to realize the control of the motor.

Next, after the motor controller is provided with the corrected reference value i× _q of the q-axis current of the motor controller, the motor controller is converted into voltages u× _q and u× _d provided to the motor by the PI controller on the right side in fig. 3, and after the voltages u× _q and u× _d are provided to the motor, the motor generates a corresponding q-axis current i _q and d-axis current i _d (d-axis current i _d =0 in fig. 3). On the other hand, q-axis current i _q as input to the neural network and reference value i× _q of q-axis current, that is, compensation value of reference value of q-axis current outputted from previous iteration training of the neural network fed back from the motor controller acts on reference value i× _q of q-axis current of the motor controller, the obtained reference value i× _q of q-axis current after compensation, and the obtained reference value i× _q of q-axis current after compensation are used for controlling q-axis current i _q of motor formed by the motor.

In addition, n on the far left in fig. 3 represents a reference value of the rotational speed n set by, for example, a computer device or a person, n representing the motor rotational speed (which is an actual value of the motor rotational speed obtained by a rotational speed sensor mounted on the motor). The q-axis current i _q and the d-axis current i _d in fig. 3 are actual values of the q-axis and d-axis currents fed back from the current sensor in the motor controller. Thus, as shown in fig. 3, in the present invention, the output nn_out of the neural network is superimposed on the output of a PI controller (for example, superimposed on the speed loop controller on the left side of fig. 3), and finally, the corrected reference value i× _q of the q-axis current is obtained to control the motor. The motor control method with feedback shown in fig. 3 is suitable for motor control under the first and second conditions described below, that is, the reference value i× _q of the compensated q-axis current will be output to realize motor control under both the first and second conditions.

The neural network in fig. 3 may be represented by the following formula (1), where x is an input sample of the neural network, which may be a state variable describing a dynamic response process of the motor controller, may be, but is not limited to, one or more of the following dynamic response parameters such as: the d-axis current i _d of the motor or motor controller, the reference value i x _d of the d-axis current i _d of the motor controller, the q-axis current i _q of the motor or motor controller, and the reference value i x _q of the q-axis current of the motor controller may be at least the q-axis current i _q including the motor or motor controller (for example, may be a case where the input sample of the neural network shown in fig. 3 includes two dynamic response parameters or a case where the input sample of the neural network shown in fig. 4 includes four dynamic response parameters).

Representing a function of the neural network itself (i.e. reflecting a functional relationship between the input and output of the neural network itself),Representing the tunable parameters in the neural network, and y ₁ represents the output of the neural network, the following formula is given:

here, the neural network has only the output of the dynamic process, which is 0 at steady state (this is achieved by training). This neural network corresponds to a feed forward quantity and therefore does not jeopardize the control stability.

In step 2, data from the dynamic process is processed and stored in order to ease the computational and storage burden. Specifically, data of the q-axis current i _q is collected from a section of dynamic response, and the collected data is preprocessed.

To ease the computational and storage burden, data from the dynamic process is processed and stored. FIG. 4 is a schematic diagram of a control system with a dynamically compensated neural network. In fig. 4, the leftmost internal box with a plurality of small circles represents a neural network whose inputs include four dynamic response parameters: the output of the neural network illustrated in fig. 4 is a compensation value for the reference value for the q-axis current i _q of the motor controller, the d-axis current i _d of the motor or motor controller, the reference value i x _d of the d-axis current of the motor controller, the q-axis current i _q of the motor or motor controller, and the reference value i x _q of the q-axis current of the motor controller. G2 is the speed loop controller, G3 is the mechanical model of the motor, and G4 is the electrical model of the motor. G3 and G4 are in practice present as physical objects of the motor. In addition, a in fig. 4 represents a PI controller, such as a current loop controller, B represents a delay, and C represents a PMSM. Finally, as shown in fig. 4, a steady state value I _{q_s} of the reference value of the q-axis current I _q is output.

On the other hand, for the processing of the q-axis current i _q, the average value of the q-axis current i _q may also be calculated every 100 control periods, for example, and the average value of the q-axis current i _q throughout the dynamic process may be stored. In a similar manner, the inputs and outputs of the neural network are averaged in a similar manner during this process, which can allow the dynamic data to be simplified.

Fig. 5 shows a schematic diagram of simplifying dynamic data by averaging.

The left hand side of fig. 5 shows the steady state set point before averaging, with a bar in the lateral direction. The right side of fig. 5 shows the averaging followed by a jagged line, where the average of a segment of the dynamic response is taken as input to the training model. By averaging, dynamic data can be simplified.

In step 3, neural network training is performed for training purposes to improve the dynamic response of motor control under the first operating condition. The load change in the one time period under the first working condition is outside a first preset range, and may specifically include a working condition of abrupt load change.

From the perspective of the neural network controller, its "control objects" (including PI controller and motor) may constitute fig. 4. Modeling according to the following formula (2), wherein i _q represents the q-axis current of the motor, f1 (y 1) represents the functional relationship between the output y1 of the first neural network and the q-axis current i _q,

iq＝f1(yl)(2)

Taking fig. 2 as an example, the load increases sharply, resulting in a sharp increase in q-axis current i _q. In such extreme conditions with significant load fluctuations (which are "first conditions" in the claims), the training purpose of the neural network is to improve the dynamic response of the controller in conditions of abrupt load changes, as a specific example of improving the dynamic response of the controller in conditions of abrupt load changes, for example to prevent said q-axis current overshoots. The dynamic response parameters of the motor controller under the first working condition are used as first input samples, the dynamic response under the first working condition is changed to serve as training purposes to construct a first objective function, the neural network is trained by the aid of the first input samples and the first objective function until a trained first neural network is obtained, and a first standard input sample and a first standard output sample of the trained first neural network are obtained.

Wherein the first objective function of the first neural network is set to J ₁ of the following formula (3), wherein I _{q_s} is a steady state value of the reference value of the q-axis current I _q, which can be read from fig. 4.

J₁＝∑(i_q-I_{q_s})2 (3)

Using the data stored in step 2, training the neural network once using a back propagation algorithm represented by the following formula (4), and then adjusting the neural network parameters as represented by the following formula (5):

Wherein, Representing a function of the first neural network itself (i.e. reflecting a functional relationship between the input x and the output y ₁ of the first neural network itself,Representing the tunable parameters in the neural network, η representing the coefficient of the back-propagation algorithm, η affecting the speed of neural network training, since J ₁ defines the sum of squares of the errors, here by modifying the tunable parameters in the neural network/>, according to equations (4) and (5)So that J ₁ is minimized.

In step 4, the weight of the first neural network is repeatedly and iteratively updated and training of the neural network is performed, thereby obtaining a trained first neural network.

Once the parameters of the neural network are adjusted, the next dynamic process may be activated and the training process may be repeated starting from step 2. Fig. 6 is a schematic diagram showing an iterative process.

In fig. 6, "G" represents the weight of the updated neural network, "H" represents the process data and the network training, wherein the up-down arrow represents the process of repeating the iteration, which is repeated in time order from left to right in fig. 6. The method comprises the steps of repeatedly and iteratively carrying out updating weight and network training as shown in fig. 6 to obtain a trained first neural network, and respectively taking a first input sample and a first output sample meeting training termination conditions in the iterative training process as a first standard input sample and a first standard output sample. In particular, a first input sample and a first output sample of a last training process may be used as the first standard input sample and the first standard output sample.

FIG. 7 shows a schematic diagram of a data collection and training process for neural network training under abrupt load conditions.

In fig. 7, S1 indicates that the result output of the neural network, that is, the reference value i× _q of the compensated q-axis current is output from the computer device (in which the neural network is implemented) to the motor controller, S2 indicates that the current dynamic response parameter is fed back from the motor controller to the neural network in the computer device, for example, the motor controller applies the compensation value of the reference value of the q-axis current output from the previous iteration training to the reference value i× _q of the q-axis current of the motor controller, the obtained reference value i× _q of the compensated q-axis current and the q-axis current i× _q of the formed motor, and feeds back the obtained q-axis current i _q and the reference value i× _q of the q-axis current to the neural network. S3 represents training of the neural network based on the objective function of the formula (3), S4 represents updating of the weight of the first neural network, NN represents the neural network implemented in the computer device, n represents the rotational speed of the motor, i _q represents the q-axis current of the motor or the motor controller, and NN_out represents the output of the neural network.

As the number of training rounds increases, the value of the first objective function, J ₁ (denoted as "L" in fig. 7, i.e. "loss", which represents the difference between the steady-state values I _{q_s} of the reference values of the q-axis current I _q and the q-axis current I _q represented by the above equation (3), will decrease, the lost value will not drop again after several rounds of training, while the overshoot of the q-axis current I _q gradually decreases and the overshoot disappears at the final position (the position indicated by the arrow in the upper right corner in fig. 7), i.e. the behavior of the dynamic process achieves the objective of "suppressing the overshoot current". Thus, after the objective is achieved, the input X ₁ and output Y ₁ of the neural network are recorded. Here, X ₁ and Y ₁ are each an array of length m1, covering the whole dynamic process. Wherein, X ₁[1]、X₁ [2] … and Y ₁[1]、Y₁ [2] … in FIG. 7 denote time sequences, and X ₁[1]、X₁ [2] … and Y ₁[1]、Y₁ [2] … denote a plurality of data elements arranged in time sequences.

In step 5, neural network training is performed for training purposes to improve the dynamic response of motor control under the second operating condition. The load change during said one period of time in said second operating condition is within a first preset range, e.g. the load remains unchanged or slightly changed, said second operating condition belonging to a normal operating condition. The dynamic response parameters of the motor controller under the second working condition are used as second input samples, a second objective function is constructed for training purposes by taking the improvement of the second dynamic response performance of the motor controller under the second working condition, the neural network is subjected to iterative training by using the second input samples and the second objective function until a trained second neural network is obtained, and a second standard input sample and a second standard output sample of the trained second neural network are obtained, wherein the second input samples and the second output samples meeting training termination conditions in the iterative training process are respectively used as the second standard input samples and the second standard output samples.

Specifically, when the load remains unchanged or slightly changed, there is always a shift request according to the operation requirement. In such normal conditions without significant load fluctuations, it is desirable that the rotational speed controller responds to the rotational speed command as quickly as possible, and therefore, here, the training purpose of the neural network is to improve the dynamic response of the controller under normal conditions (belonging to the "second condition" in the claims), as a specific example of improving the dynamic response of the controller under normal conditions, such as a response to an acceleration of the rotational speed, the data of the dynamic process of which is also recorded as in step 2, here, modeled according to the following formula (6), where n represents the rotational speed of the motor, f ₂(y₂) represents the functional relationship between the output y ₂ of the second neural network and the rotational speed n,

n＝f₂(y₂) (6)

The objective function of the second neural network is set to a second objective function J ₂ shown in the following formula (7), where n represents the rotational speed of the motor and n _ref represents a rotational speed reference value of the motor.

J₂＝∑(n_ref-n)² (7)

Here, the second objective function J ₂ is constituted based on the rotation speed n and the reference value n _ref of the rotation speed, so minimizing J ₂ means making the rotation speed n track the reference value n _ref of the rotation speed faster. On the basis of the formula (7), the algorithm of the following formulas (8) and (9) is adopted to adjust the parameters of the neural network:

Wherein, Representing a function of the second neural network (i.e. reflecting the functional relationship between the input and output of the second neural network itself),Representing the tunable parameters in the neural network, η representing the coefficient of the back-propagation algorithm, η affecting the speed of neural network training, since J ₂ defines the sum of squares of the errors, here by modifying the tunable parameters in the neural network/>, according to equations (8) and (9)So that J ₂ is minimized.

In step 6, parameters of the neural network are updated and training of the neural network is repeated, thereby obtaining a trained second neural network.

Once the parameters of the neural network are updated, the next dynamic process may be activated and the training process step 5 may be repeated. Such a repetitive training process is shown in fig. 8. FIG. 8 is a schematic diagram of a data collection and training process for neural network training under normal conditions. As shown in fig. 8, S1 represents a reference value n _ref (where the reference value n _ref of the rotational speed n is the reference value of the rotational speed n represented by n x at the leftmost side in fig. 3) given from a computer device (in which the neural network is implemented) to the rotational speed n of the motor controller, which may be set by a computer or manually, S2 represents feedback of a current dynamic response parameter from the motor controller to the neural network in the computer device, for example, the motor controller applies a compensation value of the reference value of the q-axis current outputted from the previous training to the reference value i x _q of the q-axis current of the motor controller, the obtained reference value i x _q of the compensated q-axis current, and the q-axis current i _q, of the formed motor, and the obtained reference values i _q and i x _q of the q-axis current to the neural network, S3 represents training of the neural network based on equation (6), and NN represents the neural network implemented in the computer device in S3.

After several rounds of training, the value J ₂ (denoted as "L" in FIG. 8, i.e. "loss") of the objective function is no longer decreasing. Meanwhile, compared with the situation that the neural network is not trained, the dynamic response process of the rotating speed n is accelerated, the dynamic adjustment time is shorter, the purpose of accelerating the dynamic adjustment of the rotating speed is achieved, specifically, the curve indicated by the lower left circle and the arrow in fig. 8 represents the situation that the neural network is not trained, the rotating speed n does not respond very fast in the situation, the curve indicated by the lower right circle and the arrow represents the situation that the neural network is trained, and the rotating speed n responds to the instruction of fast reaching the speed change in the situation, namely, the situation represents that the objective function J ₂ reaches the preset objective. Thus, after training is finished, the input X ₂ and the outputs Y ₂.X₂ and Y ₂ of the neural network in the dynamic process under the working condition are recorded as an array with the length of m2, and the whole dynamic process is included. Wherein, X ₂[1]、X₂ [2] … and [1], [2] … in Y ₂[1]、Y₂ [2] in FIG. 8 represent time sequence, and X ₂[1]、X₂ [2] … and Y ₂[1]、Y₁ [2] represent a plurality of data elements arranged time sequence.

In step 7, training the neural network by taking the dynamic response of motor control under the improved load abrupt change working condition and the normal working condition as a training purpose to obtain the fused neural network. Specifically, a first standard input sample and a second standard input sample are used as mixed input samples, a first standard output sample and a second standard output sample are used as mixed output samples, a third objective function is built for training purposes by improving the first dynamic response performance of the motor controller under the first working condition and the second dynamic response performance of the motor controller under the second working condition, and the neural network is trained until a trained fused neural network is obtained. This step will be specifically described below.

The control effect under normal conditions is opposite to the control effect under load abrupt conditions, which means that the neural network has "forgotten" the previous learning result. To solve this problem, the training method of fig. 9 is further used in the present invention.

FIG. 9 is a schematic diagram of a neural network training process combining two conditions, a normal condition and a load ramp condition.

In fig. 9, sa denotes a weight for updating the neural network, sb denotes a neural network training to minimize the function J, and NN denotes a neural network employed in Sb.

In fig. 9, D1 represents a data set of neural network training under a first working condition, D2 represents a data set of neural network training under a second working condition, and two sets of input-output relationships X ₁、Y₁、X₂ and Y ₂ of D1 and D2 are used to form a total training set X and Y, with a length of m1+m2, where X is a combination of the first standard input sample X ₁ and the second standard input sample X ₂, and Y is a combination of the first standard output sample Y ₁ and the second standard output sample Y ₂. The purpose of the training of fig. 9 is to enable a single neural network to function in both modes of optimally handling normal and abrupt load conditions, where the neural network with the third objective function J shown in equation (10) below is retrained using a training set:

J＝(Y-y)² (10)

wherein Y is a combination of outputs obtained by inputting the first standard input sample X ₁ and the second standard input sample X ₂ to the first neural network and the second neural network, respectively, and Y is a standard output of the fused neural network, wherein the standard output of the fused neural network includes: the first standard output sample Y ₁ and the second standard output sample Y ₂.

After several rounds of training, the third objective function J is minimized, which means that the fused neural network has fitted all the input-to-output relationships, and thus is able to complete the processing of both cases.

On the other hand, the training process of the above neural network may be training in "real time" along with the interaction between the computer and the motor controller, so that the dynamic response in a specific scene can be optimized until the preset optimization objective is met, and of course, the above neural network may also be training offline and then be executed in the CPU of the motor controller, thereby saving the CPU load and the memory of the motor controller.

FIG. 10 is a schematic diagram showing the performance of a neural network incorporating two conditions according to the present invention.

In fig. 10, the horizontal axis represents time, the vertical axis i _q represents q-axis current, and n represents rotational speed. The change in load is shown by the upper left arrow, here from 1 to 4 times the load (i.e. representing a large load), when the q-axis current i _q rises with little overshoot. Then, the rotational speed changes as indicated by the arrow at the bottom right, here from rotational speed 800 to rotational speed 400, and it can be seen that at this time, under a heavy load, the rotational speed is reduced by the reduction of the q-axis current i _q, so that it is ensured that the overshoot of the rotational speed is not significant (although the overshoot of a part of the current is sacrificed) and a balance point of both can be achieved, which also means that for the command of the rotational speed change, the balance point of the current and the rotational speed can also be found by using the fused neural network of the present invention.

Thus, as shown in fig. 10, the neural network of the present invention can not only meet the safe operation requirement of the dynamic process under the condition of abrupt load change, but also improve the dynamic response (such as increasing the adjustment speed) of the system under the normal state. In a similar manner, more cases may be used to train the neural network, and the training results may be added to the total training sets X and Y, where the neural network may cover a wider and wider range of cases after successful training of the neural network with X and Y.

As shown in fig. 11, the motor control device 100 based on the neural network of the present invention includes:

The first training module 110 uses a dynamic response parameter of the motor controller under a first working condition as a first input sample, constructs a first objective function for training purposes by improving a first dynamic response performance of the motor controller under the first working condition, and performs iterative training on the neural network by using the first input sample and the first objective function until a trained first neural network is obtained, and obtains a first standard input sample and a first standard output sample of the trained first neural network, where obtaining the first standard input sample and the first standard output sample of the trained first neural network includes: respectively taking a first input sample and a first output sample which meet a first training termination condition in an iterative training process as the first standard input sample and the first standard output sample;

The second training module 120, configured to use a dynamic response parameter of the motor controller under a second working condition as a second input sample, construct a second objective function for training purposes to improve a second dynamic response performance of the motor controller under the second working condition, and perform iterative training on the neural network by using the second input sample and the second objective function until a trained second neural network is obtained, and obtain a second standard input sample and a second standard output sample of the trained second neural network, where obtaining the second standard input sample and the second standard output sample of the trained second neural network includes: respectively taking a second input sample and a second output sample which meet a second training termination condition in the iterative training process as the second standard input sample and the second standard output sample; and

The fusion training module 130 uses the first standard input sample and the second standard input sample as a mixed input sample, uses the first standard output sample and the second standard output sample as a mixed output sample, so as to improve the first dynamic response performance of the motor controller under the first working condition and the second dynamic response performance of the motor controller under the second working condition, construct a third objective function for training purposes, and trains the neural network until a trained fusion neural network is obtained.

Wherein the first training module 110 performs the following actions:

Modeling according to the following equation (2), wherein i _q represents the q-axis current of the motor, f ₁(y₁) represents the functional relationship between the output y ₁ of the first neural network and the q-axis current i _q,

i_q＝f₁(y₁) (2)；

Setting a first objective function of the first neural network as a first objective function J ₁ shown in the following formula (3), wherein I _{q_s} represents a steady-state value of the q-axis current I _q,

J₁＝∑(i_q-I_{q_s})² (3)；

Collecting dynamic response parameters of a motor controller under a first working condition as a first input sample, performing one-time neural network training by using a back propagation algorithm shown in the following formula (4), adjusting the neural network parameters according to the following formula (5), optimizing the neural network parameters in an iterative loop to obtain the first neural network,

Wherein,Representing a function of the first neural network itself,Representing the tunable parameters in the neural network, η represents the coefficients of the back propagation algorithm.

Wherein the second training module 120 performs the following actions:

Modeling according to the following equation (6), where n represents the rotational speed of the motor, f ₂(y₂) represents the functional relationship between the output y ₂ of the second neural network and the rotational speed n of the motor,

n＝f₂(y₂) (6)

Setting a second objective function of the second neural network as a second objective function J ₂ shown in the following formula (7), where n represents the rotation speed of the motor, n _ref represents a reference value of the rotation speed of the motor,

J ₂＝∑(n_ref-n)² (7); and

Collecting dynamic response parameters of the motor controller under a second working condition as a second input sample, training the neural network based on the formula (7), adjusting the neural network parameters according to the following formulas (8) and (9), optimizing the neural network parameters in an iterative loop to obtain the second neural network,

Wherein,Representing a function of the second neural network itself,Representing the tunable parameters in the neural network, η represents the coefficients of the back propagation algorithm.

Wherein the fusion training module 130 performs the following actions:

Using the first input sample and the second input sample,

The third objective function of the fused neural network is set to a third objective function J shown in the following equation (10),

J＝(Y-y)² (10)

Wherein Y is a combination of outputs obtained by inputting the first input sample and the second input sample to the first neural network and the second neural network, respectively, and Y is a target set value of the outputs of the first neural network and the second neural network; and

The fused neural network is obtained by training the neural network based on equation (10) such that the third objective function J is minimized.

Next, a motor control system based on a neural network according to the present invention will be described.

As shown in fig. 12, the motor control system 200 based on the neural network of the present invention includes:

The computing device 210 is configured to implement the neural network training method for motor control and obtain the trained fusion neural network; and

And a motor controller 220, configured to obtain a dynamic response parameter, input the dynamic response parameter to the trained fused neural network, obtain an output of the fused neural network, correct a reference value i× _q of a q-axis current output by the motor controller using the output of the fused neural network, and control the motor with the corrected reference value i× _q of the q-axis current.

Finally, a training system for a motor-controlled neural network of the present invention will be described.

As shown in fig. 13, the training system 300 for a motor-controlled neural network of the present invention includes:

A motor 310 provided with a rotational speed sensor for measuring the rotational speed of the motor, the rotational speed sensor outputting the measured rotational speed of the motor;

a load 320 driven by a drive system formed by the motor and a motor controller, and the drive system and the load 320 are adapted to operate in a first operating condition in which a load variation in a time period is outside a first preset range or in a second operating condition in which a load variation in a time period is within a first preset range;

Computer equipment 330 for implementing the neural network training method for motor control and obtaining the trained fusion neural network; and

And a motor controller 340, configured to control the motor 310, where the dynamic response parameter is acquired, the dynamic response parameter is input to the fusion neural network to obtain an output of the fusion neural network, the output of the fusion neural network is used to modify a reference value i× _q of q-axis current output by the motor controller, and the motor 310 is controlled with the modified reference value i× _q of q-axis current.

As described above, by using the motor control method based on the neural network and the motor control device based on the neural network, the PMSM system can keep safe operation in a dynamic process under the extreme load condition of abrupt load change, and meanwhile, the dynamic response under the normal working condition can be improved by applying the neural network.

Moreover, in the present invention the training of the neural network may be performed in real time or on-line, for example, by training the interactions between the neural network and the motor controller in real time or on-line, i.e., the training performed on the neural network and the control performed on the motor controller are performed in real time or on-line, so that the dynamic response in a specific case may be optimized until a predetermined optimization objective is reached.

On the other hand, the motor control method based on the neural network and the motor control device based on the neural network according to the invention can also be used for training the neural network which finally covers all possible conditions (under the load abrupt change working condition and under the normal working condition), wherein the conditions can also be called off-line training (such off-line training can be performed by a vehicle-mounted processor, for example, the q-axis current and the rotating speed of the motor are collected by the vehicle-mounted processor and the neural network training is performed, or such off-line training can also be performed by a cloud server, for example, the q-axis current and the rotating speed of the motor are collected by the vehicle-mounted processor and are uploaded to the cloud server through the vehicle network, and the neural network training is performed by the cloud server), and then the motor control device is applied to a CPU of the motor controller, or the motor control device can also be applied to a CPU of an ECU (Electronic Control Unit, an electronic control unit), a DCU (Drive Control Unit, a driving controller) and an MCU (motor controller), and the motor controller can be saved.

The above examples mainly illustrate the neural network training method for motor control, and the motor control method based on the neural network and the apparatus thereof. Although only a few specific embodiments of the present invention have been described, those skilled in the art will appreciate that the present invention may be embodied in many other forms without departing from the spirit or scope thereof. Accordingly, the present examples and embodiments are to be considered as illustrative and not restrictive, and the invention is intended to cover various modifications and substitutions without departing from the spirit and scope of the invention as defined by the appended claims.

Claims

1. A neural network training method for motor control, comprising:

Constructing a neural network;

2. A neural network training method for motor control as claimed in claim 1,

The first training step includes:

Acquiring a first input sample, wherein the first input sample comprises a dynamic response parameter of a motor controller in a time period under the first working condition, and the load change in the time period under the first working condition is out of a first preset range;

Constructing the first objective function based on q-axis current of the motor; and/or the number of the groups of groups,

The second training step includes:

Acquiring a second input sample, wherein the second input sample comprises a dynamic response parameter of the motor controller in a time period under a second working condition, and the load change in the time period under the second working condition is positioned in a first preset range;

constructing the second objective function based on the rotational speed of the motor; and/or the number of the groups of groups,

The fusion training step comprises the following steps:

constructing a third objective function based on the mixed output samples;

And training the neural network by utilizing the mixed input sample and the third objective function until a trained fused neural network is obtained.

3. A neural network training method for motor control as claimed in claim 2,

The first training step includes: constructing the first objective function by taking the aim of preventing the q-axis current of the motor controller from overshooting under the first working condition as a training aim;

Wherein the first objective function is constructed based on a difference in steady state values of the q-axis current and a reference value of the q-axis current.

4. A neural network training method for motor control as claimed in claim 2,

In the first training step, training the neural network using the first input sample and the first objective function includes:

Acquiring a functional relationship between the output y ₁ of the first neural network and the q-axis current i _q;

the neural network is trained using a back propagation algorithm based on the first objective function, the functional relationship, a function of the neural network itself, and parameters of the neural network are adjusted.

5. A neural network training method for motor control as claimed in claim 2,

The one time period under the first working condition comprises a first dynamic response period of the motor controller under the first working condition, and the first training step comprises the following steps: repeating the steps of acquiring first input samples and training the neural network based on the acquired first input samples to optimize parameters of the neural network through a first iterative training process to obtain the trained first neural network, wherein the repeatedly acquired plurality of first input samples comprise dynamic response parameters of the motor controller in a plurality of first dynamic response periods; and/or

One time period under the second working condition comprises a second dynamic response period of the motor controller under the second working condition, and the second training step comprises the following steps: and repeating the step of acquiring a second input sample and the step of training the neural network based on the acquired second input sample to optimize parameters of the neural network through a second iterative training process to obtain the trained second neural network, wherein the plurality of second input samples repeatedly acquired comprise dynamic response parameters of the motor controller in a plurality of second dynamic response periods.

6. The neural network training method for motor control of claim 5,

The obtaining a first standard input sample and a first standard output sample of the trained first neural network includes: respectively taking a first input sample and a first output sample which meet a first training termination condition in the first iterative training process as the first standard input sample and the first standard output sample;

Obtaining a second standard input sample and a second standard output sample of the trained second neural network comprises: and respectively taking a second input sample and a second output sample which meet a second training termination condition in the second iterative training process as the second standard input sample and the second standard output sample.

7. A neural network training method for motor control as claimed in claim 2,

The second training step includes: and constructing the second objective function for training purposes by improving the rotating speed response of the motor controller under the second working condition, wherein the second objective function is constructed based on the difference value between the rotating speed of the motor and the rotating speed reference value of the motor.

8. A neural network training method for motor control as claimed in claim 2,

The fusion training step comprises the following steps: the first standard input sample and the second standard input sample are respectively input into the neural network to obtain a combination of outputs and a difference value of the standard outputs of the fused neural network to construct the third objective function;

The standard output of the fused neural network comprises: the first and second standard output samples.

9. A neural network training method for motor control as claimed in claim 3,

The first training step comprises the following sub-steps:

Modeling according to the following equation (2), wherein i _q represents the q-axis current of the motor controller, f ₁(y₁) represents the functional relationship between the output y ₁ of the first neural network and the q-axis current i _q,

_iq＝f₁(y₁) (2)；

Setting a first objective function of the first neural network as a first objective function J _1, shown in the following formula (3), wherein I _{q_s} represents a steady-state value of a reference value of the q-axis current I _q,

J₁＝∑(i_q-I_{q_s})² (3)；

Collecting the q-axis current i _q and the reference value i x _q of the q-axis current as the first input sample, or collecting the q-axis current i _q, the reference value i x _q of the q-axis current, the reference value i x _d of the d-axis current i _d and the d-axis current of the motor controller as the first input sample, performing a neural network training using a back propagation algorithm shown in the following formula (4), adjusting a neural network parameter according to the following formula (5), optimizing the neural network parameter in an iterative loop to obtain the first neural network,

Wherein,Representing a function of the first neural network itself, x being said first input sample, y ₁ being the output of the first neural network,Representing the tunable parameters in the neural network, η represents the coefficients of the back propagation algorithm.

10. The neural network training method for motor control of claim 7, wherein the second training step includes the substeps of:

Modeling according to the following equation (6), where n represents the rotational speed of the motor, f ₂(y₂) represents the functional relationship between the output y ₂ of the second neural network and the rotational speed n,

n＝f₂(y₂) (6)

Constructing a second objective function of the second neural network as a second objective function J _2, shown in the following formula (7), where n represents the rotation speed of the motor, n _ref represents a rotation speed reference value of the motor,

J2 Delta sigma (nref-n) 2 (7); and

Collecting q-axis current i _q and reference value i x _q of q-axis current of the motor as the second input sample, or collecting q-axis current i _q of the motor, reference value i x _q of q-axis current, reference value i x _d of d-axis current i _d of the motor and reference value i x _d of d-axis current of the motor as the second input sample, performing neural network training based on formula (7), adjusting neural network parameters according to the following formulas (8) and (9), optimizing the neural network parameters in an iterative loop to obtain the second neural network,

Wherein,Representing a function of the second neural network itself, x being said second input sample, y ₂ being the output of the second neural network,Representing the tunable parameters in the neural network, η represents the coefficients of the back propagation algorithm.

11. The neural network training method for motor control of claim 8, wherein the fusion training step includes the substeps of:

using the mixed input samples as described above,

Setting the objective function of the fused neural network to a third objective function J shown in the following formula (10) _,

J＝(Y-y)² (10)

Wherein Y is the first criterion? Input sample and the second criterion? Input samples are respectively input to the combination of the outputs obtained by the neural network, and y is the standard output of the fusion neural network; and

Training the neural network based on equation (10) such that the objective function J is minimized to obtain the trained fused neural network.

12. The neural network training method for motor control of claim 1, wherein the dynamic response parameters include one or more of:

the q-axis current of the motor controller, the reference value of the q-axis current of the motor controller, the d-axis current of the motor controller, and the reference value of the d-axis current of the motor controller.

13. The neural network training method for motor control of claim 1, further comprising:

collecting dynamic response parameters over a specified period of time, the specified period of time including a plurality of dynamic response periods, comprising:

Acquiring an average value of dynamic response parameters in each dynamic response period;

and taking the average value of the dynamic response parameters in each dynamic response period as the first input sample and the second input sample.

14. The neural network training method for motor control of claim 1, wherein the output of the neural network is used for compensation of an output parameter of the motor controller.

15. A motor control method based on a neural network, comprising:

A control step of inputting the dynamic response parameter to the fused neural network obtained by the neural network training method for motor control according to any one of claims 1 to 14, obtaining an output of the fused neural network, correcting a reference value i× _q of q-axis current output by the motor controller by using the output of the fused neural network, and controlling the motor by using the corrected reference value i× _q of q-axis current.

16. The motor control method based on the neural network of claim 15,

Obtaining the fused neural network in real time by using the neural network training method for motor control according to any one of claims 1 to 14, with respect to the control performed on the motor in the control step; or alternatively

The fused neural network is obtained in advance using the neural network training method for motor control according to any one of claims 1 to 14 with respect to the control performed on the motor in the control step.

17. The motor control method based on the neural network of claim 15,

The dynamic response parameters include one or more of the following:

18. A computer readable medium having a computer program stored thereon, characterized in that,

The computer program, when executed by a processor, implements the neural network training method for motor control of any one of claims 1 to 14 or implements the neural network-based motor control method of any one of claims 15 to 17.

19. A computer device comprising a memory module, a processor and a computer program stored on the memory module and executable on the processor, characterized in that the processor implements the neural network training method for motor control of any one of claims 1 to 14 or implements the neural network-based motor control method of any one of claims 15 to 17 when executing the computer program.

20. A neural network-based motor control system, comprising:

A computing device for implementing the neural network training method for motor control of any one of claims 1 to 14 and obtaining the trained fused neural network; and

And the motor controller is used for acquiring dynamic response parameters, inputting the dynamic response parameters into the trained fusion neural network and obtaining the output of the fusion neural network, correcting the reference value i- _q of the q-axis current output by the motor controller by adopting the output of the fusion neural network, and controlling the motor by using the corrected reference value i- _q of the q-axis current.

21. A training system for a motor controlled neural network, comprising:

computer equipment for implementing the neural network training method for motor control according to any one of claims 1 to 14 and obtaining the trained fusion neural network; and