An AI algorithm-based servo driver online model parameter identification method

Abstract

The application discloses a kind of online model parameter identification methods of servo driver based on AI algorithm, steps are as follows: S1. the parameter identification network of being served object is constructed, and current state and target value are input, and the parameter and the predicted state of next time of identification are output;S2. sampling strategy and parameter identification network iterative training are carried out;S3. parameter identification is carried out after the data of true machine sampling is optimized to network.The identification method uses AI algorithm for the system identification of servo motor can solve the problems such as low identification precision and low efficiency, to promote the application in industrial field.

Description

Technical Field

[0001] This invention relates to the field of servo drive parameter identification technology, and in particular to an online model parameter identification method for servo drives based on AI algorithms. Background Technology

[0002] Servo drives are essential basic industrial products, widely used in the control of various industrial motors. A servo drive mainly consists of two parts: hardware such as the drive circuit and the control algorithm. Traditional servo controllers use conventional methods such as PID control algorithms for various control tasks. However, in high-precision control tasks (such as rapid vibration suppression in point-to-point motion, high-precision track tracking, etc.), traditional servo controllers may experience insufficient control performance due to various reasons. Model-based control is an effective method to improve control accuracy. However, traditional model identification mostly uses offline calibration methods. These methods require high calibration costs and high consistency in product manufacturing and assembly before sampling calibration can be performed. This greatly limits the large-scale use of such methods in the industrial field. Due to the diversity of controlled objects, high-precision online identification methods are even more important in the field of servo drives. Online identification cannot use pre-designed system excitation signals, while simultaneously requiring high-speed and high-precision calculations. This presents significant challenges to high-precision online identification algorithms, and current traditional methods struggle to simultaneously meet both requirements.

[0003] The rapid development of modern artificial intelligence technology has already had a revolutionary impact on the field of technological applications, leading a new generation of technological revolution. Compared with traditional methods, AI methods have stronger model representation capabilities and extremely fast solution speeds.

[0004] Therefore, designing an online model parameter identification method for servo drives based on AI algorithms is of great significance. Summary of the Invention

[0005] The purpose of this invention is to provide an online model parameter identification method for servo drives based on AI algorithms, in order to solve problems such as low accuracy and low efficiency in servo drive parameter identification in the prior art, and to promote its application in the industrial field.

[0006] In view of this, the solution of the present invention is as follows:

[0007] A method for online model parameter identification of servo drives based on AI algorithms includes the following steps:

[0008] S1. Construct a parameter identification network for the servo object, taking the current state and target value as input, and the identified parameters and the predicted state at the next moment as output; the parameters include the load inertia, the stiffness of the reducer, the damping of the reducer, and the stiffness of the motor output shaft; the state includes the motor speed, the motor position, and current data.

[0009] S2. Perform iterative training of the sampling strategy and parameter identification network;

[0010] S3. After optimizing the network using data sampled from a real machine, parameter identification is performed.

[0011] In this invention, the parameter identification network includes three fully connected input layers, connected to three LSTM layers, and connected to two HEADs: a parameter identification output layer consisting of two fully connected layers, and a predicted state output layer consisting of two fully connected layers.

[0012] In this invention, the total training loss in step S2 is obtained by weighting the identification metric loss and the state prediction loss, specifically as follows:

[0013]

[0014]

[0015] loss = k1 * loss x +k2*loss s ;

[0016] Where: x represents the four predicted system identification parameters, x' represents the true values ​​of the four system identification parameters; s represents the predicted system state at the next moment, including the motor speed, motor position and current; s' represents the true value of the system state at the next moment; k1 and k2 are the weights of the identification parameter loss and the system state loss at the next moment, respectively.

[0017] In this invention, step S2 is based on the driver's control algorithm to collect training data, and the network parameters and sampled data are updated alternately until the control algorithm and the identification network simultaneously achieve stable convergence.

[0018] In this invention, after training in step S2, the inference frequency is increased by linear interpolation.

[0019] In this invention, step S2 uses a non-uniform distribution for initial sampling, thereby increasing the sampling rate at the boundaries of the identification parameters.

[0020] In this invention, step S2 uses an experience replay strategy for sampling. A continuous segment of data is randomly selected from the sampled data as training data. Multiple training data are shuffled to form a training batch for batch training.

[0021] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0022] 1. Traditional system identification algorithms heavily rely on the accuracy of modeling, and their representation and automatic identification capabilities are not easily able to meet control requirements under complex conditions. The identification method described in this invention uses AI algorithms for the system identification of servo motors, which can effectively improve the identification accuracy and does not rely on prior knowledge.

[0023] 2. Compared with existing similar technologies, the identification network provided by this invention adopts a network structure with two heads. By adding a symbolic state prediction loss, the backbone network itself can have the ability to regress the state transition equation, reduce the impact of training data noise, and effectively improve the stability of the regression process.

[0024] 3. The identification method described in this invention adopts an alternating iteration of sampling data and network updates, which can effectively improve the consistency of training data and reduce the required sampling amount. This allows the identification network to converge at a faster speed in the state space that the controller may generate, thereby reducing the sampling training time, reducing the model size, and improving the identification accuracy.

[0025] 4. The identification method described in this invention uses random sampling and experience replay training, which can improve the independence of training data and effectively improve the stability of the regression process. Attached Figure Description

[0026] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0027] Figure 1 This is a schematic diagram of the identification method described in this invention.

[0028] Figure 2 This is a diagram of the dual-HEAD network structure of the identification network described in this invention.

[0029] Figure 3 This is a schematic diagram illustrating the iterative update of the sampling strategy and identification network described in this invention. Detailed Implementation

[0030] To make the objectives, technical solutions, and beneficial effects of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described in this specification are merely for explaining the invention and are not intended to limit the invention.

[0031] This solution combines traditional control algorithms to provide an online model parameter identification method for servo drives based on AI algorithms. It can support various traditional model-based control methods, such as model reference control, MCP, and filter-based shaping control. The flowchart of the online model parameter identification method for servo drives is shown below. Figure 1 As shown, the method includes the following steps:

[0032] Step 1: Construct the parameter identification network for the controlled servo object. For a general servo control system, its characteristics are characterized by dynamics. Considering the flexibility of the reducer, the control dynamics can be expressed by the following dynamic equations:

[0033]

[0034]

[0035]

[0036]

[0037]

[0038]

[0039] In this system, the physical quantities M represent torque, J represent moment of inertia, α represent angular acceleration, B represent rotational damping, ω represent angular velocity, and θ represent angle. The subscripts of these physical quantities represent the corresponding objects: m represents a motor, ml represents a coupling, and l represents the load. Specifically, M... in K represents the input torque. in This represents the equivalent torsional stiffness of the coupling.

[0040] Dynamic analysis reveals that the main system parameters affecting control performance include four key indicators: load inertia, reducer stiffness, reducer damping, and motor output shaft stiffness. These four indicators influence the state at the next moment, derived from the current state and control actions. The state includes the motor speed, motor position, and current.

[0041] The parameter identification network takes as input a time series of {state-action} data pairs, with a sequence length of 10 in this scheme. In this case, the state specifically refers to the motor's position, speed, and actual current. The action is the driver's target value, which can be either the target current or the target speed. The backbone network consists of three fully connected input layers followed by three LSTM layers. The network outputs four main parameter indicators and a predicted state for the next time step, as additional outputs. Therefore, this network has two heads. The first is a two-layer fully connected output layer containing the four indicators, and the second is a two-layer fully connected output layer containing the predicted state. The network structure is as follows: Figure 2 As shown.

[0042] Step 2: In a simulation environment, random sampling and experience replay are used to iteratively pre-train the sampling strategy and parameter identification network. The driver's control algorithm is used to collect training data. Since the driver's control algorithm depends on the accuracy of the system parameter identification, and the training data for identification is obtained from the control algorithm, data collection and network training are alternately updated. Every 10 network parameter updates are followed by a sampling data update. Ultimately, both the identification network and the control algorithm achieve stable convergence simultaneously. Figure 3 As shown.

[0043] To improve the network's generalization ability, two measures are adopted: 1) A random distribution initialization method is used. Experiments show that the regression effect at boundary points is relatively poor, so a non-uniform distribution is used for initial sampling. The sampling probability is increased at the boundary of the identification parameters; non-uniform sampling distributions such as the Weibull distribution can be used. 2) A sampling experience replay strategy is adopted. A continuous segment of data is randomly selected from the collected samples as a training data point. Multiple training data points are shuffled to form a training batch for batch training.

[0044] This training process includes two losses: the first is the identification metric loss, and the second is the state prediction loss. The total loss is obtained by weighting the two losses. The loss is calculated using the following formula:

[0045]

[0046]

[0047] loss = k1 * loss x +k2*loss s ⑥.

[0048] Where x represents the four predicted system identification parameters, and x' represents the true values ​​of the four system identification parameters. s represents the predicted system state at the next time step, which has three values: motor speed, motor position, and current. s' represents the true value of the system state at the next time step. k1 and k2 are the weights of the identification parameter loss and the system state loss at the next time step, respectively. After the above training process, a simulation-based pre-trained model can be obtained.

[0049] Step 3: Employ linear interpolation to increase the inference frequency. This solution uses a low-frequency 4K frequency for inference, but high-frequency 16K identification data is required for algorithm control. Therefore, linear interpolation is used to obtain the high-frequency identification data, as detailed below:

[0050] y(0)=A0 ⑦;

[0051] y(1)=A0+(A3-A0) / 3 ⑧;

[0052] y(2)=A0+2(A3-A0) / 3 ⑨;

[0053] y(3)=A3 ⑩.

[0054] Where A0 represents the AI ​​inference action at the previous time step, and A3 represents the AI ​​inference action at the current time step. y represents the interpolation result at time step i. Through linear interpolation, the inference frequency can be increased by an integer multiple.

[0055] Step 4: Fine-tune the network using data sampled from a real device. The steps are the same as in Step 2, but the training data comes from the target real device, not simulation data. Generally, only a few dozen sampled trajectory data points with different loads are needed. Further fine-tuning can achieve better results.

[0056] In one embodiment of the present invention, according to the online model parameter identification method for servo drives based on AI algorithms provided above, the specific experimental results are shown in Table 1. It is easy to see that the identification errors of core indicators such as load inertia, reducer stiffness, reducer damping, and motor output shaft stiffness are less than 5%. At the same time, the prediction error of the state is less than 10%, which also demonstrates the effectiveness of the identification indicators of this method.

[0057] Table 1:

[0058] 1 Load inertia 5％ 0.2 2 reducer stiffness 3％ 0.2 3 Reducer damping 5％ 0.2 4 Motor output shaft stiffness 3％ 0.1 5 State prediction 10％ 0.1

[0059] Although embodiments of the invention have been shown and described, other advantages and modifications will be readily apparent to those skilled in the art. Therefore, the invention is not limited to the specific details, representative devices and illustrated examples shown and described herein without departing from the spirit and scope of the general concept as defined by the claims and their equivalents.

Claims

1. A method for online model parameter identification of servo drives based on AI algorithms, characterized in that, Includes the following steps: S1. Construct a parameter identification network for the servo object, taking the current state and target value as input, and the identified parameters and the predicted state at the next moment as output; the parameters include the load inertia, the stiffness of the reducer, the damping of the reducer, and the stiffness of the motor output shaft; the state includes the motor speed, the motor position, and current data; the target value is the target current or the target speed. S2. Perform iterative training of the sampling strategy and parameter identification network; S3. After optimizing the parameter identification network based on data sampled from the actual machine, parameter identification is performed. The parameter identification network comprises three fully connected input layers, three LSTM layers, and two HEADs: a parameter identification output layer consisting of two fully connected layers, and a predicted state output layer consisting of two fully connected layers. In step S2, the total training loss consists of the identification parameters. and state prediction The weighted average is as follows: ； in: x For the four predicted system identification parameters, x’ Identify the true values ​​of parameters for the four systems; s The system state at the predicted next moment includes the motor speed, motor position, and current. s’ The value represents the true state of the system at the next moment; k1 and k2 are the identification parameters. Weights and state prediction The weight.

2. The parameter identification method according to claim 1, characterized in that, Step S2 is based on the driver's control algorithm to collect training data. The network parameters and sampled data are updated alternately until the control algorithm and the parameter identification network reach stable convergence simultaneously.

3. The parameter identification method according to claim 1, characterized in that, After training in step S2, the inference frequency is increased by linear interpolation.

4. The parameter identification method according to claim 1, characterized in that, Step S2 uses a non-uniform distribution for initial sampling, increasing the sampling rate at the boundaries of the identification parameters.

5. The parameter identification method according to claim 1, characterized in that, Step S2 employs an experience replay strategy for sampling. A continuous segment of data is randomly selected from the sampled data as training data. Multiple training data are shuffled to form a training batch for batch training.

Images