阅读说明：本技术 一种数据驱动的异步电机动力学模型建模方法 (Data-driven modeling method for dynamics model of asynchronous motor ) 是由 漆星 郑常宝 于 2021-02-07 设计创作，主要内容包括：本发明涉及一种数据驱动的异步电机动力学模型建模方法,使用异步电机的实际运行数据来建立异步电机的动力学模型,并且使用数据驱动型非线性动力学稀疏表征方法来辨识动力学模型中的系数；与传统基于等效电路的异步电机动力学建模方法相比,本发明所建立的异步电机动力学模型不会受到模型误差的影响,且对噪声的鲁棒性更高；另一方面,相较于其他的数据驱动型动力学建模方法,例如神经网络、支持向量机等,本发明所建立的异步电机动力学模型结构更为简洁,且更具有可解释性。(The invention relates to a data-driven modeling method for a dynamic model of an asynchronous motor, which is characterized in that the dynamic model of the asynchronous motor is established by using actual operation data of the asynchronous motor, and coefficients in the dynamic model are identified by using a data-driven nonlinear dynamics sparse representation method; compared with the traditional asynchronous motor dynamics modeling method based on the equivalent circuit, the established asynchronous motor dynamics model is not influenced by model errors and has higher robustness to noise; on the other hand, compared with other data-driven dynamics modeling methods such as a neural network, a support vector machine and the like, the asynchronous motor dynamics model established by the method is simpler in structure and more interpretable.)

1. A data-driven modeling method for a dynamics model of an asynchronous motor is characterized by comprising the following steps: according to the method, actual operation data of the asynchronous motor are used for establishing a dynamic model dx/dt & xi theta (x, u) when the asynchronous motor operates, wherein x is a state vector of the dynamic model, u is an input vector in the dynamic model, theta (x, u) is a dictionary library formed by the state vector x and the input vector u, xi is a sparse coefficient matrix, and meanwhile, a data-driven nonlinear dynamic sparse representation method is used for identifying the sparse coefficient matrix xi in the established dynamic model, so that the final data-driven asynchronous motor dynamic model is obtained.

2. The modeling method of a dynamic model of a data-driven asynchronous motor according to claim 1, characterized in that the step of using actual operation data of the asynchronous motor to build the dynamic model of the asynchronous motor is as follows:

(1) the data acquisition stage, this stage is in asynchronous motor actual operation in-process, the operational data of gathering the motor each moment, include: collecting d-q axis voltage data of the motor at t moments, and recording the data as [ v [ [ v ]_sd1,v_sq1,…,v_sdt,v_sqt](ii) a Collecting d-q axis current data of the motor at t moments, and recording as [ i_sd1,i_sq1,…,i_sdt,i_sqt]Collecting rotor position data [ theta ] of the motor at t moments₁,…,θ_t]And speed data [ omega ] of the motor_r1,…,ω_rt]；

(2) A dynamic model establishing stage, wherein the structure of the motor dynamic model established at the stage is as follows: xi (x, u), wherein x is a state vector of the dynamic model, u is an input vector in the dynamic model, Θ (x, u) is a dictionary library formed by the state x and the input u, and xi is a sparse coefficient matrix; v is to be_sd、v_sqAnd omega_rThe input vector u as the kinetic model is denoted as u ═ v_sd,v_sq,ω_r]I is to_sd、i_sqAnd θ is a state vector x of the kinetic model, denoted as x ═ i_sd,i_sq,θ]。

3. The modeling method for the data-driven asynchronous motor dynamics model according to claim 1, characterized in that the used data-driven nonlinear dynamics sparse representation method comprises the following specific steps:

(1) constructing a dictionary repositoryWhereinDenotes all possible pairwise combinations of x and u, denoted by 23 dimensions in total;

(2) establishing a dynamics model dx/dt ═ xi Θ (x, u), wherein a sparseness coefficient matrix xi is denoted xi ═ xi [ ]₁,ξ₂,ξ₃]In which ξ_1,2,3Called the sparse coefficient vector, and ξ_1,2,3＝[0,0,0,…,a,…,0]A represents a certain value, and each xi dimension is the same as theta (x, u) and is also 23 dimensions;

(3) determining specific numerical values in the sparse coefficient vector by using LASSO regression in the statistical theory, wherein the specific calculation formula is as follows:

where k 1 … 3 denotes the dimension of the vector ξ_kRepresents the finally determined sparse coefficient vector and,representing the intermediate process sparse coefficient vector in estimation, | · | | | non-woven phosphor₂Representing a L2 norm, | · | | non-woven₁Represents the L1 norm;

(4) will estimate the completed ξ_1,2,3Substituting into the kinetic model to obtain a final model, which is recorded as:

Technical Field

The invention relates to the technical field of motor control, in particular to a data-driven modeling method for a dynamics model of an asynchronous motor.

Background

The primary problem of motor control is the modeling problem of the motor, and the mathematical model of the motor is the most important tool in the research of motor control theory and the research and development of motor control system products. In the field of motor control, scholars and engineers often build two motor mathematical models according to different application scenarios: one is a steady state model, otherwise known as a static model; the other is a dynamic model, otherwise known as a dynamic model. The steady-state model, which takes into account only the steady-state course of the motor and not the dynamic course of the motor, is generally of the form y (f), (u), where y is the output variable and u is the input variable. The modeling method of the steady-state model is often simple, and not only a mechanism-based motor steady-state model modeling method (e.g., an equivalent circuit method) can be used, but also a data-based motor steady-state model modeling method (e.g., a neural network, a support vector machine, a random forest, etc.) can be used, which is not described herein again. Compared with a steady-state model, the establishment process of the motor dynamic model is often more complicated because the dynamic model not only needs to include steady-state information when the motor operates stably, but also needs to include dynamic information when the state of the motor changes. The dynamic model of the electric machine is generally described in the academic and industrial sectors in the form of a state equation based on mechanism, that is, dx/dt is in the form of Ax + Bu, where x is a state variable of the electric machine, u is an input variable of the electric machine, and a and B are coefficient matrices of x and u, respectively. The modeling method of the motor dynamics model studied by scholars is based on mechanism, so that the modeling method is called a model driving method, and specifically, the motor dynamics model is established by using a small signal analysis method on the basis of an equivalent circuit model. The research on the data-driven motor dynamics model modeling method is still in a blank state so far. The reason for this is that research on the dynamic model modeling method of the data-driven system is still not deep enough at the present stage, and particularly, a modeling method of the data-driven dynamic model with a compact structure and interpretability cannot be established yet. For example, a differential neural network is a data-driven dynamic model modeling method proposed in recent two years, and the method can establish a neural network type dynamic model of a system, but the model structure is often too complex and lacks interpretability, and cannot meet the requirements of high real-time performance, high reliability and high robustness in the field of motors.

Disclosure of Invention

The invention solves the problems: the method overcomes the defects of the prior art, provides a data-driven modeling method for the dynamics model of the asynchronous motor, and has the advantages of simple structure, easy explanation, no influence of model errors, high noise robustness and the like.

The idea of the method of the invention is as follows: the invention relates to a data-driven modeling method for a dynamics model of an asynchronous motorThe method comprises the following steps: firstly, in the actual operation process of the asynchronous motor, the operation data of the motor at each moment is collected, and the method comprises the following steps: d-q axis voltage data v of motor_sd、v_sq(ii) a D-q axis current data i of motor_sd、i_sqRotor position data theta of the motor and rotational speed data omega of the motor_r(ii) a The motor dynamics model established by the second step is in a form of dx/dt xi theta (x, u), wherein x is a state vector of the dynamics model, u is an input vector in the dynamics model, theta (x, u) is a dictionary library formed by the state x and the input u, xi is a sparse coefficient matrix, and meanwhile v is divided into a first step and a second step_sd、v_sqAnd ω_rThe input vector u as the kinetic model is denoted as u ═ v_sd,v_sq,ω_r]I is to_sd、i_sqAnd θ as the state x of the kinetic model, denoted as x ═ i_sd,i_sq,θ](ii) a And thirdly, identifying a sparse coefficient matrix xi of the dynamic model by using the data collected in the first step and using a data-driven nonlinear dynamics sparse representation method, so as to establish a final data-driven asynchronous motor dynamic model, wherein the model can be used as a substitute model or an improved model of a traditional motor model, and thus, researchers are guided to research and develop a new motor control technology.

The specific method comprises three stages:

the data acquisition stage is used for acquiring the operation data of the motor at each moment in the actual operation process of the asynchronous motor, and comprises the following steps: collecting d-q axis voltage data of the motor at t moments, and recording the data as [ v [ [ v ]_sd1,v_sq1,…,v_sdt,v_sqt](ii) a Collecting d-q axis current data of the motor at t moments, and recording as [ i_sd1,i_sq1,…,i_sdt,i_sqt]Collecting rotor position data [ theta ] of the motor at t moments₁,…,θ_t]And speed data [ omega ] of the motor_r1,…,ω_rt]；

A kinetic model establishing stage, wherein the structure of the motor kinetic model established in the stage is in a dx/dt-xi Θ (x, u) form, wherein x is a state vector of the kinetic model, u is an input vector in the kinetic model, and Θ (x, u) is formed by the state x and the input uAnd xi in the dictionary library, wherein xi is a sparse coefficient matrix. V is to be_sd、v_sqAnd ω_rThe input vector u as the kinetic model is denoted as u ═ v_sd,v_sq,ω_r]I is to_sd、i_sqAnd θ as the state x of the kinetic model, denoted as x ═ i_sd,i_sq,θ]；

And thirdly, a sparse coefficient matrix identification stage, wherein the sparse coefficient matrix xi is identified by using a data-driven nonlinear dynamics sparse representation method. The method comprises the following specific steps:

1, constructing a dictionary baseWhereinRepresenting all possible pairwise combinations of x and u, a specific extension may be written 23 dimensions in total;

2, establishing a dynamics model dx/dt ═ xi (x, u), wherein xi is a sparse matrix denoted xi [ xi ]₁,ξ₂,ξ₃]In which ξ_1,2,3Called a sparse vector, and ξ_1,2,3＝[0,0,0,…,a,…,0]A represents a particular value, noting that each xi dimension is the same as Θ (x, u), also 23 dimensions;

3, determining a specific form of the sparse vector by using LASSO regression in the statistical theory, wherein the calculation formula is as follows:

where k is 1 … 3 ξ_kRepresents the optimal sparse vector for the image to be processed,representing sparse vectors in the estimation process, | · |. luminance₂Representing a L2 norm, | · | | non-woven₁Represents the L1 norm;

4, will estimate the completed xi_1,2,3Substituting the kinetic model in step 2 to obtain a final model, which is recorded as:

note that while Θ (x, u) has 23 dimensions, it is due to ξ_1,2,3Is sparse, i.e. ξ_1,2,3Most elements in the model are 0, so that the final result is usually only 4-5 dimensions, the complexity is greatly reduced, and the established model has the characteristics of simple structure and easy explanation.

Compared with the prior art, the invention has the following advantages: compared with the traditional model-driven asynchronous motor dynamics modeling method, the data-driven asynchronous motor dynamics model established by the invention is not influenced by model errors, and has higher simulation accuracy on actual working conditions. Compared with the existing data-driven dynamic modeling method, the data-driven asynchronous motor dynamic model established by the invention is simple in structure and easy to explain.

Drawings

FIG. 1 is a data-driven modeling method for a dynamics model of an asynchronous motor according to the present invention;

FIG. 2 illustrates the environment required during the motor operation data acquisition phase;

fig. 3 is a comparison result between the dynamic model of the asynchronous motor established by the method of the present invention and the dynamic model established by the conventional method under the same working condition.

Detailed Description

The invention is further described with reference to the following figures and detailed description.

As shown in fig. 1, the data-driven modeling method for the dynamics model of the asynchronous motor comprises the following specific steps:

1. motor operation data acquisition

The environment required by the motor operation data acquisition stage is shown in fig. 2, the tested motor is arranged on the motor counter-dragging rack, the tested motor operates in a rotating speed loop mode, and the dynamometer is used for providing load torque for the tested motor. The dynamometer can be a motor operating in a torque loop mode, and can also be a magnetic powder brake or an eddy current brake.

After the tested motor runs, a data collector is used for collecting d-q axis voltage data v of the motor_sd,v_sqD-q axis current data i of motor_sd,i_sqMotor rotor position data theta and motor rotational speed data omega_r. The motor is driven at different rotating speeds and different torques for 1000 seconds. Wherein, data is collected once every 10 milliseconds, 100000 groups of data are collected in total, and the data is marked as [ [ v ]_sd1,v_sq1,…,v_sdt,v_sqt],[i_sd1,i_sq1,…,i_sdt,i_sqt],[θ₁,…,θ_t],[ω_r1,…,ω_rt]]Wherein t is 100000. And after the data acquisition is finished, the data is sent to a computer through a data acquisition unit for algorithm calculation.

2. Motor dynamics model establishment

The motor dynamics model established at this stage has a structure of the form dx/dt · xi Θ (x, u), where x ═ i_sd,i_sq,θ]Is the state vector of the kinetic model, u ═ v_sd,v_sq,ω_r]And theta (x, u) is an input vector in the dynamic model, a dictionary library formed by the state x and the input u is formed, and xi is a sparse coefficient matrix.

3. Solution of sparse coefficient matrix

This stage uses a data-driven nonlinear dynamics sparse representation method to solve for the numerical value of the sparse coefficient matrix xi. Firstly, constructing a dictionary libraryWhereinAll possible pairwise combinations of x and u are indicated,concrete writing 23 dimensions in total; then let xi ═ xi₁,ξ₂,ξ₃]In which ξ_1,2,3Called a sparse vector, and ξ_1,2,3＝[0,0,0,…,a,…,0]A represents a particular value, noting that each xi dimension is the same as Θ (x, u), also 23 dimensions; and then each group of data acquired in the motor operation data acquisition stage is iterated by using a LASSO regression algorithm in a statistical theory, wherein the calculation formula is as follows:

where k is 1,2,3 denotes a certain dimension, x_kState vector, ξ, representing a certain dimension_kRepresents the optimal sparse vector for the image to be processed,representing sparse vectors in the estimation process, | · |. luminance₂Representing a L2 norm, | · | | non-woven₁Represents the norm L1, 0<λ<1 denotes a penalty factor.

Iterating 100000 times by using the above formula, and finally obtaining xi₁,ξ₂,ξ₃And substituting dx/dt ═ xi Θ (x, u), finally obtaining a dynamic model of the asynchronous machine as:

the dynamic model of the asynchronous motor built by the method is compared with the dynamic model built by the traditional method. The traditional model is based on an equivalent circuit of an asynchronous motor, and the specific model form is as follows:

wherein psi_rdIs d-axis flux linkage, omega, of the rotor side of the machine_rIs the mechanical speed of the motor. T is_s＝L_s/R_s、T_r＝L_r/R_rReferred to as stator and rotor time constants of the asynchronous machine, respectively, where L_s、L_rRepresenting stator and rotor inductances, R_s、R_rThe stator and rotor resistances are shown separately. The stator and rotor inductances can be divided into L_s＝L_m+L_σsAnd L_r＝L_m+L_σrWherein L is_mFor exciting the inductance, L_σs、L_σrRespectively, the leakage inductance of the stator and the rotor, and finally,referred to as the leakage flux coefficient.

Fig. 3 shows the comparison result between the dynamic model of the asynchronous motor built by the method of the present invention and the dynamic model built by the conventional method under the same working condition. The comparative working condition is the change condition of the q-axis current in the process of switching from the electric state to the power generation state of the motor. Wherein the solid line is the real data of the motor operation; the long dotted line is the q-axis current data simulated under the condition of the asynchronous motor dynamics model established by using the method; the short dashed line is the simulated q-axis current data using a conventional asynchronous machine dynamics model. It can be found that the asynchronous motor dynamic model established by the method of the invention can better describe the real operation condition of the motor, and the data simulated by the asynchronous motor dynamic model established by the traditional method has some differences compared with the real data. Specifically, the simulation accuracy of the conventional asynchronous motor dynamic model on the real condition is 78.4%, and the simulation accuracy of the asynchronous motor dynamic model established by the method is 84.8%. This is because the traditional modeling method is based on the equivalent circuit of the asynchronous motor, and because the equivalent circuit of the asynchronous motor is an approximation and linearization to the real motor model, there is a gap with the strongly coupled and nonlinear real motor model. The invention is based on the actual motor operation data, and can describe the actual condition of the motor operation more accurately.

Table 1 shows the results of the comparison of the method of the present invention with other data-driven dynamics modeling methods, including random forests, support vector machines, and differential neural networks. The comparison indexes include model complexity o (n) (the larger n represents the more complex the model), model order (the higher the order, the more complex the model), and whether the model is an explicit model or an implicit model (the explicit model refers to whether the structure of the model can be expressed in an analytic form, that is, in a form written as dx/dt ═ f (x) + g (u), and the implicit model refers to a form of expression in which the model is not analyzed, and can only be expressed by a structure such as a binary tree or a neural network). Compared with other data-driven dynamics modeling methods, the asynchronous motor dynamics model established by the method is simpler in structure, is an explicit model and is more interpretable.

TABLE 1