CN102945324A - Multi-model least square support vector machine (LSSVM) modeling method of brushless direct current motor - Google Patents

Abstract

The invention discloses a multi-model least square support vector machine (LSSVM) modeling method of a bearingless brushless direct current motor. According to the data driving principle, proper signals are utilized for exciting the bearingless brushless direct current motor, input and output samples are obtained, the samples are clustered according to the input set and the output set through the affinity propagation clustering, subclasses subjected to secondary clustering are subjected to the least square support vector machine fitting, a local LSSVM module is built, and further, the weighting and the form are adopted for constructing a system model of the bearingless brushless direct current motor. The multi-model LSSVM modeling method does not depend on the system mechanism and the specific parameters, a bearingless brushless direct current motor system is resolved, the modeling difficulty is lowered, and the modeling precision is improved.

Description

The multi-model modeling method of least squares support method of brshless DC motor

Technical field

The present invention relates to a kind of multi-model modeling method of least squares support method of DC motor without bearing and brush, be applicable to the technical field of electric drive control.

Background technology

Brshless DC motor has good speed adjusting performance, efficient height, is convenient to control, simple in structure, reliable, easy to maintenance, is widely applied at industrial circle.Yet conventional mechanical bearings can cause the problems such as motor vibrations, noise, wearing and tearing, heating, life-span weak point, and the high speed and the hypervelocity that have a strong impact on motor reliably turn round.The problem of bringing for solving conventional mechanical bearings, grow up on the magnetic suspension bearing basis without the bearing technology, utilize the similarity of magnetic suspension bearing and electric machine structure, the magnetic suspension bearing winding that produces suspending power is inserted motor stator, by the decoupling zero control to torque winding and suspending power winding, realize the stable suspersion of rotor, make rotor have simultaneously the function that produces torque and certainly suspend.In operational process, do not need the mechanical bearing support except having rotor, also inherited the advantages such as the magnetic bearing motor is unlubricated, nothing wearing and tearing, machinery-free noise.

With brshless DC motor and the DC motor without bearing and brush that combines and consist of without the bearing technology, on the basis of traditional brshless DC motor, in stator, add suspending windings, by torque winding and suspending windings electric current are controlled, realize the suspension operation of motor.Its advantage is: 1) rotating speed is high, and power density is high, and volume is little; 2) can directly link to each other with high speed prime mover, cancel reducing gear, transmission efficiency is high; 3) moment of inertia is little, and dynamic response is fast; 4) machinery-free friction, long service life, low in energy consumption.These advantages so that DC motor without bearing and brush be with a wide range of applications in fields such as Aero-Space, biological medicine, semiconductor manufacturings.

Set up accurately system model, significant to the DC motor without bearing and brush control system design.DC motor without bearing and brush has also added the suspending power winding except the torque winding, the design feature of self has determined that it is the nonlinear system of a complexity.There is coupling between the suspending power that the torque that the motor torque winding produces and suspending windings produce, and also has coupling between the magnetic levitation component of force.Therefore, DC motor without bearing and brush has multivariate, non-linear, strong coupling and the wide characteristics of condition range, is difficult to set up accurately mechanism model.

Consider contact and difference between input, the output data, and be cluster and migrate attribute more than the field samples data and by the working point in the actual industrial production, start with from inputoutput data, to the sample data cluster, adopting the multi-model process modeling is the effective way that addresses this problem.Existing clustering method has: based on the Fuzzy c-means clustering algorithm of Definite kernel function, can realize the cluster to irregularly shaped data; Based on the multi-model on-line identification algorithm of subtractive clustering, according to Least Square Recurrence algorithm online updating model parameter; Based on the clustering algorithm of particle group optimizing, can alleviate traditional clustering method responsive to initial value, easily be absorbed in the problem of local optimum.But it is given in advance that these clustering algorithm ubiquity clusters number need, and precision relies on the data distribution and speed of convergence waits problem slowly; And traditional clustering method has only been utilized the importation of sample data, because inconsistency and the incompleteness of data are unfavorable for the raising of model accuracy.

Summary of the invention

Purpose of the present invention provides a kind of multi-model modeling method of least squares support method that is applicable to DC motor without bearing and brush, for the efficient control of DC motor without bearing and brush creates conditions.

The technical solution used in the present invention is: the multi-model modeling method of least squares support method of brshless DC motor is characterized in that comprising the steps:

1) with suspending power F _α, F _βAnd electric current I _sBe input, rotor radial displacement α, β and rotational speed omega _rFor output makes up the DC motor without bearing and brush governing system;

2) adopt at random that square wave encourages the DC motor without bearing and brush governing system, gather corresponding input data set x={F _α, F _β, I _s, output data set y={ α, β, ω _r; Gather the input of m group, output data, input sample of data is X=[x ₁, x ₂, L, x _m] ^T, the output data sample is Y=[y ₁, y ₂, L, y _m] ^T, x wherein _kBe that (k is natural number to k group input data set, 1≤k≤m), x _k={ F _α, F _β, I _s} _k, y _kBe k group output data set, y _k={ α, β, ω _r} _kSelect n group data sample as training sample set X _Train, Y _Train, all the other m-n group data samples are test samples collection X _Test, Y _Test(n and m are the natural number greater than 1, n＜m);

3) parameter in the setting affine propagation clustering algorithm: damping factor λ, deflection parameter p and maximum iteration time, the X that utilizes affine propagation clustering that training sample is concentrated _Train=[x ₁, x ₂, L, x _n] ^TCarry out cluster one time, hard clustering number 1C(1C is the natural number greater than 1), with training sample set X _Train, Y _TrainBe divided into some classes: X ₁₁, Y ₁₁X ₁₂, Y ₁₂ X _1C, Y _1C

4) the subclass collection X to obtaining after the cluster _1j, Y _1j(1j is natural number, 1≤1j≤1C), press Y _iCarry out the secondary affine propagation clustering, determine that total clusters number C(C is the natural number greater than 1);

5) with { F _α, F _β, I _sBe input, { α, β, ω _rBe output, utilize each the subclass X after the secondary cluster _i, Y _iData set up local LSSVM model (i be natural number, 1≤i≤C); Select Gaussian function as the LSSVM kernel function, adopt cross-validation method commonly used to obtain suitable regularization parameter γ and nuclear parameter σ, obtain respectively LSSVM ₁..., LSSVM _C(i=1, L, C);

6) with the output weighting of each local LSSVM model, weights be sample to be tested to the fuzzy membership of each subclass, obtain the output of DC motor without bearing and brush governing system namely

F wherein _iBe the output valve of i sub-LSSVM, μ _iDegree of membership value for correspondence.

The invention has the beneficial effects as follows:

1, adopt the secondary affine propagation clustering, solved clusters number and needed the in advance given problem that distributes with nicety of grading dependence data, the secondary cluster is conducive to the Further Division of sample space.

2, least square method supporting vector machine has accurately characteristics, the counting yield height of Approximation of Arbitrary Nonlinear Function, uses it for the structure of submodel, makes each partial model can reflect well system's local feature.

3, multi-model modeling method of least squares support method does not rely on system mechanism and design parameter, and system decomposes to DC motor without bearing and brush, has reduced the modeling difficulty, has improved modeling accuracy.

Description of drawings

Fig. 1 is the structure block diagram of DC motor without bearing and brush governing system 4, has: DC motor without bearing and brush 1, Current Control PWM frequency converter 2, Current Control PWM frequency converter 3.

Fig. 2 is the multi-model LSSVM modeling process flow diagram of DC motor without bearing and brush governing system 4.

Embodiment

The present invention is according to the data-driven principle, utilize suitable signal that DC motor without bearing and brush is encouraged, obtain the input and output sample, utilize affine propagation clustering with sample by input set, output clustering, the subclass that obtains after the secondary cluster is carried out the least square method supporting vector machine match, set up local LSSVM model, then adopt the system model of weighted sum formal construction DC motor without bearing and brush.Implementation is as follows:

1, as shown in Figure 1, regard DC motor without bearing and brush 1 and Current Control PWM frequency converter 2, Current Control PWM frequency converter 3 as an integral body, by the field orientation principle suspending power frequency converter 2 is set, by the control of brshless DC motor torque frequency converter 3 is set, with suspending power F _α, F _βAnd electric current I _sBe input, rotor radial displacement α, β and rotational speed omega _rBe output, its whole equivalence is DC motor without bearing and brush governing system 4.

2, adopt the at random square wave of realistic operation to encourage DC motor without bearing and brush governing system 4, gather corresponding input data set x={F _α, F _β, I _s, output data set y={ α, β, ω _r.Gather 100 groups of inputs, output data, make that input sample of data is X=[x ₁, x ₂, L, x _m] ^T, the output data sample is Y=[y ₁, y ₂, L, y _m] ^T, m=100.Select 60 groups of data samples as training sample set X _Train, Y _Train, all the other 40 groups of data samples are test samples collection X _Test, Y _Test

3, the parameter in the setting affine propagation clustering algorithm: damping factor λ, deflection parameter p and maximum iteration time, the X that utilizes affine propagation clustering that training sample is concentrated _Train=[x ₁, x ₂, L, x ₆₀] ^TCarry out cluster one time, hard clustering number 1C is with training sample set X _Train, Y _TrainBe divided into some classes: X ₁₁, Y ₁₁, X ₁₂, Y ₁₂..., X _1C, Y _1C

4, the subclass collection X to obtaining after the cluster _i, Y _i(i=11, L, 1C) presses Y _iCarry out the secondary affine propagation clustering, determine total clusters number C.

5, with { F _α, F _β, I _sBe input, { α, β, ω _rBe output, utilize each the subclass X after the secondary cluster _i, Y _iThe data of (i=1, L, C) are set up local LSSVM model.Select Gaussian function as the LSSVM kernel function, adopt cross-validation method commonly used to obtain suitable regularization parameter γ and nuclear parameter σ, obtain respectively LSSVM ₁..., LSSVM _C

6, with the output weighting of each local LSSVM model, weights be sample to be tested to the fuzzy membership of each subclass, obtain the output of DC motor without bearing and brush governing system 4, namely

F wherein _iBe the output valve of i sub-LSSVM, μ _iDegree of membership value for correspondence.The multi-model LSSVM modeling flow process of DC motor without bearing and brush governing system 4 as shown in Figure 2.

Claims (1)

1. the multi-model modeling method of least squares support method of a brshless DC motor is characterized in that comprising the steps:

1) with suspending power F _α, F _βAnd electric current I _sBe input, rotor radial displacement α, β and rotational speed omega _rFor output makes up the DC motor without bearing and brush governing system;

2) adopt at random that square wave encourages the DC motor without bearing and brush governing system, gather corresponding input data set x={F _α, F _β, I _s, output data set y={ α, β, ω _r; Gather the input of m group, output data, input sample of data is X=[x ₁, x ₂, L, x _m] ^T, the output data sample is Y=[y ₁, y ₂, L, y _m] ^T, x wherein _kBe k group input data set, x _k={ F _α, F _β, I _s} _k, y _kBe k group output data set, y _k={ α, β, ω _r} _kSelect n group data sample as training sample set X _Train, Y _Train, all the other m-n group data samples are test samples collection X _Test, Y _Test, wherein k is natural number, 1≤k≤m, n and m are the natural number greater than 1, n＜m;

3) parameter in the setting affine propagation clustering algorithm: damping factor λ, deflection parameter p and maximum iteration time, the X that utilizes affine propagation clustering that training sample is concentrated _Train=[x ₁, x ₂, L, x _n] ^TCarry out cluster one time, hard clustering number 1C is with training sample set X _Train, Y _TrainBe divided into some classes: X ₁₁, Y ₁₁X ₁₂, Y ₁₂ X _1C, Y _1C, wherein 1C is the natural number greater than 1;

4) the subclass collection X to obtaining after the cluster _1j, Y _1j, press Y _iCarry out the secondary affine propagation clustering, determine total clusters number C, 1j is natural number, and 1≤1j≤1C, C are the natural number greater than 1;

5) with { F _α, F _β, I _sBe input, { α, β, ω _rBe output, utilize each the subclass X after the secondary cluster _i, Y _iData set up local LSSVM model; Select Gaussian function as the LSSVM kernel function, adopt cross-validation method commonly used to obtain suitable regularization parameter γ and nuclear parameter σ, obtain respectively LSSVM ₁..., LSSVM _C, i=1, L, C;

6) with the output weighting of each local LSSVM model, weights be sample to be tested to the fuzzy membership of each subclass, obtain the output of DC motor without bearing and brush governing system namely

F wherein _iBe the output valve of i sub-LSSVM, μ _iDegree of membership value for correspondence.

Images