CN109120191B

CN109120191B - Brushless direct current motor position sensing method based on LSSVM hierarchical classification

Info

Publication number: CN109120191B
Application number: CN201811180639.7A
Authority: CN
Inventors: 秦斌; 王欣; 秦羽新
Original assignee: Hunan University of Technology
Current assignee: Hunan University of Technology
Priority date: 2018-10-09
Filing date: 2018-10-09
Publication date: 2022-07-15
Anticipated expiration: 2038-10-09
Also published as: CN109120191A

Abstract

The invention provides a brushless direct current motor position sensing method based on Least Square Support Vector Machine (LSSVM) hierarchical classification, aiming at the problem of brushless direct current motor rotor position detection. The method provided by the invention takes the stator voltage and current of the brushless direct current motor as the input of the decision LSSVM, the rotor position information as the output, the rotor position of the direct current motor is divided into 6 areas, each LSSVM outputs the combination of the corresponding area classes, and the decision is gradually layered until the position of the corresponding area of the rotor is distinguished; the LSSVM network is trained through a grid optimization method to determine LSSVM optimal parameters, a trained network model is applied to motor operation, motor stator voltage and current are collected to serve as input of the LSSVM, final rotor position information is determined through layered decision, logical commutation signals are calculated through rotor positions, and on-off of corresponding switch tubes corresponding to each area, namely commutation logical signals, is determined.

Description

Brushless direct current motor position sensing method based on LSSVM hierarchical classification

Technical Field

The invention relates to a rotor position sensing method in the field of brushless direct current motors, in particular to a brushless direct current motor position sensing method based on Least Square Support Vector Machine (LSSVM) hierarchical classification.

Background

The brushless direct current motor controls the electronic commutation circuit through a rotor position signal to enable each winding of the stator armature to conduct commutation continuously, so that a stator magnetic field and a rotor permanent magnetic field always keep a space angle of about 90 degrees, and torque is generated to push the rotor to operate.

The rotor position information of the traditional brushless direct current motor is measured by a position sensor, and a position detection device needs to be installed, but the brushless direct current motor with the position detection device has the following defects: the motor volume is increased, which is not beneficial to the miniaturization of the motor; the position sensor is arranged in a limited space in the motor, so that the position sensor is difficult to install and maintain; difficult to adapt to harsh environments; the sensor is complicated to wire and is prone to introduce interference. Therefore, brushless dc motors without position sensors have become a focus of research.

The position measurement technology of the existing rotor position sensorless device mainly comprises a back electromotive force method, a current detection method and an intelligent algorithm. The back electromotive force is in direct proportion to the speed, so that a zero-crossing signal cannot be obtained by detecting the back electromotive force when the rotating speed is very low or even zero; the implementation of the current method mainly depends on the accuracy of the current sensor, and generally, the sensor can better obtain a signal of the current magnitude, and small changes of the current phase and the waveform are difficult to detect. Therefore, the counter-potential method and the current method have limitations when used alone.

Disclosure of Invention

The technical problem is as follows: the rotor signal detection methods with position sensors and various devices without position sensors have limitations, so that the method is difficult to be applied to occasions with higher requirements on the operation of the motor.

The technical scheme is as follows: in order to solve the problems, the LSSVM is used for rotor position detection in a layered classification mode, and the motor can stably operate according to requirements through current speed control. For a brushless direct current motor, a mapping model between a rotor position signal and motor voltage and current is established, and LSSVM hierarchical classification is adopted to realize the mapping. And the motor voltage and current are used as the input of the LSSVM, and the rotor position information is used as the output, so that the judgment of the motor rotor position is realized.

The invention discloses a rotor position detection system based on an LSSVM (least square support vector machine). the working principle of a brushless motor shows that the induction flux linkage of a phase winding of a winding A, B, C is a function of an electric angle theta, and the induction flux linkage of a phase winding of a winding A, B, C has a certain relation with the voltage and the current of a three-phase end, so that the electric angle theta, namely a rotor position signal, can be predicted according to the voltage and the current of the three-phase end.

The LSSVM is a binary classification model, and the purpose of the LSSVM is to find a hyperplane to segment samples, and the segmentation principle is interval maximization. The general form of the multivariate nonlinear classification model is described as follows: y is _i＝sgn(g(x_i) Therein) are provided with

i is 1,2, …, N represents the input quantity of LSSVM classification prediction model, y_iRepresenting the model target output quantity. Mapping function phi (x)_i):

And mapping the samples of the original input space into a high-dimensional feature space omega by using a kernel function, and performing linear classification on the sample data in the feature space by using the mapping function.

The LSSVM classification function may be expressed as: g (x) ═ ω · Φ (x)) + b (1)

Wherein the vector ω ∈ RⁿThe bias b ∈ R. The SVM minimizes the sample error while using the principle of structure risk minimization, the function fitting problem can be described as an optimization problem:

in the formula

Is a function that maps input data to a high-dimensional feature space; omega ∈ Rⁿ，e_iAs an error, e_i，b∈R，C>And 0 is a penalty coefficient used for controlling the smoothness of the solution, the larger the value of the penalty coefficient is, the stronger the penalty degree of the error is, and T is a transposition.

Converting the model into dual space according to equation (1) to solve, the following Lagrange function is obtained:

in the formula alpha_iE.g. R is Lagrange multiplier, for omega, e respectively_i,b,α_iCalculating partial derivatives, and making the partial derivatives be 0:

elimination of omega, e_iAnd (4) finishing to obtain a linear equation system:

wherein y is [ y ]₁…y_N]，α＝[α₁…α_N]，E^T＝[1…1]Ω is the kernel matrix, G is the kernel function satisfying Mercer's theorem:

solving the formula (5) to obtain α, b, and obtaining the nonlinear regression function as:

Determining a classification result according to the signs of g (x), wherein: optimal lagrange multiplier alpha_i(ii) a b is a bias term.

The LSSVM is a learning method based on kernels, and kernel function selection has important influence on the LSSVM performance. To this end, the present invention builds classifiers using the 3 kernel functions shown in table one, respectively.

Table one used 3 kinds of kernel functions

Name of Kernel function	Expression of kernel functions
		Polynomial kernel function	G_P(x_i,x)＝(x^Tx_i+1)^u(u∈N)
RBF kernel function	G_R(x_i,x)＝exp(-‖x-x_i‖²/σ²)(σ≠0∈R)
		Sigmoid kernel	G_S(x_i,x)＝tanh(a(x^Tx_i)+c)(a,c∈R)

In table I, G_P，G_R，G_SRespectively representing a polynomial kernel, an RBF kernel and a Sigmoid kernel. U in the polynomial kernel function is used for setting the highest term degree of the polynomial kernel function; sigma in the RBF kernel function is a width parameter of the function, and the radial action range of the function is controlled; and a and c in the Sigmoid kernel function are used for setting parameters in the kernel function.

The invention provides a brushless direct current motor position sensing method based on LSSVM (SVM) hierarchical classification, A, B phase voltage u_a(k),u_b(k) Current i_a(k),i_b(k),i_a(k-1),i_b(k-1) as the input of LSSVM, S (K) as the rotor position signal, taking the rotor position signal as the output of LSSVM, outputting the combination of the corresponding region classes by each LSSVM, and gradually making layered decision until the position of the corresponding region of the rotor is distinguished. G_j(x_iX) is the kernel function, j is 1,2, …, L are L classifiers, K is the time series, and K is the rotor position.

The invention provides a position detection algorithm based on SVM layered classification, which mainly comprises an LSSVM layered classification modeling and model operation part.

The LSSVM layered classification modeling part mainly comprises the following steps:

step 1: for the input and output detection signals of the brushless direct current motor acquisition system with the same type of the position sensor: A. phase B voltage u_a(k),u_b(k) Current i_a(k),i_b(k),i_a(k-1),i_b(k-1) serving as the input of the LSSVM, S (K) serving as a rotor position signal, serving as the output of the LSSVM, dividing an electric angle theta of 0-360 degrees of rotation of a rotor of the direct current motor into 6 areas of 60 degrees, wherein the electric angle theta is the number of each area, the positions of the rotor are represented by the serial numbers 1-6 of the areas, and the measured training data and the measured test data are normalized;

step 2: setting 5 LSSVM2 classifiers, wherein the first layer 1 classifier determines the samples of two selected categories (such as 1-2) as positive samples, the samples of the rest 4 categories (such as 3-6 categories) as negative samples, the second layer 2 classifier determines the data of 2 categories (such as 1-2 categories) as positive samples and negative samples respectively by the first layer, thereby distinguishing and identifying (category 1 and category 2), the data of 4 categories determined by the first layer is applied to the second classifier, and is divided into 2 two categories (such as 3-4 categories and 4-5 categories), the second layer 2 classifier is further distinguished by the 3 layer 2 classifier, and in this way, 5 such two categories can be obtained, wherein one of the structures is shown in the table two.

Example of a two-level classifier

Determining a kernel function G_jThe kernel function selects one of a polynomial, a radial basis function, and a sig function.

Step 3: and respectively selecting training samples of corresponding classes according to the output of the classifier, wherein the positive sample is +1, the negative sample is-1, respectively training the 5 LSSVM models by adopting a Vapnik algorithm, and verifying and testing by adopting a grid optimization method to obtain an optimal model parameter punishment coefficient C and a kernel function parameter so as to obtain 5 optimal LSSVM two classifiers.

The LSSVM hierarchical classification operation part mainly comprises the following steps:

step 1: collecting related voltage and current input signals in real time and normalizing;

step 2: inputting related voltage and current input signals into an established LSSVM classifier to obtain a classification result of a region Ki where a rotor is located;

when the classification belongs to the

classes

1, 2, … and 6, two or more results are obtained, the result of the last classification is kept unchanged; when the classification result belongs to the

classes

1, 2, … and 6, the last classification result is kept unchanged.

Step3 performs real-time control according to the classified rotor position. And calculating a logic commutation signal through the position of the rotor, and determining the on-off of each region corresponding to the corresponding switch tube, namely the commutation logic signal. The transformation relation between the rotor position information and the commutation logic signal is shown in table three, wherein 1 represents on, and 0 represents off.

Conversion relation table for three-rotor position information and phase conversion logic signal

Has the beneficial effects that: the position sensing method has the advantages of good dynamic performance, high robustness and the like. The operation speed of the algorithm is high, and the reaction speed of the controller is improved.

Drawings

Fig. 1 is a structural diagram of a brushless direct current motor position sensing based on LSSVM hierarchical classification.

The specific implementation mode is as follows:

the invention provides a brushless direct current motor position sensing method based on LSSVM layered classification, which combines a system structure diagram and a specific implementation scheme thereof to be detailed as follows:

step 1: collecting the same model brushless DC motor with position sensorSystem input/output detection signal: A. phase u of B voltage_a(k),u_b(k) Current i_a(k),i_b(k),i_a(k-1),i_b(k-1) as the input of the LSSVM, and S (K) as the output of the LSSVM, wherein the rotor position signal is obtained by dividing the electrical angle theta of 0-360 degrees of the rotor rotation of the DC motor into 6 regions of 60 degrees, and the rotor positions are represented by the region numbers 1-6.

Step 2: the method is characterized in that 5 LSSVM2 classifiers are arranged in total, the first layer 1 classifier determines samples of two selected categories 1-2 as positive samples, the remaining 4 categories 3-6 samples are determined as negative samples, the second layer 2 classifier determines the category 1-2 data and continues to determine the samples of the

Claims

1. A brushless direct current motor position sensing method based on Least Square Support Vector Machine (LSSVM) hierarchical classification is characterized in that a LSSVM hierarchical classifier is used for obtaining position signals of a rotor of a brushless direct current motor, so that errors of the motor caused by the existence of a position sensor are reduced, and the size of the motor is reduced; the method adopts an LSSVM two-classification structure, determines final classification output through hierarchical decision, and mainly comprises an LSSVM hierarchical classification modeling and model operation part:

1) The LSSVM layered classification modeling part mainly comprises the following steps:

step 1: inputting and outputting detection signals to and from a brushless direct current motor acquisition system with a position sensor of the same type, taking A, B phase voltages ua (K), ub (K) and currents ia (K), ib (K), ia (K-1) and ib (K-1) as the input of an LSSVM, taking a rotor position signal S (K) as the output of the LSSVM, wherein each LSSVM output corresponds to an area sample set, K is a time sequence, and K is the rotor position, and making decisions continuously until each category is separated; dividing the rotor of the direct current motor into 6 areas with an electric angle of 0-360 degrees of rotation, wherein each area is 60 degrees, the area where the rotor is located is represented by a class serial number of 1-6, and performing normalization processing on the measured training data and the measured test data;

step 2: setting 5 LSSVM two classifiers in total, wherein the first layer comprises six classes from class 1 to class 6, samples of the two selected classes are taken as positive samples, and samples of the remaining four classes are taken as negative samples; the second layer two second classifiers are used for respectively defining two categories in the positive samples of the first layer as positive samples and negative samples so as to distinguish the two categories; the second layer of second classifiers divides the four classes in the first layer of negative samples into two groups, each group comprises two classes, and two second classifiers in the third layer are obtained, the two classes contained in the two second classifiers in the third layer are respectively continuously distinguished, so that the remaining four classes can be distinguished, and 5 binary classifiers are obtained in total; determining a kernel function Gj, wherein the kernel function selects one of a polynomial, a radial basis and a sig function;

Step 3: respectively selecting training samples of corresponding classes according to the output of the classifiers, wherein the positive sample is +1, the negative sample is-1, training the 5 LSSVM two classifiers by adopting a Vapnik algorithm, and verifying and testing by adopting a grid optimization method to obtain an optimal model parameter penalty coefficient C and a kernel function parameter to obtain 5 optimal LSSVM two classifiers;

2) the LSSVM layered classification model operation part mainly comprises the following steps:

step 2: inputting related voltage and current input signals into the established LSSVM classifier to obtain a classification result of the region where the rotor is located;

when the classification belongs to the classes 1, 2, … and 6, two or more results are obtained, the result of the last classification is kept unchanged; when the classification result belongs to the classes 1, 2, … and 6, the classification result is kept unchanged;

step 3: performing real-time control according to the classified rotor position; and calculating a logic commutation signal through the position of the rotor to determine the on-off of the switching tube corresponding to each area.