CN101917150A - Fuzzy Neural Network Generalized Inverse Robust Controller for Permanent Magnet Synchronous Motor and Its Construction Method - Google Patents

Abstract

The invention discloses a fuzzy neural network generalized inverse robust controller for permanent magnet synchronous motors and a construction method thereof. The compound controlled object is composed of an internal model controller and a fuzzy neural network generalized inverse phase; two linear transfer functions and one An integrator is connected in series with the fuzzy neural network whose parameters and weight coefficients are determined to form the generalized inverse of the fuzzy neural network. The generalized inverse of the fuzzy neural network is connected in series with the compound controlled object to form a generalized pseudo-linear system, and the PMSM is linearized and decoupled equivalently. A second-order velocity pseudo-linear subsystem and a first-order current pseudo-linear subsystem are formed; the two pseudo-linear subsystems are respectively introduced into the internal model control method to construct an internal model controller. The invention overcomes the dependence of the optimal gradient method on the initial value and the random and probabilistic problems caused by local convergence and the simple genetic algorithm, obtains high-performance control of the motor, resistance to load disturbance and adaptability, and simplifies the difficulty of control , the structure is simple, and the system robustness is high.

Description

Fuzzy Neural Network Generalized Inverse Robust Controller for Permanent Magnet Synchronous Motor and Its Construction Method

technical field

The invention relates to a permanent magnet synchronous motor controller, which is suitable for the robust control of a permanent magnet synchronous motor driven by a voltage source inverter, and belongs to the technical field of electric drive control equipment.

Background technique

Permanent magnet synchronous motor (Permanent Magnet Synchronous Motor, referred to as PMSM) has been widely used in aerospace, weapons and national defense, CNC machine tools, industrial robots, flexible control, communication industry, oil field and chemical industry, as well as fans and pumps with long annual running time. application.

The control methods of permanent magnet synchronous motor speed control system mainly include constant voltage frequency ratio control, vector control, direct torque control and differential geometric state feedback control. Among them, the permanent magnet synchronous motor speed control system under the constant voltage frequency ratio control mode based on the steady state model has a simple structure, low cost, and is easy to implement. It can meet the general speed control requirements, but the system performance is not high, and it relies too much on system dynamic mathematics The model is an open-loop control, and the load capacity at low speed is limited. When the load or speed command is suddenly increased, it is prone to out-of-step phenomenon, and the ideal dynamic control performance cannot be obtained. The permanent magnet synchronous motor speed control system under the vector control mode based on the dynamic model has the advantages of good dynamic performance, wide speed range, and high control accuracy. Widely used, however, due to the great dependence of vector control on motor parameters, the speed and flux linkage can satisfy the decoupling relationship only when the flux linkage reaches a steady state and remains constant. It is difficult to ensure complete decoupling, and the actual control effect is difficult to achieve The results of theoretical analysis, and the vector rotation coordinate transformation used in the simulation of the DC motor control process is relatively complicated, and the robustness of the system is greatly reduced. The permanent magnet synchronous motor speed control system under the direct torque control method based on the stator flux orientation is convenient to realize full digitalization, and there is no need to make the AC motor equivalent to the DC motor, and the complicated rotation coordinate transformation and motor model are omitted. In the vector control, the control effect is affected by the change of the rotor parameters. It is only necessary to detect the stator resistance and observe the stator flux linkage of the motor. It uses the torque and flux linkage hysteresis comparison to achieve partial dynamic decoupling. Defects such as large torque ripple. The differential geometry method is a method of linearization and decoupling control of nonlinear systems developed with differential geometry as a tool. The purpose is to transform the complex system into a simple linear system after accurate linearization of the nonlinear system. In this way, the linear theory can be used to analyze and design the linear controller in a wide working range without losing the controllability and accuracy of the system. At the same time, it is required to obtain accurate mathematical models and use complex and abstract mathematical tools, so it is difficult to apply in engineering.

At present, although the neural network inverse system control method can realize the linear decoupling of the permanent magnet synchronous motor, the several integral pseudo-linear subsystems formed after decoupling are open-loop unstable, and the neural network based on the empirical risk minimization The network has defects such as local minimum points, over-learning, and the selection of structures and types relies too much on experience. At the same time, in the actual operation of permanent magnet synchronous motors, there are load mutations, many system controllable parameters, unmodeled dynamic effects, and easy out-of-step. These uncertain factors cause model mismatch and make the system deviate from the expected control target.

Contents of the invention

Because the dynamic model of permanent magnet synchronous motor control is a nonlinear, strongly coupled multivariable time-varying system, in order to overcome the deficiencies of several basic control methods in the prior art above, the present invention provides a generalized fuzzy neural network for permanent magnet synchronous motors The inverse robust controller realizes the linearized decoupling control of the permanent magnet synchronous motor, suppresses the parameter perturbation and load disturbance well, overcomes the unmodeled dynamic disturbance, and improves the dynamic response speed and stability of the speed control system of the permanent magnet synchronous motor. State tracking accuracy, to achieve high-performance robust control.

Another object of the present invention is to provide a construction method of the generalized inverse robust controller of the fuzzy neural network, which can comprehensively process the decoupled pseudo-linear subsystems to ensure the control effect of the permanent magnet synchronous motor.

The technical solution adopted by the controller of the present invention is: it is composed of an internal model controller and a generalized inverse of a fuzzy neural network; the internal model controller is composed of a speed internal model controller and a current internal model controller connected in parallel, and the speed internal model control The device is composed of a speed internal model and a speed controller, and the current internal model controller is composed of a current internal model and a current controller connection; the generalized inverse of the fuzzy neural network is connected in series with the compound controlled object to form a generalized pseudolinear system, and the generalized pseudolinear system It is equivalent to a velocity sub-linear system and a current sub-linear system; the generalized inverse of the fuzzy neural network consists of a five-layer fuzzy neural network with 5 input nodes and 2 output nodes plus two 2 linear transfer functions and a Composed of an integrator; the compound controlled object includes a current speed detection and calculation module connected with an extended inverter control part that drives the PMSM. The current speed detection and calculation module includes the current speed calculation part, Park transformation, Clarke transformation and photoelectric encoder connection, and the Clarke transformation and photoelectric encoder are connected to PMSM.

The construction method of the controller of the present invention includes the following steps in turn: first, the PMSM is equivalent to the Clarke transformation, the Park transformation and the third-order model, and the PMSM is formed as a whole through the current speed detection and calculation module and the extended inverter control part to form a composite controlled The generalized inverse of the fuzzy neural network is composed of two linear transfer functions and one integrator in series with the fuzzy neural network whose parameters and weight coefficients have been determined. system, the generalized pseudolinear system linearizes and decouples the PMSM into a second-order velocity pseudolinear subsystem and a first-order current pseudolinear subsystem; finally, the second-order velocity pseudolinear subsystem and the first-order current pseudolinear subsystem The linear subsystem introduces the internal model control method to construct the internal model controller, and combines the internal model controller with the generalized pseudo-linear system to form a fuzzy neural network generalized inverse robust controller to control the compound controlled object.

The beneficial effects of the present invention are:

1. The fuzzy neural network has the advantages of strong self-learning ability, parallel computing ability, nonlinear approximation ability and fuzzy logic's strong fuzzy reasoning ability of neural network at the same time. Combining it with the linearization and decoupling characteristics of the inverse system, Use fuzzy logic technology to improve the learning ability of neural network; use the learning ability of neural network to extract fuzzy rules or adjust fuzzy rule parameters; use neural network to realize fuzzy logic system and parallel fuzzy reasoning. This combination overcomes the influence of the unmodeled dynamics of the control system, and has strong robustness and fault tolerance. A stable pseudo-linear subsystem control problem, and further reasonable construction of a linear closed-loop controller, can obtain high-performance control of the motor, anti-load disturbance and self-adaptability, which greatly simplifies the control difficulty.

2. Using the fuzzy neural network plus transfer function and integrator to construct the generalized inverse control system of the compound controlled object, completely getting rid of the dependence of the traditional control method on the mathematical model and parameters of the controlled system of the permanent magnet synchronous motor, effectively overcoming the It solves the instability of the integral fuzzy neural network inverse system, and solves the control problem caused by the difficult measurement of the partial state of the original high-order controlled system. It is a major breakthrough in the traditional inverse system control method. The generalized inverse system is composed of the original system. The generalized pseudo-linear system can not only realize the linear decoupling of the original system, but also make the poles of the pseudo-linear subsystem formed after decoupling be reasonably arranged in the complex plane by reasonably adjusting the parameters of the linear link, so that an ideal open-loop Frequency characteristics, realize large-scale linearization, decoupling and order reduction, so that the controller can be conveniently constructed according to the linear control theory for high-precision control, which is conducive to system synthesis, simple structure, high system robustness, and easy engineering implementation.

3. The method of determining and adjusting fuzzy neural network parameters and weights is the combination of genetic algorithm and optimal gradient method. The traditional fuzzy neural network simply uses the optimal gradient algorithm. Although it is simple to implement and has strong local search ability, its online learning cycle is long, the algorithm convergence speed is slow, and it is easy to fall into local minimum. Genetic algorithm as a global search and Although optimization technology itself cannot express knowledge, it has strong learning ability and optimization ability. At the same time, genetic algorithm has global parameters and network structure, and can search for about 90% of the optimal solution at a relatively fast speed. However, its late search mutation probability is small, it is difficult to maintain population diversity, and its implementation is more complicated. It is not as good as the optimal gradient method when real-time control is required. Combining the two, learning from each other, on the one hand, the genetic algorithm ensures the global convergence of network learning, overcomes the dependence of the optimal gradient method on the initial value and the local convergence problem; on the other hand, the optimal gradient learning ensures the local search ability, The online adjustment ability is strong, and the implementation is simple. At the same time, it overcomes the random and probabilistic problems brought about by the simple genetic algorithm, and helps to improve the search probability.

4. Based on the dSPACE real-time simulation system as the experimental platform, it realizes a completely seamless connection with MATLAB/Simulink/RTW. The dSPACE real-time system has the advantages of strong real-time performance, high reliability, and good scalability. The processor in the dSPACE hardware system has high-speed computing capability and is equipped with rich I/O support, which can be combined by users according to needs; the software environment is powerful and easy to use, including automatic code generation/download and test/debugging complete set of tools. Using its powerful software and hardware experiment platform can realize high-precision speed control of permanent magnet synchronous motors, complete the control algorithm of the motor from conceptual design to mathematical analysis and testing, from the realization of real-time simulation tests to the monitoring and adjustment of experimental results A set of parallel engineering, short development cycle, resource saving, powerful and easy to implement.

5. The present invention can also be applied in other types of motors such as synchronous motors, DC motors, and asynchronous motors, and in a synchronous coordination control system with a plurality of networked AC motors (multi-motors) as power devices, the application prospect broad.

The present invention will be described in further detail below in conjunction with the accompanying drawings and specific embodiments.

Description of drawings

Fig. 1 is a connection diagram of the PMSM 1 of the permanent magnet synchronous motor body and the current speed detection and calculation module 31.

Fig. 2 is a structure and a simplified equivalent model diagram of the compound controlled object 3 composed of the PMSM 1, the extended inverter control part 32 and the current speed detection and calculation module 31.

Fig. 3 is a fuzzy neural network generalized inverse system 4 structure and its equivalent model diagram;

Fig. 4 is generalized pseudo-linear system 5 structures and two sub-linear system diagrams that are equivalent to;

Fig. 5 is a structural diagram of the fuzzy neural network generalized inverse robust controller 7 of the present invention.

Fig. 6 is a functional block diagram of the fuzzy neural network generalized inverse robust controller 7 of the present invention using the dSPACE experimental platform to implement the control system.

Detailed ways

，以下均简称定子电流。As shown in FIG. 5 , the fuzzy neural network generalized inverse robust controller 7 of the present invention controls the compound controlled object 3 . The fuzzy neural network generalized inverse robust controller 7 is composed of the internal model controller 6 and the fuzzy neural network generalized inverse 4. The internal model controller 6 is composed of a speed internal model controller 61 and a current internal model controller 62 connected in parallel, wherein the speed internal model controller 61 is composed of a speed internal model 611 and a speed controller 612; the current internal model controller 62 is composed of The current internal model 621 and the current controller 622 are connected. At the same time, the fuzzy neural network generalized inverse 4 in the fuzzy neural network generalized inverse robust controller 7 is connected in series with the compound controlled object 3 to form a generalized pseudolinear system 5, which decouples the original high-order nonlinear coupling system into a A second-order speed pseudo-linear subsystem 51 and a first-order current pseudo-linear subsystem 52 . Further, the fuzzy neural network generalized inverse 4 is based on the equivalent mathematical model of the permanent magnet synchronous motor PMSM 1, aiming at the coupling between the motor speed, voltage and stator current, on the basis of analyzing the generalized reversibility of PMSM 1, using a 5 A five-layer fuzzy neural network 41 with two input nodes and two output nodes plus two linear transfer functions and an integrator is formed. Composite controlled object 3 is composed of PMSM1, current speed detection and calculation module 31 and extended inverter control part 32. It is an extended inverter composed of inverse Park transformation and voltage source inverter in SVPWM debugging mode. The device control part 32 drives the PMSM1, and at the same time connects a whole composed of the current speed detection and calculation module 31 composed of the current speed calculation part, Park transformation, Clarke transformation and photoelectric encoder 2. PMSM1 is equivalent to Clarke transformation, Park transformation and third-order model, and the third-order model is the third-order differential equations in the dq coordinate system. Among them, the current speed detection and calculation module 31 is not only an important part of the compound controlled object 3, but also a signal feedback link for the current, speed and rotor displacement as internal model control and Park transformation. It should be noted that the current signal output by the actual controlled object is the square of the stator current

, hereinafter referred to as the stator current.

，使解耦后形成的各个伪线性子系统的极点在复平面内合理配置，实现积分型不稳定子系统到稳定子系统的转变。在此基础上，将二阶速度伪线性子系统51和一阶电流伪线性子系统52分别引入内模控制方法构造内模控制器6，合理设计内模控制器6，与广义伪线性系统5结合组成模糊神经网络广义逆鲁棒控制器7以控制复合被控对象3，实现对PMSM1的高精度鲁棒控制，使得系统克服未建模动态的干扰，具有优良的动静态控制性能，抗干扰能力和高精度跟踪性能。可根据不同的控制要求采用不同的硬件或软件来实现。具体用以下7个步骤来描述：The construction method of the above-mentioned fuzzy neural network generalized inverse robust controller 7 is: first, based on the permanent magnet synchronous motor body PMSM1, the current speed detection and calculation module 31 and the calculation module 31 composed of the current speed calculation part, Clarke, Park transformation and photoelectric encoder 2 are firstly based on the permanent magnet synchronous motor body PMSM1 The extended inverter control part 32 composed of the inverse Park transformation and the voltage source inverter under the SVPWM modulation forms a whole to form a compound controlled object 3 to drive the load. Secondly, the generalized inverse of the fuzzy neural network with 2 input nodes and 2 output nodes composed of 5 input nodes, 2 output nodes fuzzy neural network 41 (5-layer network) plus 2 linear transfer functions and 1 integrator is adopted. 4 is connected in series with the compound controlled object 3 to form a generalized pseudo-linear system 5, thereby linearizing and decoupling a multi-variable, strongly coupled high-order nonlinear system such as PMSM 1 into a second-order velocity pseudo-linear subsystem 51 and a first-order current pseudo-linear subsystem 52, by rationally adjusting the parameters a _j0 , a _j1 , . . . of the linear transfer function,

, so that the poles of each pseudo-linear subsystem formed after decoupling are reasonably arranged in the complex plane, and the transformation from an integral unstable subsystem to a stable subsystem is realized. On this basis, the second-order velocity pseudo-linear subsystem 51 and the first-order current pseudo-linear subsystem 52 are respectively introduced into the internal model control method to construct the internal model controller 6, and the internal model controller 6 is reasonably designed to be compatible with the generalized pseudo-linear system 5 Combined with the generalized inverse robust controller 7 composed of fuzzy neural network to control the compound controlled object 3, high-precision robust control of PMSM1 is realized, which enables the system to overcome unmodeled dynamic disturbances, has excellent dynamic and static control performance, and is anti-interference capability and high precision tracking performance. It can be realized by using different hardware or software according to different control requirements. Specifically described in the following 7 steps:

Step 1: As shown in FIG. 1 , construct a current speed detection and calculation module 31 . In the control signal of PMSM1, the two-phase stator currents i _sA and i _sB are obtained by Hall element detection, and the two-phase stator currents i _sA and i _sB are transformed by Clarke, and then obtained by Park transformation, i _sd and i _sq are obtained by the photoelectric encoder 2 After detecting the rotating speed signal obtained by PMSM1 and i _sd , i _sq through the calculation part of current speed, the output current signal ^{is 2} =i _sd ² +i _sq ² , rotor angular velocity ω _r and angular _displacement θ are used as the current speed The output of detection and calculation module 31. The current speed detection and calculation module 31 simultaneously serves as the output of the compound controlled object 3 described below and provides feedback signals for the internal model controller 6 .

Step 2: As shown in Figure 2, construct the compound controlled object 3 of PMSM1. The composite controlled object 3 is composed of an extended inverter control part 32 , a mathematical model of an equivalent PMSM1 and the above-mentioned current detection and calculation module 31 . Equivalent to the integral composition of the extended inverter control part 32 and the PMSM1 body, so that it is similarly equivalent to a controlled DC motor; the extended inverter control part 32 is composed of reverse Park transformation and SVPWM modulation It consists of a voltage source inverter, followed by the mathematical model of PMSM1 in series; the mathematical model of PMSM1 is composed of Clarke transformation, Park transformation and DC model in series, but the actual connection with the controller is still the PMSM1 body. The input of compound controlled object 3 is the stator voltage under the dq two-phase rotating coordinate system, that is, u=[u ₁ , u ₂ ] ^T ＝[u _sd , u _sq ] ^T , and the output is the rotor angular velocity and the two-phase stator current signals , that is, y=[y ₁ , y ₂ ] ^T ＝[ω _r , i _s ² ] ^T . Among them, u _sd and u _sq are respectively the d-axis and q-axis voltages in the two-phase rotating coordinate system, which are used as the input signal of the compound controlled object 3 and also the output signal of the system reversibility analysis; ω _r and i _s are respectively The rotor angular velocity and stator current signals output by PMSM1 are also an important part of the input signal for system reversibility analysis.

2个输出量分别为复合被控对象3的输入u＝[u₁，u₂]^T＝[u_sd，u_sq]^T。其中，

和分别为模糊神经网络广义逆系统4的两个输入量，

和

是第2步骤中转子角速度和两相定子电流信号y₁和y₂以及他们各阶导数的线性合成量，a₁₀、a₁₁、a₁₂、a₂₀和a₂₁分别为系数。Step 3: As shown in Figures 2 to 3, after analysis, equivalent and derivation, the mathematical model of the compound controlled object 3 of the entire permanent magnet synchronous motor under vector control mode is obtained as a two-phase rotating coordinate system, that is, under the dq coordinate system The third-order nonlinear differential equations, and according to the inverse system theory, it is proved that the generalized inverse system of the third-order differential equations exists, and the relative order of the vectors is {2, 1}, and then the generalized inverse of the system is deduced, and the permanent magnet synchronization is established. The motor generalized inverse system model provides a methodological basis for the fuzzy neural network generalized inverse 4, and at the same time determines its two input quantities as

The two output quantities are respectively the input u=[u ₁ , u ₂ ] ^T =[u _sd , u _sq ] ^T of the composite plant 3 . in,

and are the two input quantities of the generalized inverse system 4 of the fuzzy neural network,

and

is the linear combination of rotor angular velocity, two-phase stator current signals y ₁ and y ₂ and their derivatives in the second step, and a ₁₀ , a ₁₁ , a ₁₂ , a ₂₀ and a ₂₁ are coefficients respectively.

，f₂＝-(u_i ⁽²⁾-m_ij)²/σ_ij ²(m_ij和σ_ij分别为第i个输入语言变量的第j个项的高斯函数的中心和宽度)，每个神经元输出相应的隶属度函数，权值w_ij ⁽²⁾＝m_ij；第三层是规则层，节点数为9，用于产生模糊逻辑规则和前件匹配，即计算每条规则的适应度，本层神经元节点执行模糊与操作相应位置上“与”运算，即f₃＝min{u₁ ⁽³⁾，u₂ ⁽³⁾，……u₅ ⁽³⁾}，a₃＝f₃，权值w_ij ⁽³⁾＝1；第四层归一化层，节点数为9，网络连接定义了规则节点的结论，产生每条规则对应于输入所产生的输出，是后件匹配，执行“或”运算，即

a₄＝min{1，f₄}(p表示神经元节点的输入个数)，权值w_ij ⁽⁴⁾＝1；第五层解模糊层(输出层)，节点数为2，用于解模糊，实现清晰化计算，产生控制规则的总输出，即

权值w_ij ⁽⁵⁾＝m_ijσ_ij。其中5个输入节点中，模糊神经网络广义逆4的第一个输入

作为模糊神经网络41的第一个输入；其经二阶系统s/a₁₀s²+a₁₁s+a₁₂(二阶系统是与模糊神经网络41连接的二阶线性环节G₁(s)s，a₁₀、a₁₁、a₁₂为线性传递函数的系数)的输出为即为模糊神经网络41的第二个输入；再经1个积分器s^-1输出y₁，即为模糊神经网络41的第三个输入；模糊神经网络广义逆4的第二个输入作为模糊神经网络41的第四个输入；其经一阶系统1/a₂₀s+a₂₁(一阶系统是与模糊神经网络41连接的一阶线性环节G₂(s)，a₂₀、a₂₁为一阶环节的系数)的输出为y₂，即为模糊神经网络41的第五个输入。于是，模糊神经网络41与2个线性传递函数和1个积分器一起组成模糊神经网络广义逆4，模糊神经网络41的输出就是模糊神经网络广义逆4的输出。Step 4: As shown in Fig. 3, use the fuzzy neural network 41 plus 2 linear transfer functions and 1 integrator to construct the generalized inverse 4 of the fuzzy neural network, providing a methodological basis for the learning and training of the fuzzy neural network 41. According to the specific situation of PMSM1, reasonably adjust the parameters a ₁₀ , a ₁₁ , a ₁₂ , a ₂₀ , a ₂₁ of the generalized inverse 4 linear transfer function of the fuzzy neural network to characterize the dynamic characteristics of the generalized inverse system, so that the decoupling formed The poles of the single-input and single-output pseudo-linear subsystem are reasonably arranged in the complex plane to realize the transformation from the integral unstable subsystem to the stable subsystem, and realize the open-loop linearized stable control of the nonlinear system. Wherein the fuzzy neural network 41 adopts a five-layer network. The first layer is the input layer, the number of input nodes is 5, and the neurons are the input nodes, which represent the input language variables. This layer is only used to transmit signals to the next layer, that is, f ₁ = u _i ⁽¹⁾ , a ₁ = f ₁ (f _z and a _z respectively represent the net input and activation function of the z-th layer node, z=1, 2, 3, 4, 5; ⁽¹ ) and i in u _i (1) represent the first layer node neurons The i-th input of the following analogy), the weight w _ij ⁽¹⁾ = 1 (represents the connection weight coefficient from the i-th input language variable to the j-th neuron in the next layer); the second layer of fuzzy layer, node The number is 15, and each node represents a language variable value, which is used to calculate the membership function of each input component. The neurons in this layer select the Gaussian function as the activation function, that is,

, f ₂ =-(u _i ⁽²⁾ -m _ij ) ² /σ _ij ² (m _ij and σ _ij are respectively the center and width of the Gaussian function of the jth term of the i-th input language variable), each Neurons output the corresponding membership function, weight w _ij ⁽²⁾ = m _ij ; the third layer is the rule layer, the number of nodes is 9, which is used to generate fuzzy logic rules and antecedent matching, that is, to calculate the adaptation of each rule degree, the neuron nodes in this layer perform the "AND" operation on the corresponding position of the fuzzy AND operation, that is, f ₃ =min{u ₁ ⁽³⁾ , u ₂ ⁽³⁾ ,... u ₅ ⁽³⁾ }, a ₃ =f ₃ , weight w _ij ⁽³⁾ = 1; the fourth layer of normalization layer, the number of nodes is 9, the network connection defines the conclusion of the rule node, and the output generated by each rule corresponding to the input is the consequent matching , perform an OR operation, that is,

a ₄ =min{1, f ₄ } (p represents the input number of neuron nodes), weight w _ij ⁽⁴⁾ =1; the fifth layer of defuzzification layer (output layer), the number of nodes is 2, used for Defuzzification, realize clear calculation, and generate the total output of control rules, namely

Weight w _ij ⁽⁵⁾ = m _ij σ _ij . Among the 5 input nodes, the first input of the generalized inverse 4 of the fuzzy neural network

As the first input of the fuzzy neural network 41; it passes through the second-order system s/a ₁₀ s ² +a ₁₁ s+a ₁₂ (the second-order system is the second-order linear link G ₁ (s) connected with the fuzzy neural network 41 s, a ₁₀ , a ₁₁ , a ₁₂ are the coefficients of the linear transfer function) the output is It is the second input of the fuzzy neural network 41; then output y 1 through an integrator s ^- ₁ , which is the third input of the fuzzy neural network 41; the second input of the generalized inverse 4 of the fuzzy neural network As the fourth input of the fuzzy neural network 41; it passes through the first-order system 1/a ₂₀ s+a ₂₁ (the first-order system is the first-order linear link G ₂ (s) connected with the fuzzy neural network 41, a ₂₀ , a ₂₁ is the coefficient of the first-order link), the output is y ₂ , which is the fifth input of the fuzzy neural network 41 . Therefore, the fuzzy neural network 41 forms the fuzzy neural network generalized inverse 4 together with two linear transfer functions and an integrator, and the output of the fuzzy neural network 41 is the output of the fuzzy neural network generalized inverse 4 .

并保存；②将保存的数据信号分别离线求得速度一阶、二阶导数

和电流一阶导数

此时有：y₁＝ω_r，

进而按照第3步骤中的方法求得

并对信号做规范化处理，组成模糊神经网络41的训练样本集

③首先使用遗传算法离线训练模糊神经网络41，粗调其隶属函数的参数和输出的初始权值，其中交叉概率P_c和变异概率P_m采用自适应方式，用合适的函数来衡量算法的收敛状况(P_c＝k₁/(f_max-f)，P_m＝k₂/(f_max-f)，f_max和f分别表示群体中的最大、平均适应度，k₁、k₂为大小在0～1之间的正实系数)，终止进化代数设定为G＝300，于是得到一个全局近似解，具体训练步骤与一般遗传算法类似，粗略的确定模糊神经网络41的各个参数和权系数；然后在控制器具体运行时，采用带动量项和变学习率的误差反传最优梯度法在线实时细化调整模糊神经网络41的参数，使模糊神经网络41输出均方误差精度保持在0.0005以内。Step 5: as shown in FIG. 3 , adjustment and determination of parameters and weight coefficient values of the fuzzy neural network 41 . Combining the genetic algorithm and the optimal gradient method, the learning of the fuzzy neural network 41 is divided into two stages: off-line learning and on-line adjustment of weight coefficients. Specifically, it is divided into the following steps: ① Add the step excitation signal {u _sd , u _sq } to the two input terminals of the compound controlled object 3, and collect the rotor angular velocity ω _r and current i _sA of PMSM1 with a sampling period of 5 ms, i _sB , obtain the required data through the current speed detection and calculation module 31

and save; ② save the data signal Obtain the first and second derivatives of velocity offline respectively

and the first derivative of the current

At this time: y ₁ =ω _r ,

Then follow the method in step 3 to obtain

And normalize the signal to form the training sample set of the fuzzy neural network 41

③ Firstly, the genetic algorithm is used to train the fuzzy neural network 41 off-line, and the parameters of its membership function and the initial weight value of the output are roughly adjusted. Among them, the crossover probability P _c and the mutation probability P _m adopt an adaptive method, and a suitable function is used to measure the convergence of the algorithm Condition (P _c =k ₁ /(f _max -f), P _m =k ₂ /(f _max -f), f _max and f represent the maximum and average fitness in the population respectively, k ₁ and k ₂ are the size positive real coefficient between 0 and 1), the termination evolution algebra is set to G=300, so a global approximate solution is obtained, the specific training steps are similar to the general genetic algorithm, roughly determine each parameter and weight of the fuzzy neural network 41 coefficient; then when the controller is actually running, the error backpropagation optimal gradient method with the momentum item and the variable learning rate is used to fine-tune and adjust the parameters of the fuzzy neural network 41 in real time, so that the fuzzy neural network 41 outputs the mean square error precision to remain at Within 0.0005.

即为模糊神经网络广义逆4的两个输入量，对应的输出分别为ω_r，

即电流速度检测与计算模块31输出的电流和转子角速度，实现了将原高阶、耦合的非线性复杂系统的控制转化为简单的线性系统控制。Step 6: Construct the fuzzy neural generalized pseudo-linear system 5 as shown in FIG. 4 , and linearize and decouple the original compound controlled object 3 into a velocity sub-linear system 51 and a current sub-linear system 52 . First, two linear transfer functions and one integrator are connected in series with the fuzzy neural network 41 whose parameters and weight coefficients are determined to form the generalized inverse 4 of the fuzzy neural network, as shown in the small dotted line box in the left figure of Fig. 4; then, this The generalized inverse 4 of the fuzzy neural network and the compound controlled object 3 are connected in series to form a generalized pseudolinear system 5, as shown in the big dashed box in the left figure of Fig. 4, the generalized pseudolinear system 5 is composed of a second-order velocity pseudolinear subsystem 51 It is equivalent to a first-order current pseudo-linear subsystem 52 in parallel, as shown in the right diagram of Figure 4, which is equivalent to a second-order velocity pseudo-linear subsystem 51 and a first-order current pseudo-linear subsystem 52 inputs are

are the two input quantities of the generalized inverse 4 of the fuzzy neural network, and the corresponding outputs are ω _r ,

That is, the current and the angular velocity of the rotor output by the current speed detection and calculation module 31 realize the transformation of the control of the original high-order, coupled nonlinear complex system into a simple linear system control.

分别结合速度伪线性子系统51和电流伪线性子系统52两个伪线性子系统的性质、实际运行中所面临的干扰及参数的时变特性构造模糊神经网络广义逆鲁棒控制器7。本发明采用线性系统鲁棒控制理论中内模控制原理、Lyapunov(李雅普诺夫)理论等设计方法设计模糊神经网络广义逆鲁棒控制器7。其中，内模控制器6由线性化了的速度内模控制器61和电流内模控制器62构成。D₁(s)和D₂(s)分别为两个控制器的干扰信号，速度内模控制器61由速度内部模型611和速度控制器612组成，电流内模控制器62由电流内部模型621和电流控制器622组成。适当选择参数a₁₀，a₁₁，a₁₂，a₂₀，a₂₁，使得二阶线性速度子系统的内部期望模型611为G_1m(s)＝1/(a₁₀s²+a₁₁s+a₁₂)＝1/(s²+1.414s+1)，于是设计得到相应的速度控制器612为一阶电流线性子系统的内部期望模型621为G_2m(s)＝1/(a₂₀s+a₂₁)＝1/(s+1)，同样可设计得到相应的电流控制器622为

其中，a₁₀、a₁₂、a₁₁为速度内部期望模型611的传递函数G_1m(s)的系数，取值为a₁₀＝a₁₂＝1，a₁₁＝1.414，此时内部模型G_1m(s)为典型的二阶稳定线性系统；F₁(s)为相应速度控制器612的一型低通滤波器，F₁(s)＝1/(0.5s+1)²；a₂₀、a₂₁为电流内部期望模型621的传递函数的系数，取值为a₂₀＝a₂₁＝1；F₂(s)为相应电流控制器622的一型低通滤波器，F₂(s)＝1/(2s+1))。整个模糊神经网络广义逆鲁棒控制器7的结构及连接情况如图5所示。Step 7: Construct fuzzy neural network generalized inverse robust controller7. According to the second and third steps, the relative order of the system is {2, 1}. According to the sixth step, the input of the generalized pseudolinear system 5 composed of the generalized inverse 4 of the fuzzy neural network and the compound controlled object 3 is

The fuzzy neural network generalized inverse robust controller 7 is constructed by combining the properties of the two pseudo-linear subsystems of the velocity pseudo-linear subsystem 51 and the current pseudo-linear subsystem 52, the disturbances faced in actual operation and the time-varying characteristics of parameters. The present invention adopts the internal model control principle in the linear system robust control theory, Lyapunov (Lyapunov) theory and other design methods to design the fuzzy neural network generalized inverse robust controller 7 . Wherein, the internal model controller 6 is composed of a linearized velocity internal model controller 61 and a current internal model controller 62 . D ₁ (s) and D ₂ (s) are the interference signals of the two controllers respectively, the speed internal model controller 61 is composed of the speed internal model 611 and the speed controller 612, and the current internal model controller 62 is composed of the current internal model 621 And current controller 622 composition. The parameters a ₁₀ , a ₁₁ , a ₁₂ , a ₂₀ , a ₂₁ are properly selected so that the internal expectation model 611 of the second-order linear velocity subsystem is G _1m (s)=1/(a ₁₀ s ² +a ₁₁ s+a ₁₂ )=1/(s ² +1.414s+1), so the corresponding speed controller 612 is designed as The internal expectation model 621 of the first-order current linear subsystem is G _2m (s)=1/(a ₂₀ s+a ₂₁ )=1/(s+1), and the corresponding current controller 622 can also be designed as

Among them, a ₁₀ , a ₁₂ , and a ₁₁ are the coefficients of the transfer function G _1m (s) of the speed internal expectation model 611, and the values are a ₁₀ =a ₁₂ =1, a ₁₁ =1.414. At this time, the internal model G _1m ( s) is a typical second-order stable linear system; F ₁ (s) is a type I low-pass filter of the corresponding speed controller 612, F ₁ (s)=1/(0.5s+1) ² ; a ₂₀ , a ₂₁ is the coefficient of the transfer function of the current internal expectation model 621, and the value is a ₂₀ =a ₂₁ =1; F ₂ (s) is a type-1 low-pass filter of the corresponding current controller 622, F ₂ (s)=1 /(2s+1)). The structure and connection of the whole fuzzy neural network generalized inverse robust controller 7 are shown in Fig. 5 .

The schematic diagram of the implementation of the permanent magnet synchronous motor speed control system based on the fuzzy neural network generalized inverse robust controller 7 on the dSPACE real-time simulation and test system experimental platform is shown in Figure 6. There are PMSM1 and dSPACE 81 in Figure 7, and the attached modules include analog input ADC module, analog output DAC module, signal detection part, photoelectric encoder 2, Hall element, magnetic powder braking unit, industrial control display module 83 and intelligent power module IPM 82 ; The software environment mainly includes real-time code generation download software RTI, comprehensive experiment and test environment software ControlDesk and Simulink simulation software. The fuzzy neural network generalized inverse robust controller 7 is realized by dSPACE to control the compound plant 3 . The experimental control program is downloaded from the host computer to the dSPACE control board, and the experiment start signal is sent through the ControlDesk visual control interface, and the control system operates independently; the 6-way PWM control signal output by the control board is sent to the intelligent power module to drive the motor; the detection part collects current, voltage, The speed and protection signals are fed back to the control board and stored for control effect analysis. The parameters can be modified offline or online to control the motor to achieve high-precision and stable operation and shorten the system development cycle.

The present invention realizes the linearized decoupling control of the multivariable, strongly coupled time-varying nonlinear system of the permanent magnet synchronous motor by constructing the generalized inverse of the fuzzy neural network, and the control problem of the complex system that couples the stator current, voltage and speed with each other It is transformed into a simple control problem of the second-order speed linear stability subsystem and the first-order current linear stability subsystem, and combined with the internal model control principle, a robust controller is conveniently and reasonably designed to achieve high precision of the permanent magnet synchronous motor speed Robust control overcomes the unmodeled dynamic disturbance of the system, so that the system has excellent dynamic and static performance, anti-interference and high-precision tracking performance.

Claims (3)

1. A fuzzy neural network generalized inverse robust controller of a permanent magnet synchronous motor is characterized in that: the fuzzy neural network generalized inverse robust controller (7) is formed by combining an internal model controller (6) and a fuzzy neural network generalized inverse (4); the internal model controller (6) is composed of a speed internal model controller (61) and a current internal model controller (62) which are connected in parallel, the speed internal model controller (61) is composed of a speed internal model (611) and a speed controller (612), and the current internal model controller (62) is composed of a current internal model (621) and a current controller (622) which are connected; the fuzzy neural network generalized inverse system (4) and the composite controlled object (3) are connected in series to form a generalized pseudo-linear system (5), and the generalized pseudo-linear system (5) is equivalent to 1 speed sub-linear system (51) and 1 current sub-linear system (52); the generalized inverse fuzzy neural network (4) consists of a five-layer fuzzy neural network (41) with 5 input nodes and 2 output nodes, two linear transfer functions and 1 integrator; the composite controlled object (3) comprises a current speed detection and calculation module (31) and an extended inverter control part (32) for driving the PMSM (1) in a connected mode, the extended inverter control part (32) is formed by connecting an inverse Park conversion and a voltage source type inverter in an SVPWM debugging mode, the current speed detection and calculation module (31) comprises a current speed calculation part, a Park conversion, a Clarke conversion and a photoelectric encoder (2) in a connected mode, and the Clarke conversion and photoelectric encoder (2) is connected with the PMSM (1).

2. A construction method of a fuzzy neural network generalized inverse robust controller of a permanent magnet synchronous motor is characterized by sequentially comprising the following steps:

1) clarke transformation, Park transformation and a third-order model are equivalent to form a PMSM (1), and the PMSM (1) forms an integrally-formed composite controlled object (3) through a current speed detection and calculation module (31) and an extended inverter control part (32);

2) the method comprises the steps that 2 linear transfer functions, 1 integrator and a fuzzy neural network (41) with determined parameters and weight coefficients are connected in series to form a fuzzy neural network generalized inverse (4), the fuzzy neural network generalized inverse (4) and a composite controlled object (3) are connected in series to form a generalized pseudo-linear system (5), and the generalized pseudo-linear system (5) linearizes and decouples the PMSM (1) and enables the PMSM to be equivalent to 1 second-order pseudo-linear speed subsystem (51) and 1 first-order pseudo-linear current subsystem (52);

3) and respectively introducing a second-order velocity pseudo-linear subsystem (51) and a first-order current pseudo-linear subsystem (52) into an internal model control method to construct an internal model controller (6), combining the internal model controller (6) with a generalized pseudo-linear system (5) to form a fuzzy neural network generalized inverse robust controller (7), and controlling a composite controlled object (3).

3. The method of construction of claim 2 wherein:

in the step 1), the current obtained by Clarke conversion of two-phase stator current of the PMSM (1) and Park conversion and the current, rotor angular speed and angular displacement output after the rotation speed detected by the photoelectric encoder (2) is calculated by a current speed calculating part are used as the output of a current speed detecting and calculating module (31); the input of the composite controlled object (3) is the stator voltage under a d-q coordinate system, and the output is the rotor angular velocity and the two-phase stator current;

in the step 2), 2 input quantities of the generalized inverse (4) of the fuzzy neural network are respectively the rotor angular velocity, the two-phase stator current signals and the linear synthesis quantity of each order derivative thereof, and 2 output quantities are respectively the input of the composite controlled object (3); the method for determining the parameters and the weight coefficient values of the fuzzy neural network (41) comprises the following steps: phi step excitation signal u_sd，u_sqRespectively adding the signals to 2 input ends of a composite controlled object (3), and acquiring the rotor angular speed omega of the PMSM (1) in a sampling period of 5ms_rAnd current i_sA，i_sBObtaining the required data through a current speed detection and calculation module (31)

And storing; ② data signal to be storedRespectively off-line obtaining first and second derivatives of speed

First derivative of sum current

Forming a training sample set of a fuzzy neural network (41); thirdly, the fuzzy neural network (41) is trained off line by using a genetic algorithm, a global approximate solution is obtained by roughly adjusting parameters of membership functions of the fuzzy neural network and output initial weights of the membership functions, all parameters and weight coefficients of the fuzzy neural network (41) are roughly determined, and then an optimal gradient method is adopted by error back propagation of driving variables and variable learning rates during specific operationRefining and adjusting parameters of the fuzzy neural network (41) to keep the output mean square error precision of the fuzzy neural network (41) within 0.0005;

in the step 3), the input of the 1 second-order velocity pseudo linear subsystem (51) and the input of the 1 first-order current pseudo linear subsystem (52) are two input quantities of the generalized inverse (4) of the fuzzy neural network respectively, and the output is the current output by the current velocity detection and calculation module (31) and the rotor angular velocity respectively; the generalized inverse robust controller (7) of the fuzzy neural network adopts dSPACE to realize to control the compound controlled object (3).

Images