Aiming at the problem of strong dependence on prior data in neural sliding mode control system, a fractional-order speed control system of permanent magnet synchronous motor based on fuzzy RBF neural network is proposed, which combines the generalization ability and self-learning ability of RBF neural network and the strong adaptability of fuzzy inference algorithm. The introduction of fuzzy reasoning provides an effective guidance for the uncertainty of neural network. Meanwhile, the introduction of fractional calculus operator increases the degree of freedom of the traditional sliding mode controller, thermostatic element which further optimizes the controller. The simulation results show that the proposed fuzzy RBF neural fractional sliding mode controller has better control performance than RBF Neural Sliding Mode controller. Permanent magnet synchronous motor (PMSM) has been widely used in robots, CNC machine tools, medical equipment and other fields. However, the speed control performance of PMSM is affected by the parameters change and load disturbance in the system. The strong robustness of the sliding mode control technology to the internal and external disturbances of the system provides an effective way for the high performance speed regulation of the motor. However, considering that the ideal output of the sliding mode controller is the switch of high frequency switching, the time delay of the actuator in the motion control system will lead to the trajectory of the system state on the sliding mode surface. Traces will not occur accurately on the set switching manifold surface, and chattering will occur consequently [1], which undoubtedly limits the application scope of sliding mode control technology. Through the research of buffeting reduction methods, scholars at home and abroad have obtained a lot of achievements [2?10].
Among them, literature [2?4] proposed a high-order sliding mode control algorithm, but this method is more complex, the output signal of the controller is coupled with its derivatives, which is not conducive to the design of sliding mode control law. Literature [5?7] observes the load torque based on disturbance observer, and designs a kind of integral sliding mode controller to suppress the disturbance, but it is not conducive to the design of sliding mode control law.
Under the action of these methods, the dynamic performance of the system will be affected to a certain extent. In reference [8?10], the intelligent algorithm is introduced into the optimal design process of sliding mode controller. RBF neural network and fuzzy inference are used to tune the switching gain of sliding mode controller respectively, but there will be static error in the system under the action of these methods. This paper synthetically considers the application of fuzzy reasoning algorithm and RBF neural network in the optimization design of sliding mode controller, uses the strong adaptability of fuzzy reasoning algorithm to adjust the weight of RBF neural network, and then uses RBF neural network to train the actual output of fractional sliding mode controller, which achieves better comprehensive control. Yes. In the formula, [ud, uq] is the stator voltage in the two-phase rotating d? Q coordinate system, [id, iq] is the stator current, [lambda, lambda] is the stator flux, [Rs] and [Ld, Lq] are the stator resistance and inductance, [_f, _r] is the electrical angle of the motor and the given speed, [Lmd] is the stator phase inductance, [Idf] is the equivalent current, [np] is the logarithm of the poles. In the formula, [Te, TL] is the electromagnetic torque and load moment, [J] is the moment of inertia, [Bm] is the friction factor. In the formula, [x=theta_T], [A=010-BmJ], [B=0kpJ], [u=iq], [d=0-TLJ]. In formula [xk = theta K kT]. In the formula, [theta? K] is the position instruction, [_? K] is the change rate of the position instruction, [ek] is the position error, [dek] is the error of the change rate of the position instruction.
[fk=-theta?K-_?K theta?K 1BmJ?K _?K 1]. Formula: There is no systematic theoretical deduction method to determine the order [_] of fractional calculus operator. In this paper, through repeated tests, the value is determined to be [_=0.14]. In the formula [unk] is the output of the fuzzy RBF neural network. Each node in the fuzzification layer has the function of membership function, and the Gauss function is used as the membership function. [cij] and [bj] are the mean and standard deviation of the first input variable on the Gaussian function of the j-th fuzzy set. In formula [W = w1, w2,… WNT] is the weight vector between the fuzzy inference layer and the output layer.
In the formula, [_] is the learning rate, [alpha] is the inertia coefficient. In this paper, the simulation platform shown in Fig. 1 is adopted with the simulation tool of MATLAB software. The parameters of PMSM are as follows: [Rs = 1.5, Ld = Lq = 8.
5 *] [10-3H, NP = 4, J = 2.5 * 10-3 kg? M2, Bm = 0.8 * 10-3 N?M?S]. The input signal is a sinusoidal signal: [theta? K = 0.5sin6pi k].
The structure of fuzzy RBF neural network is selected as two input layers, 36 fuzzification layers, 36 fuzzy reasoning layers and one output layer. The initial value of network weight [W] is chosen as the random value of [-1,1]. The learning rate and inertia coefficient were selected as follows: [_=0.6, a=0.05].
To verify the superiority of the proposed algorithm, sliding mode controller based on RBF neural network optimization is used for comparative analysis.
In this paper, the construction process of fuzzy RBF neural fractional sliding mode controller is discussed in detail, and a new method for updating the weight coefficients between layers, the center vectors and the baseline vectors of nodes in the network structure is given. The superiority of the proposed algorithm is demonstrated by comparing with RBF Neural Sliding Mode controller. The simulation validates the algorithm from three aspects: speed response, trajectory of system state convergence to sliding mode surface and control variables.
The simulation results show that the proposed algorithm has high integrated control performance.