CAJ | 학술논문

针对传统模型参考自适应控制存在的鲁棒性问题和神经网络结构庞大因而计算量膨胀的问题。提出了一种变结构神经网络L1自适应控制方法，其中变结构神经网络用于在线辨识系统存在的未知非线性函数，该网络通过对节点进行唤醒与催眠以动态调节结构，以最少的节点数进行有效的逼近，降低计算复杂度；L1自适应控制用于网络权值学习与系统非线性补偿，反馈回路中设有一个低通滤波器，只要满足L1增益条件，就能确保系统的输入输出信号的瞬态响应和稳态跟踪性能与一个期望的线性时不变系统的响应保持一致。通过对四旋翼飞行器进行仿真，验证了该方法的有效性。
침대전통모형삼고자괄응공제존재적로봉성문제화신경망락결구방대인이계산량팽창적문제。제출료일충변결구신경망락L1자괄응공제방법，기중변결구신경망락용우재선변식계통존재적미지비선성함수，해망락통과대절점진행환성여최면이동태조절결구，이최소적절점수진행유효적핍근，강저계산복잡도；L1자괄응공제용우망락권치학습여계통비선성보상，반궤회로중설유일개저통려파기，지요만족L1증익조건，취능학보계통적수입수출신호적순태향응화은태근종성능여일개기망적선성시불변계통적향응보지일치。통과대사선익비행기진행방진，험증료해방법적유효성。
L1 adaptive control method based on variable structure neural network is proposed to solve the problems that the robustness exists in the traditional model reference adaptive control and the neural network structure is so bulky that the computation expands. The neural network is used as an identification generator for unknown nonlinear functions in the system, and the structure of the network is adjusted dynamically by activating and hypnotizing the nodes to approximate the functions with minimum nodes and reduce the computation procedures. The L1 adaptive control is used to obtain the weights of the network and compensate the nonlinearity of the system. A low pass filter is adopted in the feedback loop, so long as the L1 gain requirement is fulfilled, the transient response of input/output signals and the steady-state tracking performance of the system will be coincident with the specifications of the desired linear time-invariant system. The simulation results of quadrotor show the efficiency of the novel method.