Abstract: This paper uses neural networks to identify the model of a pneumatic servo system. The principle of RBF neural network identification and its difference from the parameter estimation method are analyzed. A neural network identification model is designed for the actual system. Experiments show that the model established by this method is correct. Keywords: Neural network identification; Pneumatic servo system Introduction The compressibility, low viscosity, and thermosensitivity of gases make it difficult to grasp the characteristics of pneumatic servo systems. In summary, pneumatic servo systems have the following characteristics: (I) Time-varying: that is, the system parameters are not constant, but change with time, and the parameters are related to the position during the motion. (2) Thermosensitivity: the system characteristics are greatly affected by temperature. (3) Pressure sensitivity: the system characteristics are greatly affected by the pressure fluctuation of the gas source. (4) Nonlinearity: the high compressibility of the gas and the friction of the actuator make the system characteristics exhibit severe nonlinearity. These characteristics make it difficult to obtain the model of the pneumatic servo control system, which brings many difficulties to the control of the pneumatic servo system. The mathematical model of the controlled system is very important for the analysis and control of the system. There are two methods for establishing the mathematical model of the dynamic system: mechanical modeling and experimental modeling. Identification and modeling methods include the step response method, frequency response method, correlation analysis method, and parameter estimation method. In recent years, advancements in neural network research have provided new methods for dynamic system identification, with many successful applications reported. 2. Differences between Neural Network Identification and Parameter Estimation Methods Essentially, both are the same; both utilize input-output information to obtain a model reflecting the system characteristics through a certain algorithm, and both employ fitting methods. The difference lies in that neural networks are more flexible than parameter estimation methods, capable of approximating arbitrary nonlinear functions and reflecting the characteristics of arbitrary nonlinear systems. They possess self-learning and memory capabilities. Neural networks use various nonlinear functions, while parameter estimation methods only use polynomials. Therefore, in a sense, neural network identification can be considered a generalization and improvement of parameter estimation methods. The effectiveness of neural network identification depends on the selected model and the method of calculating the weights. Among various neural network models, the RBF (Radial Basis Function) network has a simple structure, fast learning convergence, and high accuracy, making it very suitable for identifying and modeling single-output systems. Therefore, this paper selects this model to identify a switch-valve-controlled pneumatic servo system. 3 RBF Network Model The RBF network model has only one hidden layer and one output. Each hidden node outputs according to the RBF rule, as shown in Figure I. Network output: 4 Dynamic Identification Principle Based on RBF Network Any SISO nonlinear system can be described by equation (3). When using RBF network for identification, the first n values ​​of the system input time and the first f values ​​of the output time f are usually selected as the input vector of the RBF network. The RBF network is trained with samples using a learning algorithm. The weights after training represent the inherent characteristics of the system being identified. During training, the index function is: The weight correction algorithm is: Where: N—— Number of samples; r — RBF network output value; η —— Learning rate factor 5 Identification of Switch Valve Controlled Pneumatic Position Servo System Based on RBF Network Model As can be seen from the mechanism analysis modeling, the switch valve controlled pneumatic position servo system is a third-order system. Therefore, the RBF neural network model shown in Figure 2 is designed to identify this type of system. In the figure, the number of hidden nodes m is 21; Cj is taken at equal intervals between the maximum and minimum values ​​of the input and output. The weights of each hidden node are obtained by using the algorithms (4) and (5). Figure 3 shows the actual system output and model simulation output using the same controller. Curve 1 is the actual output and curve 2 is the model simulation output. It can be seen from the figure that the two are basically consistent, which shows that the model is correct. 6 Conclusion This paper uses RBF neural network theory to conduct identification research on the switch valve controlled pneumatic position servo system and draws several conclusions: (1) The number of hidden nodes in the RBF neural network should not be too small, otherwise the identification model will be significantly different from the actual object. (2) When sampling training samples in real time, the samples should be preprocessed to improve the accuracy of the identification results. At the same time, the learning rate factor has a great influence on the convergence of the RBF network. (3) It is appropriate to use RBF neural network to identify and model the switch valve controlled pneumatic position servo system. [References] [1] Gu Zhongwen. Industrial System Modeling. Hangzhou: Zhejiang University Press, 1995 [2] Xu Yaoling et al. Application of Artificial Neural Network in System Identification. Acta Automatica Sinica, 1991, 17(1): 91-94 [3] Ri Jianguo. RBF network modeling method and its application for dynamic systems. Xi'an: Proceedings of the Chinese Neural Network Conference. 1993(10) [4] Wang Xuanyin. Modeling research of PCM pneumatic system. Machine Tool & Liquid Bed, 1997 [5] Wang Xuanyin. Research on control of pneumatic driven robot. Postdoctoral research report of Zhejiang University. 1997(11) Please click here to download the original text