ANFIS-based modeling of nonlinear motor systems
1 Introduction fruit. Modeling identification and control of nonlinear systems is an important application direction. Studies have shown that for nonlinear systems, the use of traditional analysis methods can only be applied to specific applications without a universally applicable method. Artificial neural network provides powerful insights for the modeling of nonlinear systems with its excellent nonlinear mapping approximation capability and self-learning ability. It can also adjust parameters according to a given data set to obtain a good fuzzy model. In recent years, how to integrate the fuzzy system into the network has various forms. 314. We can use some neuro-fuzzy systems to model the control object in order to obtain better results than simply applying the species technology. In this paper, we will use a model based on a high-algorithm model, which is modeled on an adaptive neuro-fuzzy inference system, such as a 5-ton press, such as an adaptive neuro-fuzzy inference system. Compare the experimental results with the back propagation network. This research, funded by the key project of the Chinese Academy of Sciences, gives experimental results and analysis of modeling; 2,5 Introduction From Miao said, the fuzzy inference system based on adaptive network proposed by Tu Teng, also known as adaptive neuro-fuzzy inference system, is a kind of fuzzy inference system based on Takagi Suzuki model TakagiSugenoModel w. When the fuzzy set employs non-trapezoidal non-angled membership functions, the Dingye 35å©63 fuzzy system is more economical than the 4å’–13 fuzzy system, ie, the required fuzzy rules and the number of input fuzzy sets are less. 2.1 The Structure of Liuxuan 3 There are two simple rules of the 38388, the fuzzy system of the sixteenth structure is as follows: The corresponding ANFIS structure l connection only the flow of the signal, there is no weight associated with it; Square node with adjustable parameters Nodes, circular nodes without nodes with adjustable parameters. From parameters. The function of each layer is as follows: The first layer fuzzifies the input variable and outputs the membership degree of the corresponding fuzzy set. The transfer function of each node may be in accordance with the form of the selected membership function, and corresponding parameter sets may be obtained, which are called conditional parameters. . For example, a commonly used Sri Lankan membership function is a set of all conditional parameters. The second layer implements the operation of the fuzzy set of the conditional part, and outputs the degree of applicability of each rule corresponding to formula 1, usually using multiplication. The third level naturalizes the application of rules. The transfer function of each node of the CO layer 4 is a linear function, a local linear model, and the output of each rule is calculated. By all, 6, the set of parameters is called the conclusion parameter. The 5th floor calculates the sum of the outputs of all the rules. From the input and output relations of the above network, it can be seen that the network is completely equivalent to the fuzzy inference system of the formula. The learning of the fuzzy inference system boils down to the adjustment of the linear parameters of the nonlinear parameter and the conclusion parameter of the conditional parameter. 2.2 Hybrid Learning Algorithm For all parameters, the gradient-descent back propagation algorithm can be used to adjust the parameters; however, a hybrid algorithm can be used. Improve the speed of learning. In the hybrid algorithm, the back propagation algorithm is still used for the conditional parameters, and the parameters of the conclusions are adjusted using the linear least-multiplier estimation algorithm. The conclusion part can use the least-multiplier estimation algorithm because the final output of the ANFIS is linear to the f-conclusion. The output is still in the ff system. The output is first in each iteration of the hybrid learning algorithm. The input signal is along the network's forward direction. Pass up to layer 4, at this time fixed the condition parameters, use the least multiplication estimation algorithm to adjust the conclusion parameters; then, the signal continues to pass along the positive direction of the network until the output layer is the 5th layer. Thereafter, the obtained error signal propagates back in the network so that the condition parameters can be adjusted. Using the hybrid learning algorithm, for the given conditional parameters, the global best of the conclusion parameters can be obtained. This not only can reduce the dimension of the search space in the gradient method, but also can greatly improve the convergence speed of the parameters. Detailed description of the hybrid learning algorithm References 51. 3 Nonlinear motor system modeling Next, we modeled a DC motor system with nonlinear friction effects. The actual control system used to collect the input and output signals includes the 12-bit, 6-payload, 1200-bit microcomputer built-in computer, the conversion board power amplifier circuit DC torque motor, and the DC tachometer generator for speed feedback. The analog voltage input range and output control voltage range are both 5,5 volts, and the analog-digital converted digital range is 2048, 2048. For convenience, both input and output units use digital quantities. The model of the controlled system includes a collection of all components except the computer. At a sampling period of 5 milliseconds, the input signal lasts 10 seconds, that is, 2000 sampling periods. In order to adapt the model to different frequencies, the training input signal is a composite of sinusoidal signals with multiple frequency components. The signal of type 9 is taken as the input of the actual system, and the true output of the motor, that is, the rotational speed signal is obtained. Actual motor input and output 2. 3.1 Modeling on the basis of 15 In Section 2, the structure and algorithm of 5 have been introduced, but it is not mentioned that the initialization of parameters in 18 asks for the partition of the input space, thereby determining the number of fuzzy rules. Usually the input space can be segmented using the average segmentation method or Tian Kan, Julong method. First, the false motor model is the order system, that is, the input and output relations can be used for us to use the average segmentation method, and the input interval for each of 1 and 1 is equal. Divided into three parts, the fuzzy membership functions are Gaussian. In this way, the two inputs form a total of 9 combinations, ie there are 9 fuzzy rules. According to the description in Section 2, it can be seen that there are 12 32+32 condition parameters, and there are 27 conclusion parameters. 93. The error indicator is the MSE of the actual output and the model output. After 100 iterations, the ANFIS identification results are 3 , The mean squared error of training results, =4.9854. 3.2 Verification of the identification model Firstly, a multi-frequency component sinusoidal signal with variation in amplitude and frequency is also used as a test signal to test the nonlinear model. The test result is 4, and the mean-square error is 5=6.5836. Second, we test two typical nonlinear characteristics of the modeled system with saturation and deadband. We use two single-frequency sine signals as the test signals 123 and 12 respectively. The input is to the model and input to the actual system. The output and input signals obtained from the actual system and the built test model are respectively 56. From the output of the actual system and the output of the model, it can be seen that the sum of the squared error of the 2000 sampling points is within 7 and the accuracy can be high. From this we can also see that the actual output and the model output almost coincide with the curve. From the above tests, it can be seen that the established ANFIS model can not only adapt well to changes in amplitude and frequency, but also can well contain the nonlinear characteristics of the motor system. The established model has strong adaptability to changes in amplitude and frequency within the range of training signals, and achieves the purpose of dynamic accurate modeling. 4 summarize the dynamic model. Compared with Japan, the main advantage of 15 is that the convergence speed is very fast, and the convergence time is only 8 tenths. When faced with a complex system, the parameters required for modeling increase rapidly, and the convergence speed is very important. At this time, Can 15 showed more advantages than 8 measurements. At the same time, in the analysis of the experimental results, the reason why ANFIS guarantees a faster convergence rate than BPNN from both initialization and learning algorithms is clarified. In addition, force. In practical applications. ANFIS provides a powerful tool for modeling and time series analysis of nonlinear systems. Zhao Zhenyu Xu Yongyu, foundation and application of fuzzy theory and neural network. Tsinghua University Press, 996.6. University Press, 3337, 1998.7. Barbecue wire Grill Grate is made of high quality 304 stainless steel, never rusting and durable. BBQ Wire mesh does not have any coating or chemical ingredients, making food safer. BBQ Grate,BBQ Mesh,BBQ Grill, grill grate,grill mesh Shenzhen Lanejoy Technology Co.,LTD , https://www.ccls-vaccine.com
Multi-functional Grill Cooking Grid Grate: This wire mesh is mainly used for BBQ Grill Mat for outdoor cooking, it can also be used as a cooling and baking rack. Or you can develop other uses for it.