1. Input and output of neural network
The analysis of meshing force of parallel coupled planetary gear includes five factors: power flow value, input speed ratio, input torque ratio, output speed ratio and output load ratio. Only the input speed ratio and output load ratio are considered, as shown in the figure
The curve obtained by the above two factors. As the curve in the figure is nonlinear, it can be seen that the input speed ratio and output load ratio are not comprehensive. Therefore, five influencing factors are selected as the input factors of the neural network prediction model, and ∑ FB in P10 row and P20 row is selected as the output value of the prediction model.
2. Determine the structure of RBF neural network model
RBF neural network is divided into regularized neural network and generalized RBF neural network. Because the training samples of regularized network and “basis function” are one-to-one correspondence. When the number of samples P is large, the calculation of the network is large. In addition, when the number of samples P is large, the weight matrix is also large, so it is easy to produce ill conditioned problems when solving the weights of the network. The design of generalized RBP network includes structure design and parameter design. The structure design mainly solves the problem of how to determine the number of hidden nodes in the network. To solve this problem, we can reduce the number of hidden layer neuron nodes m, that is n < m < p (n is the sample dimension, P is the number of samples), so as to obtain the generalized RBF network. In this paper, the generalized RBF network is used.
Using the newrb function library in MATLAB, the parameters include input vector group, target classification vector group, target mean square error, radial basis function expansion speed, the maximum number of neurons and return value. The larger the spread speed of radial basis function is, the smoother the fitting is. However, excessive expansion speed requires more neurons to adapt to the rapid change of function. If the expansion speed is set too small, it means that many neurons are needed to adapt to the slow change of the function, and an appropriate expansion speed needs to be set.
3. K-means clustering of training samples
The data center of each basis function should be considered in parameter design. In the self-organizing selection of data center, various dynamic clustering algorithms are often used to select the data center. In the learning process, the position of the data center needs to be dynamically adjusted. The commonly used method is K-means clustering. Its advantage is that it can determine the expansion constant of each hidden node according to the distance between the clustering centers. Since the number of hidden nodes of RBF network has a great influence on its generalization ability, this paper uses K-means clustering method to cluster the training data.
Adams is used to establish the dynamic model of the parallel planetary gear mechanism, and the simulation experiment is designed, and the training data samples are calculated. There were 90 groups of training data, and the estimated number of centers m was 30. The Euclidean distance between the training sample points and the clustering center points is calculated. The Euclidean distance between the training sample points and the clustering center points is classified according to the minimum Euclidean distance between the training sample points and the clustering center points. The kmeans function in MATLAB is used to achieve the clustering combination of the training samples, and the expansion constant of the corresponding radial basis function is determined according to the distance between the centers, The calculated expansion constant is 2.497.