2 0。005 0。575 75
3 0。006 1 100
4 0。007 2 125
4。2 Construction of BP neural network
The oil consumption model is accurate for calculation, but is time-consuming。 A trained neural network model can greatly reduce the computation time and improve the efficiency without sacrificing the accurac。 Therefore, the cylinder oil consumption model is reconstructed using the BP neural network model。 The built BP network is pided into three layers: input layer, hidden layer and 论文网
output layer。 Three neurons of the input layer are the e/B of the top ring, the closed gap of the second piston ring and the stretch force of oil ring。 One neuron in the output layer is the cylinder oil consumption after removing the cylinder liner wall evaporation。 BP network with a single hidden layer can approximate an arbitrary continuous nonlinear function。 So a single hidden layer is adopted。 The neuron number in the hidden layer directly affects the nonlinear predictive performance of the network。 According to the Kolmogorov theoremand after a lot of trials, the neuron number in the hidden layer of network is decided to be 7。 The linear function purelin is chosen as the transfer function of the neurons in
the hidden layer。 A linear function purelin is also chosen as the transfer function of the output neurons。 In order to expedite the training convergence of the network and
prevent falling into a local minimum, the gradient descent algorithm with momentum is used as the learning and training algorithm。 The neural network model is shown in Fig。 13。
64 groups of calculated data are obtained from 64 schemes given in Table 3, which will be used as the required samples of the neural network。 64 samples are pided into two parts, in which 58 samples are used to train the BP network and the other six groups are used to test it。
Table 3。 Comparison between calculated and predicted
values of the BP network model
Order Cc/(g • h−1) Cp/(g • h−1) Er/%