Gear Fault Diagnosis Based on BP Neural Network

Gear transmission is more complex, widely used in machinery fields, which form of fault has some nonlinear characteristics. This paper uses BP neural network to train the gear of four typical failure modes, and achieves satisfactory results. Tested by using test data, test results have an agreement with the actual results. The results show that the BP neural network can effectively solve the complex state of gear fault in the gear fault diagnosis.

surface near the section line. The reason is that the surface fluctuation cycle contact stress exceeds the material's ultimate stress of tooth, its performance in the form of initial pitting corrosion and expansion. Tooth wear is due to the working surface of metal particles, dust and sand into the teeth. Tooth surface rough, poor lubrication is another cause of tooth wear. In addition, the misalignment, shaft wear and tensional resonance, can cause large changes in the torque of the gear meshing point, or to make more impact, will accelerate the gear wear .After gear wear, tooth thickness becoming thinner, tooth profile becoming deformation, lateral gap becoming large, will cause the gear dynamic load increase, not only to increase the vibration and noise, but also likely to lead to broken teeth. Tooth bonding (scratches) is due to the rupture of oil film in sliding contact tooth surfaces, direct contact with the tooth surface contact area, the friction force and the pressure produced by the action of high temperature instant, local adhere and stripping the metal surface.

The principle and algorithm of BP neural network
BP neural network (Back-Propagation Network), is a kind of back propagation network, it consists of the back propagation and error information communication two processes , proposed by Rumelhart and McCelland in 1986. BP neural network includes three layers, input layer, hidden layer and output layer. Input layer mainly accepts signal input, the middle layer is information processing and conversion layer, mainly processes input signal, it can include one or more intermediate layers.
Through a large number of data training to find a more reasonable weight, can correctly identify a variety of signals, and can correctly output the results.
(1)Output node model The hidden nodes and output model: (2) F is a non linear function; Q neural unit threshold.
(2)Function model Function is to reflect the lower input function to the upper node stimulation intensity and stimulation function, general admission for the (0,1) values of Sigmoid continuous function.
The error model is a reflection of the error function between the desired output and the output of the neural network calculation.
As the expected output value of the i node； pi O To calculate the output value for the i node.
(4)Self learning model Neural network learning process is the process which sets the weight matrix ij W and updates the error connected between the lower and upper node. BP network learning mode -the need to set expectations and unsupervised learning -just enter the mode. Self learning model is

Gear fault diagnosis of BP neural network
Gear fault diagnosis in the application of BP neural network needs to choose some parameters. In this paper selection main parameters are peak value, mean value ,the absolute mean value, RMS, RMS, amplitude, RMS amplitude, variance, degree of skew, Waveform index Peak index, pulse index, margin index, skew, steepness, total of 15 parameters. (1) The input layer. Through the analysis of the above 15 parameters to determine the fault type, and selecting 8 training samples to train the neural network, the training sample data are shown in table 1.
(2) The hidden layer The number of hidden layer nodes is the key of solving the problems. It has a direct impact on the BP network model identification ability of sample. The selection method is not yet conclusive, general according to experience and learning through training, combined with the number of times and the sample to determine. According to the experience in this paper, in this case because the input parameters of the eye have 15, so the hidden nodes are 31.
According to the output of the network, there are 4 typical fault conditions, to encode these typical fault, using the normalized, select the network output layer nodes is 4, respectively 1 Table 2 for all types of faults in the desired output.

Conclusion
In this paper, based on the theory of neural network, the gear signal as the input signal, , to overcome the previous multi input, single output neural network system, the paper designs the neural network fault diagnosis system 15-31-4. Through a large number of samples of the input verified by the experimental samples to train the neural network, , the neural network proves that the design has high recognition ability, better diagnostic capability in fault diagnosis of gears.