Spiral bevel gears are widely used in automobile, aircraft, ship and other equipment due to their advantages of strong bearing capacity, high transmission efficiency and stable transmission ratio. Due to its long-term operation in the harsh conditions of heavy load and enclosed space, it is easy to fail. Once it fails, it will affect the normal operation of the entire transmission chain and even the entire equipment, resulting in huge economic losses, even casualties. Therefore, the research on fault diagnosis method of spiral bevel gear is very important to ensure its safe, efficient and reliable operation.
The spiral bevel gear has a wider meshing surface than the spur gear. The interference noise component of the vibration signal caused by the continuous impact of the gear teeth during the meshing process is more complex. The fault vibration signal presents strong background noise, strong nonlinearity, and non-stationary characteristics. The signal bispectrum based on higher-order statistics can completely retain the amplitude, frequency, and phase of the signal, and has the characteristics of time-shift invariance, scale variability, and phase retention, It can suppress colored Gaussian noise and is suitable for feature extraction of strong background noise, strong nonlinear and non-stationary signals. Li Xuejun et al. used the bispectral distribution area as the characteristic quantity for the fault diagnosis of spur gears; Jiang Yonghua et al. proposed a fault feature extraction method combining empirical mode decomposition (EMD) and bispectral analysis, and applied it to rolling bearing fault diagnosis; Chen Zongxiang and others effectively diagnose the early fault of motor bearing by combining the improved empirical mode decomposition and bispectral analysis.
Traditional fault pattern recognition methods such as BP (Back Propagation) neural network, Support Vector Machine (SVM), KNN (K-Nearest Neighbor) are shallow machine learning methods, which are difficult to effectively build complex mapping relationship between fault and signal 275-282. Convolution Neural Network (CNN), as a kind of deep learning model, has strong self-learning ability of features, which can avoid the disadvantages of time-consuming feature filtering and non-universal features of traditional algorithms. It has been applied in speech recognition, image classification and other fields with good results. In recent years, it has also been preliminarily studied and applied in rolling bearing and spur gear fault diagnosis.
Although the existing cases have proved that CNN is an effective means to solve the problem of fault diagnosis, the research and application in the fault diagnosis of spiral bevel gears are less. The main reason is that it is difficult to construct efficient fault input samples of spiral bevel gears for training accurate CNN diagnosis models. Although some input sample construction methods have been proposed in the existing research, it is easy to lose fault information based on a single time and frequency domain signal construction method, resulting in low generalization ability and poor universality of CNN fault diagnosis model, and it is also difficult to construct CNN samples using time-frequency signals. Due to its unique advantages, bispectrum is expected to construct efficient fault input samples of spiral bevel gears for training accurate CNN diagnosis models.
Because the bispectrum is a symmetric spectrum, the fault information contained in the bispectrum is redundant. Taking the full bispectrum directly as the input sample of CNN will reduce the training efficiency of CNN. Replacing the global bispectrum with the local bispectrum containing the global information can not only reduce the redundancy of fault information, but also reduce the dimension of CNN training samples and improve the training speed of CNN. Therefore, the local bispectrum of the spiral bevel gear signal is used to construct the input sample of CNN, enhance the generalization performance of the diagnosis model, and apply it to the spiral bevel gear fault diagnosis, which is expected to achieve the expected diagnosis effect.