Residual life prediction of gear based on deep belief network integration

Due to the influence of alternating stress, the gears in many equipment often have the faults of tooth surface wear and broken teeth, which lead to long downtime, high damage and high maintenance cost. Through the remaining useful life (rul) prediction, the maintenance strategy can be formulated in advance to avoid fatal faults and reduce losses. At present, researchers at home and abroad have done a lot of research on this. Zhou Zhigang and others established a model by observing the change of dynamic meshing force of gear pair with time and the change of dynamic contact force of corresponding bearing, and applied the model to predict the corresponding life of gear and bearing; Feng y x et al. Studied a method of fault diagnosis for the gearbox of fan by using decision tree and expert system. Shi Hui et al. Established a model of random filtering combined with kernel density estimation to predict the remaining life by using the collected gear vibration acceleration and noise data; By decomposing the vibration acceleration signal layer by layer, Zhang Xinghui extracted the energy of different frequency bands to represent the corresponding state of gear, and finally completed the identification of gear from normal to degradation to failure and the prediction of rul under different states by Gaussian hidden Markov method; Wang Jinhai et al. Considered the phenomenon of gear separation, and then established a model based on gear contact geometry and S-N curve to analyze the relationship between the contact fatigue degradation process and rul of gear system under different rotation speeds.

But these methods need to extract the characteristic value manually; The first mock exam is two. The single model based on data collection is often not able to display good generalization performance under other circumstances. The deep belief network can learn the relationship between the data by itself and extract the features more related to the final task without special feature extraction. Moreover, the complementarity of different neural network individuals for the same decision in ensemble learning can improve the generalization performance of neural network ensemble. Ensemble learning needs to be composed of individual networks with large accuracy and difference. Negative correlation learning adds the penalty term of multiple networks to the error function of individual networks, which can increase the difference between individual networks. Liu y pointed out that the negative correlation learning combined with genetic algorithm can better predict and classify, so it is often combined with genetic algorithm to build an integrated model for prediction.

Based on the analysis of gears, a method combining genetic algorithm and negative correlation learning is proposed to construct a deep confidence network integration model, which can simulate the nonlinear relationship between the gear eigenvalues and the running time, so as to predict the fault time and deduce the remaining life.

On the basis of analyzing the influencing factors of gear life and collecting the corresponding data, an integrated model based on deep belief network is established to predict the remaining life of gear. The remaining life is not only related to the current state of the gear, but also related to the historical state. The deep confidence network can explore the rules in the data and build the corresponding model to predict the remaining life of the gear. The integrated model can make the model have better generalization ability, so it can be applied to the prediction of other data sets. The next step is to study how to increase the diversity of individual networks, and consider adding niche technology to build the integration model.