Study On Gear Residual Life Prediction Method Based On Singular Value Decomposition And Deep Recurrent Neural Network

Gear is the most important part in the automobile gearbox,and the automobile gearbox is the main assembly of the automobile transmission system,which takes on the important mission of transferring power from the engine to the wheel.Once the gear of the gearbox occurs abnormal or malfunction in the running car,it will lead to the failure of the automobile gearbox and cause traffic accidents,which endanger the safety of people's life and property.Therefore,the prediction of gear remaining life has important research value and practical significance to ensure the safe driving of vehicles.So far,the use of cumulative fatigue damage theory and the combination of multi model co-simulation or finite element analysis based on the cumulative fatigue damage theory is the most common method to predict the automobile gear residual life.In addition,some scholars have also studied and proposed the method of predicting the automobile gear residual life by using complex mathematical models and artificial intelligence.However,the current theory of cumulative fatigue damage is not very perfect,and there are various defects,for example,the order of loading automobile gear is sometimes neglected and the S-N curve of corresponding material is nonexistent.The precision of the model based life prediction method depends very much on the used mathematical model and the accuracy of this method will be greatly affected if the mathematical model is not in conformity with the actual situation.The previous residual life prediction method based on artificial intelligence only uses a single feature or a few sensitive features or at most the principal component analysis for feature fusion as a criterion,and the used artificial intelligence algorithms are generally back propagation neural network,radial basis function neural network and support vector machine to predict the residual life.However,the single feature is greatly influenced by the experiment sample and the working condition,and the trend of several sensitive features may be different,so that the prediction accuracy will be greatly influenced.Therefore,the feature fusion is needed.Principal component analysis is a linear dimension reduction algorithm,while the whole life cycle vibration data of a automobile gear is usually nonlinear.The output of feedforward algorithm is only related to input and has nothing to do with its order,and it is not suitable to deal with the life-cycle characteristics associated with time.Therefore,this paper selects a manifold learning algorithm suitable for nonlinear data and a deep recurrent neural network suitable for time-correlation features to predict the automobile gear residual life.Because of the existent random noise in the automobile gear vibration signal,it is necessary to be denoised by removing the noise and leaving the fault shock signal as much as possible.In this paper,the singular value decomposition and differential spectrum theory are introduced,and the noise reduction method of segmented adaptive singular value decomposition is proposed and has some improvements in comparison with the existing maximum differential singular value decomposition.Finally,this method is used to reduce the noise of the fault simulation signal of the automobile gearbox.The comparison results show that this method has a higher noise reduction performance than the traditional maximum differential singular value decomposition method.The time-domain and frequency-domain characteristics of the automobile gear vibration signal can,to a certain extent,reflect the change of the automobile gear state.However,since the trend of multiple features may vary different,and a single feature is greatly affected by experimental sample and operating condition,it is often not accurate to use these characteristics to judge the operating condition of the automobile gear.Therefore,these features need to be reduced and merged.This paper introduces some typical manifold learning algorithms.In order to select a manifold learning method suitable for the full life cycle characteristics of automobile gears,a series of gradual vibration signals of the automobile gearbox gear changing from normal to minor faults to complete failure are simulated according to the fault vibration signal model,and timedomain and frequency-domain characteristics of these vibration signals are extracted.Then use various manifold learning algorithms to reduce the dimensions of these features.By comparing the simulation results,it is found that the isometric mapping algorithm has the best dimensionality reduction effect.According to the automobile gear trend fusion characteristics obtained by dimension reduction,this paper will use deep long short-term memory neural network to predict its trend and indirectly predict the automobile gear life.By comparing and analyzing a variety of typical recurrent neural networks,this paper uses a deep long short-term memory neural network to predict the automobile gear residual life,and studies the method of data normalization,network parameter initialization and learning algorithm.In addition,this paper presents the specific flow of the automobile gear residual life prediction method of the combination of segmented adaptive singular value decomposition,isometric mapping algorithm and deep long short-term memory neural network.In this paper,segmented adaptive singular value decomposition,isometric mapping algorithm and deep long short-term memory neural network are combined to predict the automobile gear residual life.This method is used to analyze the gradual vibration signals of the automobile gearbox gear changing from normal to small failure to complete failure and the whole life cycle vibration signal measured in the gear contact fatigue test.Both simulation and experimental results show the effectiveness of the proposed method.