Research On Rolling Bearing Health Management Based On Graph Information And Deep Learning

In the modern industrialised production system,machines and tools are widely used in different industries.The reliable operation of machinery and equipment has an irreplaceable role to play in safeguarding modern industrial production and life.However,bearings and other rotating machinery are often subject to heavy loads,high temperatures,humidity and other environmental factors when they are in operation,making them more prone to failure.When the bearings of machinery and equipment fail,it often directly affects the efficiency of work and even causes greater social losses or generates accidents.Therefore,the health condition of the bearings is of vital importance.Traditional diagnostic methods are not obvious in the extraction of fault signal features,the process of building predictive models requires a certain amount of a priori knowledge and performs relatively weakly when dealing with large data.The development and application of deep learning neural networks provide new ideas for condition detection and life prediction of bearings.This paper is based on convolutional neural networks and long and short term memory networks for the operation and maintenance of bearings to exhibit health management open research,the details of which are as follows.(1)for the bearing fault features time-frequency image features are not obvious,difficult to identify the problem.In this paper,a fault diagnosis method based on image enhancement and local maximum synchronous compression is proposed.First,the acquired time-domain signal is transformed by local maximum synchronous compression to obtain the instantaneous frequency image of the signal.Then,the proposed local histogram cropping equalisation image enhancement algorithm is used to image process the time-frequency image to obtain a more clearly textured time-frequency image as the training and test set.Secondly,in order to extract fault features from the enhanced time-frequency images,an MDCNet network is proposed,which consists of a multi-size convolutional kernel module,a dual-channel pooling layer and a cross-stage partial network.Finally,the fault signals were collected on a bearing fault test bench for prediction with an accuracy of 99.7%.A comparison with other machine learning methods was also made.The results show that the prediction accuracy of the proposed method is higher than that of other methods,indicating that the method is feasible in engineering practice and can meet the needs of practical engineering.(2)For the current problem of low accuracy and poor predictability of bearing life prediction,a new method of rotating machinery life prediction is proposed.Firstly,the rolling bearing vibration signal is collected,the time domain and frequency domain features of the signal are extracted,and a feature matrix is created.Then unsupervised classification and feature selection are carried out by improving the self-organising feature mapping method.Sensitive features are used to construct rolling bearing health indicators.Next,the important information is extracted through macro and micro attention mechanisms to achieve weight amplification.The fused features are used to achieve rolling bearing life prediction.Finally,the proposed method is experimentally validated.The experimental results show that the method is able to predict bearing life with limited measurement data,outperforms other prediction methods,and can provide new ideas for health state prediction and management of mechanical components.