Research On Bearing Failure Classification Based On Deep Learning

Rolling bearings are the main components of mechanical equipment,with high operating accuracy and low prices.They can be widely used in equipment in fields such as machinery,heavy industry,and transportation systems.The faults of rotating machinery and motors in mechanical equipment are all related to rolling bearings.Therefore,conducting research on the diagnosis and classification of rolling bearing faults has important practical significance for improving the reliability of mechanical equipment.This project proposes two bearing fault classification methods based on deep learning by studying the low accuracy of traditional fault diagnosis classification methods.One method is the Gaussian Bernoulli Deep Belief Network(GB-DBN)combined with Attention Mechanism(ATT)for bearing fault diagnosis and classification.This method utilizes attention mechanism to dynamically weight bearing fault features,improving the accuracy of bearing fault diagnosis and classification.Another method for bearing fault diagnosis and classification is the combination of Deep Convolutional Neural Networks with Wide First-layer Kernel(WDCNN)and Long Short-Term Memory network(LSTM).This method can directly classify raw data and achieve good classification results.The main research design and innovative content of the two models are as follows:1.A new bearing fault diagnosis classification model is designed using Gaussian Bernoulli deep belief network combined with attention mechanism.The improved deep belief network adopts the Gaussian Bernoulli Restricted Boltzmann Machine(GB-RBM)model,which solves the problem that the traditional Restricted Boltzmann Machine(RBM)has poor reconstruction and fitting effect for non binomial distribution data,and can better adapt to data of various distributions.The new method proposed in this project utilizes attention mechanism to solve the problem of over focusing on certain signal fault features in traditional fault classification problems,thereby improving the generalization ability of the model.In addition,this topic has improved the loss function,using 1-cosine similarity as the loss function to better preserve the differences between categories,reduce the sensitivity of the model to signal strength,and improve the accuracy of bearing fault classification.2.Deep Convolutional Neural Networks with Wide First-layer Kernel with Long Short-Term Memory network for bearing fault diagnosis(WDCNN-LSTM)can achieve end-to-end fault detection methods for real-time monitoring of industrial machinery,thereby improving production efficiency.This method uses a convolutional neural network with a first layer wide convolutional kernel to directly input one-dimensional vibration signals without any transformation,greatly improving the ability to extract fault features.The network also has certain denoising characteristics,which can reduce the impact of noise on diagnostic classification results.Long Short-Term Memory networks are used to learn long-term dependencies of signal features and better handle global features.WDCNN-LSTM solves the problem of low classification of raw data by traditional classification methods and effectively suppresses high-frequency noise.In order to verify the effectiveness of the new method proposed in this project,two commonly used benchmark real vibration datasets were used for experimental verification,and compared with other deep learning fault diagnosis classification methods.The experimental results showed that the two deep learning based bearing fault classification methods proposed in this project can achieve relatively advanced classification results in the field of bearing fault diagnosis.