Research On Bearing Fault Diagnosis Based On Data Feature Analysis And Deep Learning Algorithm

In the production process of various industries,the use of rotating machinery and gear systems is usually involved.Industrial production has higher and higher requirements for mechanical equipment,and machinery equipment is moving towards integration,high degree of automation,high efficiency and diversified loads.Mechanical components are more prone to structural and functional failures under such complex production conditions,which not only cause safety risks but also economic losses.In order to ensure the long-term and reliable operation of mechanical equipment and ensure the safe and effective operation of the production system,it is necessary to perform fault diagnosis and health monitoring on the key components of the mechanical system.With the development of artificial intelligence technology,mechanical fault diagnosis methods based on deep learning are gradually enriched,related algorithms mainly extract high-dimensional features of data through the framework of deep learning.These methods not only does not require too much expert knowledge,but also greatly improve the accuracy of fault diagnosis.The raw data measured by sensors in the working environment of mechanical equipment often contains a lot of noise,which not only affects feature extraction,but also affects the convergence speed and diagnostic accuracy of deep learning algorithms.For this reason,this paper proposes three denoising strategies based on the signal decomposition algorithm.They are the denoising algorithm with non-local means improved traditional threshold,the denoising algorithm based on peak statistical threshold and the denoising algorithm based on variational modal decomposition and adaptive chirp mode pursuit.These proposed denoising algorithm has its own advantages,through simulation experiments and real bearing data experiments,the effectiveness of the proposed algorithm is verified,and it can be concluded that the denoising algorithm based on variational modal decomposition and adaptive chirp mode pursuit is more suitable for processing the vibration data of rotating machinery.In order to more effectively analyze the time-frequency characteristics of the vibration data of rotating machinery,this paper proposes a new time-frequency analysis method based on the variational modal decomposition and adaptive chirp mode pursuit algorithm.According to simulation experiments and real data experiments,the proposed timefrequency analysis method is superior to the traditional time-frequency analysis method based on short-time Fourier transform and wavelet transform.For early bearing faults,the traditional Fourier transform is difficult to effectively extract the fault characteristic frequencies.In this paper,envelope spectrum,Teager energy spectrum and multi-point optimal minimum entropy deconvolution are used to extract the characteristic frequencies of early bearing faults successfully.At the same time,this paper combines the denoising algorithm based on variational modal decomposition and adaptive chirp mode pursuit to improve the performance of the multipoint optimal minimum entropy deconvolution adjusted to extract the fault frequency.Based on traditional fault feature extraction algorithms and traditional machine learning algorithms,and inspired by multi-scale convolutional neural networks this paper proposes multi-modal multi-scale convolutional neural networks.The coarse-grained data of the fault data,the time domain features,the frequency domain features and the fault feature frequency extracted by the multipoint optimal minimum entropy deconvolution adjusted are used as different modal to put into the convolutional neural networks to form a multi-modal multi-scale convolutional neural network.The effectiveness of the algorithm is verified in the bearing failure and gearbox failure data sets.In order to improve the efficiency of the deep learning algorithm for fault classification,this paper proposes a deep learning framework based on multivariate variational modal decomposition and multi-modal convolutional neural networks.The time domain features and the frequency domain features of mode components obtained by multivariate variational modal decomposition algorithm and the characteristic frequency obtained by the multipoint optimal minimum entropy deconvolution adjusted are used as the input of the neural network.This deep learning framework not only improves the accuracy of fault recognition but also has application potential in multi-sensor data fusion.