Research On The Fault Diagnosis Method For Rotating Machinery Using Deep Convolutional Neural Network

Rotating machinery is widely used in national economy industries.It is of great significance to detect faults of rotating machinery effectively.Feature extraction and information fusion of the faulty signals of rotating machinery are the two key factors that affects the accuracy of damage detection.However,the conventional methods of feature extraction use only the manual selected features and the information fusion usually focus on single fusion-level,which may have a negative effect on the effectiveness and automation of mechanical fault diagnosis.Meanwhile,deep convolutional neural network(DCNN)has been attracting increasing attentions in recent years and shows a great potential for feature extraction and information fusion of the faulty signals of rotating machinery.Nevertheless,DCNN is designed based on visual system structure,which may cause the unsuitability of the application of DCNN in the fault diagnosis of rotating machinery.With the analysis of the research background,this paper focuses on the research of DCNN based fault diagnosis methods for rotating machinery to solve the problems above,including manual feature extraction,information fusion with single fusion-level,and the unsuitability between DCNN and fault diagnosis of rotating machinery.An adaptive feature extraction method using Fourier-DCNN for vibration signals is investigated to solve the problems of the feature extraction in fault diagnosis of rotating machinery.Vibration signal contains rich information of rotating machinery and is sensitive to various faults,which makes vibration signal become one of the most commonly used signals in fault diagnosis as well as multi-sensory monitoring.According to the characteristics of the vibration signal,a Fourier-DCNN based adaptive feature extraction method is proposed to overcome the weaknesses of manual feature extraction.Meanwhile,the conventional model of DCNN is modified and improved.Trough the Fourier transform layer with multi-segments,the information in frequency domain can be extracted effectively,and with the enhanced Dropout-pooling algorithm,the training speed and generalization of the model are improved.Finally,the proposed method is validated through the experiment of the fault diagnosis of planetary gearbox and PHM2009 challenge data.The results demonstrate that the proposed method is effective and has a better performance in fault diagnosis than methods with artificial feature extraction or classic DCNN without improvement.A combined DCNN based adaptive information fusion method for multi-vibration signals is proposed to solve the problems of the information fusion in fault diagnosis of rotating machinery.Based on the research of adaptive feature extraction of vibration signals,the combined information fusion method is proposed when the features are extracted layer-by-layer.The multi-vibration signals based fault diagnosis method has a more convincing diagnostic accuracy than single vibration signal based method.Nevertheless,the corresponding signal processing method for multi-vibration signals is also more complicated.According to the characteristic of multi-vibration,a combined DCNN based adaptive information fusion method is provided to overcome the weaknesses of information fusion with single fusion-level.Through the structure of combined-information fusion,the interference among multi-vibration signals is limited and the performance of information fusion is improved.The effectiveness of the proposed method is evaluated through experiments and PCA analysis.The results show that the proposed method can achieve a better diagnosis result than methods with single fusion level or the classical DCNN model.A large-size DCNN based feature extraction and information fusion method is provided to process the multi-type signals for fault diagnosis of rotating machinery.With the development of industrial intelligence and big data technology,current,speed,noise,temperature,pressure and other equipment information are also increasingly applied in the fault diagnosis of rotating machinery,which forms a multi-type signals based fault diagnosis method.The speed signal,current signal and noise signal are selected as the research objects,and with the result of the vibration signal in previous chapters a large-size DCNN based feature extraction and information fusion method is explored.A more common structure of the DCNN model is adopted to extract and fuse features effectively from both the time domain and the frequency domain.Adam optimization method and an adaptive learning rate are also added to improve the DCNN model.Through theoretical analysis and experimental verification,it is proven that the proposed method can process multi-type signals effectively and achieve a better result than methods with single fusion level or using Fourier-DCNN for vibration signal.In practical application of the DCNN based diagnosis method,especially for some large or special rotating equipment,the proposed method may suffer from few data samples.It is difficult to train DCNN,a supervised model,completely with few data samples,which may cause a negative effect on the diagnosis result.The transfer learning method is introduced and combined with DCNN to solve the problem.Through the pre-training and the transfer learning with limitation parameters,DCNN model can be trained completely even with very few data samples.The effectiveness of the proposed method is validated by two groups of experimental data.The results demonstrate that the proposed method can obtain a better diagnostic accuracy than methods without the transfer learning.