Research On Feature Learning And Fault Recognition Of Oil And Gas Equipment Based On Deep Networks

With the continuous deepening of industrial intelligence,many new problems and challenges appear in the domain of mechanical health management,such as massive data,high-dimensional features and information hiding.Traditional fault diagnosis mode relies on signal processing technology and artificial feature extraction,which cannot meet the need of processing massive multi-source heterogeneous monitoring data in the context of big data.It is urgent to introduce deep learning and other advanced data analysis tools and innovate existing diagnostic ideas to further improve the accuracy,adaptability,intelligence and robustness of fault diagnosis.Therefore,this paper takes the oil and gas dynamic equipment as the research object and deep networks as the diagnostic tool,carrying out in-depth investigations in the aspects of sensitive feature mining,crossworking condition diagnosis,background noise filtering,self-adaptive construction of diagnostic model,multi-source heterogeneous data processing and fusion to solve inherent problems of fault diagnosis in practical applications.The main researches are as follows.(1)Self-learning method of fault-sensitive features based on convolutional neural network(CNN): the extraction of fault-sensitive features relies on signal processing and expert experience,which is difficult and unreusable.To solve this problem,a onedimensional(1D)CNN-based feature learning and fault diagnosis method is proposed,which can directly process the original vibration monitoring and automatically mine the feature representation for fault recognition.Based on 1D CNN,the end-to-end intelligent diagnosis can be realized.The fault diagnosis experiment demonstrates that the diagnostic accuracy of the proposed method is improved by 5 percentage points,and the standard deviation is reduced by 1 percentage point by comparing with the traditional method of “signal processing + feature extraction + pattern recognition”.(2)Fault recognition method across working conditions based on deep domain adaptation: the working condition fluctuation of dynamic equipment will cause the distribution difference among monitoring data,which further results in the inadequate applicability and low recognition accuracy of the diagnostic model.To solve this problem,a deep domain adaptation is proposed,which can project monitoring data from different working conditions into a common feature space by multi-layer non-linear mapping,eliminating working condition disturbance and realizing cross-working condition diagnosis.The fault diagnosis experiment demonstrates that the proposed method has stronger robustness to the working condition fluctuation and the accuracy of crossworking condition diagnosis is increased by 10 to 30 percentage points by comparing with traditional shallow transfer learning methods.(3)Vibration denoising method based on deep network under strong noise disturbance: in real-world applications,the background noise of vibration signal is too strong to distinguish the fault feature information,leading to the difficulty of fault diagnosis.To solve this problem,a deep denosing network is proposed,which establishes the complex mapping relationship between the noisy signal and the pure vibration by data driven method.While filtering out the background noise,fault features in the vibration signal are retained and restored to the greatest extent.Fault diagnosis experiments demonstrate that the proposed method can significantly enhance the signal-to-noise ratio of sample signals,thus improving the accuracy of fault diagnosis under strong noise interference.(4)Adaptive construction method of deep network based on genetic algorithm(GA)optimization: the structure design and parameter setting of deep networks are difficult,relying on expert experience and repetition test.To solve this problem,a GA-based adaptive construction method of deep network is proposed,which can construct the best deep diagnostic model under the current task by automatic iteration optimization.In two fault diagnosis experiments,the proposed method achieves 100% diagnostic accuracy.Compared with relevant literatures,the proposed method is more comprehensive and thorough in parameter optimization.(5)Deep fusion diagnosis method based on multi-source heterogeneous monitoring parameters: existing fault diagnosis strategies adopt a single monitoring parameter with incomplete information and lack effective utilization of multi-source heterogeneous monitoring parameters.To solve this problem,a deep fusion diagnosis method is proposed.A multi-stream CNN is proposed to synchronously process one-dimensional vibration signals and two-dimensional infrared images and automatically mine deep features.In addition,the feature fusion technology is investigated to retain complementary information and remove redundant data.The fault diagnosis experiment demonstrates that the accuracy of fusion diagnosis is 98.87%,which is 10 percent points higher than the traditional single monitoring diagnosis method.Furthermore,through the interaction of multi-source heterogeneous monitoring data and deep networks,the training difficulty of deep models is reduced and the efficiency of intelligent diagnosis is improved.