Deep Sparse Learning And Its Application In Predictive Maintenance Of Rotary Machinery

Driven by the integration innovation of intelligent manufacturing,industry 4.0,made in China 2025 and industrial big data,modern industry is undergoing a new revolution from traditional manufacturing to the intelligent manufacturing industry.Rotary machinery,as one of the most important components of this revolution,is constantly being innovated to improve production and economic efficiency.However,the key component faults of rotating machinery will inevitably occur in the long-running operation.Once the fault occurs,it would cause economic loss and even catastrophic accidents.Therefore,it is of great importance to developing corresponding fault diagnosis methods based on sparse representation learning theory to accurately judge and timely respond to early faults of rotating machinery.The main contribution of this work is as follows:Firstly,to overcome the underestimation deficiency of traditional sparse regularization,a weighted dual regularization sparse representation approach is proposed.The sparse learning model was constructed via generalized minimax-concave penalty function and norm regularization,the feature weights were introduced into the sparse learning model.The aim of removing irrelevant components and extracting features was achieved by smoothing weighted short-time Fourier coefficients.The validity of the proposed method is verified by bearing fault feature extraction experiments.Secondly,to avoid the underestimation and inaccurate estimation of fault feature components in the original sparse low-rank model,an enhanced sparse low-rank representation method is proposed.The generalized minimax-concave penalty function is used to overcome the deficiency of the original regularization underestimation,the truncated nuclear norm regularization is used for self-similarity and periodicity protection of weak fault impact components.The effectiveness and superiority of the proposed method are verified by rolling bearing fault feature extraction experiments.Thirdly,since the mismatch issue between the noise prior assumptions and the actual situation in the sparse representation model,divergence is used to measure the difference between the real noise distribution and fitted noise distribution,the mismatch between the prior assumptions of Gaussian noise and the outliers and outliers could be solved by minimizing the difference.Finally,the fault impact characteristics are retained and outliers and outliers are removed.The superiority of the proposed method is verified by the experiment of planetary gear fault feature extraction.Fourthly,the traditional time-frequency analysis methods suffer from blurry energy and low resolution.A new unified sparse time-frequency analysis(STFA)framework is proposed to concentrate the blurry energy,restrain noise,separate condition-related components and also retain the signal reconstruction property.The STFA framework leverages the weighted Elastic Net(EN)sparse regularization for sparsity-inducing and energy concentration and uses the reconstruction error term for condition-related component separation and signal reconstruction,which bridges the gaps among sparsity,decomposition,transformation and reassignment.Theoretical analysis and comprehensive investigation of the proposed framework are performed in practical cases.Fifth,sparse representation theory is combined with convolutional neural network to establish a sparse deep learning model.A wavelet convolutional kernel is designed to replace the original randomly initialized convolutional kernel and improve the interpretability of convolutional neural network.Sparse regularization was introduced to simplify the model and avoid over-fitting.Meanwhile,the channel weighted residual module is used to learn the importance of each feature adaptively to enhance the directivity of feature extraction.The proposed network shows strong ability and high identification accuracy in defect detection of additive manufacturing process.Finally,the main content of the research is summarized and the future work is discussed.