Research On Fault Diagnosis Algorithm Based On Multi-Label Learning

With the high-speed and high-quality development of information and technology and productivity,the deep integration of the new generation of artificial intelligence technology and industrial manufacturing has become the core driving force of the new round of the industrial revolution.Rolling bearing is the key component of rotating machinery and equipment,which is an important production tool in modern industry.Its health status has greatly affected the efficiency,safety and life cycle of the equipment.Therefore,it is of great practical significance to achieve accurate intelligent bearing fault diagnosis.At this stage,the diagnosis method of failure still has limitations.Firstly,most algorithms still use the multi-classification method of single-label to diagnose single faults and composite faults,which not only regards different degrees of fault as unrelated and completely different fault categories,but also regards compound fault as a new fault category different from all faults in the combination,which breaks the correlation between faults and their development;Secondly,existing algorithms fail to make full use of multi-scale features and multi-sensor data.Using a single network to extract all the features of the whole model will ignore the features on multiple scales,resulting in the absence of important features.Moreover,vibration signals from different sensors contain different degradation information.Ignoring the correlation between these degradation information will limit the prediction accuracy and generalization ability of the algorithm.In view of the above problems,this thesis introduces multi-label learning and deep learning in the field of fault diagnosis,and carries out the following two aspects of research:(1)In order to solve the problem that the existing single-label algorithms cannot deal with the correlation between faults,this thesis applies multi-label learning to the field of fault diagnosis,improves the existing multi-label algorithms,and proposes a multi-label fault diagnosis algorithm based on correlation and redundancy.The proposed method gives the corresponding weight to the k-nearest neighbor of each instance according to the cosine similarity,uses the weighted sum of its label vectors to expand the original feature set to combine the instance correlation,removes redundant features by calculating mutual information,and takes the label correlation into consideration to select the label specific feature set with the most discrimination and low redundancy for each class label.The experimental results on ten benchmark multi-label data sets and one compound fault diagnosis data set show that the proposed algorithm is a more effective and accurate multi-label classification algorithm,and has good application performance in fault diagnosis.(2)In order to solve the problem that the existing algorithms do not make full use of multi-scale features and multi-sensor data,and do not consider the correlation between fault labels,this thesis proposed a fault diagnosis algorithm based on multi-scale separable convolution.The proposed algorithm applies deep learning and multi-label learning to the field of fault diagnosis at the same time,and uses the original one-dimensional vibration signal data of multiple sensors as the input of multiple channels.Separable convolution was used instead of standard convolution to map the channel correlation between different sensor data and the time correlation between the same sensor data in multi-sensor data.Two networks with different kernel sizes and depths are constructed by stacking separable convolution linear bottleneck blocks with inverted residuals and pooling layers to extract multi-scale features sufficiently.After feature fusion,they are sent into LSTM to extract long-term time features.The proposed algorithm is verified in a single fault data set and a compound fault data set respectively.The results show the effectiveness and superiority of the algorithm in fault diagnosis.