Research On Component-level Intelligent Fault Diagnosis Method Under Nonideal Data-Driven

As the old saying goes," pull one hair and the whole body is affected." Mechanical systems are like dominoes.A failure at the level of a component can bring the whole project to a halt,even cause property damage and casualties.In the early days of our industrial development,the maintenance efficiency of equipment was never efficient enough due to the backward infrastructure conditions of the time.In recent years,cloud computing,artificial intelligence and other cutting-edge technologies have sparked a new generation of industrial revolution in the industrial world,and "data is king" has become the consensus of the times.Since condition monitoring data of machinery equipment carries implicit knowledge of its health status,data-driven intelligent fault diagnosis methods have received a lot of attention.However,fault diagnosis models built on ideal data conditions in the lab cannot be fully adapted to the complex environment of the engineering site,making the traditional deep learning paradigm out of touch with reality.However,the fault diagnosis model trained based on the ideal laboratory data fails to fully adapt to the complex environment of the engineering site,resulting in the traditional deep learning paradigm divorced from the reality.To this end,this paper takes key components of mechanical equipment,such as gearboxes,as research objects,and simulates a variety of close to reality engineering scenarios for common problems driven by non-ideal data.There are variable excitation sources in the working environment of many mechanical equipment,and the process of component damage is random,so many monitoring data with complex failure modes are generated.Deep neural network(DNN)has feature inference and nonlinear mapping capabilities to learn representative features in signals in an end-to-end adaptive manner.However,general DNN have poor sensitivity to the fault degree of the samples,limited ability to identify compound faults,and the intermediate learning process is invisible,which leads to DNN becoming an uninterpretable black-box model.To address the above problems,a Stacked Residual Multi-Attention Network(SRMANet)with interpretability is proposed as a feature extraction method for vibration signals.The Squeezeexcitation Residual(SE-Res)block in SRMANet generates additional features with minimal redundancy and sparsity.It uses attention fusion units to enhance the interpretability of the model and eventually obtain representative features for different types of samples.SRMANet visualizes the learning process of the model from multiple perspectives.Finally,the interpretability,accuracy and adaptability of the model under different operating conditions are verified on complex datasets from planetary gearboxes,respectively.Fault diagnosis methods based on traditional DNN paradigms can appear helpless when targeting non-ideal data conditions in industrial sites.Transient speed variation can cause a drift in the distribution of the sample space,which seriously hinders the adoption and diffusion of data-driven fault diagnosis methods.Meanwhile,weakly supervised learning conditions,such as the lack of sufficient labeled samples,further deteriorate the training conditions for intelligent diagnostic models.To overcome the interference of transient speed variation and weakly-supervised learning conditions,the Complementary-Label Adversarial Domain Adaptation Network(CLADAN)is constructed in this paper.First,joint cost-friendly complementary label learning and Conditional Adversarial Domain Adaptation Network(CDAN)for cross-domain fault diagnosis under weakly-supervised conditions.At the same time,the "multi-peak distribution" of complementary labeled classifiers is avoided by the technique of discretizing the category probability mapping.Also,the "multi-peaked distribution" of the prediction probabilities of complementary labeled classifiers is avoided by the technique of discrete category probability mapping.In the adversarial training process,this paper proposes a virtual adversarial regularization(VAR)strategy to further weaken the interference of transient speed variation from the perspective of improving the robustness of the domain adaptation model,so that the diagnostic model learns to domain invariant features.The superiority of CLADAN was demonstrated on the gearbox variable speed dataset and the University of Ottawa bearing dataset,respectively.Although CLADAN has good performance,its process of overcoming transient speed variation condition interference is implicit;meanwhile,its sample label space in the source and target domains is completely symmetric,i.e.,the sample type in the target domain must be visible.In real industry,the asymmetric problem of the domain label space is inevitable,because it is often difficult to fully obtain the prior information of the target domain.To address the above challenges,this paper proposes a fault diagnosis method based on improved universal domain adaptation to reduce the dependence on the spatial symmetry of domain labels under transient speed variation.Combining the a priori knowledge of signal theory,the angular domain resampling technique is used to reduce the disturbing components of the raw signal that are related to the speed but not to the fault.An improved objective function with weighted contrastive domain discrepancy loss is designed and embedded into the universal domain adaptive network,further constraining the decision boundary of the diagnosis model from the perspective of data condition distribution.Based on the sample potential label information,the improved loss function can adaptively reduce the distribution distance of the same types of samples and enlarge the distribution distance of various types of samples.Most cross-domain fault diagnosis methods often require that the fault database can open access authority to all fault modes during the model training phase,which are unrealistic.In industrial sites,the types of samples in the fault database are gradually increasing,and this change requires that the diagnostic model can continuously learn new failure modes.However,the continuous learning process can easily lead to catastrophic forgetting of the standard fault diagnosis model.To adapt to the dynamic changes of fault types,this paper proposes a continual learning model based on weight space meta-representation(WSMR)to implement class-incremental fault diagnosis on a multimodal dataset of natural faults of switch machine plunger pumps.Among them,the success of Modified Wavelet Kernel Net(MWKN)demonstrates that improving the structure of the base model can reduce the forgetting of old knowledge.WSMR maintains the diagnostic performance of the model for both new and old faults in a cost-friendly manner by inferring about and preserving the meta-representation distribution of the task-specific base model.The results show that the combination of MWKN and WSMR can effectively reduce the forgetting rate of the class incremental fault diagnosis model,which is better than the existing methods.