Data-driven Health State Diagnostic Method For Typical Planar Parallel Mechanisms

Modern high-end mechanical equipment,categorized as ’non-living organisms’,have lifecycles similar to ’living organisms’.As equipment systems become more intelligent,complex and precise,key mechanisms responsible for core operational functions are required to maintain excellent health state.The data-driven intelligent health monitoring and diagnostic technology is essential for ensuring safe and reliable operation of key mechanisms and preventing major accidents.Planar parallel mechanisms in modern high-end precision mechanical equipment,which are widely used in semiconductor production,precision machining,and aerospace manufacturing,offer advantages of high stiffness,precision,and fast response.The performance of these mechanisms directly affects output smoothness and accuracy.Therefore,this thesis focuses on the intelligent health state diagnosis of planar parallel mechanisms,utilizing operational performance data as the information source and leveraging deep learning and transfer learning techniques.The research contents include the following four aspects:Firstly,a multi-scale graph convolutional network diagnostic model is proposed for complete label space scenarios,incorporating an unsupervised convolutional auto-encoder module.By studying the data correlation in the high-dimensional feature space,a correlation model is constructed to establish the relationship between samples and embed relational information in the feature mining process.At the same time,a novel multi-scale feature extraction architecture is proposed by combining the differential perception performance of graph convolutional networks to achieve adaptive fusion of deep representation features at different scales,thereby promoting the mining and utilization of effective information in performance data.The simulation and experimental cases of a planar parallel mechanism validate the excellent feature extraction performance and health state diagnosis accuracy of the proposed model.Secondly,a graph convolutional auto-encoder diagnostic model is proposed for the imbalanced label space scenario.Through studying the complementary relationships among samples,complementary information in homogeneous attribute samples is deeply mined and evaluated.Moreover,the information richness of imbalanced class samples is improved by integrating and fusing complementary information,which enhances the completeness of feature space in imbalanced scenarios.The experimental cases under multiple working conditions demonstrate that the proposed model can effectively alleviate the problem of model performance degradation caused by data imbalance,and has excellent comprehensive diagnostic accuracy and robustness.Subsequently,a clustering representation-guided unsupervised graph convolutional network diagnostic model is proposed for the non-complete unlabeled space scenario.This study introduces a technique to extract vectorized deep encoding features,enabling maximum retention of high-quality information in service performance data without the guidance of health state label information.Additionally,a hierarchical information fusion strategy is proposed to efficiently utilize information in auxiliary samples and promote accurate allocation of virtual health state labels by clustering algorithms.The model’s exceptional feature extraction performance and complete feature space construction ability are validated through experimental cases.Finally,a cascade transfer network diagnostic model guided by a backward mask update mechanism is proposed for multiple cross-domain scenarios.Based on the differential perception field of the dilated convolutional neural network,a deep representation feature cascade strategy is constructed,which endows the model with dynamic adjustment performance.Additionally,Gaussian kernel mapping is used to quantitatively evaluate the contribution of common knowledge from the source domain,enhancing the interpretability of the model during the transfer process.The multiple transfer cases established by the planar parallel mechanism demonstrate that the model has excellent transfer stability and universality,effectively alleviating the limitations of the ”block” transfer strategy.This research effectively promotes the application of intelligent maintenance technology in the field of planar parallel mechanisms,expands the applicable scope of existing diagnosis theories,and provides new modeling ideas for the complex environments and special scenarios.