Incomplete Data-Driven Industrial Process Fault Diagnosis

With the advent of smart manufacturing,manufacturing industry is based on the value demand from the user side,putting forward higher requirements for product quality,production efficiency,energy consumption,environmental pollution,production and personnel safety,and driving high quality,efficient,green and safe operation of the whole industrial process.At the same time,the further liberation and upgrading of productivity leads to the rapid advancement of integrated automation systems in the direction of scale,integration,and intelligence.Thus,building a positive connection between data and production activities through data-driven fault diagnosis to meet the actual needs of industrial processes for stability,reliability,and safety,will become a key foundation for manufacturing industry enhancing competitiveness and innovation,and is also the core of smart manufacturing.Data is the most important medium to connect industrial processes and users,but in some specific scenarios,the dilemma of incomplete elements such as fault source,label,and category hinders traditional data analysis methods from fully describing operation status and leads to the problem of insufficient robustness of fault diagnosis results.Overcoming the challenges posed by incomplete elements and achieving timely and high-performance fault diagnosis is conducive to early elimination of potential risks brought about by faults and prevention of major accidents.In this paper,we focus on three core steps of fault detection,fault isolation,and fault identification in fault diagnosis,and take deep learning technology as the core to explore fault feature enhancement and fault pattern learning,which is driven by incomplete data from the specific situation faced by each.Meanwhile,the complex characteristics of modern industrial processes,such as strong nonlinear coupling of variables,long and dynamical changing of procedures,and inconspicuous fault magnitude,are also fully considered.The main research of this paper is summarized as follows:(1)For fault detection and isolation of complex nonlinear industrial processes in scenarios with incomplete prior information and unknowable fault sources,we propose an unsupervised and integrated method based on global-local inter-variable structures.First,a convolutional operator based on global-local inter-variable structures is proposed to extract global and local nonlinear features between variables to improve the accuracy of modeling complex nonlinear industrial processes.Then,a stacked autoencoder is constructed based on the feature enhancement to represent the normal operation state by an unsupervised reconstruction learning paradigm,so that reconstruction error can be used to effectively monitor the operation state and detect faults in time.In addition,a self-attentive variable selection module is embedded in this integrated method to drive the interpretability of fault detection results,and relative attention and its control limit are proposed to achieve nonlinear fault isolation quantitatively.Experimental results show that a comprehensive and adequate characterization of global and local inter-variable structures can effectively enhance nonlinear features and thus improve the accuracy of fault detection,while the end-to-end fault isolation is advantageous in mitigating redundant fault variables and smearing effects.(2)For fault detection and isolation of complex dynamic industrial processes in scenarios with incomplete prior information and unknowable fault sources,we propose an unsupervised and integrated method based on deep slow feature analysis.First,a convolutional long short-term memory module is introduced to extract long-sequence spatiotemporal features from highdimensional time-series data to improve the accuracy of modeling complex dynamic industrial processes.Then,based on the high-order spatiotemporal features,an unsupervised deep slow feature analysis learning paradigm with strong coupling to dynamic fault detection is proposed to enhance the representation of the normal operation state and highlight abnormal operation states by extracting slow features with extremely slow-changing patterns.Compared with the state-ofthe-art reconstruction learning paradigm based on adversarial encoders,the deep slow feature analysis learning paradigm can improve fault detection performance in a lightweight manner based on long-sequence spatiotemporal features.In addition,the variable selection module is also embedded in this integrated method,obtaining clearer and more accurate backtracking of fault variables based on the inherent causality revealed by long-sequence temporal features.(3)For semi-supervised fault identification in scenarios with incomplete label information,we propose a method based on a self-attention convolutional ladder network.First,we propose a module that adaptively fuses global and local inter-variate structures by self-attention weights to learn features with strong representation ability while relying on only a few labeled fault samples,especially for incipient faults.Then,to overcome the underfitting problem caused by the insufficiency of labeled fault samples,a convolutional ladder network is constructed based on the feature enhancement to fully learn the information from a few labeled fault samples and some unlabeled fault samples.In addition,a skipped connection structure is introduced in the ladder network to balance the different obj ectives between supervised and unsupervised learning in a single-stage semi-supervised training process.Experimental results show that the proposed semi-supervised approach can effectively improve the robustness of fault identification in the presence of incomplete data by maximizing the use of both labeled and unlabeled fault samples,and the visualization of the self-attention vector verifies the contribution of self-attention fusion mechanisms to represent incipient faults.(4)For fault identification in scenarios with incomplete category information,we propose an open-set fault identification task and a method based on hypersphere manifold-guided reciprocal point learning.The open-set fault identification task not only needs to identify known class faults,but also needs to distinguish unknown class faults from known class faults.On the one hand,in order to optimize the empirical classification risk of known class faults,an angular threshold penalty is proposed to explicitly enhance the intra-class compactness and inter-class diversity of known class faults in the hypersphere manifold space.Thus,it is possible to extract fault features with strong discrimination and improve the performance of known class fault identification.On the other hand,in order to optimize the potential open-set space risk in the absence of unknown class fault samples,the distribution of unknown class faults in the hypersphere manifold space is estimated by reciprocal point and manifold mixup in indirect and direct ways to improve the distinguishability between known and unknown class faults,respectively.Compared with the state-of-the-art open-set object identification methods,the proposed method enhances the optimization of empirical classification risk and open-set space risk in reciprocal point learning by hypersphere manifold,which can improve the detection accuracy of unknown class fault while maintaining the high identification performance of known class fault.