Research On Fault Diagnosis Approaches For Industrial Processes Based On Explainable Deep Networks

Industrial process fault detection and diagnosis technology are essential for safe production and efficient operation.Deep networks are widely used in industrial process fault diagnosis modeling because of their multilayer nonlinear structure that can effectively capture the behavior of complex nonlinear systems and extract high-dimensional coupled big data features layer-wise.However,the black-box nature of deep networks hinders their safe use and decision trust.Although the interpretable artificial intelligence theory developed in recent years can enhance the transparency of deep networks,it needs more relevant research in the field of fault diagnosis for straightforward application.For this reason,there is an urgent to study industrial process fault diagnosis methods based on explainable deep networks to enhance the application of explainable deep networks in fault diagnosis and overcome the challenges of designing deep network structures with guaranteed diagnostic performance.In this paper,the research connotation and feasible ideas of explainable deep network fault diagnosis methods are analyzed in depth.The research of explainable deep network methods for generating fault detection indices is driven by diagnosable performance and network knowledge throughout the three basic tasks of detection,isolation,and estimation.The main research results and innovations are summarized as follows:(1)To address the problems of poor sample interpretability,insufficient model robustness,and high demand for offline modeling samples of existing deep network-based fault detection methods,the minimum threshold-driven sample norm density-weighted deep autoencoder(DAE)fault detection method is proposed.By analyzing the representative samples in offline modeling of fault detection and its impact on network training,the reasons for the high threshold of DAE test statistics are explained.Then,the minimum thresholddriven norm density-weighted explainable DAE monitoring model is proposed,which successfully suppresses the “tail lifting” phenomenon of test statistics under sample norm imbalance conditions,improves the interpretability of the training samples,and the robustness of the model against representative samples? the second step is to generate fault detection indices.A new threshold learning framework based on Freedman’s martingale difference inequality is constructed to derive the minimum required sample size for offline modeling,and a fast method for setting the statistical threshold based on the ranking is proposed,which achieves satisfactory fault detection performance,significantly reduces the offline modeling sample requirement,and improves the threshold learning speed.(2)For the existing deep network fault detection model lacking the correct network structural design guidance and secured fault detectability,performance-guaranteed variational autoencoder(VAE)reconstruction error and latent variable parameter fault detection method are proposed.The ideal spatial mapping of fault detection indices is constructed to analyze the conditions that significantly differ between normal and abnormal samples in fault detection indices.Based on Taylor expansion,the fault detectability of the VAE reconstruction error model and latent variable parameter model are analyzed,then constructing the first-order and secondorder fault detection indices that keep the fault-affected terms from disappearing.Based on this,fault detectability-related propositions are proposed to guide the structural design of deep networks,ensuring the detectability of VAE-based fault detection approaches.In addition,the optimal activation function combination design for the VAE latent variable parameter model is proposed,which achieves the best fault detection performance.(3)The fault attribution methods based on layer-wise contributionfiltered propagation and layer-wise incremental expectation propagation are proposed to address the problems that existing explainable deep network attribution methods do not consider the influence of negative activation contribution,have poor attribution accuracy or efficiency,and cannot be applied to statistical attribution.By considering the absolute activation contribution of the network layer input to the output under different activation functions,two deep network attribution methods based on layer-wise contribution-filtered propagation and layer-wise incremental expectation propagation are proposed by defining the positive/negative contribution and accurately quantifying the absolute activation increment caused by input variables,respectively?on this basis,a statistical equivalent network is proposed.It builds a new framework of interpretable attribution applicable to fault detection index models,constructing the bridge between deep network attribution and statistical attribution.The proposed methods enhance the interpretability of deep network-based fault attribution and effectively improve fault isolation’s accuracy.(4)The fault estimation methods based on the deep decoupling transfer network and interpretable prototype representation are proposed to address the problems in traditional deep network fault estimators,such as lack of decoupled structure design,under-mined network knowledge of detection model,and insufficient fault estimation accuracy.By using parallel forward propagation of sample variables and the trick of sharing network parameters,the fault impact decoupling deep network is designed to realize the decoupling of fault effects on detection indices,which guarantees the fault isolability of the model? on this basis,the deep decoupling transfer network is constructed based on the reconstruction loss and domain transfer loss,which allows the transmission of fault samples to the normal domain thereby realizing fault estimation purposes.On the other hand,interpretable prototype representation is introduced to develop our fault estimation method.It realizes fault estimation objectives by guiding the learning of faulty samples-corresponded normal prototypes according to the input gradient descent direction of a well-learned fault detection model.The proposed method develops a deep network fault decoupling framework,which extends the fault estimation function of the detection model using the learned network knowledge and effectively improves the fault estimation accuracy.