Application Of Time-series Deep Networks In Online Fault Diagnosis In Process Industry

Deep learning and process industry trouble-shooting are combined together because of the rapid development of sensing technology and machine learning.These models have better complex modeling capabilities compared to shallow neural networks.In real industrial processes,process dynamic information is very important because the dynamic implementation of monitoring systems is the guarantee of stable operation of chemical processes and should be retained during the model building process.While traditional encoder-decoder networks exhibit the ability to extract time-dependencies,the architecture encodes the input sequence into a fixed-length internal representation.This limits the performance of this network when relatively long input sequences are considered.Although the long-short time neural network can effectively capture long time sequence features but because the network input layer compresses the real domain information into a fixed length,the network still suffers from gradient disappearance and gradient explosion as the time sequence grows.To address this problem,we first propose a spatio-temporal feature extraction method based on a deep convolutional bidirectional encoderdecoder representation network with gated recurrent units for dynamic fault diagnosis.For the traditional recurrent neural networks with large problems in extracting spatial dependencies,a design of spatio-temporal convolutional operation is designed The structure can effectively extract fault-sensitive spatial-temporal domain local features in the sequence.The system dynamics is characterized by a bidirectional encoder-decoder network combining the before and after sequences to extract the representations from future time steps.In addition,gated cyclic units are deployed in the encoder and decoder to obtain a more compact network topology than conventional cyclic networks by reducing the gating structure.The resulting deep network not only generalizes the importance of temporal features,but also allows for simultaneous interpretable feature representation and classification.Then,an attentional mechanism is introduced to propose an attentional enhancement mechanism for fault diagnosis of complex chemical process data.Unlike traditional fault diagnosis and classification methods,the attention layer is able to detect and focus on local temporal information.The resulting deep network not only retains the importance and contribution on each local instance,but also allows for simultaneous interpretable feature representation and classification.A detailed comparative study shows that the developed model is competitive with several other approaches during the Tennessee Eastman benchmarking process.Finally,we propose a dynamic fault detection and diagnosis technique based on a deep encoder-decoder network with a self-attentive mechanism.The self-attentive mechanism is used to weight the local feature vectors and retain the correlation between the local information of the signal and the process operation state,so as to extract the effective feature vectors.The extracted features are then fed into a bidirectional encoder-decoder network.The resulting deep network not only generalizes the importance of local temporal features,but also allows for interpretable feature representation and classification at the same time.Experiments on the benchmark Tennessee Eastman process show that the proposed model outperforms classical diagnostic methods based on convolutional layerbased long short-term memory（LSTM）networks in terms of receiver operating characteristic（ROC）and precision recall（PR）curves.