Research On Chemical Fault Diagnosis Methods Based On Deep Learning

Modern chemical processes tend to be more complex that lead to the closer links between chemical components and a slight mistake may cause the chain reaction.Therefore,chemical industries have always been concerned about methods for reducing the risk of accidents because they may commonly in extreme environments,such as extraordinarily high temperature or pressure,which may result in public damage and large economic losses.Therefore,chemical fault diagnosis has become one of the hotspots in the field of chemical industry and possesses significant practical significance.With the rapid development of wireless communication,computer and sensor technology,large amounts of data are being produced and acquired.The data can be analyzed to distinguish whether a fault has occurred in chemical processes,while determining significant potential in chemical fault diagnosis.This method also called data-driven fault diagnosis which mining potential information from the large amounts of collecting data and become a hot issue in recent years.As a novel type of data-driven method,deep learning demonstrates significant ability in data expression and able to learn the feature of input patterns adaptively,which has been widely studied in the field of pattern recognizize.Considering the defferent characteristic of chemical fault data,including the lack of labeled samples,data imbalance and data stream.We then employ the reaerch to solve these problems based on the deep neural network,which is a typical model of deep learning.A deep neural network(DNN)with a novel active learning criterion for inducing chemical fault diagnosis is presented in this study.DNN with deep architectures could be developed through deep learning to learn a suitable feature representation from raw sensor data in an unsupervised manner using stacked denoising auto-encoder(SDAE)and work through a layer-by-layer successive learning process.Considering the expensive and time consuming labeling of sensor data in chemical applications,in contrast to the available methods,we employ a novel active learning criterion for the particularity of chemical processes for further fine-tuning of diagnosis model in an active manner rather than passive manner.The result shows the the proposed method can obtain superior diagnosis accuracy and provide significant performance improvement in accuracy and false positive rate with less labeled chemical sensor data by further active learning compared with existing methods.Since the different faults occur with different frequencies while this issue has scarcely been addressed in developing a diagnosis model in chemical processes.A novel imbalanced modified deep neural network(IMDNN)is proposed to promote the fault diagnosis for imbalanced data.The first step in designing the data processing method to solve the imbalance in the view of data.Then the research employs an imbalanced modified processing method combined with active learning for the extraction and generation of valuable information in view of model feedback rather than blindness.Deep learning is utilized in this method as a basic diagnosis model to excavate potential feature automatically.The simulation results indicate that IM-DNN considerably better than existing method and possesses significant robustness and adaptability in chemical fault diagnosis.Then for the continuously arriving of new fault modes,DNN is promoted in an incremental hierarchical way.Unlike the traditional model that trained on a static snapshot of data,instead,we inherit the existing knowledge in “clone” and hierarchically expand the diagnosis model by similarity of faults.The similar faults that judge by fuzzy clustering merge to a superclass and every sub-model share the same architecture succeed in the previous knowledge that can be trained in parallel.The result shows the adaptability and efficiency for fault diagnosis in chemical data stream.