Research On Fault Diagnosis Method Based On Incremental Structure

In recent years,the rapid development of science and technology,the modern industrial chemical process has become increasingly complex and structured,and the operating conditions and operating procedures have become more and more variable.While increasing the productivity of the enterprise and bringing great benefits,it has also made production.The probability of equipment and production process failure gradually increases.A small fault at one place may cause widespread failure or even stop the production process.Fault diagnosis and detection technology is the most commonly used method for handling abnormal working conditions in industrial chemical processes,and it provides safety guarantees for the entire industrial chemical production process.As a data-driven deep neural network,with its powerful feature extraction capabilities,it performs well in processing modern industrial process data with complexity,multivariability,and strong coupling,and has achieved many results.Deep neural networks provide a new direction for fault diagnosis and detection technology.However,with the development of industrialization,industrial data generated by chemical processes has grown geometrically,and 80% of all global industrial data has been generated in the past five years.In the face of such a huge non-static data flow,traditional deep neural networks can no longer meet the requirements for processing data changes that change over time in a given time,resulting in more and more unprocessed data being accumulation.However,traditional deep neural networks process data in batches,in which the categories of all objects are known in advance,and all training data can be accessed simultaneously in any order.But already trained deep neural network models often completely forget all the knowledge they have learned after relearning new categories of knowledge.This phenomenon is called catastrophic forgetting.In order to solve this problem,this paper proposes a fault diagnosis method based on incremental structure,which can continuously learn new knowledge from new category data while retaining the old knowledge that has been learned,thereby avoiding catastrophic forgetting.This thesis first proposes a DSResnet network structure as a feature extractor.By introducing the idea of dense jump connections,the feature maps of each residual block are combined on the channel structure,and then passed to the last 1 × 1convolution layer.In this way,a combination of low-level features and high-level features is provided to improve the performance of network feature extraction.Finally,the network structure proposed in this thesis is compared with a convolutional neural network(CNN)with the same number of layers and a residual neural network(Resnets)without dense jump connections to prove the superiority of DSResnet feature extractor.Secondly,this thesis proposes an enhanced loss function formed by combining distillation loss and AM-Softmax loss function,and combining a small sample set with a new class of sample set.Finally,the incremental structure-based fault diagnosis method proposed in this article thesis uses the DSResnet network as a feature extractor,and adds an enhanced training set and an enhanced loss function,so that the network can continuously learn new knowledge from new data while retaining what has been learned Old knowledge.Incremental comparison experiments were performed with three classification algorithms of "Fine tune","iCaRL" and "Fixed representation" on the TE process fault data set of 10 categories.It is verified that the fault diagnosis method based on incremental structure proposed in this thesis has higher classification accuracy than the existing incremental classification algorithms,and can successfully avoid the catastrophic forgetting problem.