Research And Application Of Anomaly Detection Method In Chemical Process

Nowadays,with the development of science and technology,chemical products can be seen everywhere in people’s daily life.At the same time,the emergence of chemical products also has an important impact on people’s lives.In this context,the number of chemical enterprises has also increased,and the scale of chemical enterprises has become larger and larger.Safe chemical production process is the primary issue that chemical enterprises need to guarantee.In recent years,chemical accidents at home and abroad have occurred frequently.At about 5:10 on March 3,2019,a hydrogen sulfide gas poisoning accident occurred in the phosphoric acid filling area of the logistics department of Weng Fudazhou Chemical Co.,Ltd.,resulting in 3 deaths and 3 injuries.At around 16:20 on January 14,2021,Henan Shunda New Energy Technology Co.,Ltd.,located in Zhumadian High-tech Industrial Development Zone,had a suffocation accident during the extraction of the protective agent from the hydrolyzed protective agent tank,resulting in 4 deaths.3 people were injured,and the direct economic loss was about 10.1 million yuan.At present,the inspection of safety conditions in the vast majority of chemical companies is still based on the traditional,inefficient and poor real-time method of regular inspection.Under the background of polypropylene digital twin project,this paper studies univariate anomaly detection algorithm,multivariable anomaly detection algorithm and anomaly tracing algorithm,and verifies them based on the real data of polypropylene plant.The research contents are as follows:1.Experiments are carried out in three data preprocessing algorithms,Fourier transform,wavelet transform and Hilbert-Huang transform.Combined with the actual needs of the enterprise,the best data preprocessing algorithm under the current project is obtained.2.Due to the steady-state of polypropylene plant,a single variable anomaly detection algorithm based on statistical process control is proposed to monitor the whole plant.And can complete the monitoring of production process capacity through the principle of Six Sigma.3.Aiming at the problem that the PCA algorithm cannot highlight the key points of the feature vectors obtained by reducing the data dimension,the SPCA algorithm using the L1 regular term is proposed to sparse the feature vectors.And through the experimental comparison,it is confirmed that the SPCA algorithm has a higher detection rate for faults.4.When using Bayesian network for traceability and using the previous K2 algorithm to model the network structure,it has a strong dependence on the input order of nodes.This paper proposes to use AHP to pre-sort the order of input nodes,which makes the constructed network structure more scientific.