Broad Learning-based Complex Industrial Process Monitoring

With the rapid development of science and technology in today’s era,modern industries are developing in the direction of complexity,intelligence and refinement.If the industrial process monitoring system does not detect the failure of the process industry,it will cause huge economic losses to the process industry,and even casualties.Therefore,real-time and effective monitoring of the operating status of complex industrial processes and ensuring the safe and reliable operation of industrial processes is an important guarantee for the sustainable development and sustainable profitability of complex industrial processes.This article focuses on the time-varying characteristics and data space of modern complex industrial processes.Structural characteristics and imbalances are not analyzed and modeled,and the focus is on the monitoring of complex industrial processes based on breadth learning.The specific content is summarized as follows:(1)Aiming at the time-varying characteristics of complex industrial processes and the spatial structure characteristics of different data in industrial processes,a learning method that can realize the global dispersing but neighbor preserving-broad learning(GDNP-BL)of monitoring data in different working conditions is proposed.It can realize long-term online monitoring of continuous time-varying complex industrial processes.First,an improved manifold regular constraint method is proposed,which is introduced into the traditional broad learning to capture the global spatial structure characteristics and local neighborhood structure characteristics of complex industrial data.After obtaining the best expression result of industrial process monitoring data in low-dimensional manifold feature space based on the GDNP-BL model,the corresponding process statistics are constructed for process fault monitoring.At the same time,in order to obtain the optimal process monitoring model and meet the needs of online monitoring of time-varying industrial processes,an online incremental learning strategy that can realize the process monitoring data characteristics of the GDNP-BL system is proposed,which can realize the process monitoring amount based on changes in working conditions.And online update of control limits.Finally,experimental verification was performed on the data set of non-linear numerical simulation,TE process and actual Baosteel industrial process.Experimental results show that the proposed method can effectively improve the accuracy of fault detection and reduce the false alarm rate of fault detection.It ensures the sensitivity and real-time performance of process monitoring,and is suitable for long-term online monitoring of continuous time-varying complex industrial processes.(2)Aiming at the problem of data imbalance in the industrial process due to the inconsistent frequency of failures in the industrial process,a data distribution-based cost-sensitive-BL(DDb Cs-BL)method based on data distribution characteristics is proposed.To solve the category pattern classification tasks with uneven distribution and different misclassification costs.DDb Cs-BL searches for the best classification boundary of the cost-sensitive BL learner on the basis of fully considering the statistical distribution characteristics of the sample data,ensuring that the minority sample information is not lost,thereby ensuring the pattern classification of BL on various data sets performance.Experiments are carried out in the process of unbalanced data set and TE process,the results show that the proposed method can effectively improve the accuracy of fault diagnosis.