Effectiveness of different machine learning models for state recognition and fault diagnosis in fluidized beds

FAN Jian; DOU Mingying; ZHAO Fei; MA Cunzhi; ZHANG Ruxin; JIN Yunfeng; LI Shuyue; ZHANG Yongmin

doi:10.13732/j.issn.1008-5548.2025.06.009

Abstract：Objective The complex gas-solid flow in fluidized beds often leads to operational failures such as particle agglomeration,distri-bution plate clogging,gas short-circuiting,and hydrodynamic instability,jeopardizing production safety and economic benefits.Compared to the machine learning(ML)methods,manual monitoring and signal processing in traditional approaches have limi-tations in signal extraction and multi-type fault differentiation.This study evaluates the feasibility of ML methods for state identi-fication and fault diagnosis in gas-solid fluidized beds,benchmarking and optimizing typical ML models using the built-in AI model of DTEmpower.The research results provide insights into the development of artificial intelligence-based stabilization techniques for industrial fluidized bed systems. Methods A small cylindrical bubbling bed cold mold was constructed with transparent glass walls to visualize flow dynamics.The flow characteristics of fluid catalytic cracking(FCC)particles under normal working conditions,particle agglomeration(10％～50％by volume),and distribution plate blockage(12.5％～62.5％by area)were simulated.Transient pressure pulsa-tion signals were collected,and bed eigenvalues were calculated based on pressure sensor data.For fault diagnosis,four ML models,k-nearest neighbors(KNN),random forest(RF),support vector regression(SVR),and radius nearest neighbor(RNN),as well as the AI-agent of DTEmpower,were selected based on their characteristics and model features.Five types of models were constructed and their model performance was validated and compared. Results and Discussion As particle agglomeration worsened,the bed pressure drop and density decreased steadily,while the distribution plate pressure drop fluctuated,showing an overall downward trend.Distribution plate clogging increased plate pres-sure drop but decreased bed pressure drop and density.The trend variations could be used to determine the operational status of fluidized beds.An experienced operator could identify simple faults from trend data but not complicated faults.To achieve accu-rate identification of fault category and severity under conditions of particle agglomeration and distribution plate blockage,ML methods should be used.For known fault severities(limited domain),the AI-agent model and KNN achieved the highest accu-racy(88.89％),followed by RF(88.67％)and SVR(19.44％).For unknown fault severities(unqualified domain),RF had the highest accuracy(100％),followed by AI-agent(95.84％),SVR(87.5％),and RNN(87.5％).Overall,the KNN model demonstrated superior accuracy(93.75％)and differentiation rate(81.25％),while the AI-agent also excelled in fault diagno-sis under comparison. Conclusion This study offers insights into the development of AI-based technologies in operational stability solutions for indus-trial fluidized bed systems.Through a comprehensive comparison of five mainstream ML models,the optimal model for condition monitoring and fault diagnosis in fluidized beds was identified.The proposed method can distinguish the differences between par-ticle agglomeration and distribution plate blockage while quantitatively assessing failure severity,significantly enhancing opera-tional safety.It provides a foundation for industrial-scale state identification and fault diagnosis systems based on actual indus-trial data.However,it should be noted that this study was conducted at a laboratory scale with simplified operating conditions and limited parameter variations.Future work should focus on applying these methods to industrial-scale applications.Histori-cal DCS data should be employed,and specific reaction conditions and mechanisms should be considered,which will require specialized model adjustments.