Practical applications of data-driven fault detection(FD)are limited by their portability.The costs of model training and validation are extremely high when each system requires a model retrained on its own fault and ...Practical applications of data-driven fault detection(FD)are limited by their portability.The costs of model training and validation are extremely high when each system requires a model retrained on its own fault and fault-free data.Therefore,this paper proposes a statistical-based online cross-system FD method to address the problem of model portability.The proposed FD model can be cross-utilized between building chillers with various specifications while it only needs to update the original fault detection model by the normal operation data of the new chiller system,thus saving huge fault experimental costs for the fault detection of new chiller.First,a theoretical basis for the proposed cross-system fault detection method is presented.Then,experiments were conducted on three building chillers with different specifications.Both fault and fault-free data were collected from the three chillers.The development and validation of the proposed cross-system fault detection method are then conducted.Results show that the cross-system fault detection models perform well when used with different chillers.For instance,when the fault detection model of system#1 was cross-utilized to system#2,the detection accuracies of refrigerant leakage,refrigerant overcharge,and reduced evaporator water flow were 99.73%,90.17%,and 96.94%,respectively.Compared with original models,the detection accuracies were improved by 33.78%,84.07%,and 65.56%,respectively.Therefore,the proposed cross-system fault detection method has potential for online application to practical engineering FD.展开更多
基金This work was supported by the Natural Science Foundation of Chongqing(No.cstc2019jcyj-msxmX0537)the China Postdoctoral Science Foundation(No.2021M693714)+3 种基金the Chongqing Postdoctoral Science Foundation(No.cstc2020jcyj-bshX0073)the National Natural Science Foundation of China(No.51906181)the Excellent Young and Middle-aged Talent in Universities of Hubei(No.Q20181110)the Graduate Research and Innovation Foundation of Chongqing(No.CYS20013).
文摘Practical applications of data-driven fault detection(FD)are limited by their portability.The costs of model training and validation are extremely high when each system requires a model retrained on its own fault and fault-free data.Therefore,this paper proposes a statistical-based online cross-system FD method to address the problem of model portability.The proposed FD model can be cross-utilized between building chillers with various specifications while it only needs to update the original fault detection model by the normal operation data of the new chiller system,thus saving huge fault experimental costs for the fault detection of new chiller.First,a theoretical basis for the proposed cross-system fault detection method is presented.Then,experiments were conducted on three building chillers with different specifications.Both fault and fault-free data were collected from the three chillers.The development and validation of the proposed cross-system fault detection method are then conducted.Results show that the cross-system fault detection models perform well when used with different chillers.For instance,when the fault detection model of system#1 was cross-utilized to system#2,the detection accuracies of refrigerant leakage,refrigerant overcharge,and reduced evaporator water flow were 99.73%,90.17%,and 96.94%,respectively.Compared with original models,the detection accuracies were improved by 33.78%,84.07%,and 65.56%,respectively.Therefore,the proposed cross-system fault detection method has potential for online application to practical engineering FD.
