【正文】 me unit (DTU) estimates the process time delay. The membership function unit (MFU) suggests w hether a new local model should be inserted. It estimates an additional local model in the middle of the interval betw een the tw o neighboring local models that are the most excited. The model is submitted to the MIA if the global validation of the resulting fuzzy model is sufficiently improved, pared to the original fuzzy model. . Model Information Agent (MIA) The task of the MIA is to maintain the active MFM and its status information. Its primary activity is p rocessing the online learning results. Whe n a new local model is received from the O LA, it is accepted i f it passes the stability test and its confidence index is sufficient. If it is accepted, a “ready for tuning” flag is set for the CAA. Another flag ind icates w hether the local model has been tuned since start up or not. If the model confidence index is very low , the automatic mode may be disabled. The MIA contains a mechanism for inserting additional local models into the MFM. This may occur either by re quest or automatically, using the MFU of the OLA. The MIA may also store the active MFM to a local database or recall a previously stored one, w hich is useful for changing of modes. At initial configuration, the MIA is filled w ith default local models base d on the initial estimation of the process dynamics. They are not exact but may provide reliable (although sluggish) control performance, similar to the Safe mode. Using online learning through experiments or normal operating records (w hen the conditions a re appropriate for closed loop identification),an accurate model of the plant is estimated gradually by receiving identified local models from the OLA. . Control Algorithm Agent (CAA) The CAA prises an industrial nonlinear control algorithm and a procedure for automatic tuning of its parameters from the MFM. Several different CAAs may be used in the controller and may be interchanged in the initial configuration phase. The follow ing modes of operation are supported: _ Manual mode: openloop operation (actuator constraints are enforced). _ Safe mode: a fixed PI controller w ith conservatively tuned parameters. _ Auto mode (or several auto modes w ith different tuning parameters): a nonlinear controller. The CAAs share a mon interface of interaction w ith the OS and a mon modular internal structure, consisting of three layers: 1. The control layer offers the functionality of a local linear controller (or several local linear controllers simultaneously), including everything required for industrial control, such as handling of constraints w ith anti w indup protection, bump less mode sw itching, etc. 2. The scheduling layer performs real time sw itching or Scheduling (blending) of tuned local linear controllers, so that in conjunction w ith the control layer a fixedparameter nonlinear controller is realized. 3. The tuning layer executes the automatic tuning procedure of the controller parameters fro m the MF M w he n the MIA reports tha t a new local model is generated if auto tuning is enabled. The parameters of the control layer and the scheduling layer are replaced in such a manner that real time control is not disturbed. Three CAAs have been developed and each has been proved effective in specific applications: the Fuzzy parameter scheduling controller (FPSC), the deadtime pensation controller (DTCC), and the rule based neural controller (RBNC). In this paper, only the concept of the FPSC is described briefly in the follow ing subsection. . Fuzzy parameterscheduling controller An overview of the FPSC is show n in Fig. 4. ARTICLE IN PRESS The control layer of the FPSC includes a single PID controller in the form suitable for controller blending using velocity based linearization. It is equipped w ith anti w indup protection and bump less transfer. The scheduling layer of the FPSC performs fuzzy blending of the controller parame ters (in the case of Ti, its inverse value) according to the sched uling variable s(k) and the membership functions bj(k) of the local models. The instrument of velocitybased linearization enables the dynamics of the blended global controller to be a linear bination of the local controller dynamics in the entire operating region, not just around equilibrium operating points. This provides the potential to improve performance w ith few local models and more transparent behaviors in off equilibrium operating points (Leith amp。 , 2021).The tuning layer of the FPSC is based on the magnitude optimum (MO) criterion implemented using the multiple integration (MI) method (Vrancˇ ic180。 if so, it terminates processing. Otherw ise, it filters the signals y and v and performs a low level analysis. The SC scans the pre processed buffer for the last recognizable event that may be evaluated, or is otherw ise important, ., a step change of the reference signal, a step change of the measured disturbance signal, an o ccurrence of an unmeasured disturbance, or the presence of oscillation. If a n eve nt that ma y be eval uated is detec ted a nd the conditions for feature estimation are fulfilled (there is no excessive oscillation, the signal to noise ratio is sufficient, the p rocess response has settled after the event and there w as a period of steady state before the event), the corresponding buffer interval is sent to the PE, otherw ise the execution is terminated. The PE may extract the follow ing features of detected events: overshoot, settling time, rise time, oscillation decay rate, and tracking error measure or regulation error measure. Using a fuzzy evaluation procedure, an overall performance index (PI) is also calculated fro m the fea tures. T he CP M res ults are se nt to the huma n–mac hi ne i nte