|
By the term ICAM Agent we mean an encapsulated MATLAB function, which contains the best MATLAB routines to perform a given task, plus ancilliary logic and routines to manage and diagnose the task, e.g., set up the task optimally or determine if the task has been done successfully or not. The output of an ICAM Agent is not only the best-possible task result, but diagnostic information about that result. In some cases we use a very high-level MATLAB function for the task at hand -- an example is the prediction error minimization or PEM algorithm for linearized model identification or LMId; in other cases we build our own world-class routine -- an example is the FDIA routine, which has been published in the best conference proceedings and well received for its unique capability to handle faults in all sensors and actuators in a multi-loop control system (demonstrated for the pilot plant model, which has five loops; ten faults can be handled). The encapsulation and logic for the FDIA Agent is shown in some detail on its web page. Therefore, we focus on the LMId Agent for this discussion. LMId Agent set-up: The simplest and most common way to provide "sufficiently exciting inputs" to a process to be modelled (inputs must "excite the dynamics", according to model identification theory) is to use independent pseudo-random binary sequences (PRBSs) as inputs to each of the plant inputs. We found that this did not reliably produce good linearized models. We then investigated generalized binary noise (GBN) sequences -- since their frequency content can be matched to the bandwidths of the process loops we obtain optimal excitation; PBRSs are deficient in their low frequency content, which explains the superiority of GBNs. Thus, the agent performs simple step-response tests to estimate the loop bandwidths and uses GBN sequences accordingly. LMId Agent diagnosis: We illustrated our approach for determining the goodness of fit for the resulting linearized model on the LMId web page. The percentage of fit for the first three plant outputs is good to excellent (90.6 to 99.2%), the fifth output is an 84.4% fit but the fourth output is only 77% fit -- these results are passed to the FDIA Agent, so it is "aware" that FDIA for loop 4 may be less certain than for loops 1, 2, 3. The excitation provided by GBN inputs and the goodness of fit can be seen in the Linearized Model Identification (LMId) Agent web page. We plan to implement ICAM agents as function blocks (under standard IEC 61499) for use in industrial process control applications. This will facilitate the creation of commercial products to be licensed to a vendor for sale to end users. Return to the UNB PAWS Home Page. Information supplied by: Jim Taylor Last update: 10 March 2010 Email requests for further information to: Jim Taylor (jtaylor@unb.ca) |