Presented at the IEEE International Measurement Technology Conference, Baltimore, Maryland, May 1–4, 2000 Diagnosis of a Continuous Dynamic System from Distributed Measurements Eric-J. Manders 1 and Lee A. Barford 2 1 Department of Electrical Engineering and Computer Science, Vanderbilt University, P.O. Box 1824 Station B, Nashville, Tennessee 37235. Phone: +1 615 322-2771, fax: +1 615 343-6702, email: manders@vuse.vanderbilt.edu 2 Agilent Laboratories, 1501 Page Mill Road, MS 4AD, Palo Alto, California 94303-0889. email: lee barford@labs.agilent.com Abstract A diagnosis application has been built for a three- tank fluid system. The tank system is equiped with a distributed measurement and control system based on smart transducer nodes with embedded computing and networking capabilities that use the IEEE 1451.1 ob- ject model to provide high level sensing and actuation abstractions. The diagnosis system operates on-line on a workstation that appears on the network as another transducer node. The diagnosis methodology has sev- eral aspects that allow distribution of the monitoring and diagnosis functionality on a network of embedded processors. The current application represents the ini- tial phase in building a truly distributed monitoring and diagnosis application. 1 Introduction Smart transducers with computing and networking capabilities have started to become commercially avail- able. The embedded processor on a transducer sup- ports complex sensing and actuation tasks, and other high level applications. Combined with the networked communication, this facilitates the construction of dis- tributed measurement and control systems. The ad- vantages of such an approach include increased scala- bility and potentially also improved robustness. This paper discusses monitoring and diagnosis of complex continuous dynamic systems in the context of exploiting smart transducer technology. In complex industrial processes, operational safety and reduced down-time are typically ensured by hardware redun- dancy and localized hardware safety mechanisms (e.g., check valves). To reduce cost, hardware redundancy is increasingly being replaced by functional redundancy techniques. The use of functional redundancy requires a model of the system under scrutiny and uses function- al relations between system variables to infer discrep- ancies in measured variables. When such discrepancies occur a fault is detected. Fault detection followed by a fault isolation stage to accurately locate the failing physical component, establishes the fault detection and isolation (FDI) paradigm. In addition to the benefits of distributed measure- ment and control systems mentioned earlier, an FDI application may exploit the ability to run application code on an embedded processor. Local signal process- ing and analysis enables data reduction on the sensor, and thus a decrease in the necessary network band- width. In addition, as distributed measurement and control systems grow in size and complexity it will be- come increasingly important to detect and isolate faults both in the system under observation as well as in the measurement and control system itself. This opens the possibility for diagnostic analysis on the transducer. Transcend is a framework for model-based diagno- sis of continuous dynamic systems based on functional redundancy techniques [8]. The fault isolation algo- rithms apply qualitative constraint analysis methods that effectively realize a parameter estimation scheme. Model parameters correspond directly to system com- ponents and when estimated parameter values deviate from their expected values, the associated components are implicated. The qualitative approach avoids dif- ficulties in the convergence, precision, and computa- tional complexity of established numerical parameter estimation methods, especially when system behavior is non-linear. Because qualitative methods process in- reprint