Copyright@ IFAC Fault Detection. Supervision and Safety for Technical Processes. Budapest, Hungary. 2000 ARCHITECTURE AND SELECTED ALGORITHMS OF A PROCESS CONDITION MONITORING MICROSYSTEM U. Marschner l , I. Jossa l , G. Elender 2 , T. Herrmann 2 , W.-J. Fischer l I Dresden University of Technology, Semiconductor and Microsystems Technology Laboratory, 01062 Dresden, Germany, Tel.: +49(0)351 463-5399, FAX: -7021 , email: marschner@ihm.et.tu-dresden.de 2 Loher Aktiengesellschaft, Research and Development Department, Postfach 1164.94095 Ruhstorfa.d. Rott, Tel.: +49 (0)853139-0, Fax: 39-212 Abstract: A new microsystem is presented which is able to process measurements including feature extraction and classification for local process monitoring. The heart of the micro- system is a 16-bit digital signal processor which can operate at up to 100 MIPS (million in- structions per second). The extracted features are stored in a large non-volatile memory and are used for a long term trend analysis. Current and predicted faults are displayed locally and announced to the staff or a host computer via field bus or internet. The microsystem has been applied to an autonomous bearing fault diagnosis. Copyright © 2000 IFAC Keywords: Local fault diagnosis, monitoring, hybrid microsystem, self-organizing neural network INTRODUCTION Present diagnosis systems are spatially splitted: be- side the sensors they require a separate computer and so transmission of the sensor signals to it. These systems are expensive such that a diagnosis is often done manually and periodically by especially trained personnel. The new approach is to miniaturize the system that extreme that it can be mounted directly nearby a technical process. This construction leads also to an improved electronic system with respect to the mechanical and electromagnetic reliability. The fulltime availability provides an additional advan- tage: the system state can be tracked permanently and developing faults can be detected. The microsystem presented includes a digital signal processor (DSP) and is therefore freely programma- ble. Necessary algorithms are implemented as remote controllable software programs. Thus the microsys- tem has a wide application range and marks a new quality in local high-performance signal processing. Previous DSP-solutions required a large printed cir- cuit board, not only because of the DSP housing geometry but also because of the power consumption and the necessary DSP cooling. The construction of the microsystem benefits from new semiconductor technologies which reduce the power consumption below a threshold crucial for extreme miniaturization without loss of computational speed. 411 The high computational power is needed in order to compensate for architectural restrictions of low-cost general-purpose DSPs, e.g. to increase the numerical precision. To implement floating-point arithmetic on the fixed-point DSP used, operands must be con- verted to fixed-point numbers and then back to floating-point numbers. Fixed-point values are con- verted to floating-point values by normalizing the input data. Nevertheless, this software overhead and also additional hardware needed by the DSP encour- aged in the past the development of application spe- cific integrated circuits (ASIC's), e.g. control system processors (lones et al.. 1998) or dynamic spectrum analyzers (Marschner et aI., 1998). Signal-based and model-based algorithms have been implemented towards a process condition monitoring microsystem based on high-level behavioural (Fron- tier, 1998) and low-level technology oriented de- scriptions. The paper discusses these algorithms, the hardware realization and the application of the mi- crosystem for bearing diagnosis. 2 DIAGNOSIS MODEL Fig. I depicts the signal flow within the microsystem. Following a general approach (Frank, 1996), (Hoff- mann, 1999) the signals are at first processed by an