Abstract: In this paper, fault detection methods for hydraulic systems based on a parity equation approach with neural net models are presented。 Hydraulic systems are used in manifold applications in industry。 They are however not yet the subject of intense research in the area of fault detection and diagnosis, which can be mainly attributed to their strong nonlinear behavior, which exacerbates the physical modeling extensively。 To avoid the difficulties associated with the physical modeling, a data-driven modeling approach based on the LOLIMOT neural network will be presented in this paper。 Different subsystems of the hydraulic servo axis will be modeled using different sensor configurations。 Experimental data from a real testbed allow to compare the model fidelity of the different resulting neural nets and can also be used to illustrate the capabilities of the parity-equation based fault detection approach, which in general allows the detection of tiny faults, such as sensor offset faults in the area of a few percent of the maximum sensor readout。79609

Keywords: Fault detection, fault diagnosis, hydraulic servo systems, modeling, neural nets。

1。INTRODUCTION

The ongoing trend to integrate mechanic and electronic components together with information processing by micro-controllers allows to offer entirely new functions, such as e。g。 improved control, automatic commissioning or integrated fault management, see e。g。 Isermann [2005a]。 The key driver for this development is an effort to ex- ploit cost reduction potentials。 Harnessing modern fault management methods can unlock enormous savings as the maintenance work can depart from the (traditional) maintenance with its regular maintenance schedules and employ condition-based maintenance cycles, where only those components are serviced that are likely to fail  in the near future, Isermann [2005b]。

While fault detection and fault management have been investigated quite thoroughly for electro-mechanic con- verters, such as e。g。 induction motors, see Wolfram [2002], much less fault detection and diagnosis approaches have been reported for hydraulic systems。 An overview of fin- ished and ongoing research work in the area of fault de- tection and fault diagnosis for hydraulic systems is given in the survey paper by Bredau et al。 [2008]。 There are two different trends: One trend is to use special sensors that allow to detect changes in the material properties of the hydraulic fluids。 Special condition monitoring sensors shall not be discussed in this paper。 Instead, the reader is referred e。g。 to the paper by Dieter and Bauer [2007]。

The use of neural nets for fault detection and  diagnosis in hydraulic systems as reported in the literature has   so

far been limited to few applications。 A fault detection ap- proach governing a complete hydraulic system consisting of pressure supply, three valves and three cylinders has been developed by Ramdén [1998]。 Here, a neural net is utilized to detect valves which are stuck close or stuck open, see Ramdén et al。 [1995]。 A nonlinear optimization scheme was used to detect valves, which were stuck par- tially open, i。e。 at an opening of 25% and above。 These approaches have been tested with a simulation model only and are not very sensitive to small, incipient faults。

In this contribution, new and advanced fault detection methods for the supervision of hydraulic components, such as proportional valves and cylinders based on neural net models are developed and their performance is evaluated at a testbed。 Sensor configurations close to the series instrumentation are investigated with respect to their fault detection performance。 The use of signals provided by non- series sensors has been handled in a very restrictive way。 Since there is always a trade-off between the level of fault detection and economical constraints, one tries to employ as few additional sensors as possible, see e。g。 the experts discussion organized by Backé and Bork [2004]。