| Summary: | Safety and functionality of a fluid power control system can considerably be increased by implementing
predictive maintenance routines. Modern predictive maintenance practices are based on automatic
condition monitoring and fault diagnosis of the system components. In most cases, low-quality raw sensor
data are directly monitored for constraint violations or threshold crossings. Subsequent fault diagnosis
is often performed by a knowledge-based expert system based on "order-of-magnitude reasoning".
This means that quantitative sensor data are first transformed into more understandable "linguistic
terminologies" such as "low", high", etc., and are then assessed by production rules in order to diagnose
system (or component) faults. A major problem with this technique is that it is not usually feasible
to directly measure the desired quantity, e.g., the flow rate inside a valve. Another problem is the
association of noise and variations with directly measured signals, which might be misinterpreted as
faults, especially in highly dynamic systems.
In practice, failure modes often involve a change in the model structure, which may be interpreted as
change(s) in one or several system parameters. The theme of this thesis is on automatic generation of
fault symptoms in the form of qualitative variation of system physical parameters by on-line processing
of low-quality raw sensor data. To accomplish this, a novel model-based methodology has been proposed
that has integrated four levels of information processing in a structured hierarchy:
1. State/parameter estimation of the hydraulic system components using state-space models, stochastic
signal processing techniques such as Kalman filtering, and raw sensor data from the hydraulic
system.
2. Monitoring and change detection in the identified parameters of the system components, using
statistical tests, such as sequential probability ratio test.
3. Generation of fault symptoms in the form of qualitative changes in the physical parameter values,
such as "increased", "decreased", etc.
4. Fault recognition by fault symptom classification using neural network pattern classifiers,
5. Fault diagnosis maintenance aiding using knowledge-based expert systems.
By using a second-order linear system as an example, we have shown how each element of the proposed
hierarchical methodology effectively processes the lower quality data received from the previous element
and provides higher quality information for the next element in the hierarchy, so that an incipient fault or
an abrupt failure can be successfully detected and diagnosed. The proposed fault detection and diagnosis
(FDD) technique has also been applied on a real hydraulic test rig which has been built in the Robotics
and Control Laboratory, at UBC.
The hydraulic test rig has a two-stage proportional directional flow control valve, which has been
thoroughly modelled for simulation of faults. A step-by-step methodology has been adopted to obtain
the physical valve parameters from static measurements, as well as through numerical search techniques
using dynamic measurements.
In order to estimate the system parameters and states in real-time, nonlinear state-space models have been
developed for various hydraulic components, including the two-stage servovalve, a hydraulic cylinder,
and a manipulator. Extended Kalman Filtering (EKF) is applied on the state-space models to get the
parameter estimates. Only low-cost robust sensors such as pressure transducers and position sensors
have been used for this purpose. More expensive or hard-to-measure states such as flow rates and orifice
areas are predicted using novel state-space models.
One of the major achievements of this thesis has been incorporation of a novel state-space model for a
valve orifice area that allows us not only to obtain accurate estimates of the flow rate through the valve,
but also to detect several incipient faults and abrupt failures in the valve and its connecting ports. The
valve orifice area is considered as a nonlinear unknown function of the valve spool position. No a priori
knowledge about the orifice profile or the spool deadband size is assumed. The functional relationship,
along with the deadband size are automatically revealed during the on-line estimation process, while the
decision as to which port is open to the flow is made internally. Experimental results were promising
and showed that the identified valve orifice area is an excellent measure in quick detection and diagnosis
of incipient or gradual faults, as well as and abrupt failures, in servovalves and servo-actuator systems. === Applied Science, Faculty of === Mechanical Engineering, Department of === Graduate
|
|---|
