|
Abstract:
|
By framing the structural health monitoring and control problem as being one
of enhancing structural system intelligence, novel solutions can be achieved through
applications of computational strategies that mimic human learning and attempt
to replicate human response to sensory feedback. This thesis proposes several new
methods which promote adaptive, intelligent decision making by structural systems
relying on sensory feedback and actuator compensation. Four significant contributions
can be found in this thesis study. The first method employs an adaptable subclass of
Artificial Neural Networks (ANNs), called Radial Basis Function Networks (RBFNs)
for robust control in the presence of sensory failure. The second method exploits
this computationally efficient network to detect and isolate system faults in real time.
The third algorithm utilizes an RBFN to effectively linearize the nonlinear actuator
dynamics of a Magnetorheological (MR) damper, thereby improving control of the
semiactive device. Lastly, an open loop observer is implemented experimentally to
both detect damage and act as a trigger for control of the newly developed Adaptive
Length Pendulum-Smart Tuned Mass Damper (ALP-STMD).
Some limitation of existing algorithms in the field of real time structural health
monitoring and control are that they rely heavily on fixed parameter methods, assume
standard linear time invariant assumptions, or mandate accurate modeling of system
dynamics. By embedding the proposed reasoning and decision making algorithms into
the feedback methodology and design, greater generalization and system adaptivity
is possible. Specifically, the proposed methods develop novel solutions for adaptive
neural control, fault (sensor failure) tolerant control, real time damage detection,
adaptive dynamic inversion, and control applications for STMDs.
The neural network adaptive control formulation is successful in rejecting first
mode disturbances despite online sensor failure. It is also capable of improving the
performance of a baseline Hoc controller in the presence of sensor failure and earthquake
ground motion. The proposed fault tolerant controller is validated on a two
degree of freedom shear frame subjected to six earthquake records. Furthermore, this
application involves the use of piezoelectric patches as sensors and actuators.
The RBFN algorithm in combination with an open loop observer is capable of both
detecting and isolating stiffness degradation and recovery in multi-degree of freedom
systems in real time. The method is validated on experimental data taken from online
damage tests using the Semi-Active Independent Variable Stiffness (SAIVS) device.
Other validations involve simulations on a two degree of freedom system and a ten
degree of freedom system with both independent and coupled damage case scenarios.
In all scenarios, the RBFN is capable of identifying the length of time and degree of
freedom in which stiffness variation occurred.
A neural network formulation is developed to perform dynamic inversion for semiactive
control of an MR damper. The MR damper acts as a base isolator in a scaled
two story building. Both the building and damper models were based on tests performed
at Rice University. The control performance of the adaptive RBFN dynamic
inversion method is compared to both passive-off and passive-on methods of semiactive
control for MR dampers.
The last contribution serves to combine both real time structural health monitoring
and control in a proof of concept experimental study. An open loop observer is
used to trigger an ALP -STMD device in the presence of base excitation and stiffness
damage. The stiffness damage is generated from strategically regulating the current
applied to Shape Memory Alloy (SMA) braces in a two degree of freedom shear frame.
Once damage exceeds a predefined threshold, the ALP-STMD uses a another SMA
to adjust its pendulum length to tune in real time to the dominant pulse present in
the base excitation. |
