Keywords: active control, intelligent structures, stochastic search, case-based reasoning, tensegrity structures.

This paper reports on the construction of a unique active tensegrity structure containing five modules and ten actuators on telescopic bars in a closely coupled configuration. The control scenario under study is a constant slope requirement for quasi-static control. Controlling for serviceability requirements is more likely to be of practical use than controlling for structural safety events having long return periods since for the latter application, reliability requirements on controllers would make costs prohibitive.

The first part of the paper contains a proposal for a framework for computational control. The principal tasks of the system are to: (i) compare the current structural performance to the desired performance, (ii) reason about the goals not satisfied (diagnosis) and find a set of control actions that improve the performance (search), (iii) execute the set of actions proposed in step 2, (iv) measure the behaviour of the new structural state, and (v) use the current event as an instance for learning and planning. This process iterates continuously over the life of the structure. Since the aim of many intelligent structures will be to adapt over periods of minutes and hours rather than seconds (for example, serviceability control), computational control is a feasible alternative to predetermined control commands. The desired output of the system is intelligent performance over the life of a structure. Many combinations of reasoning, search and learning methods can be used within this framework.

Controlling coupled features, such as adaptive geometry of complex structures, requires reasoning and advanced search techniques since local minima are usually present in the search space. Techniques can be combined to simulate the effects of simultaneously executed control actions and find a set of control actions that best improves system performance.

Tensegrity (tensile-integrity) structures are lightweight, reusable structures. Behaviour is geometrically nonlinear [1] and this is modelled using a combination of dynamic relaxation analysis with a neural-network based correction [2]. There is no direct solution for bar movements given required slope. Furthermore, the entire search space is so large that a brute force method of generate, analyse and test would take up to years to find a control solution.

Three stochastic search algorithms have been compared for this task [3]; simulated annealing, genetic algorithms and PGSL - a probabilistic algorithm developed at EPFL [4]. PGSL was selected for this task because of its performance and because of the relatively few number of search parameters that required calibration for specific tasks. Using this algorithm, good control commands were identified after one to two hours of calculation.

Using feedback from measurements on the structure creates possibilities for improving performance during service. For each control solution carried out, a record of the initial structural state, corrective actions, and a success metric are stored for future retrieval. A "distance metric" is used to select the best case. This case is then adapted using a stochastic search algorithm. Further tests on the three search algorithms revealed that PGSL was much faster than the others for case adaptation. In this approach, cases are employed to come close to the solution and stochastic search is used to identify a good adaptation. This application of case based reasoning reduced calculation time to the order of tens of seconds.

Addition of cases when good control solutions are identified allows the case base to grow during service. In this way, the structure learns with its experience. Work underway involves removing an element from the structure. Through small control movements, we are studying whether the control system can correctly identify the position of the damaged element. Once this is done, the next step is to provide control commands so that the structure can become stronger than in its damaged state. Other work underway is related to control of structural vibrations through geometrical changes. Such control is expected to be particularly useful for non terrestrial applications.

The combination of the three attributes of learning, self diagnostics and self repair are bringing us closer to demonstrating the feasibility of intelligent structures. This will lead to new possibilities for structural engineers in areas of innovative structures, applications in outer space, and new ways to achieve damage tolerant design.

1

Fest, E., Shea, K., Domer, B. and Smith, I.F.C. "Adjustable tensegrity structures", J of Structural Engineering, Vol 129, No 4, 2003, pp 515-526. doi:10.1061/(ASCE)0733-9445(2003)129:4(515)

2

Domer, B., Fest, E., Lalit, V. and Smith, I.F.C. "Combining the dynamic relaxation method with artificial neural networks to enhance the simulation of tensegrity structures", J of Structural Engineering, Vol 129, No 5, 2003. doi:10.1061/(ASCE)0733-9445(2003)129:5(672)

3

Domer, B., Raphael, B., Shea, K. and Smith, I.F.C. "A study of two stochastic search methods for structural control", J of Computing in Civil Engineering, Vol 17, No 4, 2003. doi:10.1061/(ASCE)0887-3801(2003)17:3(132)

4

Raphael, B. and Smith I.F.C. "A direct stochastic algorithm for global search", Applied Mathematics and Computation, Elsevier, to appear, 2003. doi:10.1016/S0096-3003(02)00629-X

purchase the full-text of this paper (price £20)

go to the next paper
return to the table of contents
return to the book description
purchase this book (price £82 +P&P)