Damage identification using inductive learning
A damage identification method incorporating the use of inductive learning is presented. Inductive learning is the process of learning from examples. The method utilizes as much dynamic-response data as is available, ordering this information to find the best data with which to discriminate among a set of damage states available for dynamic testing. This method takes into account the inherent variabilities in the damage identification problem. These inherent variabilities include but are not restricted to sensor noise, changes in environmental conditions, slight changes in boundary conditions, and manufacturing differences. The method statistically isolates changes in the dynamic-response characteristics due to damage from these inherent variances. This method is model-independent and can be used to accommodate any sensors, actuators, and data type.
In order to demonstrate the method, an experiment was performed on a 12” x 12” x ⅛” aluminum plate hung horizontally from the corners to simulate free-free boundary conditions. The plate was sensed and actuated by two piezoelectric patches mounted diagonally symmetric from one another. A small test mass (2% of the mass of the plate) was placed at four discrete locations, changing the physical properties of the structure. The structural impedance-responses were measured for all of the damage cases for both sensors. This information was processed by the damage identification algorithm to generate rules to which a small amount of data, extracted from a single set of structural impedance-response information, can be applied. The method was able to successfully discriminate all of the damage states from one another as well as to detect the existence of a change in physical properties due to a damaged state of which there was no prior knowledge.