In contemporary times, new developments and technological methods are being used as
part of a process called Strutural Health Monitoring (SHM). SHM is the development of
strategies for detection, prevention and characterization of undesirable damages in civil
and mechanical structures of static behavior (i.e., bridges, railways) and dynamics (i.e.,
satellites, vehicles, industrial equipment). A large number of sensors collect information over
a period of time, which can generate a high amount of data that needs to be transmitted
and stored. However, failure or other malfunctions can cause data loss, which directly
impacts analysis and decision making. To work around this problem, a new technique
appears: A Data Imputation. An imputation process basically replaces lost data with
substituted values and “fills” the missing application data with plausible values. This
imputation is a practice of filling in missing data and avoids the complexity generated
by the missing data. For this, this work will proceed to a comparative study of several
imputation techniques referring to imputation by means, fashion, regression, knn and
recurrent neural networks. Based on this, this work proposes an evaluation method that
compares the error rate generated in the detection of damages. The methods were tested
using data sets from a monitoring system installed on the Z-24 bridge (Switzerland), which
was subjected to conditions of varying variability as well as progressive damage trials. The
occurrence of missing data was done artificially. The results show that recurrent neural
networks imputation provides the best results.
|
