Detection and quantification of temperature sensor drift using probabilistic neural networks

Temperature effects are a major driver of strain and deformations in weather-exposed civil infrastructure, such as bridges and buildings. For such structures, long-term temperature data holds the potential for data-driven prediction of expected structural behavior, which in turn enables the detection of anomalous structural behavior. For this reason, structural health monitoring (SHM) strategies typically employ temperature sensors. However, the success of SHM is contingent on the quantity and quality of the available temperature data. Hence, accurate automatic methods for the detection and quantification of anomalies in temperature data are needed. In particular, gradual temperature sensor drifts are difficult to detect and can introduce errors into the thermal compensation of strain sensors, which can be erroneously confounded with time-dependent structural behavior. Current data-driven methods use air temperature as predictor because it exhibits good correlation with temperatures in the structure. Sensor drift is typically quantified by analysis of the prediction residuals; however, these methods are not robust to outliers and can be affected by seasonal biases. In this work, a probabilistic neural network is used as the nonlinear data-driven temperature prediction model which enables the introduction of a sensible threshold to mitigate seasonal model bias. Furthermore, a novel drift detection method based on the evolution of parameters of a trinomial probability distribution is introduced, together with a robust drift quantification method. The performance of this method is assessed using real temperature data from a pedestrian bridge, spanning over seven years of the structure's life.