50 Y. Reuland et al. based on vibration characteristics [7]. However, most of the research in this domain has focused on using a high-density sensor layout, which is unrealistic for monitoring systems of most smart-buildings and which might induce unnecessary costs. Current human detection and localization methodologies are model-free, meaning that they rely only on processing and analyzing measurement data. Coupling the measured response with a structural behavior model of the building has the potential to reduce the number of sensors necessary for human localization. However, predictions of the structural response for vibrations induced by human footsteps are prone to multiple sources of uncertainty including material properties, geometry and boundary conditions. Therefore, a model-based data-interpretation methodology that is robust in presence of uncertainty is needed to perform human localization using vibration measurements from a sparse sensor configuration. In this paper, a procedure for human detection and localization is presented using floor vibration data. First a methodology to detect the presence of an occupant is described. Once the presence of an occupant is established, localization is carried out using error-domain model falsification (EDMF) [8], which is a model-based data interpretation methodology that has already been employed successfully for improving knowledge of the behavior of bridges [9, 10], water-supply-networks [11] and wind around buildings [12]. 6.2 Methodology 6.2.1 Human Detection Detection of human presence on a floor is achieved through on a two-step procedure [4]. In the first step, structural characteristics of the floor slab are derived from prior ambient vibration measurements. Two types of information are obtained from these baseline measurements, the dominant frequencies of the slab at the measured locations as well as the level of vibrations that characterize ambient conditions (due to factors such as outside traffic). With the ambient signal taken as a Gaussian white noise process, the signal statistics that are obtained from the initial measurements, mainly the standard deviation of the signal, are calculated. With knowledge of the ambient vibration level, thresholds are set and then used to detect “anomalies” such as human steps. As a human step can be conceptualized to be an impact load, measurable levels of energy are carried by the shockwaves that are close to the dominant frequencies of the structure. Based on this observation, and in order to improve the signal-tonoise ratio for human detection, the recorded signals are bandpass-filtered around the dominant frequencies. A classic sixthorder Butterworth filter is used to filter the signals. Various recordings of several minutes taken on several days are then used to compute detection thresholds based on the standard deviation of the bandpass-filtered ambient-vibration accelerations. Human detection is achieved when two consecutively measured samples exceed the detection threshold. 6.2.2 Error-Domain Model Falsification Popper [13] asserted that data cannot be used to validate correct models, rather to falsify wrong models. Motivated through this assertion, error-domain model falsification was proposed by Goulet et al. [8] as a multiple model methodology, where wrong models are falsified using thresholds, which are defined by the uncertainties associated with the system. The EDMF methodology has been applied to fourteen full-scale systems [14]. In this paper, the application of this methodology is extended to human localization using acceleration data. Assume g( ) is a model with n identification parameters i. The response of such a physics-based model is subjected to uncertainty from many sources such as model fidelity, geometric simplifications, material property and boundary conditions. Many of these uncertainty sources cannot be approximated using zero-mean independent Gaussian distributions. Moreover, some of these uncertainties are systematic in nature with unknown correlation between measurement locations. Let the combination of all these uncertainties at a measurement location i be "mod,i. If msuch measurement locations are present, thenyD[y1, y2 : : : ym] T is the vector of measured structural responses at mlocations. The uncertainty associated with each measurement is "meas,i. If Qi is the true (unknown) structural response at a measurement location then it can be represented as the difference between model prediction and model error or the difference between measurement and measurement error as shown in Eq. (6.1) Qi Dgi "mod;i Dyi "meas;i (6.1)
RkJQdWJsaXNoZXIy MTMzNzEzMQ==