An extended Kalman filter (EKF) based displacement estimation method for nonlinear SDOF systems under seismic excitation is proposed. In this method, time intervals where the system experiences significant nonlinearity or not are firstly distinguished. For a time period when the system is in an elastic phase, available observations for EKF are acceleration, displacement obtained via double integration of acceleration, and residual displacement. During a time period with significant nonlinearity, acceleration and virtual displacement measurement are employed as observations, and the displacement is estimated along with time-variant stiffness using an augmented state vector in EKF. The results are further smoothed by extended Kalman smoother (EKS). The proposed method is studied on an SDOF system with a bi-linear hysteresis model in detail and further verified considering various hysteresis models and earthquake excitations. The estimated displacements are shown to be accurate.
This study develops a Convolutional Neural Network model to predict the likelihood of accident occurrence in an inter-city expressway. The model utilizes the temporal and spatial information of traffic states of past one hour as an input for predicting the accident occurrence in two hours ahead from the prediction start time. In order to efficiently learn the traffic features prior to the accident occurrence, the input data is arranged in three-dimensional tensor form, analogous to image data. The results based on the ROC curve showed that the proposed model was able to identify the accident occurrence with good accuracy. Further, in addition to the capability of binary classification, the result demonstrated that the likelihood calculated in the output layer could be interpreted as the probability of accident occurrence.
In this research, 500 high resolution rubber bearing images with damages on them are collected and manually labeled to build the data set. Then the data set is adopted to train a Fully Convolutional Network (FCN) model, aiming to predict damages on the rubber bearings from a large amount of high-resolution images. The method is called Cropping Segmentation, which uses cropped image with size 224×224 as input images to train the FCN model, instead of traditional Squashing Segmentation. However, even though Cropping Segmentation has high accuracy, there are a lot of noises in the background. To solve the problem, Context Detection is carried out to exclude noises in the background. By intersecting the outcomes of Cropping Segmentation and Context Detection, the predicted damages on the rubber bearing are retained, while the noises outside the bearing are removed. Context Detection includes CNN-based Context Classification and FCN-based Context Segmentation. By testing and comparing, Context Segmentation has a better performance on finding rubber bearings pixels.