This study approaches the problem of quantifying the network sensor errors as a supervised learning problem and leveraging deep neural networks to map observed traffic flow counts to the systematic errors in the sensors. The author aims at building a model that could reconstruct the erroneous flow irrespective of the level of random noise in the sensors, which is unknown in the real-world. By reconstructing the erroneous flow with high accuracy, the transportation planners could gauge the true traffic flow demand in the network and can make informed infrastructure related decisions.
The study begins by simulating the traffic network under dynamic flow assignment settings to generate the base flow that is treated as the ground truth. The authors then introduce measurement errors to the base flow to generate the observed flow which is transformed into a multi-dimensional time-series tensor data, where each time step has dimension equal to the number of sensors in the network. Next, they introduce deep neural network comprising of 1-Dimensional Convolutional Neural Networks (1-D CNNs) to extract high-level spatial-temporal features from the observed flow time series data. To understand the generalization capability of the deep learning model, they deploy it against numerous test cases with varied levels of random errors and proportion of malfunctioning sensors in the network. Results indicate that the flow reconstructed using the deep learning model is very close to the ground truth flow and that the model predicts the systematic errors in the test cases with high accuracy.
The major advantages of this study are that, firstly, the model is robust to the flow imbalance in the network unlike most of the network sensor health studies in the past. Secondly, the approach escapes dealing with complicated flow-density relations one might encounter while modeling dynamic flow using traditional analytical statistical approaches.