This article proposes a robust decoupling network named RDNet to provide stable traffic predictions even when perturbations exist in historical data. A decoupling block is designed in the RDNet for dividing traffic data into the invariable component (IC) and variable component (VC). The IC of historical or future data is estimated through the invariable block without historical data and thus would not be perturbed. The variable block is developed to forecast the VC of future data using the VC of historical data. Besides, the robust graph neural network and smoothing loss are designed to reduce the effects of perturbations. The RDNet fuses the obtained IC and VC of future data to produce the predictions, and the invariable and decoupling losses are developed for stabilizing the prediction. The results on six open datasets have demonstrated that the RDNet can achieve a 15.62% average improvement in accuracy compared with the state-of-the-art predictor.

The graph-based traffic forecasting is generally realized on the assumption of sufficient data, which could be impractical in the regions without well-deployed mobile sensors or data-processing facilities. Recent studies have developed a solution with the cross-region transfer learning, i.e. transferring traffic knowledge from the source regions to target ones, whose traffic data and computing resources are limited. Nevertheless, relevant issues, including initialization selection and domain adaptation, have not been effectively tackled in the cross-region graph-based traffic forecasting. This paper proposes the cluster-granularity spatiotemporal transfer (CGSTT), which transfers the cluster-granularity knowledge from the source region to target one for the cross-region graph-based traffic forecasting, as not all source knowledge is positive to the target region. Additionally, the domain adaptation is achieved by the dual alignment consisting of the covariate alignment and label alignment of the source/target data, making the proposed CGSTT adapt to the target region efficiently. The superiority of the proposed method over ten compared baseline methods for both short-term and long-term predictions is demonstrated by the conducted experiments on four tasks, which show that it outperforms the state-of-the-art method by achieving an 8.89% average improvement in forecasting accuracy. The PyTorch implementation of the CGSTT is available athttps://github.com/canyangguo/CGSTT.

To achieve accurate traffic forecasting, previous research has employed inner and outer aggregation for information aggregation, and attention mechanisms for heterogeneous spatiotemporal dependency learning, which results in inefficient model learning. While learning efficiency is critical due to the need for updating frequently the model to alleviate the impact of concept drift, limited work has focused on improving it. For efficient learning and accurate forecasting, this study proposes the dynamic spatiotemporal straight-flow network (DSTSFN). Breaking the aggregation paradigms employing both inner and outer aggregation, which may be redundant, the DSTSFN designs a straight-flow network that employs bipartite graphs to learn directly the dependencies between the source and target nodes for outer aggregation only. Instead of the attention mechanisms, the dynamic graphs/networks, which outdo static ones by possessing time-varying dependencies, are designed in the DSTSFN to distinguish the dependency heterogeneity, making the model relatively streamlined. Additionally, two learning strategies based on respectively the curriculum and transfer learning are developed to further improve the learning efficiency of the DSTSFN. Our study could be the first work designing the learning strategies for the multi-step traffic predictor based on dynamic spatiotemporal graphs. The learning efficiency and forecasting accuracy are demonstrated by experiments, which show that the DSTSFN can outperform not only the state-of-the-art (SOTA) predictor for accuracy by achieving a 2.27% improvement in accuracy and requiring only 8.98% of the average training time, but also the SOTA predictor for efficiency by achieving a 9.26% improvement in accuracy and requiring 91.68% of the average training time.

The graph convolution network (GCN), whose flexible convolution kernels perfectly adapt to the complex topology of the road network, has gradually dominated the spatiotemporal dependency learning of traffic flow data. Defining and learning the spatiotemporal characteristics and relationships of the traffic network efficiently and accurately, which are the important prerequisites for the success of the GCN, have become one of the most burning research problems in the field of intelligent transportation systems. This paper proposes a fast spatiotemporal learning (FSTL) framework containing the fast spatiotemporal GCN module, which reduces the computational complexity of the spatiotemporal GCN from${\cal O(k^{2})}$to${\mathcal{ O(k)}}$, where$k$is the number of time steps of data learned in each GCN operation. To mine globally and fast the correlations of road node pairs, a correlation analysis based on the normal distribution with the complexity of${\mathcal{ O(N)}}$, where$N$is the number of nodes in the traffic network, is proposed to construct the global correlation matrix. Besides, the multi-scale temporal learning is integrated into the FSTL to overcome the receptive field constraints of the spatiotemporal GCN. The experimental results on four real-world datasets demonstrate that the FSTL achieves 48.88% and 5.26% reductions in the training time and mean absolute error, respectively, compared with the state-of-the-art model.

Wheelset fault detection with high accuracy is challenging due to poor image quality. Specifically, the wheelset images are collected dynamically outdoors and suffer from diffuse reflection and environmental interference. Thus, the images contain light stripe adhesions (light flairs) and local fractures to be inpainted. The existing inpainting models are inapplicable to restore grayscale wheelset images. They are also too heavy to be deployed in an embedded wheelset monitoring equipment. In this paper, we propose a lightweight high-precision inpainting model that consists of a recurrent similarity network with the ghost convolution (RSG-Net) to remove light flairs and repair local fractures. RSG-Net replaces standard Pconv (partial convolutional) layers with soft-coding ones that can improve the feature representational ability. To reduce the influence of the background region features on image restoration, an asymmetrical similarity measure is designed to calculate not only the angle difference between the target and the source feature vectors but also the activation of the source ones. The multi-scale structural similarity (MS-SSIM) loss term is introduced to precisely guide the structural information restoration, such as the stripe edges. Moreover, the ghost convolution is introduced in RSG-Net to realize the model compression that can retain the core features of wheelset images and remove the redundant features. We conduct three groups of experiments that demonstrate the accuracy superiority of the proposed RSG-Net over the baseline methods, and the number of parameters is reduced by about 50%.

Accurate travel time prediction (TTP) is a significant aspect in the intelligent transportation system (ITS) . Travel times of certain road segments explicitly reflect the traffic conditions of those sections. Effective TTP of road segments is instrumental in route planning, traffic control, and traffic management. However, the accuracy of TTP is greatly affected by the intricate topological structure of traffic network and the dynamics of traffic flow over time. This paper develops a TTP method based on the spatial-feature-based hierarchical clustering (SFHC) and deep multi-input gated recurrent unit (DMGRU) . The proposed two-stage method is capable of capturing the spatial-temporal features of traffic network. Specifically, the SFHC divides the road segments into several clusters having similar traffic features, and then the clustered data is fed into the DMGRU for TTP. Our experiments conducted on the practical dataset demonstrate that the designed prediction method can achieve the mean absolute percentage error (MAPE) of 3.3109% and mean absolute error (MAE) of 2.5658, which outperform various combinations of baseline clustering algorithms and prediction models.

Sensing navigational environment represented by navigation marks is an important task for unmanned ships and intelligent navigation systems, and the sensing can be performed by recognizing the images from a camera. In order to improve the image recognition accuracy, this paper combined a contour accentuation algorithm into a multiple scale attention mechanism‐based classification model for navigation marks. Experimental results show that the method increases the accuracy of navigation mark classification from 95.98% to 96.53%. Based on the classification model, an intelligent navigation mark recognition system was developed for the Changjiang Nanjing Waterway Bureau, in which the model is deployed and updated by the TensorFlow Serving.

In recent years, machine learning is applied in the bioinformatics and medical fields to analyze relationships among biological features and behaviors. However, it is difficult to discover the significant features of large-scale datasets. A novel feature extraction method called bioinformatic explainable deep neural network (BioExpDNN) is proposed to filter the critical features with strong influences on the dataset and to explain the interaction of features. In the practical experiments, this study adopted three biomedical science datasets from the UCI (University of California, Irvine) Machine Learning Repository: (1). Cryotherapy Data Set (CDS) contains 6 attributes and 2 classes (i.e., recovery and non-recovery); (2). Cervical Cancer Behavior Risk Data Set (CCBRDS) consists of 18 attributes and 2 classes (i.e., cervical cancer patient and healthy body); (3). Heart Failure Clinical Records Data Set (HFCRDS) includes 12 clinical attributes and 2 classes (i.e., death and life). In comparison results, extracted features were considered as inputs of a classifier based on deep neural network for classification. The classification accuracy was selected as an evaluation factor to evaluate the performance of feature extraction methods. The experimental results showed that the classification accuracies of CDS, CCBRDS, and HFCRDS were 92.59%, 100%, and 78.9%, respectively.

The increasing demand for on-device deep learning necessitates the deployment of deep models on mobile devices. However, directly deploying deep models on mobile devices presents both capacity bottleneck and prohibitive privacy risk. To address these problems, we develop a Differentially Private Knowledge Distillation (DPKD) framework to enable on-device deep learning as well as preserve training data privacy. We modify the conventional Private Aggregation of Teacher Ensembles (PATE) paradigm by compressing the knowledge acquired by the ensemble of teachers into a student model in a differentially private manner. The student model is then trained on both the labeled, public data and the distilled knowledge by adopting a mixed training algorithm. Extensive experiments on popular image datasets, as well as the real implementation on a mobile device show that DPKD can not only benefit from the distilled knowledge but also provide a strong differential privacy guarantee (ε=2$) with only marginal decreases in accuracy.

Blockchain and deep learning have been important techniques for intelligent applications. This study surveys relevant hot research topics from January 2018 to August 2020. Furthermore, five topics of blockchain and deep learning which include (1) infrastructure, (2) finance and trade, (3) transportation and logistics, (4) smart contract, and (5) information security are discussed in this study.

The global environment has become more polluted due to the rapid development of industrial technology. However, the existing machine learning prediction methods of air quality fail to analyze the reasons for the change of air pollution concentration because most of the prediction methods take more focus on the model selection. Since the framework of recent deep learning is very flexible, the model may be deep and complex in order to fit the dataset. Therefore, overfitting problems may exist in a single deep neural network model when the number of weights in the deep neural network model is large. Besides, the learning rate of stochastic gradient descent (SGD) treats all parameters equally, resulting in local optimal solution. In this paper, the Pearson correlation coefficient is used to analyze the inherent correlation of PM2.5 and other auxiliary data such as meteorological data, season data, and time stamp data which are applied to cluster for enhancing the performance. Extracted features are helpful to build a deep ensemble network (EN) model which combines the recurrent neural network (RNN), long short-term memory (LSTM) network, and gated recurrent unit (GRU) network to predict the PM2.5 concentration of the next hour. The weights of the submodel change with the accuracy of them in the validation set, so the ensemble has generalization ability. The adaptive moment estimation (Adam) an algorithm for stochastic optimization is used to optimize the weights instead of SGD. In order to compare the overall performance of different algorithms, the mean absolute error (MAE) and mean absolute percentage error (MAPE) are used as accuracy metrics in the experiments of this study. The experiment results show that the proposed method achieves an accuracy rate (i.e., MAE=6.19 and MAPE=16.20 %) and outperforms the comparative models.

This letter proposes random neural networks (RNNs) to randomly train several neural network (NN) models for the promotion of traditional NN. Moreover, an arrival time prediction method (ATPM) based on RNNs is proposed to predict the stop-to-stop travel time for motor carriers. In experiments, the results showed that the average accuracies of RNNs are 94.75% for highway and 78.22% for urban road, respectively. Furthermore, the accuracies of the proposed ATPM are higher than previous data mining methods. Therefore, the proposed ATPM is suitable to predict the stop-to-stop travel time for motor carriers.

Abstract The control signals of cellular networks have been used to infer the traffic conditions of the road network. In particular, consecutive handover events are being used to estimate the traffic speed. During traffic congestion, consecutive handover events may be rare because vehicles move slowly, and thus very few or no speed reports would be generated from the congested area.However, the traffic speed report rate during traffic congestion has not been investigated in the literature. In this paper, we present an analytic model to estimate the speed report rate from cellular network signaling in steady traffic conditions, that is, the traffic speed and flow are assumed constant. Real field trial data were used to validate our analytic model. In addition, computer simulations were conducted to study how speed reports are generated in dynamic traffic conditions when traffic speed and flow change rapidly. Our study indicates that in a typical cell of length 1.5 km with a typical expected call holding time of 1 min, no speed report was generated from a congested three‐lane highway. Our study demonstrates that the lack of speed reports from consecutive handover events during rush hours indicates severe traffic congestion, and new methods that can estimate traffic speed from cellular network data during severe traffic congestion need to be developed. Copyright © 2013 John Wiley & Sons, Ltd.

The real-time traffic information estimation is an important issue for Intelligent Transportation System (ITS). Compared to traditional methods, the traffic information estimations from cellular network data are readily available, more cost-effective, and easier to deploy and maintain. In this paper, a vehicle speed estimation method is proposed to analyze the voice call signals (i.e., Call Arrival (CA) and Handover (HO)) from cellular networks. Some numerical models are proposed to estimate traffic flow according to HOs and estimate traffic density according to CAs. Then the estimated traffic flow and estimated traffic density are adopted to calculate the estimated vehicle speed. The experimental results show the accuracy of proposed vehicle speed estimation method is 89.75%. Therefore, this method is suitable to be adopted to obtain real-time traffic information for road user.

Information and communication technologies have improved the quality of Intelligent Transportation Systems (ITS). The real-time traffic information has traditionally been collected via stationary vehicle detectors or GPS-based probe cars. Compared to the traditional ways, estimating traffic information from Cellular Floating Vehicle Data (CFVD) is more cost-effective, and easier to acquire. In this paper, this study proposes a novel approach to evaluate the relation of call arrival, handover, traffic flow, and traffic density. Moreover, the traffic speed is estimated by the proposed approach according to CFVD. Through the analytical analysis, this study analyzes the effects of traffic information (e.g. traffic flow and vehicle speed) and communication behaviors (e.g. call arrival rate and call holding time) on handovers and call arrivals. In the simulation, this study compares the estimated traffic information with the real traffic information. The experiment results show that the accuracy of traffic speed estimation is 89.75%. Therefore, the proposed approach can be used to estimate traffic speed from CFVD for ITS.

Conventional temporary telephone number (TTN) service provides a secondary telephone number to a service subscriber. The subscribers can protect privacy by sharing their TTNs without revealing their private numbers. However, TTN service is rarely provided because it requires a large amount of temporary telephone numbers. In this paper, we present a restricted form of TTN service - temporary call-back telephone number (TCN) service -- to reduce the amount of telephone numbers needed. A TCN is assigned only for the communications between a subscriber and the correspondent parties specified by the subscriber. Other users, except for the specified correspondent parties, cannot reach the subscriber through the assigned TCN. A TCN assignment can be specified by a telephone number mapping record that consists of a subscriber's number, a TCN, and a correspondent party's number. A TCN can be assigned to more than one subscriber, as long as the subscribers' specified correspondent parties are different. Analytic models using Markov chains have been developed to show that the TCN service significantly reduces the amount of temporary numbers required. The TCN service can be implemented using the Intelligent Network (IN) service architecture as a value-added service. TCN Subscribers can initiate calls to and receive calls from the specified correspondent parties without revealing their private telephone numbers. However, the TCN service is not a replacement of the conventional TTN service because only the specified correspondent parties can reach the subscriber by the assigned TCN.