Zhenghua Zhang,∗ Jiafeng Zhang,∗∗ Rui Gao,∗ Chongxin Fang,∗∗∗ and Jin Qian∗∗∗∗


  1. [1] M. Chen, X.H. Yu, and Y. Liu, PCNN: Deep convolutional networks for short-term traffic congestion prediction, IEEE Transactions on Intelligent Transportation Systems, 19(11), 2018, 3550–3559.
  2. [2] S.S. Anjum, Modeling traffic congestion based on air quality for greener environment: An empirical study, IEEE Access, 7, 2019, 57100–57119.
  3. [3] H. Kim and S. Han, An efficient sensor deployment scheme for large-scale wireless sensor networks, IEEE Communications Letters, 19(1), 2015, 98–101.
  4. [4] K. Zheng, S. Zhao, Z. Yang, X. Xiong, and W. Xiang, Design and implementation of LPWA-based air quality monitoring system, IEEE Access, 4, 2017, 3238–3245.
  5. [6] in this paper. It makes data transmission more efficient and effective. After getting monitoring data, they are preprocessed first. Here, the methods of eliminating the error of loss and adaptive weighting
  6. [7] are adopted. These can provide the optimal data samples for the subsequent fusion method. In
  7. [8] introduced a method of correlation analysis. This paper has studied this method of correlation analysis. In
  8. [9] used particle swarm optimization to model traffic pollution. For the fusion of large amounts of traffic data
  9. [10], In
  10. [11] and
  11. [12] proposed a data fusion scheme based on the BP neural network. Although the BP neural network is widely used, it still has the disadvantage of slow convergence and local minimum. In recent years, data fusion theories and prediction models have been widely studied in intelligent transportation. In
  12. [13], a hybrid algorithm (STL − LSTM) which combines the addition mode of Seasonal-Trend decomposition based on Loess (STL) and the LSTM neural network was proposed. Chen et al.’s goal was to mitigate the influences of irregular fluctuation and improve the performance of short-term metro ridership prediction. Pan and Zhou 489
  13. [14] proposed a differential evolution back propagation (DE − BP) neural network traffic prediction model applicable for a smart cities’ network to predict the network traffic. Wang and Fang
  14. [15] used a multiple regression model and a back propagation (BP) neural network model. The prediction model is optimized by combining the two methods. Document
  15. [16] proposed a simulation model based upon the traffic conditions. One of the performance indicators relates to traffic pollution. To predict carbon dioxide emissions from roads, second-order regression models were employed to fit the exact carbon dioxide (CO2) emissions in
  16. [17]. In
  17. [18], a mixed forest prediction method considering the spatiotemporal correlation characteristics of urban road traffic state was constructed by improving the existing random forest algorithm. In summary, the works mentioned above are mainly based on traffic flow prediction and traffic pollution monitoring. The content of these studies is relatively single and not representative. They did not link traffic congestion to pollution. How to combine the urban traffic pollution and congestion information is quite an interested topic. The main contributions of this paper are concluded as follows: First, we combined the BP neural network with correlation coefficients, analysed the relationship between traffic pollution and congestion and studied the potential for potential traffic. Further, the wireless sensor network and ZigBee wireless communication network technology were used to collect traffic data. Then, this paper adopts a three-layer data fusion scheme to improve the validity of sample data. Finally, this paper optimizes training samples by correlation analysis and optimizes BP network for convergence problems. The rest of this paper is organized as follows. In Section 2, the deployment of the sensor monitoring network used in this paper is described. Section 3 explains the preprocessing process of traffic pollution data in detail, that is, the method of eliminating the blunder error and adaptive weighting. In next Section 4 we can see, the correlation coefficient is analysed, the initialization weight is optimized and optimized BP neural network algorithm suitable for traffic data is derived. In Section 5, the experiment is verified. Finally, Section 6 gives conclusions and future work. 2. Collection of Traffic Pollution Monitoring Data The traffic pollution data monitoring and acquisition system
  18. [19] used in this paper is mainly composed of three units: sensor unit, microcontroller processing unit and ZigBee wireless communication module. The data are finally passed to the monitoring centre server through the embedded gateway as shown in Fig. 1. The existing air quality monitoring system is to monitor urban traffic pollution on both sides of urban roads. Urban air pollutants mainly include carbon monoxide, hydrocarbons, sulphur dioxide, nitrogen oxides, ozone, respirable particulate matter, etc. The traffic pollution monitoring system mainly consists of three parts. At the front end of the system, the traffic pollution monitoring and acquisiMonitoring acquisition module ZigBee Embedded gateway ZigBee 4G Monitoring sensors Microprocessors ZigBee 4G Monitoring center server Embedde d gateway Figure 1. System hardware structure. Intelligent gateway nodes 1 4 2 3 Monitoring acquisition nodes Figure 2. Road layout of system monitoring acquisition nodes and gateways. tion module combine with the ZigBee wireless transceiver module to form a data acquisition module. The middle end of the system is the embedded ARM intelligent gateway module. 4G module and server system form the back end. The road layout of the system monitoring collection node and intelligent gateway is shown in Fig. 2. This model is built at large crossroads in the centre of Yangzhou City. Collect traffic data of four points at the intersection and upload it to the monitoring centre server through the network. Then the system backend for big data fusion. Due to various factors, it is difficult to directly measure traffic congestion data. An idea of calculating the intersection waiting time by the vehicle timing device is more suitable for this. It is the time from the stop to the exit, recorded as t. In this way, it can be judged that the congestion at each intersection is at each moment, imitating the rules of the traffic operation index.Set the red light time to 60 s and the congestion index as p, p = t 60 × 2. The congestion index is abbreviated as TPI. Its value is divided into five stages, namely, 0 ∼ 2, 2 ∼ 4, 4 ∼ 6, 6 ∼ 8 and more than 8. They are defined as “unblocked”, “basic 490 Correlation analysis Various sensors Monitoring data inspection Data weighting BP and correlation coefficient combined Crossroads First layer fusion Second layer fusion Third layer fusion Monitoring results Figure 3. Overall system information fusion structure. smooth” and “light”. Degree of congestion, “moderate congestion” and “serious congestion”. A 3-layer fusion scheme was proposed. The specific process is shown in Fig. 3. Through the data fusion scheme shown in Fig. 3, the traffic pollution and congestion dataafter processing are more effective. The following begins to analyse the specific process of data fusion for each layer. 3. Traffic Pollution Data Preprocessing We first use the distribution map method to eliminate the error. Then, the optimized monitoring data are obtained by the adaptive weighting algorithm to provide high-quality training samples. 3.1 Distribution Map Method to Eliminate Blunder Error Sensors can be subject to interference in specific environments. This can cause errors in the measurement, i.e., blunder error. Blunder errors can affect the consistency of measurement data. Therefore, it is necessary to remove the blunder error here. The data obtained from the urban traffic pollution monitoring and acquisition system are tested by the method of distribution map. Its algorithm
  19. [20] is as follows: Let the n measurement results measured by a sensor to be sorted in ascending order. Get the sequence: x1 ∼ xn. x1 is called the lower limit of the measurement and xn is called the upper limit. Record xm as the median. The upper quartile au is the median value of [xm, xn]. Lower quartile al is the median value of [x1, xm]. The dispersion of the quartile is d = au − al. The judgment interval of valid data is [ρ1, ρ2], ρ1 = al − β 2 d, ρ2 = au + β 2 d. Where β is a constant and is usually a value of 1.0 or 2.0. This constant can be chosen according to the accuracy of the system requirements. This is used for error elimination of traffic pollution data. Due to the complexity of the traffic data and the high possibility of change, the deviation between the data is large. Therefore, β can take a value greater than or equal to 2 to make the data range wider. If the measurement data are within the interval [ρ1, ρ2], it is considered to be valid consistency measurement data and the effective interval can be used to eliminate 50% of the error. Data outside the interval [ρ1, ρ2] can be considered as a loss error and will eventually be eliminated. The remaining data are valid data after the consistency check. The consistent test of the measured data by the distribution method is not limited by the data distribuTable 1 PM 2.5 Concentration Measurement Data (μg/m3) Various 1 2 3 4 5 6 7 Sensors 7:00 133.1 132.7 54.8 132.0 131.9 131.7 131.6 9:00 91.7 131.8 132.0 131.8 132.7 0 131.9 11:00 161.7 131.8 131.8 131.9 132.0 172.5 131.1 13:00 132.7 133.7 133.8 134.0 134.1 132.3 132.4 15:00 132.3 132.4 131.5 130.8 132.1 187.8 132.2 Table 2 Measurement Data of PM 2.5 Concentration after Treatment (μg/m3) Various 1 2 3 4 5 6 7 Sensors 7:00 133.1 132.7 132.0 131.9 131.7 131.6 9:00 131.8 132.0 131.8 132.7 131.9 11:00 131.8 131.8 131.9 132.0 131.1 13:00 132.7 133.7 133.8 134.0 134.1 132.3 132.4 15:00 132.3 132.4 131.5 130.8 132.1 132.2 tion because the quartile dispersion and median value in the distribution method only depend on the distribution position of the data, regardless of the size of the extreme points, thus enhance the robustness of data processing. Take the PM 2.5 concentration data measured by a given sensor at different times in the same place as an example. As shown in Table 1, the average value of the first group of data is calculated to be 121.11, the upper quartile is 132.90 and the lower quartile is 131.80. According to the distribution graph method, the data 54.8 is the divergence data, which is the error of the deviation. To reduce the measurement error, you need to reject it. Similarly, after the consistency detection process of the programme, the PM 2.5 pollution concentration data of each group is shown in Table 2. As shown in Table 2, the blunder error in the data has been removed and adaptive weighting can be performed below. 491 Start Read the No data collected by the sensor Yes Calculate mean and variance End Initiative value Figure 4. Weighted fusion flow chart. 3.2 Traffic Pollution Data Adaptive Weighting The adaptive weighted data fusion algorithm
  20. [21] has different weights for different sets of measurement data. Under the optimal condition that the total mean square error is the smallest, the corresponding weights are searched adaptively according to each set of measurement data, so that the fusion the latter value is optimal. The weighted fusion flow chart is shown in Fig. 4. In a monitoring system, there are n sensors that monitor and sample a subject. The monitoring values are x1 ∼ xn. The variance of each sensor node is σ2 1 ∼ σ2 n. The corresponding weighting factors are w1 ∼ wn. The final fusion value of the multi-sensor data should be ˆx = n i=1 ωixi. n i=1 ωi = 1, so the total mean square error is σ2 = n i=1 ωiσ2 i . The σ2 is a function of each weighting factor ωi. According to the extremum method of the multivariate function, the weighting factor when the mean square error is minimum can be found as ωi = 1 σ2 i n i=1 1 σ2 i . At this time, σ2 is the minimum value and σ2 min = 1 n i=1 σ2 i . The parameter values of the measured object are objectively existing constants, and thus the estimation can be made based on the arithmetic mean of the existing monitoring data. Let the i group of sensors make the k measurements, then: ¯xi(k) = 1 k k j=1 xi(j) i = 1, 2 · · · n (1) Get an estimate ˆ¯x = n i=1 ωi ¯xi(k). The total mean square error is ¯σ2 = 1 k n i=1 ω2 i σ2 i . According to (2): ¯σ2 = 1 k n i=1 1 σ2 i (2) Obviously ¯σ2 < σ2 min, as k increases, ¯σ2 gradually decreases. The adaptive algorithm can adaptively select a more suitable weighting factor according to different measured values. Let us take a set of PM 2.5 concentration data as an example. After the above adaptive weighted fusion algorithm programme, the mean, variance and weighting factors corTable 3 Mean, Variance and Weighting Factors of PM 2.5 Concentration Data at Each Time Time 7:00 9:00 11:00 13:00 15:00 Mean 132.42 132.02 132.10 133.29 132.00 Variance 0.247 0.126 0.260 0.530 0.354 Weighting 0.197 0.387 0.187 0.092 0.137 responding to the PM2.5 concentration data at each time are shown in Table 3. The final fusion result was 132.228. In the same way, the data fusion value of other traffic pollution parameters can be obtained by the same method and the obtained data are closer to the true value of the measured parameter than the arithmetic average value, which greatly improves the accuracy of the collected data. It provides technical support for the accurate detection of urban traffic pollution conditions. 4. Correlation Coefficient and BP Neural Network Traffic pollutants mainly include solid suspended particles, carbon oxides, hydrocarbons, lead and sulphur oxides. The correlation coefficient between each pollutant and the congestion index is analysed in turn, so that the degree of correlation with traffic congestion can be obtained. This screens out contaminants with a large correlation coefficient. In practical applications, the connection between variables is irrelevant, so we use the simplified Spearman correlation coefficient, where the correlation coefficient r is defined as r = 1 − 6 n i=1 d2 i n(n2 − 1) (3) where di = xi − yi, i = 1, 2, . . . , n, n is the sample size, xi and yi are two sets of data and r is in the range between −1 and 1, it can be seen that the correlation coefficient is divided into positive correlation and negative correlation. The correlation coefficient r was first proposed by Karl Pearson. It is generally believed that the absolute value of r is above 0.8, and the two sets of data have strong correlation. Between 0.3 and 0.8, a weak correlation can be considered. Below 0.3, there is basically no correlation. Therefore, six kinds of highly correlated pollutants are selected in this paper, namely PM 2.5, PM 10, CO, SO2, NO2 and O3. The correlation coefficients between them and TPI are recorded as r1, r2, r3, r4, r5, and r6. In the BP neural network, the traditional initialization weight problem is randomly initialized with the standard normal distribution. If the quality of the data samples is poor, the weight update amount may be small, the update speed is slow and the hidden layer is saturated. To solve this problem, we envisage that the weight can be initialized according to the correlation analysis result and the initial weight wi is recorded as wi = ri r1 + r2 + r3 + r4 + r5 + r6 × 100%, i = 1, 2, . . . , n (4) 492 The weights obtained in (4) are used as initial evaluation weights. Therefore, the robustness of the sample values and initial weights of this paper is improved. It is envisaged to use the three-layer BP network to realize traffic flow time domain information fusion calculation. Under the condition of the traditional BP network algorithm, the learning rate setting is large, the oscillation does not converge, the learning rate setting is small and the convergence speed is slow. To overcome the shortcomings of the BP algorithm, we use a strategy of combining the momentum term with the adaptive adjustment learning rate
  21. [22]. Increasing the momentum term method allows the network to not only consider the effect of the error on the gradient but also consider the influence of the trend on the error surface, thus effectively suppressing the network from falling into the local minimum state
  22. [23]. In the process of training, the learning rate adaptively adjusts the force map to make the algorithm stable, while at the same time making the learning step size as large as possible, and the learning rate is adjusted according to the local error surface. Record the gradient momentum in the gradient descent method as ΔG, ΔG = (1 − γ)Dk + γDk − 1. The Dk = − ∂E ∂ωk is the negative gradient at step k, η is the learning rate, γ is themomentum factor, 0 = γ < 1, and adding the momentum factor is equivalent to adding the damping term, which can reduce the oscillation tendency of the learning process and effectively improve the convergence. The weight adjustment algorithm for learning rate adaptive adjustment is ηk = 2γηk − 1, γ = sign(DkDk − 1), wk + 1 = wk + ηkDk. When two iterations are repeated, the gradient direction is the same, indicating that the falling speed is too slow. At this time, the learning rate is automatically doubled. When the direction is opposite, the falling is excessive and the learning rate is automaticallyhalved. Combining the above two methods, we can obtain the weight correction algorithm of momentum-adaptive learning rate BP neural network; weight update formula is wk + 1 = wk + ηk((1 − γ)Dk + γDk − 1). 5. Experiment Based on the traditional BP neural network, this experiment optimizes the BP network for the problem of local convergence. Moreover, experiments were carried out using the idea of combining correlation coefficients and BP networks. 5.1 Experimental Programme The BP neural network designed in this paper has six input nodes and one output node. The parameters collected by the urban traffic pollution monitoring system in this experiment mainly include: PM 2.5 pollutant concentration parameter, PM 10 pollutant concentration parameter, traffic exhaust gas emissions CO, SO2, NO2, O3 and other concentrations. The design scheme of BP neural network is shown in Table 4. The monitoring module of this experiment is arranged at a certain intersection in Yangzhou City. As shown in Table 4 Design of BP Neural Network Network Type: Number of nodes: 1 BP Neural Network Layers: 3 Input node Output node X1 PM 2.5 pollutant Y1 Congestion index concentration X2 PM 10 pollutant concentration X3 CO concentration X4 SO2 concentration X5 NO2 concentration X6 O3 concentration Figure 5. PM 2.5 and TPI. Figure 6. PM 10 and TPI. Figs. 5 and 6, a number of monitoring and collecting nodes are arranged at the intersection of the main roads and is equipped with a fixed ID number. The intermediate ARM gateway is used to receive and upload data. A total of 400 sets of data were collected, of which 350 were used as training samples for BP network and 50 were used as test samples. Selected training and test samples are shown in Tables 5 and 6. 493 Table 5 Part of Training Samples Samples PM 2.5 PM 10 CO SO2 NO2 O3 Y 1 47 107 1 18 46 44 5.64 2 20 40 0.5 7 15 60 6.42 3 11 16 0.4 5 20 47 4.47 4 50 72 0.8 9 29 69 2.78 5 86 147 1.2 16 43 72 6.68 6 32 47 0.9 8 40 35 4.07 7 79 140 1.1 9 29 77 3.99 8 17 64 0.5 8 24 73 6.3 9 29 88 0.6 9 32 64 5.27 10 41 104 0.7 9 41 62 3.35 Table 6 Partial Test Samples Samples PM 2.5 PM 10 CO SO2 NO2 O3 Y 1 18 66 0.2 10 30 108 3.9 2 44 87 0.4 13 47 133 1.75 3 63 105 0.5 15 66 105 3.19 4 66 94 0.4 10 30 150 5.98 5 68 85 0.6 12 44 132 4.07 6 65 92 0.6 15 43 176 7.7 7 56 92 0.7 16 42 162 2.2 8 64 103 0.8 16 42 140 5.36 9 68 109 0.9 15 33 161 6.26 10 16 32 0.5 6 16 91 4.8 The traffic pollution data in the table are different for each unit, and the data size is quite different, which is not conducive to the training experiment. Therefore, all training data will be normalized before training, and the data will be transformed into interval [0, 1]. 5.2 Experimental Results We divided this experiment into two parts, which are correlation coefficient analysis and BP network verification. 5.2.1 Relevance Experiment After the data preprocessing is completed, 100 groups of traffic pollution data at the same time every day are selected and 100 groups of data obtained after pretreatment are correlated. A total of 9 pollutants are measured in this experiment, and analyze the correlation between them and the traffic congestion index in turn. As can be seen from Fig. 7, the concentration of PM 2.5, PM 10, SO2, CO, NO2 is positively correlated with TPI. From a trend perspective, the higher the concentration of pollutants, the greater the TPI. As can be seen from Fig. 7(b), the concentration of O3 exhibits a negative correlation with TPI. Their trends are reversed. This may be understood to be unstable in a higher temperature environment. Table 7 is a statistical table of the correlation coefficients of all measured pollutants and congestion index in this experiment. To more intuitively see the relationship between them, Fig. 8 is drawn. As shown in Fig. 8, we conclude by summarizing the correlations of the four regional points: In the experimental study, we found that among all the traffic pollutants mentioned in this paper, the correlation coefficients of HC, CO2 and NO with TPI are <0.1. This correlation is basically negligible. To improve the training efficiency of the BP neural network, these three pollutants can be removed and the subsequent experiments will be verified. The BP network weights are initialized according to the correlation coefficients in (4) and Table 7. In general, CO, NO2 and TPI have the greatest correlation. In addition, the higher the congestion index of a region, the greater the correlation. 5.2.2 BP Neural Network Verification The previous correlation analysis can roughly get the relationship between traffic pollution and traffic congestion, and it can also explain the effectiveness of the data preprocessing fusion process. The following data can be obtained through BP neural network training before fusion, and further training fusion can be seen. The result is shown in Fig. 9: Figure 9(a) shows the prediction results of the correlation coefficient and the optimized neural network. The results show that after the optimized BP neural network training, the output is basically consistent with the actual situation and the trend of traffic congestion is almost consistent. This can explain the high reliability of BP neural network and further demonstrate the effectiveness of BP neural network based on data fusion of urban traffic pollution monitoring system. There are certain errors in individual points, which can be explained by some sudden traffic factors and errors caused by instrument failure. Figure 9(b) shows the prediction results of the traditional BP neural network. The predicted values are the same as the measured values, but the errors are large. It can be seen from Table 8 that different methods cause different errors. The MAE and RMSE generated by the conventional BP neural network are 0.0994 and 0.1556, respectively. The optimized BP network is 0.086 and 0.1293, respectively. The result shows that MAE and RMSE of the optimized BP network are both smaller than the conventional BP network. It is proved that the fusion scheme based on data preprocessing, correlation analysis and BP neural network modelling can make the prediction model more accurate. 494 TPI(0-10)SO2 200 100 0 200 100 0 20 10 0 10 5 0 0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100 time (days) (a) 2 1 0 100 50 0 200 100 0 10 5 0 0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100 time (days) TPI O 3 NO 2 CO TPI SO 2 PM10 PM2.5 TPI(0-10)O3 NO2COmg/m3 g/m3 10 g/m3 (b) g/m3 Figure 7. These are statistical charts comparing various pollutants with TPI. The images are labelled in an alphabetical order (a) PM 2.5, PM 10, SO2 and TPI, (b) CO, NO2, O3 and TPI. Table 7 Correlation Coefficient r PM 2.5 PM 10 SO2 HC CO CO2 NO NO2 O3 TPI 16.83% 10.22% 19.37% 3.03% 33.46% -6.25% 5.24% 22.74% −16.37% 40 30 20 10 0 -10 -20 PM2.5 PM10 SO 2 HC CO CO 2 Variouspollutants NO NO O 2 3 Correlationcoefficient(%) Figure 8. Correlation analysis. It is effective to analyse the relationship between traffic congestion and pollution. Through the three-layer data fusion, the traffic congestion data can be used to judge the current traffic congestion. The personnel of the monitoring centre can check the pollution prediction results of the traffic pollution monitoring system to know the urban road pollution and road traffic congestion in advance. The situation provides a basis for urban environmental protection and rational allocation of transportation resources and can provide a plan for traffic management and take the lead. 6. Conclusion and Future Works To study the connection between traffic pollution and congestion, we first use the combination of wireless sensor network and ZigBee technology to collect traffic data. Then, data monitoring, preprocessing, correlation analysis and BP network are combined to use in this paper. According to the experimental results, traffic pollution and congestion have a great correlation. Among them, CO and NO2 have higher correlation with congestion index. It is innovative to use the method of data fusion to study the connection 495 Measurement data Fusion data TPI 8 8 7 7 6 6 5 5 4 4 3 3 2 2 1 1 0 0 10 20 30 40 50 Fusion sample (a) 0 0 10 20 30 40 50 Fusion sample (b) Measurement data Fusion data TPI Figure 9. These are graphs comparing the predicted results with the original values. The images are labelled in an alphabetical order (a) optimized prediction and (b) traditional prediction. Table 8 The Performance Comparison of Two Methods Method MAE RMSE Optimized BP network 0.0860 0.1293 Conventional BP network 0.0994 0.1556 between traffic pollution and congestion. Experimental results show that this method makes effective. Because of the limited conditions, this study only monitored a small amount of data in a region. In the future, we can collect large amounts of data for analysis and comparison in multiple regions. This can improve the effectiveness of the experiment. It is also important to optimize the neural network to better adapt to the training of traffic data. In-depth study of this topic has great significance for the induction of traffic and the protection of urban environment. References [1] M. Chen, X.H. Yu, and Y. Liu, PCNN: Deep convolutional networks for short-term traffic congestion prediction, IEEE Transactions on Intelligent Transportation Systems, 19(11), 2018, 3550–3559. [2] S.S. Anjum, Modeling traffic congestion based on air quality for greener environment: An empirical study, IEEE Access, 7, 2019, 57100–57119. [3] H. Kim and S. Han, An efficient sensor deployment scheme for large-scale wireless sensor networks, IEEE Communications Letters, 19(1), 2015, 98–101. [4] K. Zheng, S. Zhao, Z. Yang, X. Xiong, and W. Xiang, Design and implementation of LPWA-based air quality monitoring system, IEEE Access, 4, 2017, 3238–3245. [5] J. Guevara, F. Barrero, E. Vargas, J. Becerra, and S. Toral, Environmental wireless sensor network for road traffic applications, Intelligent Transport Systems, 6(2), 2012, 177–186. [6] K. Chen, Y. Chen, Y. Tsai, and N. Chen, A novel wireless multifunctional electronic current transformer based on ZigBeebased communication, IEEE Transactions on Smart Grid, 8(4), 2017, 1888–1897. [7] W. Choi and T. Sarkar, Minimum norm property for the sum of the adaptive weights for a direct data domain least squares algorithm, IEEE Transactions on Antennas & Propagation, 54(3), 2006, 1045–1050. [8] M. Blagoiev, I. Gruicin, M. Ionascu, and M. Marcu, A study on correlation between air pollution and traffic, 2018 26th Telecommunications Forum (TELFOR), 2018, 420–425. [9] D.Y. Xu Ning, Feature-adaptive GM(1,1) model and prediction model for traffic pollution emissions in China, Systems Engineering-Theory & Practice, 38(1), 2018, 187–196. [10] P. Wang, C. Wang, J. Lai, Z. Huang, J. Ma, and Y. Mao, Traffic control approach based on multi-source data fusion, IET Intelligent Transport Systems, 13(5), 2019, 764–772. [11] J.H. Chang and C.Y. Tseng, Analysis of correlation between secondary PM2.5 and factory pollution sources by using ANN and the correlation coefficient, IEEE Access, 5(99), 2017, 22812–22822. [12] X. Pan, W. Zhou, Y. Lu, and N. Sun, Prediction of network traffic of smart cities based on DE-BP neural network, IEEE Access, 7, 2019, 55807–55816. [13] D. Chen, J. Zhang, and S. Jiang, Forecasting the short-term metro ridership with seasonal and trend decomposition using loess and LSTM neural networks, IEEE Access, 8, 2020, 91181–91187. [14] X. Pan, W. Zhou, Y. Lu, and N. Sun, Prediction of network traffic of smart cities based on DE-BP neural network, IEEE Access, 7, 2019, 55807–55816. [15] J. Wang, J. Fang, and Y. Zhao, Visual prediction of gas diffusion concentration based on regression analysis and BP neural network, The Journal of Engineering, 2019(23), 2019, 19–23. [16] S.S. Anjum, R.M. Noor, N. Aghamohammadi, I. Ahmedy, L. Mat Kiah, N. Hussin, M.H. Anisi, and M.A. Qureshi, Modeling traffic congestion based on air quality for greener environment: An Empirical study, IEEE Access, 7, 2019, 57100–57119. [17] Z. He, W. Zhang, and N. Jia, Estimating carbon dioxide emissions of freeway traffic: A spatiotemporal cell-based model, IEEE Transactions on Intelligent Transportation Systems, 21(5), 2020, 1976–1986. [18] Z. Chen, Y. Jiang, and D. Sun, Discrimination and prediction of traffic congestion states of urban road network based on spatio-temporal correlation, IEEE Access, 8, 2020, 3330–3342. [19] H. Wang, L. Dong, W. Wei, W.S. Zhao, K. Xu, and G. Wang, The WSN monitoring system for large outdoor advertising boards based on ZigBee and MEMS sensor, IEEE Sensors Journal, 18(3), 2018, 1314–1323. [20] R.P. David, R.S. Neto, and C.A. Medina, Nonassisted adaptive algorithm for data fusion in distributed detection systems, IEEE Transactions on Aerospace & Electronic Systems, 52(3), 2016, 1486–1493. 496 [21] Y. Song, W. Wei, L. Zheng, X. Yang, and L. Wei, An adaptive pansharpening method by using weighted least squares filter, IEEE Geoscience & Remote Sensing Letters, 13(1), 2016, 18–22. [22] Y. Liu, X. Zhu, and J. Yang, Fault diagnosis of PV array based on optimised BP neural network by improved adaptive genetic algorithm, The Journal of Engineering, 2017(13), 2017, 1427–1431. [23] T. Guo, H. Wang, and L. Liu, Vision-based mobile robot leaderfollower control using model predictive control, International Journal of Robotics and Automation, 34(5), 2019, 206–231.

Important Links:

Go Back