Mobile Systems and Analytics
The content in this webpage complements the paper “NB-IoT Random Access: Data-driven Analysis and ML-based Enhancements”. The paper has been accepted for publication in IEEE Internet of Things Journal, and an early access version can be found here.
The complete dataset is available below. Please consider to cite the paper if you use the dataset: G. Caso et. al. “NB-IoT Random Access: Data-driven Analysis and ML-based Enhancements,” IEEE Internet Things J., to appear, 2021.
For the NB-IoT measurements in Oslo and Rome, we used the Rohde&Schwarz (R&S) TSMA6 toolkit, along with an Exelonix Narrowband (NB) USB device and a global position-ing system (GPS) antenna.
A description of the dataset attributes collected by the TSMA6 toolkit is available here.
A description of the dataset attributes related to Random Access (RA) and jointly collected by TSMA6 and Exelonix device is available below:
The dataset contains the above RA-related data collected in 21, 69, and 24 sub-campaigns in Deep Indoor (DI), Indoor (I), and Outdoor (O) scenarios, respectively. O is divided in Outdoor Walking (OW) and Outdoor Driving (OD) sub-campaigns. Overall, two NB-IoT operators (Op1 and Op2) have been detected.
A sample of the dataset related to RA, consisting of one complete OD sub-campaign for Op2 is available below:
We configured the Exelonix module via ATtention (AT) Commands, and created the following three test cases. For each operator, we run the test cases in parallel with the scanner measurements.
Test Case 1: The device performs repeated RA executions spaced out by short waiting times
Test Case 2: The device performs, after a successful RA, a connectivity test via Internet Control Message Protocol (ICMP) ping composed by 4 pings, toward the Google Domain Name System (DNS) server located at 8.8.8.8.
Test Case 3: The device performs, after a successful RA, a 1 KB data upload via File Transfer Protocol (FTP), toward a proprietary server located in Oslo.
Statistical validation of coverage results.
We perform additional statistical significance tests to pinpoint differences between the means of the distributions. In particular, we examine the following cases:
We perform the Kruskal-Wallis non-parametric analysis of variance test to assess whether the data come from the same distribution. For the scenario comparison, since we have more than two groups, we also leverage the Dunn’s test to identify significant difference between the means of two or more distributions for a given confidence interval.
We report the p-values below: (Significant codes: ‘*’ 0.05 , ‘**’ 0.01, ‘***’ 0.001)
Parameter | DI | I | O |
RSRP | 0.1594 | 0.00033 *** | 0.0403 * |
SINR | 0.3824 | 0.00054 *** | 0.2224 |
Operator/Parameter | DI-I | DI-O | I-O |
Op1/RSRP | 1.47e-07 *** | 1.2e-07 *** | 0.2948 |
Op2/RSRP | 2.3e-06 *** | 9.2e-10 *** | 0.006 ** |
Op1/SINR | 0.00178 ** | 0.94066 | 0.00034 *** |
Op2/SINR | 0.0282 * | 0.0282 * | 0.6942 |
Out-of-bag accuracy for the Random Forest classifier, as a function of the number of trained trees.
The results show that after about 10/20 trees the Out-of-bag accuracy converges in DI and I scenarios. For O, it slightly increases after 20 trees, in particular when more features are used for the classification task. Hence, 50 trees represents a good rather conservative choice for the classifier in all scenarios.
Confusion matrix for each Operator/Scenario combination. The Random Forest classifier uses 50 trees and {RSRP, RSRQ, SINR} feature set to predict the RA outcomes.