Teleoperation serves as the fallback solution to autonomous driving but reliable functions of the teleoperation require a certain amount of mobile network resources, which cannot be guaranteed at all times. Therefore, predictive quality of service (pQoS) is introduced as a concept to increase the resilience of the teleoperation. In this paper, based on a data measurement campaign, we propose a prediction framework to prediction two important network KPIs of teleoperation: uplink data-rate and round-trip latency. Furthermore, we introduce a method to alleviate the performance degradation of machine-learning-based prediction models on previously unseen data due to concept drift by incorporating historic data into the prediction pipeline. Additionally, we introduce the metric of critical scenario detection to evaluate the prediction performance specifically for teleoperation.
Figures & tables
Fig. 1: Data measurement routes with different odd
Fig. 2: pqos framework for teleoperation that uses historic data from rem and real-time measurement from teleoperated vehicles
RSRQ
RSRP
SINR
Throughput
Lat.
Long.
UL
0.32
0.45
0.34
/
0.23
0.15
Latency
-0.06
-0.04
-0.04
0.02
-0.03
0.02
TABLE I: Correlation between Application Layer and phy kpi
Fig. 3: Empirical Cumulative Distribution Function (ECDF) for the UL data-rate and round-trip latency of the two routes
PHY
RSRP
Reference Signal Received Power - indicates the strength of the received signal from the base station.
RSRQ
Reference Signal Received Quality - reflects the quality of the received reference signal and accounts for interference.
SINR
Signal to Interference plus Noise Ratio - measures the signal quality considering both interference and noise.
CQI
Channel Quality Indicator - measures the quality factors of the radio signal and the radio channel between the end user equipment and the base station
Application Layer
Throughput
Actual ul throughput that are used by the software on vehicle
TABLE II: Input Feature kpi for Online Inference
ML Model
Value
Description
lag
60
Number of past data frames included in the prediction framework
step
60
Number of time horizons predicted into the future
max depth
4
Maximum tree depth for base learners
number of estimators
100
Number of gradient boosted trees
REM
Value
Description
distance weights
[0.8, 0.2, 0.6]
The weight vector for the three distances in knn algorithm
TABLE III: Hyper-parameters for Online Inference
Fig. 4: Mean absolute error (MAE) for uplink data-rate and round-trip latency prediction over prediction horizons for route 1
Fig. 5: Comparison of ul data-rate and round-trip latency prediction between with historic data and without historic data for route 1 (same route as where the training data are recorded) and route 2 (completely unseen data). The prediction horizon is for the next step, i.e., second. The data are smoothed with a window of 5 seconds and the outliers are removed for better visualization. True positive ( TP ), false positive ( FP ), and false negative ( FN ) showcase the performance of critical scenario detection.
Route 1
Condition
F1
Precision
Recall
UL with historic data
< 14 Mbps
0.467
0.549
0.406
UL without
< 14 Mbps
0.505
0.688
0.399
Latency with historic data
> 100 ms
0.503
0.465
0.548
Latency without
> 100 ms
0.471
0.456
0.488
Route 2
Condition
F1
Precision
Recall
UL with historic data
< 14 Mbps
0.489
0.333
0.917
TABLE IV: F1, precision, and recall for critical scenario detection for teleoperation