The repository contains the code to run the simulations used for the paper "Advancing Spectrum Anomaly Detection through Digital Twins" which can be found here. In the paper, a novel approach for identifying spectrum anomalies is proposed, which employs a digital twin of the radio environment. The simulations are based on ray tracing using Sionna. Blender is used to create the model of the scenario.
If you use the code or parts of it, please cite
@article{schoesser24,
author={Schösser, Anton and Burmeister, Friedrich and Schulz, Philipp and Khursheed, Mohd Danish and Ma, Sinuo and Fettweis, Gerhard},
journal={IEEE Communications Magazine},
title={Advancing Spectrum Anomaly Detection through Digital Twins},
year={2024},
month={11},
pages={1-7},
doi={10.1109/MCOM.001.2400221}
}- Status at the time of final paper submission.
- So far, there only have been visual changes to plots.
The code was developoed Python modules 3.10.10 and the following module versions have been used:
- Numpy: 1.23.5
- TensorFlow: 2.13.0
- Mitsuba: 3.2.1
- Sionna: 0.15.1
- Drjit: 0.4.1
- Hydra: 1.3.2
For successful export of the scene, the Mitsuba Add-on for Blender needs to be installed first. The scenario for ray tracing is created using Blender and a Python script. The script create_scenario_with_obstacles.pyis provided in the folder blender-python. To execute the script, open Blender and got to the Scripting tab. Then, open the script and run it. The script will create a scenario with obstacles and save it in the folder scenes. The scene is exported in the mitsuba format which can be processed by sionna.
Adjustments to the scenario can be specified in the file blender-python\conf\scene_attributes.yaml. Please note, that the properties transmitters, jammers and base_stations are only used for visualization purposes.
To visualize the scenario, the script create_scenario_visualization.py can be used. The script runs creates the scenario and visualizes in addition the radio devices. After the script has run, select the camera and click with the right mouse button on the camera. Select Set camera active and afterwards press F12 to render the scene. For rendering, Eevee provides fast results, whereas the results with Cycles are more realistic. When finished, the file needs to be saved manually.
Please note, that the texture files are not part of this repository. They are only needed for visualization purpose, but not for the ray tracing. If needed, please download them:
and insert the files in the folders blender-python\textures\ConcretePoured and blender-python\textures\MetalGalvanizedSteelWorn, respectively.
Generating the dataset to process consists of three steps. In a first step, a huge number of path loss maps for random transmitter locations are generated using ray tracing. Secondly, according to the number of legitimate transmitters and jammers and their associated transmit powers, they are combined (the powers are added up) to obtain the physical twin (PT) radio maps. In a third step, for each sample created in the last step, for each legitimate transmitter, the original locations of the legitimate transmitters are slightly assigned with a random offset to mimic localization inaccuracy and new pathloss maps are generated. The new resulting pathloss maps are then combined to obtain the digital twin (DT) radio maps.
They are generated using the script src\dataset_generation\pathloss_map_generation.py'. The configuration is contained in the file src\dataset_generation\conf\pathloss_map_generation.yaml. Here, for example the scene number and the number of generated pathloss maps as well as the transmitter and receiver heights and ray tracing parameters can be configured.
In this step, the pathloss maps are combined to obtain the PT (original) which might also contain the jammer. To execute the radio map generation, run the script src\dataset_generation\radio_map_generation.py. The configuration is contained in the file src\dataset_generation\conf\radio_map_generation.yaml. Here, for example the scene number and the number of generated radio maps as well as the number and transmit powers of the transmitters and jammers can be configured.
The last step of the dataset generation is denoted as measurement generation. In this step, for each PT radio map which was created in the previous step, its DT counterpart is generated. Therefore, the orginial locations of the legitimate transmitters is shifted by a random offset to mimick localization inaccuracy. Moreover, the jammer is not known by the DT and therefore not modeled. Two approaches are compared in this work to generate the DT. The script is located at src\dataset_generation\measurement_generation.py.
The original scene file is used together with the shifted transmitter locations to generate the DT radio maps. From those, the estimated path loss values between the estimated transmitter location and the sensing units is extracted.
New path loss maps are generated to train an ML model to estimate the path loss values between the estimated transmitter location and the sensing units. After the model is trained on the newly generated path loss maps, the model is used to infer the path loss values between the estimated transmitter location and the sensing units.
Currently, random forest (RF) is used as a multi-output regressor, i.e., for each sensing unit a separate RF model is trained.
As input, either only the estimated transmitter location coordinates can be used or additionally the distance between the estimated transmitter location and the sensing unit can be used. The second case is applied in the paper, as it slightly reduces the estimation error.
After those steps are executed, the final dataset file is generate, which contains for eeach sensing unit the difference between the actual received power from the PT and the estimated received power from the DT as well as a binary label indicating whether a jammer was present or not.
Currently, the following number is used to identify the different scenarios:
- PL dataset:
- 0: Sensing unit height: 1.5m, Transmitter height: 1.5m, used for PT generation
- 1: Sensing unit height: 1.5m, Transmitter height: 1.5m, used for train the ML model for path loss estimation
- 2: Sensing unit height: 5.0m, Transmitter height: 1.5m, used for PT generation
- 3: Sensing unit height: 5.0m, Transmitter height: 1.5m, used for train the ML model for path loss estimation
- RM dataset:
- 0: Sensing unit height: 1.5m, Transmitter height: 1.5m, used for PT generation
- 2: Sensing unit height: 5.0m, Transmitter height: 1.5m, used for PT generation
- Measurement dataset:
- 0: Sensing unit height: 1.5m, Transmitter height: 1.5m, using ray tracing to create the DT
- 1: Sensing unit height: 1.5m, Transmitter height: 1.5m, using ML to create the DT
- 2: Sensing unit height: 5.0m, Transmitter height: 1.5m, using ray tracing to create the DT
- 3: Sensing unit height: 5.0m, Transmitter height: 1.5m, using ML to create the DT
The file number needs to be adjusted in the file src\dataset_generation\conf\measurement_generation.yaml.
Evaluating different anomaly detection algorithms on the generated datasets can be done with the script src\anomaly_detection\sionna_anomaly_detection.py. The configuration can be done with the configuration file src\anomaly_detection\conf\sionna_anomaly_detection.yaml. Particularly, the dataset and the employed algorithm can be specified there. The results are saved in the folder datasets\results. To draw the ROC curves, the output of the anomaly detection must be a soft output, therefore probability must be set to True in the configuration file.
The ROC curves can be drawn then in the notebook notebooks\roc_curves_sionna.ipynb.