Method and apparatus for anomaly detection for individual vehicles in swarm system
A method for detecting anomalies in a swarm system comprises: collecting first movement data from multiple vehicles moving as a swarm in a first scenario; generating first training data based on positioning data and second training data based on multi-channel inertial sensor data from the first movement data; training a first learning model using the first training data and multiple second learning models using the second training data for each vehicle; receiving real-time second movement data from vehicles moving as a swarm in a second scenario; generating first input data based on positioning data from the second movement data; inputting the first input data into the first learning model to detect abnormal vehicles in real-time; generating second input data for abnormal vehicles based on inertial sensor data from the second movement data; and inputting the second input data into the corresponding second learning model to identify abnormal channels in the inertial measurement unit of abnormal vehicles.
1 . A method for detecting anomalies in a swarm system, wherein the method is implemented on a computing device comprising a processor, and a memory storing instructions or programs executable by the processor, and comprises:
collecting a plurality of first movement data from multiple vehicles that moved as a swarm according to a first scenario, wherein the plurality of first movement data includes positioning data and multi-channel inertial sensor data;
generating a plurality of first training data corresponding to each of the vehicles based on the positioning data included in the plurality of first movement data, and training a first learning model using the plurality of first training data;
generating a plurality of second training data corresponding to each of the vehicles based on the multi-channel inertial sensor data included in the plurality of first movement data, and training a plurality of second learning models corresponding to each of the vehicles using the second training data generated for each of the vehicles;
receiving in real time a plurality of second movement data from the vehicles moving as a swarm according to a second scenario, wherein the plurality of second movement data includes positioning data and multi-channel inertial sensor data;
generating a plurality of first input data corresponding to each of the vehicles based on the positioning data included in the plurality of second movement data, and inputting the plurality of first input data into the first learning model to detect in real time any abnormal vehicle estimated to be in an abnormal state among the vehicles; and
generating a second input data corresponding to the abnormal vehicle based on the multi-channel inertial sensor data included in the second movement data corresponding to the abnormal vehicle, and inputting the second input data into the second learning model corresponding to the abnormal vehicle to identify the abnormal channel among the multiple channels of the inertial measurement unit of the abnormal vehicle.
2 . The method according to claim 1 , wherein the second learning models are trained using a supervised learning technique, a semi-supervised learning technique, or an unsupervised learning technique with the second training data based on a labeling status of the second training data.
3 . The method according to claim 2 , wherein only a portion of the second training data is labeled as being in a normal state, and the second learning models are trained using a semi-supervised learning technique with the second training data.
4 . The method according to claim 3 , wherein the second learning models are Generative models.
5 . The method according to claim 4 , wherein the second learning models are any one of a VAE (Variational Autoencoder), AnoGAN, GANomaly, or Skip-GANomaly.
6 . The method according to claim 4 , wherein the second learning models generate multi-channel reconstructed data for each of the multi-channel inertial sensor data included in the second input data, identify the abnormal state of the abnormal vehicle based on reconstruction errors between each of the multi-channel inertial sensor data and each of the multi-channel reconstructed data, and identify the abnormal channel based on anomaly probability values calculated for each of the multiple channels.
7 . A method for detecting anomalies in a swarm system, wherein the method is implemented on a computing device comprising a processor, and a memory storing instructions or programs executable by the processor, and comprises:
collecting a plurality of first movement data from multiple vehicles that moved as a swarm according to a first scenario, wherein the plurality of first movement data includes multi-channel inertial sensor data;
generating a plurality of training data corresponding to each of the vehicles based on the plurality of first movement data, and training a plurality of learning models corresponding to each of the vehicles using the training data generated for each vehicle;
collecting in real time a plurality of second movement data from the vehicles moving as a swarm according to a second scenario, wherein the plurality of second movement data includes multi-channel inertial sensor data;
generating a plurality of input data corresponding to each of the vehicles based on the plurality of second movement data, and inputting the plurality of input data into the corresponding learning models to detect anomalies in each of the vehicles; and
identifying the abnormal channel among the multiple channels of the inertial measurement unit of the vehicle detected to be in an abnormal state.
8 . The method according to claim 7 , wherein the learning models are trained using a supervised learning technique, a semi-supervised learning technique, or an unsupervised learning technique with the training data based on a labeling status of the training data.
9 . The method according to claim 8 , wherein only a portion of the training data is labeled as being in a normal state, and the learning models are trained using a semi-supervised learning technique with the training data.
10 . The method according to claim 9 , wherein the learning models are Generative models.
11 . The method according to claim 10 , wherein the learning models are any one of a VAE (Variational Autoencoder), AnoGAN, GANomaly, or Skip-GANomaly.
12 . The method according to claim 10 , wherein the learning models generate reconstructed data for each of the multi-channel inertial sensor data included in the second movement data, identify the abnormal state of the abnormal vehicle based on the reconstruction errors between the inertial sensor data and the reconstructed data for each channel, and identify the abnormal channel based on the anomaly probability values calculated for each channel.
13 . A non-transitory computer-readable recording medium storing a program for executing the method of claim 1 on a computer.