Anomaly-Detection-Using-IMU-Data

This is a repository created for the ICAS Competition. This was created because the original repository was private. You can access the original repository if you want.

---

In here we use given IMU(Inertial Measurement Unit) data of a vehicle to detect anomalies. The data was in rodbag file. ROS(Robot Operating System)

---

Initial Requirements

Python

Bagpy

Pandas

Numpy

Matplotlib

Sklearn

Tensorflow

Keras

---

Steps

• you have to extract the data from given rosbag file. There are so many data in it. Then filterout the IMU data. You can use this method fot that.
• Save the extracted data in csv file for further processing techniques. You can see something like this for IMU data

Time	header.seq	header.stamp.secs	header.frame_id	orientation.x	...	orientation.y	linear_acceleration_covariance_7	linear_acceleration_covariance_8
1.6167e+09	123648	1616766027	fcu	-0.036490	...	0.021663	0.0	9.000000e-08
1.6167e+09	123649	1616766027	fcu	-0.036553	...	0.021804	0.0	9.000000e-08
1.6167e+09	123650	1616766027	fcu	-0.036523	...	0.022071	0.0	9.000000e-08
1.6167e+09	123651	1616766027	fcu	-0.036695	...	0.022406	0.0	9.000000e-08
............	...........	.................	...............	.............	...	.............	................................	................................
............	...........	.................	...............	.............	...	.............	................................	................................
............	...........	.................	...............	.............	...	.............	................................	................................

---

• After that you have to crate a deep learning model for forcasting... You can simply follow the code for more understanding

  • For forecasting I only used 6 IMU data features

  • angular_velocity.x angular_velocity.y angular_velocity.z linear_acceleration.x linear_acceleration.y linear_acceleration.z

  • Preprocess the data set and reshape the data set. No of steps = 3
    • 6 features are split to 3 consecutive data samples. Take it as the Input(X) and consecutive 4th Data as output(Y)
  • Split as Train & Test data, Scale to between 0 and 1
  • Build a LSTM Forcaster model and Check for the accuracies after fitting the model
  • Results...
    • Training Loss ~ 0.15% Validation Loss ~ 1.18% Training Accuracy ~ 81.59% Validation Accuracy ~ 44.83%

---

• Capturing the Abnormalities...