This document summarizes a study that detects anomalies in engine coolant temperature sensors using one-class classifiers. It describes how data is acquired from vehicle sensors and preprocessed, features like temperature, load, and their statistics are extracted. Different one-class classifiers like k-NN, OC-SVM, and Gaussian methods are evaluated on normal and artificially injected anomalous data. The results show k-NN and OC-SVM perform best for idle and motion modes respectively, demonstrating these techniques can detect sensor malfunctions of different severity levels.

1. Detecting Anomalies in the Engine Coolant
Sensor using One-Class Classifiers
Eronides F. da Silva Neto, Allan R. S. Feitosa, George D. C.
Cavalcanti and Abel G. Silva-Filho
2019 IEEE 90th Vehicular Technology Conference

2. 2
Agenda
- Introduction
-
- Proposed Anomaly Detection System
-
- Experiments
- Results
- Conclusion
-

3. 3
Introduction
- New services and drive experience provided by the combination of
Information and Communication Technologies (ICTs).
- Some papers have evaluated the detection of anomalies in intra-vehicular
signals as a classification problem.
- The most recent contribution (Theisller, 2017) presents an approach to
detect known and unknown anomalies in the Engine Coolant Temperature
(ECT) sensor.

4. 4
Overview of the Proposed Anomaly Detection System
- Data Acquisition
- Feature Extraction
- Feature Selection
- Classifier Evaluation
-

5. 5
Feature Extraction and Selection
- The data acquisition system connected to the OBD-II interface captures 27
different parameters (attributes) at a rate of one sample per second (1Hz).
- This work is based on the detection of contextual anomalies in the Engine
Coolant Temperature (ECT) sensor.
- Carloop development kit allow us to try to get some specific PIDs from
vehicle.
-

6. 6
Feature Extraction and Selection
- The feature extraction process is based on a sliding window with a
configurable width (N).
- Driving cycles definition
-

7. 7
Feature Selection
- Pre-processing phase normalizes the data:
- From a correlation index analysis, Engine Load is also selected as an
attribute. A driving cycle of k seconds can be represented by:
- In addition to the ECT sensor value and Engine load, the standard
deviation and variance of ECT sensor are also used:
- The final data instance of the system is:

8. 8
One Class Classifier Evaluation
The different techniques evaluated separated by their category are:
- Instance-based techniques: k-NN;
- Statical Methods: Gaussian data description, Parzen and Naive Parzen
window density estimator and Extreme Value Analysis;
- Neural Networks: Self-organizing Map (SOM);
- Rule-based: Minimum Spanning Tree (MST) and Mahalanobis distance;
- Support Vector Machines: SVDD and OC-SVM.
-

9. 9
Experiments
- To evaluate the proposed system, anomalies are artificially injected to the
ECT sensor signal based on three different anomaly levels:
- The artificial component inserted representing ECT sensor malfunctioning
is composed by a significant flicker noise component in the ECT sensor.

10. 10
Experiments: intra-vehicular database
- Two different modes of vehicle operation are recorded: Idle and Motion.
-
- Splitting each vehicular trip (30% used to outliers) , we can define some
data sets:
as target data training (normal operation)
as target data testing data set
as outlier data testing data set
- Classifier Performance Evaluation
- True Positive Rate (TPR)
- Precision
- F2-score

11. 11
Summary of the Results
- The training and testing process are executed 30 times and then the results
evaluated:

12. 12
Results
- For anomaly level II:

13. 13
Results
- Friedman test hypothesis not confirm that the two best techniques, k-NN
and OC-SVM, are statistically different.
- Nemenyi post-hoc test confirms the best techniques based on rank sums

14. 14
Conclusions
- This work has shown the need to evaluate vehicular anomalies in different
malfunctioning or degeneration degrees.
- Experiments has shown k-NN and OC-SVM as the best techniques for the
respective idle and motion modes.
- The proposed approach of injecting anomalies can be applied in other
sensors analysis.

15. Mahalo!
(Thank You)