Designing recurrent neural networks for monitoring embedded devices

Embedded systems play an important role in various tasks in many areas of our lives. In the case of safety-critical applications, e.g., in the fields of autonomous driving, medical devices or control of unmanned aerial vehicles (UAV), the correct system operation must always be guaranteed. Standard methods for monitoring an embedded application, i.e., detecting erroneous behavior at run-time, require a detailed system understanding during development which increases the design effort significantly.Our approach uses Artificial Neural Networks (ANN), specifically recurrent Long Short-Term Memory (LSTM) architectures, to realize cost efficient monitoring for embedded systems. We propose extensions to existing ANN-based monitoring approaches and investigate suitable ANN architectures, which facilitate fault detection on the open-source UAV flight stack PX4. We demonstrate that our ANN-based approach can detect faults on a raw data stream coming from the monitored system, thus minimizing the need to engineer curated data streams in order to adapt the approach to a different device.

Subjects

Anomaly Detection

Artificial Neural Network

Long Short-Term Memory

Monitoring

PX4

Unmanned Aerial Vehicles

Options

Designing recurrent neural networks for monitoring embedded devices