Speaker
Description
Spiking Neural Networks (SNNs) are emerging as a disruptive technology in artificial intelligence, offering a biologically inspired approach that enables energy-efficient and event-driven computation. Their potential impact on spatial applications, particularly for anomaly detection in satellite data, is significant. Traditional deep learning models often require extensive computational resources, making them impractical for deployment on edge devices such as satellites. In contrast, SNNs leverage sparse, asynchronous processing to efficiently handle high-dimensional time-series data, such as satellite data. Their inherent temporal processing capabilities make them particularly well-suited for detecting anomalies in dynamic and resource-constrained environments, where rapid and reliable decision-making is crucial.
A key enabler of SNNs in space applications is the integration with FPGAs, which offer significant advantages in terms of low power consumption, real-time processing, and adaptability to mission-specific requirements. By deploying SNNs on FPGAs, satellites can perform onboard anomaly detection with minimal energy overhead, reducing reliance on ground-based processing and lowering data transmission loads.
To demonstrate the potential of this approach, our team developed a spiking autoencoder and applied it to the ESA Anomalies Dataset, a collection of satellite telemetry data. The solution comes from a deep study of the biological models at the basis of neuronal function combining consolidated approach from the literature to the anomaly detection on time series data problem. Autoencoders are one of the most used and consolidated architectures for anomaly detection on time series. Their working principle is based on learning how to represent in a compress manner the characteristics of an input sample (encode) and then reconstruct the input itself from the compress representation (decode). The idea is to measure the error between the input and the reconstruction and discriminate between normal and anomal data using the reconstruction error. The SNN-based autoencoder was trained to reconstruct normal operational patterns, detecting anomalies as high values in the reconstruction error value. Preliminary results show that this method effectively captures complex temporal dependencies while maintaining high detection performance with significantly reduced energy consumption compared to conventional deep learning models. It achieves very good performance with very few data with respect to benchmark approaches enabling future implementation of on-chip (and maybe online) learning. In particular we arrived at 99.7% accuracy score with a train dataset of only 3 months length tested on the full mission dataset (84 months). This fusion of SNNs, FPGA-based neuromorphic hardware, and real-world satellite data paves the way for intelligent and autonomous anomaly detection in future space missions.
Affiliation of author(s)
Politecnico di Milano
| Track | Artificial Intelligence/Machine Learning |
|---|
