SPIKE-Dx : A Low-Power High-Throughput Fault Diagnostics Tool
using Spiking Neural Networks for Constraint Systems
Chetan Kulkarni
1
, Johann Schumann
1
, and Anupa Bajwa
2
1
KBR Inc, NASA Ames Research Center, Moffett Field 94035 CA, USA
chetan.s.kulkarni@nasa.gov, johann.schumann@nasa.gov
2
NASA Ames Research Center, Moffett Field 94035 CA, USA
anupa.bajwa@nasa.gov
ABSTRACT
Diagnostic systems are important for many aerospace systems,
which are severely limited in available power, like cubesats or
UAVs. Therefore, traditional diagnostics systems cannot be
used due to their substantial footprint and constraints. In this
paper, we present our very low power diagnostic tool SPIKE-
DX to monitor critical systems with constrained computational
and energy resources. This is made possible through spiking
neural networks (SNNs), which are executable within opti-
mized simulation environments and further implemented on
on cutting-edge neuromorphic hardware.
Based upon Failure Mode and Effect Analysis (FMEA) frame-
work, Diagnostic Bayesian Networks (DBNs) can be con-
structed that provide powerful means for diagnostic reasoning.
In this paper, we describe such DBNs and a method to auto-
matically translate the DBN into highly structured networks
of spiking neurons for execution in SPIKE-DX.
1. INTRODUCTION
Diagnostic tools are critical for many aerospace systems,
which are severely limited in available power. Typical ex-
amples include CubeSats or battery-operated UAVs. They
have numerous subsystems and sensors and need up-to-date
system health information that must be provided by a diagnos-
tics system. However, traditional diagnostics systems have a
substantial computational and power footprint.
For this paper, we have set the following goals: Develop a
powerful on-board diagnostic reasoning system, which con-
sumes minimal electrical power. Use and evaluate Spiking
Neural Networks (SNNs), executed on modern neuromorphic
hardware for efficiency and low power.
Build a diagnostic system based on FMEA models featuring
Chetan Kulkarni et al. This is an open-access article distributed under the
terms of the Creative Commons Attribution 3.0 United States License, which
permits unrestricted use, distribution, and reproduction in any medium, pro-
vided the original author and source are credited.
probabilistic reasoning and confidence metrics. Design the
diagnostic system for certification and V&V.
In this paper, we propose to translate the diagnostic BN into a
Spiking Neural Network (SNN), which consists of groups of
neurons which are sparsely interconnected. Excitatory-type
neurons are used to propagate information, inhibitory neurons
weaken signals and form the basis for a principled translation,
which implements Bayes Belief Reasoning. The translation of
the BN is performed in a pattern-oriented manner: each small
subgroup of connected Bayesian nodes (Markov Blanket) is
translated into individual groups of neurons and weighted con-
nection patterns. These groups of neurons are then composed
into a bigger SNN, which models the original BN
(
Paulin &
van Schaik, 2014; Rao, 2004; Yu, Huang, & Liu, 2018). The
SNN is then compiled onto the neuromorphic hardware using
existing, compiler-based techniques.
We will evaluate our approach using a NASA-relevant Case
Study on monitoring the power system of an autonomous UAS
at LaRC
(
Corbetta & Kulkarni, 2019; Hogge et al., 2018).
Based on existing system and diagnostic models, we propose
to develop the diagnostic Bayesian Network, and use it to test
and validate our BN-to-SNN translation framework.
Application in safety-critical environments requires that di-
agnostic models and reasoning algorithms can be verified,
validated, and certified. In this research work, we combine
the power and expressiveness of Bayesian Belief Networks
(BNs) with the ultra-low power requirements of Spiking Neu-
ral Networks (SNNs) on modern neuromorphic hardware
(
Christensen et al., 2022). We will develop a translation of
BNs to SNNs, which is efficient and amenable to certification.
This is in stark contrast to Machine-learning based approaches
with Deep Neural Networks (DNNs)
(
Zuo, Zhang, Zhang, Luo,
& Liu, 2021), which lack traceability and are not certifiable
with state-of-the art technology
Developing an efficient diagnosis procedure involves two
main steps (see Figure 1): (i) build the BN structure and (ii)
1
