Fault Detection in Control Software via Program Flow Monitoring using Spiking Neural Networks

Modern machinery depends on embedded software controllers operating reliably and precisely across a wide range of processors. As integrated circuits continue to shrink in scale and grow in complexity, they become increasingly susceptible to faults from a variety of sources. These faults raise the likelihood of controller failure, which can in turn compromise the systems they regulate. This paper presents an approach for early detection of such failures by exploiting the temporal classification capabilities of Spiking Neural Networks (SNNs). The proposed method identifies patterns in processor activity caused by program faults and determines whether a controller is deviating from expected behaviour. To evaluate the approach, we implement a test system consisting of a Proportional–Integral–Derivative (PID) controller regulating a Van der Pol oscillator. Faults are injected at run time by corrupting instructions after they are fetched from memory. Using this setup, we generate a dataset of processor signal recordings captured during both normal and fault-injected execution. These recordings are then converted into spike trains representing the instruction stream. We describe a preliminary SNN-based fault detection method and present experimental results demonstrating its ability to detect faulty controller behaviour within the generated dataset.