In traditional engineering, we often assume that if a component breaks, it stays broken. This "persistence assumption" simplifies troubleshooting: if a system fails, you look for the broken part; if a test passes, you "exonerate" the components involved and move on.
However, real-world systems—from deep-space probes to nationwide communication networks—frequently suffer from intermittent faults. These are the "ghosts in the machine" that appear and disappear without intervention. As Johan de Kleer notes in Diagnosing Multiple Persistent and Intermittent Faults, these faults are notoriously difficult to isolate because they violate the basic logic of binary diagnosis. When a component only fails sporadically, a passing test is no longer proof of health.
The Flaw of Binary Exoneration
The core challenge of intermittent diagnosis is the "No Fault Found" (NFF) phenomenon. Jing Qiu and colleagues explain in Intermittent fault’s parameter framework and stochastic petri net based formalization model that NFF outcomes occur when a system fails in the field but appears functional during maintenance.
Traditional diagnostic engines use a process of elimination. If a software test passes, every line of code executed during that test is assumed to be correct. But in the presence of "flaky tests" or race conditions, a faulty component might produce a correct output 90% of the time. Rui Abreu 0001, Peter Zoeteweij, and Arjan J. C. van Gemund addressed this in their work on BARINEL (A New Bayesian Approach to Multiple Intermittent Fault Diagnosis). Instead of exonerating components, they treat a passing test as probabilistic evidence.
BARINEL uses maximum likelihood estimation to calculate a "failure intermittency parameter"—the exact probability that a faulty component will "act healthy" during any given observation. This allows the diagnostic engine to maintain a suspicion level for a component even if it just passed a test, preventing the algorithm from prematurely ruling out the true culprit.
Modeling the Temporal Dimension
Intermittent faults are often tied to time or specific system states. A loose connection might only cause data corruption during high-vibration maneuvers, or a transistor might only fail when it reaches a specific thermal threshold.
To catch these, researchers have moved toward temporal and state-based modeling:
- Sequential Logic: Alexander Feldman and Ingo Pill developed efficient ways to diagnose sequential circuits by treating failures as independent events across time steps (Efficient Model-Based Diagnosis of Sequential Circuits). This is critical for high-radiation environments like space, where "soft errors" are fleeting.
- State Transitions: Franz Lackinger and Wolfgang Nejdl introduced DIAMON, which tracks synchronized observations over time to identify faults that only manifest during specific operational transitions (Integrating Model-Based Monitoring and Diagnosis of Complex Dynamic Systems).
- Stochastic Simulation: Early work like the DEPEND environment by Kumar K. Goswami allowed engineers to simulate clusters of transient failures that occur in close temporal proximity, capturing how repeated recovery cycles can destabilize a system (DEPEND: A design environment for prediction and evaluation of system dependability).
Why Intermittency is More Dangerous Than Failure
A permanent failure is often "safe" because it triggers binary monitors. If a motor dies, the system stops. However, as Homa Alemzadeh and Zbigniew Kalbarczyk demonstrate in the context of robotic surgery, intermittent faults can bypass these safety triggers (Systems-Theoretic Safety Assessment of Robotic Telesurgical Systems).
A transient data glitch might cause a robotic arm to "jump" or vibrate violently without ever triggering an emergency stop, because the system's monitors are looking for a total hardware shutdown, not a momentary deviation. This makes the ability to diagnose and route around intermittent faults essential for cyber-physical resilience, a theme echoed by Xu Zheng in the study of smart city infrastructures (Special Issue on Selected Technologies and Applications in Wireless Communications for Smart City Computing).
Scaling Diagnosis to Big Data
When dealing with billions of web requests or nationwide networks, simple logic models break down. John C. Platt and Emre Kiciman pioneered the use of fast variational inference to handle this scale (Fast Variational Inference for Large-scale Internet Diagnosis). By using a "noisy-OR" graphical model, they could probabilistically link intermittent failures across a massive bipartite graph of users and services, identifying root causes in hours rather than days.
Similarly, in software testing, the SEQUOIA algorithm (Spectrum-Based Sequential Diagnosis) and the TDP paradigm by Tom Zamir and Roni Stern use Markov Decision Processes to intelligently choose the next test to run. By explicitly modeling the "flakiness" of a bug, these systems can reduce the number of tests needed to find a fault by up to 80% (Using Model-Based Diagnosis to Improve Software Testing).
Summary of the Intuition
The shift in this research area is from deterministic elimination to probabilistic accumulation.
- Don't Exonerate: A passing test is just a "maybe."
- Model the Flicker: Every component has a probability of failing and a probability of "acting healthy" while failed.
- Use Context: Look for faults that only appear during specific state transitions or time windows.
- Update Beliefs: Use Bayesian inference to constantly adjust the suspicion level of every component based on a stream of noisy, intermittent observations.