When a complex engineering system malfunctions, we rely on sensors to tell us what went wrong. However, a fundamental bottleneck frequently arises: different physical faults can produce identical sensor signatures. This phenomenon, known as diagnostic ambiguity or fault indistinguishability, occurs because the mapping from a system's internal state space to its observable measurement space is often many-to-one.
If two entirely distinct failures manifest as the exact same sequence of alerts or data points, a passive monitoring system cannot uniquely isolate the root cause. Understanding why this ambiguity happens—and how to systematically break it—is a core focus of modern reliability engineering and automated diagnosis.
The Many-to-One Mapping of System Failures
To understand why faults become indistinguishable, consider a system as a hidden internal state that we can only glimpse through a limited window of sensors. As explained by [Daniel Dvorak] et al. in [Model-Based Monitoring of Dynamic Systems], insufficient sensor density creates a strict observability bottleneck. When a physical asset is under-instrumented, multiple distinct physical malfunctions inevitably project onto identical observable data streams.
Internal Fault Space Sensor Observation Space
┌───────────────────┐ ┌──────────────────────┐
│ [Fault A: Valve] ├────────┐ │ │
└───────────────────┘ │ │ │
├────────►│ [Sensor Signature] │
┌───────────────────┐ │ │ │
│ [Fault B: Pump] ├────────┘ │ │
└───────────────────┘ └──────────────────────┘
This signature overlap spans across domains. In large-scale networks, [Irina Rish] et al. demonstrate in [Optimizing Probe Selection for Fault Localization] that completely different internal node failures can generate identical sets of end-to-end error signals if the network lacks sufficient path diversity. Similarly, in software systems, [Nuno Cardoso] et al. show in [A Kernel Density Estimate-Based Approach to Component Goodness Modeling] that traditional diagnostic engines fail when multiple distinct software components exhibit identical execution patterns in hit data, masking which specific module is actually exhausting resources or failing.
The Geometry of Blind Spots: Transmission Zeros
In continuous and dynamic systems, fault indistinguishability is not just an unfortunate coincidence; it is a structural property governed by system geometry. A powerful explanation for this is provided by [Mohammad Amin Sheikhi] et al. in [Data-Driven Fault Isolation in Linear Time-Invariant Systems: A Subspace Classification Approach].
The authors frame fault isolation geometrically by analyzing how fault signals propagate through a system's transfer functions. A system contains structural "blind spots" known as transmission zeros—frequencies or directional vectors where the system's internal dynamics completely suppress or cancel out an incoming signal before it can reach the output sensors.
When a specific component failure aligns perfectly with these transmission zeros, its unique distinguishing features are entirely masked at the output. As a result, the residual signature of that fault either disappears entirely or aligns identically with the signature of a completely different failure mode. Under these conditions, distinguishing between specific sensor or actuator failures becomes theoretically impossible from passive output data alone.
Cyber-Physical and Environmental Confounding
The diagnostic challenge intensifies in modern, integrated systems where physical hardware is tightly coupled with digital layers or volatile environments. Here, ambiguity arises not just between two physical parts, but between entirely different layers of the system architecture:
- Hardware vs. Software/Sensing: In power grids and autonomous vehicles, [Ajay Chhokra] et al. note in [Hierarchical Reasoning about Faults in Cyber-Physical Energy Systems using Temporal Causal Diagrams] that a digital sensor glitch can mimic a severe physical infrastructure failure. Misinterpreting a software anomaly as a physical rupture can lead to catastrophic, unnecessary shutdowns.
- Intrinsic vs. Extrinsic Causes: In energy storage, [Joachim Schaeffer] et al. demonstrate in [Gaussian process-based online health monitoring and fault analysis of lithium-ion battery systems from field data] that passive monitoring of rising internal resistance cannot reliably differentiate between intrinsic chemical degradation inside a battery cell and extrinsic mechanical issues like a corroded external connector.
Living with Ambiguity: Characterization and Ambiguity Zones
When a system is fundamentally non-diagnosable via passive observation, how do diagnostic engines avoid making arbitrary, incorrect predictions?
One robust strategy is to formally map out and accept these limitations by defining ambiguity zones. As proposed by [Wlamir Olivares Loesch Vianna] et al. in [Health Monitoring of an Auxiliary Power Unit Using a Classification Tree], when sensor data is insufficient for precise differentiation, indistinguishable failure modes should be explicitly grouped together. Rather than guessing a single incorrect diagnosis, the system maintains reliability by outputting the entire set of plausible candidates for maintenance crews to investigate.
To find these zones before a system is even built, researchers use formal verification. [Anika Schumann] et al. in [A Scalable Jointree Algorithm for Diagnosability] and [Marco Bozzano] et al. in [SMT-Based Validation of Timed Failure Propagation Graphs] leverage automated reasoning and Satisfiability Modulo Theories (SMT) to exhaustively check whether a system model will produce identical observation sequences for both faulty and non-faulty states.
Furthermore, understanding how diagnostic algorithms behave when these boundaries are crossed is crucial. [Xingyu Su] explored this in [Time Decomposition for Diagnosis of Discrete Event Systems (Extended Abstract)], analyzing how window-based real-time monitoring algorithms perform when forced to operate within inherently non-diagnosable systems containing ambiguous fault scenarios.
Breaking the Symmetry via Active Probing
If passive observation leaves a diagnostic engine trapped inside an ambiguity zone, the only way to isolate the true root cause is to change the state of the system. This is the core intuition behind active testing or active diagnosis.
Passive Observation (Ambiguous) Active Probing (Resolved)
┌───────────┐ ┌───────────┐
│ Fault A ├───────┐ │ Fault A ├─────────────► Trajectory X
└───────────┘ │ └───────────┘
▼ Auxiliary
Identical Signal
Trajectory Injected
▲ │
┌───────────┐ │ ┌───────────┐ ▼
│ Fault B ├───────┘ │ Fault B ├─────────────► Trajectory Y
└───────────┘ └───────────┘
Instead of waiting for symptoms to manifest under normal operating conditions, an active diagnostic engine injects small, deliberate perturbations into the system to force the indistinguishable fault trajectories to diverge.
[Alexander Feldman] et al. formalize this in [FRACTAL: Efficient Fault Isolation Using Active Testing], utilizing submodular greedy algorithms to select the most cost-effective system inputs that will rapidly narrow down a massive list of potential failure causes.
A vivid physical application of this concept is demonstrated by [Sicheng Zhou] et al. in [Fault Separation Based on An Excitation Operator with Application to a Quadrotor UAV]. In autonomous drones, passive systems cannot differentiate between motor aging, sudden thrust loss, or a shifting payload, because all three look like an unexpected drop in altitude. By implementing an "excitation operator" that injects brief, targeted auxiliary signals directly into the quadrotor's motors, the system forces each fault type to respond with a unique aerodynamic signature. This active maneuver breaks the symmetry of the overlapping symptoms, allowing the drone to precisely isolate its internal health status mid-flight.
As autonomous platforms take on higher stakes, integrating these active diagnostic maneuvers directly into onboard control loops—as detailed by [Gabor Karsai] in [Model-based On-board Decision Making for Autonomous Aircraft]—allows machines to autonomously probe their own environments and architectures, transforming ambiguous, unresolvable data into clear, actionable recovery decisions.
Key Foundational Work
Model-Based Sensor Diagnosis: When Monitoring Should Be Monitored (1994) — contributed 1 concept elements to this direction.
Model-Based Monitoring of Dynamic Systems (1989) — contributed 1 concept elements to this direction.