Most fault diagnosis tools, and most fault-injection testing tools, assume that when something goes wrong there is one cause. That assumption is safest when faults add up: two small defects produce roughly the sum of their individual symptoms. It is weakest in systems where a fluid and a flexible structure push on each other continuously. Examples include a pipe carrying slurry, a rubber-lined water bearing, or an elastic blood vessel. In these systems a slightly thinned wall changes the flow, the changed flow loads the wall differently, and a second, harmless-looking defect elsewhere can tip the loop into resonance or overload. We propose a test generator that searches directly for such pairs. It perturbs the fault configuration and the operating condition, runs a two-way coupled simulator on the nominal case, each fault alone, and both faults together, and keeps the inputs whose joint response departs most from the sum of the individual responses. Each fault alone must pass the safety property while the pair fails it, or each alone must be visible to sensors while the pair is hidden.
The method comes from differential fuzzing, which generates inputs to maximize disagreement between several executions and treats that disagreement as a bug signal without needing a hand-written oracle. Here the disagreement is between the physical system and its own additive, single-fault approximation. The output is useful in two ways. It is a search method that finds dangerous fault interactions at a fixed simulation budget. It is also a benchmark of hard multi-fault cases, built by construction, on which single-fault diagnosers can be tested. Such cases are almost never observed in operating machinery.
Core Claim
Fault-injection testing for cyber-physical systems injects one fault at a time. Fuzzing-based mutation testing of C/C++ software in cyber-physical systems and its tool MOTIF run the original and a single mutated program side by side under AFL++ and search for inputs that make them diverge. Property-Based Mutation Testing counts a mutant as killed only if it violates a formal temporal-logic safety property. That criterion cut the reported mutation score on an automatic transmission controller from 100% to 53.19%, so many single faults are individually benign. Neither work asks whether benign single faults become dangerous in combination. Neither runs against physics with two-way feedback. We claim that in bidirectionally coupled fluid-structure systems, the interesting failures sit in the interaction term between faults. That term can be targeted directly as the divergence signal a fuzzer maximizes.
Concretely, let $u$ be an operating condition (inlet pressure pulse, flow rate, eccentricity), let $a$ and $b$ be two parameterized faults (site, type, severity), and let $y(\cdot)$ be the simulated output trajectory of monitored quantities such as wall stress, pressure at sensor locations, and dominant modal frequency. The interaction residual is
$$\Delta_{ab}(u) = y(a \oplus b; u) - y(a; u) - y(b; u) + y(\varnothing; u),$$
which is zero exactly when the two faults act additively. The fuzzer maximizes the normalized score
$$S = \frac{\lVert \Delta_{ab}(u) \rVert}{\lVert y(a;u) - y(\varnothing;u) \rVert + \lVert y(b;u) - y(\varnothing;u) \rVert + \epsilon},$$
where $\epsilon$ is a noise floor set from repeated nominal runs. An input is recorded as a finding when $S$ is large and one of two conditions holds. In an emergent violation, $a$ and $b$ each satisfy the safety property but $a \oplus b$ violates it. In a masking case, $a$ alone is detectable from sensors but $a \oplus b$ is not. The new element is not fuzzing a simulator, and not multi-fault diagnosis as such. It is using a second difference across four coupled simulations as the search signal, with property checks deciding what counts as a finding.
Research Program
The work has three parts: build the search, filter out numerical artifacts, and use the findings to stress diagnosers.
Building the search. The fault space for each testbed combines structural faults and fluid-side faults. Structural faults are local wall-stiffness loss, local thinning, and liner delamination represented as a stiffness drop. Fluid-side faults are partial occlusion, inlet pulse distortion, and pressure-sensor bias, each placed at one of $K$ sites with continuous severity. The search follows the energy-based seed scheduling used by CureFuzz, which ranks scenario seeds by a mix of novelty and reward when testing reinforcement-learning agents. Our reward is $S$. Our novelty term is a distance-based signature, in the style of TensorFuzz, over a cheap descriptor of the coupled response: the first few modal frequencies, the per-step count of coupling sub-iterations, and the interface displacement peak. The novelty term stops the search from collapsing onto one resonance. The nominal run $y(\varnothing;u)$ is shared across all fault pairs at the same $u$, and single-fault runs are cached. The marginal cost of a new pair is therefore often one simulation, not three. All budgets are reported in simulations.
Filtering numerical artifacts. Two-way partitioned solvers pass boundary displacement and load back and forth at every time step and iterate to agreement. They are known to lose stability when structure and fluid densities are close. Partitioned Deep Learning of Fluid-Structure Interaction describes this iterative quasi-Newton exchange through the preCICE coupling library. A second difference like $\Delta_{ab}$ amplifies solver-tolerance noise, and a coupling divergence can look like a physical interaction. This mirrors the timing-jitter false positives that MOTIF had to filter out with dedicated drivers. Every candidate finding is therefore re-simulated under a coupling tolerance tightened by two orders of magnitude and under a different coupling scheme (serial versus parallel implicit). It is kept only if $\Delta_{ab}$ changes by less than a set fraction. This differential across solver configurations is the variance-guided analogue of cross-framework consensus in XAMT. XAMT uses disagreement between frameworks to separate real bugs from implementation quirks. We use agreement between solver configurations to certify that a finding is physics, not numerics.
Stressing diagnosers. The certified findings form a multi-fault test set. We run two families of single-fault diagnosers on it. The first is a feature-based classifier trained only on single faults, following the trend-plus-frequency features and XGBoost classifier of A Novel Machine Learning Technique for Fault Detection of Pressure Sensor. The second is interventional root-cause diagnosis in the style of Robust Root Cause Diagnosis using In-Distribution Interventions. That method restores a candidate component to normal values and checks whether downstream metrics recover. With a simulator in hand, that intervention can be run exactly. The IDI paper notes that its formulation assumes one root cause per causal path and a directed chain of dependencies. Two-way coupling violates the second assumption by construction. We therefore expect IDI's fix condition to fail on emergent pairs, since restoring either fault alone may already restore nominal behavior. That failure is itself a measurable finding.
Testbeds. The primary testbed is the 1D elastic-tube hemodynamic case coupled through preCICE in Partitioned Deep Learning of Fluid-Structure Interaction. It is cheap enough for tens of thousands of runs. The secondary testbed is a reduced-resolution viscoelastic pipe with two-way coupling, set up following Analysis of Vibration Characteristics of Viscoelastic Slurry Pipe Considering Fluid–Structure Interaction Effects. That study reports that dynamic coupling lowers natural frequencies by about 54%, breaks symmetry-induced frequency degeneracy, and combines sub-additively with added-mass effects. This is direct evidence that non-additive behavior exists even in healthy pipes, and it motivates looking for it between faults. Its reported frequency values also serve as a sanity check that our reduced model is in the right regime. A third, optional testbed follows the laminated water-lubricated bearing in Fluid–Structure Interaction Mechanisms of Layered Thickness Effects on Lubrication Performance and Energy Dissipation in Water-Lubricated Bearings, which reports strong nonlinear amplification above eccentricity 0.8. That makes it a natural place to test whether interactions cluster near a known nonlinear regime.
Baselines
The naive reference points are uniform random sampling and Latin hypercube sampling over (fault pair, severities, operating condition), run at the same simulation budget and with the same single-fault caching. These establish how rare certified interactions are. In a different domain, TensorFuzz found that random search failed to trigger a NaN bug within 10,000,000 mutations while guided search found it immediately. That only shows random search can be hopeless in some spaces. It says nothing about our space, so the comparison has to be run.
The strongest single-fault reference is independent single-fault property falsification, followed by pairing the individually most severe faults. This is the natural extension of MOTIF and PBMT. MOTIF kills 40% to 83% of mutants on satellite software. PBMT reports a 53.19% property-based mutation score on the transmission controller with adaptive random testing, and 42% to 57% with falsification testing. Neither number measures interaction discovery. No reported result was found on Lacuna for multi-fault interaction discovery in any coupled physical simulator. We implement this baseline by first spending half the budget ranking single faults by property robustness margin, then spending the rest pairing the top-ranked faults. Beating it means finding emergent violations among faults that this baseline marks as individually safe, which it by definition does not prioritize.
The strongest general-purpose competitor is Bayesian optimization of $S$ over the same continuous-plus-categorical space, with a Gaussian-process or tree-based surrogate, the same caching, and a hyperparameter budget equal to ours. No reported result was found on Lacuna for Bayesian optimization on this task. Beating it means discovering more distinct certified findings per 1,000 simulations. Distinct findings are clusters in the (site pair, fault type pair) space, so that repeated hits on one resonance are not counted twice.
Two ablations isolate our ingredients. The first replaces $S$ with plain output divergence $\lVert y(a\oplus b)-y(\varnothing)\rVert$. This is standard differential fuzzing between the nominal and doubly faulted system, and it shows whether subtracting the single-fault responses is what matters. The second removes the novelty term, and shows whether diversity is needed for distinct findings rather than just a high peak score.
Why not just…?
Why not just run a Bayesian optimizer or a dense sweep over the fault parameters? With two fault sites and one operating variable, a sweep or a Bayesian optimizer will probably match us, and we will report that honestly. The threat depends on dimension. The number of fault pairs grows as $K^2$ times the number of fault-type pairs, and emergent violations occupy thin regions near resonance and threshold boundaries, which smooth global surrogates tend to smooth over. The experiment varies $K$ from 2 to 16 at a fixed budget. If Bayesian optimization matches our distinct-finding rate at large $K$, the contribution reduces to the objective $S$ and the certification step, and we will present it that way.
Why not compute interactions analytically from mixed second derivatives of the output with respect to the two fault severities? Adjoint or finite-difference mixed sensitivities give the interaction for small faults around the nominal state, and they are a real competitor for mild defects. They are local, though. The failures we target are threshold crossings, where each fault alone moves the system partway toward resonance and both together cross it. A local second-order term evaluated at zero severity can badly underestimate that. We test this directly by ranking pairs by nominal mixed sensitivity and comparing the ranking against certified findings. A high rank correlation would make the fuzzer unnecessary for screening.
Why not fuzz a cheap neural surrogate of the coupled system instead of the solver? Surrogate screening would cut cost by orders of magnitude. However, Multiphysics Bench reports that learned solvers (PINNs, Fourier neural operators, DeepONet, and diffusion models) degrade severely on transient coupled tasks. The partitioned surrogate paper reports long-term error accumulation. Because $\Delta_{ab}$ is a second difference, surrogate error enters it roughly four times. We run a surrogate-screened variant, trained on single-fault data, that proposes candidates for the true solver to verify. We report its precision. If precision is high, it becomes the recommended front end.
Why not use one-way coupling, where fluid loads the structure but the structure does not reshape the flow, since it is much cheaper? This question is also our key scientific test. We run the identical search on the one-way version of each testbed. If certified interactions are as common there, then two-way feedback is not the source of the difficulty and the cheap model suffices. The slurry-pipe results suggest otherwise, but we measure rather than assume.
Proposed Experiments and Decisive Checks
The main experiment compares all methods on both primary testbeds at budgets of 2,000, 10,000, and 50,000 simulations and at several values of $K$, with ten seeds each. It reports certified emergent violations, certified masking cases, and distinct findings per 1,000 simulations. The certification check is decisive for credibility. We report what fraction of raw high-$S$ candidates survive a tighter tolerance and a different coupling scheme. A low survival rate would mean the search mostly finds solver artifacts. That would be a useful negative result about partitioned coupling, but it would not be a result about faults.
The one-way versus two-way comparison is decisive for the physical claim. We expect a markedly higher density of certified interactions under two-way coupling. If the densities are similar, the fluid-structure framing is unjustified.
The diagnosis experiment is decisive for the single-fault claim. We evaluate the feature classifier and the interventional diagnoser on three test sets: single faults, randomly sampled fault pairs at matched severity, and fuzzer-found pairs. We report the recall of the true fault set at top-1 and top-2, following the Recall@1 metric that IDI reports at 0.90 to 1.00 on the PetShop microservice benchmark. The prediction is that accuracy on random pairs stays close to single-fault accuracy, because most random pairs are additive, while accuracy on fuzzer-found pairs falls sharply. That gap is the evidence that the single-fault assumption breaks at identifiable, searchable places in coupled physics, and that current diagnosers are evaluated on the wrong cases. If the gap does not appear, the finding would be that single-fault diagnosers are more robust to coupled interactions than the literature on signal masking suggests, and that is worth reporting too.
References
- Fuzzing-based mutation testing of C/C++ software in cyber-physical systems
- MOTIF: A tool for Mutation Testing with Fuzzing
- Property-Based Mutation Testing
- Curiosity-Driven Testing for Sequential Decision-Making Process
- TensorFuzz: Debugging Neural Networks with Coverage-Guided Fuzzing
- Partitioned Deep Learning of Fluid-Structure Interaction
- XAMT: Cross-Framework API Matching for Testing Deep Learning Libraries
- A Novel Machine Learning Technique for Fault Detection of Pressure Sensor
- Robust Root Cause Diagnosis using In-Distribution Interventions
- Analysis of Vibration Characteristics of Viscoelastic Slurry Pipe Considering Fluid–Structure Interaction Effects
- Fluid–Structure Interaction Mechanisms of Layered Thickness Effects on Lubrication Performance and Energy Dissipation in Water-Lubricated Bearings
- Multiphysics Bench: Benchmarking and Investigating Scientific Machine Learning for Multiphysics PDEs