← All research

Research 01 · Revised manuscript · under review

SB-QGWO

Finding the right sensors and network settings for jet-engine failure prediction, with far fewer training runs.

My role
First author: concept, method, code, analysis, writing
Status
Revised manuscript · under review · 2026
Data
NASA C-MAPSS turbofan degradation (FD001–FD004)
Authors
Aaditya Dev Sharma, Shlok Chakravarthy, Rik Das, Prasun Nandy, Saranik Mahapatra
16–51%fewer evaluations than TPE to match its final quality
11.08capped test RMSE on FD004 (11.67 on FD002)
93–97%coverage of 95% prediction intervals after conformal calibration

The problem

Predicting the remaining useful life of a turbofan engine depends on two coupled choices: which of the 21 sensors feed the model, and how the network is configured. Both are usually picked by hand.

Searching them jointly is the right idea, but every candidate means training a network. The real bottleneck is sample efficiency under noisy results, not raw accuracy.

What we built

  1. 01mRMR-QUBO sensor prior

    Score sensors for relevance and redundancy as a binary optimisation problem.

  2. 02Simulated bifurcation

    Solve the QUBO. Recovers the certified optimum in 74 of 75 cases (greedy: 57).

  3. 03Q-bit Grey Wolf search

    ε-bounded Q-bits, leader rotation, NOT-gate mutation and an elitist archive.

  4. 04Conformal uncertainty

    A model-agnostic layer that turns point predictions into calibrated intervals.

  5. 05Maintenance decision

    A lower-quantile replacement rule that removes residual failures.

Results

Evaluations needed to reach TPE’s final quality

Two controlled benchmarks. TPE uses its full budget; SB-QGWO matches it after 55 of 96 and 94 of 192 evaluations.

What we found

A naive quantum-inspired port has three failure modes

Saturating Q-bits, arcsine-biased decoding and a final measurement that throws away the best evaluated solution. The last one alone costs about 1.5 RMSE on the real task. SB-QGWO is designed around all three.

The sensor prior cleans up the inputs

Selected models keep about one uninformative sensor on average, against three to four without the prior, and come out smaller than TPE’s (124–752k vs 1.03–1.34M parameters).

Random-window validation picks the wrong model

Its model ranking is negatively correlated with the leakage-free protocol (ρ = −0.57), a pitfall that applies well beyond this paper.

Uncertainty makes it decision-ready

Conformal intervals reach 93–97% coverage, and the replacement rule lowers the maintenance cost rate by 5–21%. On FD001 and FD003, TPE and a CNN edge it out slightly, and the paper says so.

Making the search work meant first proving why the obvious version breaks. And a strict, leakage-free evaluation protocol changed which model wins.

Full titleSB-QGWO: sample-efficient joint sensor and hyper-parameter search with a simulated-bifurcation-seeded quantum-inspired Grey Wolf optimizer for decision-ready turbofan prognostics

PythonPyTorchMetaheuristicsQUBOConformal predictionPredictive maintenance
Next · Research 02QIH-RAG →

An honest, controlled study of which parts of a biomedical retrieval pipeline actually matter.