Propose, Check, Revise: Role-Separated Bearing Diagnosis Under Mixed Noise

Neural fault classifiers for rotating machinery are accurate when the background noise is of one kind, but they break when several kinds of noise are mixed and the shaft speed varies. On the University of Ottawa variable-speed bearing data, one of the strongest attention networks falls to about 73.9% accuracy at −9 dB under mixed noise, while it reaches 92.02% at −9 dB on single-type noise (Attention activation network for bearing fault diagnosis under various noise environments). The failure is structural. One set of weights has to filter noise, track speed, and name the fault all at once, and nothing checks whether its answer is physically possible. Multi-agent language-model systems faced a similar problem, where a single model solving and checking its own work keeps confirming its own mistakes. They fixed it by splitting the work into roles. One role proposes, and a separate role with its own evidence accepts or rejects (MACM, Self-collaboration Code Generation via ChatGPT).

We propose to carry over the part of that design that matters and leave out the rest. A fast neural detector proposes a fault hypothesis, for example "outer-race fault." A speed-tracking component estimates the instantaneous shaft frequency. An auditor then runs a statistical test aimed at that single hypothesis. It asks whether the envelope spectrum (the spectrum of the signal's amplitude envelope, where repeated bearing impacts show up as peaks) has energy at the defect frequencies that bearing geometry predicts for this speed. If the auditor rejects, a planner chooses the next hypothesis and a different denoising tool, and the auditor tests again. The claim we will test is narrow. We expect the gain at negative signal-to-noise ratio to come from two sources: the auditor sees evidence that is independent of the detector's, and a test aimed at one hypothesis is statistically easier than classifying blind. We do not expect it to come from language-model reasoning, and we design the experiments so they can show that if it is true.

Core Claim

Classifying a noisy window blind means choosing among roughly ten classes with no prior about where in the spectrum to look. Checking one hypothesis is a different problem. Once a fault type $h$ and a shaft speed are fixed, kinematics give a characteristic frequency $f_h$. The auditor only needs to decide whether a comb of harmonics at $m f_h$ rises above the local noise floor:

$$T_h = \frac{1}{\hat\sigma^2}\sum_{m=1}^{M} \big|E(m f_h)\big|^2 .$$

Here $E$ is the envelope spectrum of the signal after denoising, $M$ is the number of harmonics tested, and $\hat\sigma^2$ is the noise-floor power estimated from neighboring bins. If the envelope bins behave like complex Gaussian noise when no fault is present, $T_h$ is roughly chi-square with $2M$ degrees of freedom. This gives a false-alarm threshold that rests on a stated assumption instead of a learned one. Because mixed noise violates the Gaussian assumption, we will also calibrate the threshold empirically on healthy windows. Detection power grows with $M$ and with window length, and this is where the matched-test advantage comes from.

The second half of the claim is about independence. A verifier that reads the detector's own softmax or features will tend to fail on the same windows the detector fails on. The auditor's errors should be only weakly correlated with the detector's errors, because it works from a different view of the signal: envelope demodulation of a separately denoised signal, and, where available, a second sensor. This is the signal-processing counterpart of the independent Tester in Self-collaboration, which raised edge-case pass rates by up to 34.6% where self-checking gave little. We will measure that correlation directly instead of assuming it.

Relation to Prior Work

Two systems already place language models in diagnosis pipelines, and both differ from this proposal on exactly the point we test. FD-LLM fine-tunes a single Llama-3 model on CWRU vibration windows written out as text strings of FFT values or summary statistics. It reports above 99% on drive-end data but generalizes poorly to fan-end data, and it never tests noise. Its result motivates this proposal, because it shows that reading spectra as text does not carry a physical check. It also constrains the design: we never feed raw numbers to a language model. BatteryAgent chains physics features into a gradient-boosted tree, explains the tree with SHAP attributions, and passes them to one DeepSeek-R1 instance that recalibrates hard cases (AUROC 0.986 on 347 unseen vehicles). Its "sanity check" works from the same features the classifier used, its information flows one way, and it is never stress-tested with noise. The auditor here instead re-reads the raw signal through a different transform and can trigger new denoising. MAM splits medical diagnosis into specialist roles, but it works on text and images, not on physical signals.

The role structure comes from MACM. Its Thinker proposes conditions, a Judge admits only verified ones, and an Executor computes; this raised GPT-4 Turbo on Level-5 MATH from 54.68% to 76.73%. Our detector, auditor, and planner correspond to these three roles, with one deliberate change: our Judge is a statistical test, not a language model. That change is made because of a known constraint. Multi-agent debate spreads errors when agents agree with each other too readily (Peacemaker or Troublemaker; survey). A test with a calibrated false-alarm rate cannot be talked into agreeing.

On the signal side, the methods are reused tools rather than competitors. The planner's toolbox includes adaptive variational mode decomposition with wavelet thresholding (HPO-VMD) and time-varying-filter decomposition (TVF-ITD). These are too slow to run on every window, which is exactly why we call them only when the auditor rejects a hypothesis. Physics-informed transfer learning builds mechanical laws into the training loss. That fixes physical plausibility at training time, whereas our check runs at test time and adapts to the estimated speed of each window.

Research Program

The detector is MLSCA-CW, retrained on our noise protocol, with MHSNet (MHSNet) used on PU and MCC5-THU, where torque channels are available. Each window produces a ranked list of hypotheses with scores. Under the default policy, a window whose top hypothesis passes the auditor is emitted immediately, so the system's latency floor matches that of MLSCA-CW (reported 0.515 ms per sample).

The speed tracker estimates instantaneous shaft frequency from the vibration signal alone, using spectral ridge tracking near the running-speed harmonic. For UofO, the encoder channel is hidden at test time and kept only to score the speed estimate. The auditor resamples the signal to the angle domain using that estimate, computes the envelope spectrum, and evaluates $T_h$ for the hypothesized fault. For healthy hypotheses, it checks that every fault comb is absent. When two sensor locations exist (drive end and fan end on CWRU), the auditor runs the test on each and requires agreement. This addresses the cross-sensor gap that FD-LLM documents, without a model ever needing to transfer across locations.

On rejection, the planner chooses among the next hypothesis, a different denoising tool, a different demodulation band, or abstention. We will compare three planners at the same budget of at most three re-tests per window: a fixed rule that walks the detector's ranked list using a fixed tool, a small learned policy trained on a validation split, and a language-model planner. The language-model planner receives only structured auditor outputs (hypothesis, $T_h$, noise-floor estimate, speed confidence) and never raw numbers. We will treat the language-model planner as a hypothesis to falsify, not as the contribution.

The noise protocol mixes Gaussian, Laplacian, Brownian, and violet noise at random proportions per window. SNR is swept from 0 to −12 dB, and test-time mixtures are held out from the training mixtures. Because MLSCA-CW's exact mixing recipe is not available to us, we retrain every baseline under our protocol and report the published figures only as reference points.

Baselines

The naive reference point is the detector alone. MLSCA-CW is retrained with mixed-noise augmentation and outputs its top-1 class with no checking. This is also the strongest existing method on the CWRU and UofO question. Its paper reports 92.02% accuracy on CWRU at −9 dB and about 73.9% under mixed noise and variable speed on UofO. On PU and MCC5-THU, the strongest reported competitor is MHSNet, at 98.24% on PU at −5 dB and 74.40% on MCC5-THU at −5 dB. Every neural baseline gets the same training data, noise protocol, and hyperparameter search budget (equal trials and GPU hours) as the detector inside our system. To count as beating a baseline, the full system must show higher accuracy under our mixed-noise UofO protocol at −9 dB, outside seed variance across five seeds. It must also show a lower false-alarm rate on healthy windows at matched coverage.

The ablation that isolates the key ingredient is a confidence-threshold self-check. It uses the same detector and the same propose-and-revise loop, but the "auditor" accepts or rejects based on the detector's own calibrated confidence instead of the envelope test. If this matches the full system, independence of evidence is not doing the work. A second ablation runs the auditor on the detector's input view, with no separate denoising and no angle resampling, to separate the effect of the matched test from the effect of the different view.

The strongest agentic competitor is a single language-model agent in a ReAct-style loop. It has the same tool library, the same number of tool calls, the same structured tool outputs, and no separate auditor role. Following FD-LLM, we also include a LoRA-tuned Llama-3-8B fed serialized FFT features, retrained under our noise protocol. No result for it under noise has been reported on Lacuna, since FD-LLM evaluates only clean CWRU data.

Why not just…?

Why not just train a bigger end-to-end network with mixed-noise augmentation? This is the main threat, and our detector baseline is exactly that network, so the experiment settles it directly. The evidence that capacity alone may not be enough is the gap between 92.02% under single-type noise and about 73.9% under mixed noise and variable speed in MLSCA-CW. Fluctuating speed moves the fault frequencies, which a fixed set of weights must absorb, while angle resampling removes that shift by construction. If the augmented network closes the gap by itself, we will report that the auditor adds false-alarm control but not accuracy.

Why not just run a one-shot envelope check after the classifier, with no agents? This is a real threat, and it is close to our system without the loop. We therefore include it as the rule-based planner with a single re-test. If one check-and-abstain step captures most of the gain, the honest conclusion is that what helps is hypothesis-conditioned checking, not iterative decomposition, and the note will say so. We expect re-tests to matter only at the lowest SNRs, where the first denoising choice fails.

Why not just give one language-model agent the tools? The single-agent ReAct baseline tests this at the same tool budget. Prior evidence leans against it. Single agents rarely catch their own errors (Self-collaboration), and ReAct reached 0.631 F1 against 0.666 for role-separated LEGO on WIKIWHY (LEGO). Those results come from text tasks, though, so the comparison has to be run here.

Why not just ensemble two neural detectors? An ensemble trained on the same view shares the same blind spots. We will include a two-model ensemble and report the correlation of its members' errors alongside the correlation between detector and auditor errors.

Proposed Experiments

The benchmarks are CWRU (drive end and fan end, four loads), UofO variable speed with the encoder hidden, PU with vibration and current, and MCC5-THU, where the auditor uses gear-mesh sidebands instead of bearing combs. The metrics are accuracy and macro-F1. We also report selective accuracy against coverage, false-alarm rate on healthy windows, speed-estimation error, the fraction of windows escalated to re-testing, and median and 99th-percentile latency per window.

The first decisive check is the error-correlation measurement. We expect the auditor's errors to be much less correlated with the detector's than the self-check's are, and we expect this gap to predict where the full system's accuracy gain appears across SNR. If the correlations are similar, the central claim fails.

The second check is the SNR sweep under held-out mixtures. If the matched-test argument is right, the system's advantage over MLSCA-CW should widen as SNR falls toward −12 dB.

The third check compares the three planners at equal re-test budgets. If the language-model planner does not beat the learned policy, we will report that the decomposition, not language reasoning, carries the result.

Finally, we will compare escalation rates with latency. This shows whether the expensive decomposition tools stay confined to the small fraction of windows the auditor rejects, which is what makes edge deployment realistic.