Sideband Background Modeling for Label-Free Incipient Fault Detection

Machines worth monitoring rarely break, so a team that wants to detect early damage usually has thousands of hours of healthy vibration or process data and almost no labeled faults. Particle physicists face a structurally similar problem when they search for a rare signal buried in overwhelming background. One of their working answers is to never model the signal at all. They learn the background very precisely from nearby regions where the signal cannot be. They use that model to predict what pure background should look like in the region of interest. Then they train a classifier to tell the observed data apart from the predicted background. If the classifier can tell them apart, something extra is present. On CMS Open Data, this recipe (CATHODE) raised the significance of a hidden dimuon resonance from 1.6σ in the raw mass spectrum to 6.4σ, using no simulated or labeled signal (Isolating Unisolated Upsilons with Anomaly Detection in CMS Open Data).

This note proposes carrying that recipe into industrial condition monitoring. The physicist's "mass" variable becomes the machine's measured operating condition: shaft speed, load, temperature, or drive current. The "sidebands" become healthy history recorded at neighboring operating conditions. A conditional density model learns how healthy behavior changes with operating condition. It then generates synthetic healthy data matched to the conditions of a recent monitoring window. A small classifier trained to separate the real recent window from that synthetic healthy reference becomes a label-free fault detector. Its output also gives a population-level evidence statistic. The claim to be tested is specific. When faults are a small fraction of recent data and operating conditions have shifted, this approach beats both supervised few-shot classifiers and the obvious one-class alternative of thresholding a density model fit to healthy data.

Core Claim

Industrial few-sample methods train on fault labels. REDNet maps each segment to fitted reservoir readout weights and reaches over 90% F1 with 10 samples per class (Fault Diagnosis in REDNet Model Space). LSFConvformer fuses time and frequency features and reaches 98.37% with 20 samples per class under variable speed (LSFConvformer). SCCAM uses supervised contrastive learning and reaches 72.16% on the long-tail Tennessee Eastman task (SCCAM). DAC-FCF tries to manufacture more fault labels with a conditional GAN, and its own authors report instability below roughly 10–20 samples per class (An Advanced Convolutional Neural Network for Bearing Fault Diagnosis under Limited Data). All four of these papers motivate the proposal: each spends its effort on the scarce class.

The proposal spends its effort on the abundant class instead. It differs from CATHODE in three ways. First, it applies the method to continuous sensor streams rather than collision kinematics. Second, the conditioning variable is a physically measured operating condition rather than an invariant mass. Third, the "signal region" is a time window at given operating conditions rather than a mass window. The third difference has a consequence: the method answers a population question ("does this window contain more non-healthy events than its operating conditions explain?") alongside a per-event score.

The theoretical reason to expect an advantage over plain density thresholding is the following. Let $x$ be the feature vector of one event, such as one impact transient or one process snapshot. Let $m$ be the operating condition. Let $p_b(x\mid m)$ be the healthy density, $p_s(x\mid m)$ the fault density, and $\epsilon$ the unknown fault fraction in the window. Then the window's data density is $p_d = (1-\epsilon)p_b + \epsilon p_s$, and

$$\frac{p_d(x\mid m)}{p_b(x\mid m)} = 1 + \epsilon\left(\frac{p_s(x\mid m)}{p_b(x\mid m)} - 1\right).$$

This ratio increases with the fault-to-healthy likelihood ratio for any $\epsilon>0$. The Neyman–Pearson lemma says that likelihood ratio is the optimal test statistic for a fixed false-alarm rate. A classifier trained to separate data from a perfect background sample therefore learns the optimal discriminant, with no fault labels required. Thresholding $p_b$ alone does not have this property. It also flags healthy events that happen to fall in low-density regions, and those regions are exactly where load and speed changes put healthy data.

Research Program

The pipeline has three stages.

The first stage builds features. It borrows physics pulse-shape discrimination ideas so that the density model operates in a low-dimensional space where flows are reliable. Each detected impact transient is described by a few morphological quantities: rise time, peak-to-energy ratio, and the slope of the decaying tail. The tail-slope feature is motivated by the delayed-charge-recovery discriminator of Arnquist et al., which separates surface alpha events from bulk events by how slowly their waveform tails relax. The feature set also includes the 7-dimensional bottleneck of a 1D convolutional autoencoder trained only on healthy data, following Holl et al.. Holl et al. showed that such a bottleneck matches the classical $A/E$ discriminator and catches multi-site pulses it misses. For process data such as Tennessee Eastman, the features are short multivariate windows compressed the same way.

The second stage fits a normalizing flow — an invertible neural network that gives an exact, sampleable density — to $p_b(x\mid m)$ on healthy history. Data in the target window's operating range are excluded during fitting. The flow is then sampled at the observed $m$ values of the target window.

The third stage trains an ensemble of boosted decision trees, as in CATHODE, to separate real window events from synthetic ones. The classifier's held-out scores give per-event fault scores. A test on the classifier's output distribution gives a window-level significance.

The main technical risk is the one CATHODE names: sculpting. Imperfect interpolation, or a residual dependence of features on $m$ that the flow misses, can make the classifier "find" a difference that is only a modeling error. In industry this is the variable-speed false-alarm problem under a new name. We address it the way physicists do. Every result is paired with null controls: healthy windows run through the identical pipeline, the analog of CATHODE's same-sign dimuon check. A method that cannot hold its nominal false-alarm rate on nulls is reported as failed regardless of its detection power. We also measure how null false-alarm rates grow as the target window moves from interpolation (healthy data on both sides in $m$) to extrapolation (healthy data on one side only). The method is only claimed for the first regime.

Baselines

The naive reference point is envelope analysis at the characteristic bearing fault frequencies, with a threshold set on healthy data. It needs no learning at all and is what a plant engineer would deploy first. Beating it means a higher detection rate at matched null false-alarm rate, at fault fractions below a few percent.

The strongest competitor on the label-free question is a one-class detector built from the same parts. It uses the same features and the same conditional flow, trained on the same healthy data, but scores each event by its negative log-density under $p_b(x\mid m)$ and skips the classifier. This baseline isolates the claim itself. If it matches the full method, the classifier stage adds nothing and the idea reduces to standard density-based anomaly detection. A second one-class competitor is autoencoder reconstruction error. All one-class variants get the same tuning budget, and all thresholds are set on the same healthy calibration split. No reported result was found on Lacuna for flow-likelihood or autoencoder one-class detectors on CWRU, Tennessee Eastman, or CH-BatteryGen, so these numbers will be produced here.

The strongest supervised competitors are LSFConvformer and REDNet. They are given $k \in \{5, 10, 20\}$ labeled fault samples per class from the same machines. Their published numbers set the bar they will be held to: LSFConvformer reports 98.37% mean accuracy at 20 samples per class on its variable-speed bearing set (LSFConvformer), and REDNet reports over 90% F1 at 10 samples per class on CWRU and GFD (REDNet). On Tennessee Eastman, SCCAM's 72.16% long-tail accuracy is the bar (SCCAM). On CH-BatteryGen, the bars are the CNN's 0.9280 classification F1 and 0.8813 severity-grading F1 (Battery Fault). We do not expect a label-free method to beat these at their own task. The success criterion is that our detector, with zero labels, reaches the detection AUC that the supervised models reach at some $k>0$. We report the label count it is worth. It counts as a clear win when it outperforms them on fault types or operating conditions absent from their labeled set.

Two ablations isolate the key ingredients. Replacing the conditional flow with an unconditional one tests whether conditioning on operating condition is what controls false alarms. An oracle that trains the classifier against held-out real healthy data from the target operating conditions bounds how much performance the interpolation step loses.

Why not just…?

Why not just threshold the likelihood of a density model trained on healthy data? This is the real threat, and the proposal is designed around it. The argument above says a density threshold is not the optimal statistic when operating conditions vary, because rare-but-healthy events look anomalous. The data-versus-background classifier targets the ratio, which is. That argument holds only if the flow is accurate enough. If the one-class baseline matches the full method at matched null false-alarm rates across fault fractions and operating-condition shifts, we will report that the classifier stage is unnecessary. The ablation comparison settles this directly.

Why not just train a few-shot supervised classifier, since 10–20 labeled faults already give over 90%? REDNet and LSFConvformer show this works when labels exist for the specific fault type and conditions. The incipient-fault setting is different: the first faults have no labels, and novel fault modes are by definition unlabeled. DAC-FCF's own difficulties below 20 samples suggest label-hungry pipelines degrade there (DAC-FCF). The experiments test supervised models on held-out fault types and shifted loads, where we expect them to fail and our method to need nothing new.

Why not just use envelope analysis or stochastic resonance at known fault frequencies? These methods work for clean periodic faults at known speeds. The file's evidence suggests they struggle when background noise shares the fault band (Luo et al.), when load fluctuations mimic degradation (Li et al.), and when nonlinear amplifiers saturate or need retuning per noise level (Hu et al.). Envelope analysis is run as the naive baseline. Its frequency-band energies are also offered to our feature set, so if the frequency information alone suffices, that will show up.

Proposed Experiments

The first experiment injects faults into healthy streams at controlled rates. CWRU bearing data provide healthy and faulted recordings at several motor loads. Load serves as $m$. Each test uses a target load as the signal region, with the neighboring loads' healthy data as sidebands. Synthetic monitoring windows mix healthy events with fault events at fractions $\epsilon$ from 0.1% to 10%. We plot window-level significance and per-event AUC against $\epsilon$ for every method. Null false-alarm rates are measured on purely healthy windows. We repeat this on the variable-speed bearing set and the Ottawa set used by LSFConvformer.

The second experiment uses Tennessee Eastman for process data. Faults are injected at the 4% long-tail fraction used by SCCAM and at lower fractions. The plant's setpoint and disturbance-mode variables serve as $m$.

The third experiment tests whether the learned ratio tracks severity. It uses CH-BatteryGen, where drive current and charge/discharge state serve as $m$ and severity comes in mild, moderate, and severe grades. Because the classifier approximates a monotone function of the fault likelihood ratio, its mean window score should order severities without severity labels. We check the rank correlation against the CNN's 0.8813 supervised grading F1 as context.

The decisive check across all three is a single comparison: at fault fractions below a few percent, under shifted operating conditions, with null false-alarm rates held to nominal, does the full sideband-plus-classifier method beat the same flow used as a likelihood threshold? If yes, the physics recipe earns its place in industry. If not, we have a clean negative result. It would show that, for industrial streams, conditional density modeling on healthy data carries the benefit and the likelihood-ratio classifier adds nothing.