Garbage In, Garbage Learned: How Aliasing in Time-Series Data Is Systematically Misleading Production Machine Learning Systems

There is a quiet assumption embedded in nearly every machine learning pipeline that processes time-series data: that the data arriving at the model's input layer faithfully represents the physical phenomenon it was collected to describe. That assumption is frequently wrong. And unlike a software bug that produces an obvious crash, aliasing errors in training data produce something far more insidious—a model that learns with confidence from a corrupted signal, then deploys into production with no idea what it is actually measuring.

The problem is not new to signal engineers. Shannon's sampling theorem established the mathematical boundary decades ago: to accurately represent a signal, you must sample it at a rate exceeding twice its highest frequency component. Violate that boundary, and frequency content folds back into the sampled spectrum, masquerading as lower-frequency energy that was never present in the original signal. What is new is the scale at which this classical failure mode is now propagating through data science workflows, carried upstream by sensor networks and SCADA systems that were designed by instrumentation engineers and inherited, uncritically, by ML teams.

The Inheritance Problem in Industrial ML

In a typical industrial predictive maintenance deployment, vibration sensors on rotating machinery may be sampled at rates set years or decades ago—often at whatever rate the original data historian was configured to support. Those rates were selected for trend visualization, not spectral fidelity. A motor bearing developing a defect may produce characteristic fault frequencies in the range of several hundred hertz. If the historian is logging data at two samples per second, those fault signatures are not merely absent from the dataset—they are aliased down into the low-frequency band where they interfere with legitimate process trends.

When a data scientist builds an anomaly detection model on that historian data, the training corpus contains these folded artifacts. The model learns them. It may even achieve impressive cross-validation scores, because the aliasing is consistent and repeatable across the historical record. The problem surfaces only when the model misses a developing fault—or, equally damaging, when it flags healthy equipment as anomalous because a process change altered the aliasing pattern without altering the underlying physics.

This is not a hypothetical scenario. Field engineers at several large US manufacturing facilities have reported exactly this failure mode following post-incident investigations, where ML-based monitoring systems failed to flag bearing failures that were clearly visible in high-rate vibration data collected by portable analyzers during the same operational window.

Frequency Folding and What Neural Networks Actually Learn

Understanding why neural networks are particularly vulnerable to aliasing requires stepping back from the architecture and thinking about what these models are actually optimizing against. A recurrent network or transformer trained on time-series data is learning temporal patterns—correlations between values at different time lags. When aliasing is present, those correlations are partially synthetic. The network cannot distinguish between a real low-frequency oscillation and an aliased high-frequency artifact; both appear as structured variation in the input sequence.

Convolutional architectures applied to spectrograms face a related problem. If the spectrogram is computed from undersampled data, the frequency axis of the spectrogram is itself corrupted. Features the convolutional filters learn to detect may correspond to aliased content that shifts position as operating conditions change, producing models with poor generalization despite strong in-sample performance.

The deeper issue is that aliasing is not random noise—it is structured interference. Random noise averages out over large training sets. Aliasing does not. It is a deterministic function of the signal content and the sampling rate, which means it is perfectly reproducible and will be faithfully learned by any model with sufficient capacity.

Diagnosing Sampling Adequacy Before You Train

The practical intervention is straightforward in principle, though it requires organizational discipline to execute consistently. Before any time-series dataset is admitted to a training pipeline, the following checks should be treated as mandatory:

Establish the signal bandwidth of the target phenomenon. For rotating machinery, this means consulting the fault frequency calculations for the specific equipment—bearing defect frequencies, gear mesh frequencies, imbalance harmonics. For electrical systems, it means identifying the relevant harmonic content. The maximum frequency of interest must be documented before the sampling rate can be evaluated.

Verify anti-aliasing filter presence and cutoff. Raw sensor data from many industrial systems passes through no anti-aliasing filter whatsoever before digitization. If the data acquisition hardware does not include an analog low-pass filter with a cutoff below the Nyquist frequency, aliasing is not a possibility—it is a certainty for any signal content above that boundary.

Cross-validate with a higher-rate reference. Where possible, collect a parallel dataset using a calibrated high-rate acquisition system and compare the spectral content. Significant energy appearing in the high-rate dataset above the Nyquist frequency of the primary logger is direct evidence of aliasing contamination in the primary record.

Examine autocorrelation structure for artifacts. Aliased content introduces characteristic periodicity into the autocorrelation function of the sampled signal. A data scientist familiar with the expected physics of the system can often identify suspicious periodicities that have no plausible physical explanation—a signal that the aliasing problem itself generated.

The Organizational Gap Enabling the Epidemic

Perhaps the most consequential observation is that aliasing in ML pipelines is fundamentally an organizational problem as much as a technical one. Signal integrity and sampling theory are core competencies in RF engineering, instrumentation, and control systems disciplines. They are not, historically, part of the standard curriculum for data scientists or ML engineers.

When a data science team is handed a CSV export from a plant historian and asked to build a predictive model, no one in that workflow may have the background to ask whether the sampling rate is adequate for the target application. The instrumentation engineer who configured the historian may have retired. The data scientist may never have encountered the Nyquist criterion outside a footnote in a statistics textbook.

The solution requires deliberate cross-disciplinary collaboration at the data governance layer—before model development begins. Organizations deploying ML on time-series data from physical systems should require a sampling adequacy review, conducted by an engineer with signal processing background, as a prerequisite for dataset certification. This is not bureaucratic overhead. It is the minimum due diligence required to ensure that the model being trained corresponds to physical reality rather than a distorted shadow of it.

Raising the Standard

The machine learning community has invested enormous effort in model architectures, regularization techniques, and training methodologies. The implicit assumption underlying all of that investment is that the data being fed into those architectures is a faithful representation of the world. In time-series applications drawn from physical systems, that assumption must be earned through rigorous signal validation—not inherited uncritically from whoever configured the data logger.

Aliasing does not announce itself. It does not produce obvious artifacts or error messages. It simply teaches your model the wrong things, quietly and persistently, until a bearing fails on a Monday morning and the anomaly detection system reports everything nominal. At that point, the question is not whether the model was sophisticated enough. The question is whether anyone checked the sampling rate.