Ghost Frequencies in the Loop: How Aliasing Is Silently Destabilizing Industrial Control Systems
There is a particular category of control system instability that engineers describe with a mixture of frustration and bewilderment: the oscillation that appears on the plant floor but vanishes the moment a technician connects diagnostic equipment. It returns after the technician leaves. It defies every frequency-domain model. It survives controller retuning. In a surprising number of documented cases, the root cause is aliasing — not in the data acquisition system, but embedded deep inside the feedback sensor chain, operating below the threshold of conventional diagnostic visibility.
Aliasing, in this context, is not the coarse, obvious distortion that results from catastrophically undersampling a signal. It is something subtler and therefore more dangerous: the systematic folding of frequency content that sits just outside the control loop's intended bandwidth into the passband where the controller responds to it as if it were legitimate process information.
Why Nyquist Assumptions Break Down in Real Control Loops
The Nyquist-Shannon sampling theorem is unambiguous in its mathematical statement: to reconstruct a signal faithfully, the sampling rate must exceed twice the highest frequency component present in that signal. What the theorem does not address — because it is a mathematical idealization, not an engineering specification — is the practical difficulty of guaranteeing that no frequency content exists above the Nyquist limit in a live industrial environment.
Real sensor chains are not bandlimited in any strict sense. A pressure transmitter mounted on a hydraulic manifold will faithfully transduce mechanical vibrations from nearby pumps, pipe resonances, and solenoid switching transients. These components may extend well above the intended measurement bandwidth. If the anti-aliasing filter preceding the ADC is insufficiently steep, or if it is positioned after a signal conditioning stage that itself introduces nonlinear gain, frequency content above fs/2 will fold back into the sampled spectrum.
The situation is compounded by a feature of modern distributed control architectures: variable loop execution timing. In systems where the control task shares processor resources with communication handlers, logging routines, or safety monitors, the effective sampling interval is not constant. Jitter in the sampling epoch — even at the level of a few microseconds — is sufficient to generate spectral artifacts that a time-invariant Nyquist analysis will not predict.
Case Study: A Paper Mill's Unexplained Tension Oscillation
Consider a tension control system installed on a continuous web process at a paper manufacturing facility in the southeastern United States. The system used load cells to measure web tension and fed that measurement into a PID controller operating at a 100 Hz update rate. By design, the anti-aliasing filter had a cutoff frequency of 40 Hz — comfortably below the 50 Hz Nyquist limit.
Eighteen months after commissioning, operators began reporting a low-amplitude, 7 Hz oscillation in web tension that appeared intermittently during high-speed production runs. Frequency-domain analysis of the control error signal showed a clean spectrum with no energy at 7 Hz during normal operation. The oscillation was present in the physical process but absent in the sampled data — which should itself have been a diagnostic flag.
The eventual diagnosis required time-domain reconstruction of the raw load cell output at high temporal resolution, captured via a separate data acquisition system running at 10 kHz. This revealed a 93 Hz mechanical resonance in the web span — excited by a variable-speed drive operating near that frequency during high-throughput conditions. The 40 Hz anti-aliasing filter had insufficient attenuation at 93 Hz. With a 100 Hz sampling rate, 93 Hz aliases to 7 Hz. The controller was faithfully responding to a phantom signal, introducing energy at 7 Hz that was physically real even though its origin was a frequency the system was never designed to observe.
The Bode Plot's Blind Spot
Frequency-domain loop analysis — Bode plots, Nyquist diagrams, sensitivity functions — operates on the assumption that the system under analysis is linear, time-invariant, and properly sampled. Aliasing violates all three assumptions simultaneously. The aliased component appears at a frequency where the loop may have significant gain and phase margin, driving the controller to act on information that does not correspond to any real process state at that frequency.
This is why conventional stability analysis fails to detect the problem. The Bode plot of the controller and plant will show adequate margins at 7 Hz. It will show nothing at 93 Hz because the sampled data contains nothing at 93 Hz. The instability exists in a frequency space that the standard analytical tools cannot access.
A Systematic Methodology for Detection and Elimination
Addressing aliasing-induced instabilities in control systems requires a diagnostic approach that operates outside the sampled data stream. The following methodology has proven effective across several industrial applications.
Step 1: Characterize the physical sensor chain at high bandwidth. Before trusting the control system's own data, instrument the sensor output directly using a wideband acquisition system — ideally one capable of capturing to at least 10× the control loop's sampling frequency. This reveals frequency content that the anti-aliasing filter may be passing inadequately.
Step 2: Map the anti-aliasing filter's actual response. Datasheet filter specifications are measured under ideal conditions. In-circuit response may differ due to component tolerances, loading effects, and temperature. Measure the installed filter's attenuation at fs/2, fs, 2fs, and 3fs. Insufficient attenuation at any integer multiple of the sampling frequency is a potential aliasing source.
Step 3: Assess sampling epoch consistency. Log the actual timestamps of controller execution using a hardware timer with nanosecond resolution. Compute the distribution of inter-sample intervals. Standard deviation greater than 0.1% of the nominal period is sufficient to generate spectral spreading artifacts.
Step 4: Reconstruct the alias map. For any identified out-of-band frequency component at frequency f, compute its alias frequency as |f − round(f/fs) × fs|. If that alias frequency falls within the controller's response bandwidth, it is a candidate instability source.
Step 5: Apply targeted filtering or increase sampling rate. Once candidate aliasing sources are identified, the remediation options are straightforward: increase anti-aliasing filter order to achieve adequate stopband attenuation, shift the sampling rate to move the alias outside the controller bandwidth, or — if the mechanical source can be addressed — eliminate the excitation at the physical level.
The Time-Domain Imperative
The diagnostic methodology described above is fundamentally a time-domain exercise. It depends on capturing raw, high-bandwidth signal records and examining them for temporal structure that the sampled control system cannot see. This is not a workflow that frequency-domain analysis alone supports.
Control engineers who rely exclusively on Bode plots and spectrum analyzers connected to the control system's own data outputs are working with a filtered view of reality. That filter — the anti-aliasing filter and the sampling process itself — is precisely the element under investigation. You cannot diagnose a sampling artifact using the sampled data alone.
As industrial control systems migrate toward higher update rates, more complex sensor chains, and shared-resource computing architectures, the conditions that generate aliasing-induced instabilities are becoming more common, not less. The engineers who will identify and resolve these failures fastest are the ones who understand that the time domain, captured at sufficient bandwidth and temporal precision, is the only diagnostic space where ghost frequencies cannot hide.