An Industrial AI Sample Stream Can Hide Two Different Input Frequencies

Establish what each delivered row represents

An industrial AI dataset may arrive as timestamps and amplitudes. To decide whether it supports a frequency-based task, establish the relationship between those values and the continuous input. A continuous signal x(t) has a value throughout time. Uniform sampling retains its values at particular instants: x[n] = x(n/fs), where n is an integer sample index and fs is the sampling rate. The rate supplies the time spacing; the sequence supplies the retained amplitudes.

NI's digitizer explanation separates the analog input path, which conditions the waveform, from the ADC's conversion cadence. The bandwidth statement concerns the input path's frequency response. It does not independently establish how often the conditioned signal is sampled.

Also distinguish acquisition cadence from export cadence. A claim of 1,000 delivered rows per second leaves open whether rows contain original samples, repeated values or processed summaries. Obtain the provider's description of the underlying per-channel sampling and any transformations before applying a sampling interpretation to an export. The example below explicitly supplies one original sample per row, with known uniform timing and no intervening processing.

References: [NI1]

AD-42 starts with two different continuous signals

AD-42 is an original mathematical example with two alternative inputs. They are separate possible recordings, not two tones added together. Let t be time in seconds. The inputs are x150(t) = cos(2π × 150 × t) and x850(t) = cos(2π × 850 × t). Each has unit peak amplitude, zero mean and zero phase relative to t = 0. Their amplitudes are dimensionless normalized values, not measured volts or vibration units.

Stipulate an ideal path with unity gain and no phase shift for either candidate, followed by instantaneous samples at fs = 1,000 samples/s. The sample instants are t = n/1,000 seconds, separated by 1 ms. There is no noise, timing jitter, quantization, distortion, missing sample or later transformation. These assumptions isolate the sampling mechanism from equipment performance.

The continuous inputs differ. At t = 0.0005 seconds, halfway between the first two sample instants, x150(t) is approximately 0.891006524 while x850(t) is approximately -0.891006524. A record containing only the stipulated sample instants has no observation at that intervening time. The question is what its retained values can distinguish.

The cosine identity makes every retained value equal

At sample index n, the 850 Hz argument can be written as 2π × 850 × n/1,000 = 2πn - 2π × 150 × n/1,000. For integer n, removing 2πn does not change a cosine. Cosine is also even, so cos(-θ) = cos(θ).

Consequently, cos(2π × 850 × n/1,000) = cos(2πn - 2π × 150 × n/1,000) = cos(2π × 150 × n/1,000). This is an identity at every integer n. It is stronger than a visual resemblance between two short plots.

The first five calculated sample pairs are below, rounded to nine decimal places. For example, at n = 1 the arguments are 0.3π and 1.7π radians; both cosines give 0.587785252 after rounding. The equality comes from the identity, rather than from rounding the displayed values.

These five rows demonstrate the calculation. They do not establish that five samples determine an arbitrary continuous waveform. The all-integer identity establishes that these two particular alternatives remain indistinguishable even if the ideal sequence continues indefinitely.

An apparent 150 Hz component leaves its physical origin open

For AD-42, the two frequencies differ in their continuous behavior while sharing the same discrete behavior. The sequence can be written as cos(0.3πn). With the supplied time spacing, that admits a 150 Hz baseband interpretation. The alternative 850 Hz input supplies exactly that sequence too: 1,000 Hz - 850 Hz = 150 Hz.

Analog Devices' MT-002, Rev. A, October 2008, explains aliasing and the first Nyquist zone. At AD-42's rate, fs/2 = 500 Hz. Interpreting the input as baseband needs a supported restriction of admitted content below that boundary.

Finding a 150 Hz component in these samples therefore cannot alone establish that the original input was 150 Hz. Nor does it prove that an unwanted 850 Hz tone existed. Both alternatives are compatible with this record. The useful conclusion concerns the information available, not a diagnosis of the source.

Retaining more samples at the same instants extends the same ambiguous sequence. More precise amplitude representation also leaves the ideal equality intact. Additional acquisition information could change the set of admissible inputs, but neither a longer copy of this stream nor additional decimal places supplies that information.

References: [AD1]

Two offers can supply the same cadence with different frequency meaning

Suppose two conceptual offers each supply one raw, uniformly timed value per channel every 1 ms. Compare them for a task that must distinguish AD-42's two possible original tones. Keep the example's ideal response assumptions wherever a tone is admitted.

Offer A declares that content reaching the sampler is confined to 0-400 Hz and that content outside that range is excluded in this ideal comparison. Offer B permits content anywhere from 0-900 Hz. Those round limits are invented information assumptions, not recommended hardware specifications or filter choices.

For the two stipulated candidates, A's declared restriction rules out 850 Hz and leaves 150 Hz admissible. B admits both. B can return the same sample values for either candidate, so its wider analog admission does not fulfill this particular distinction. Its equal row cadence establishes no additional ability to identify which original tone occurred.

The comparison does not establish that A is generally better. A would exclude an 850 Hz signal that another task might require. B's broader admission could be useful under a different, adequately specified interpretation. The buyer's decision depends on the physical information required and the basis for interpreting the delivered values.

In an actual proposal, a bandwidth headline alone is insufficient evidence for either offer's declared restriction. Ask the acquisition owner to identify the admitted signal bands, the basis for controlling competing content, the actual channel cadence and the transformations before delivery. A supported response description matters more than treating a bandwidth number as a perfect pass-or-stop boundary.

A known higher input band can support intentional bandpass sampling

MT-002 also describes intentional bandpass sampling: a known higher input band can map into the first zone without overlapping competing spectral content. Interpretation retains the original band location and any spectral reversal. The analog path must support the original frequencies. This is why half the sampling rate is not a universal upper limit on useful analog input frequency.

For a separate ideal illustration, stipulate that only positive-frequency content in the 800-900 Hz band is admitted at 1,000 samples/s. The mapping f to 1,000 - f places 800 Hz at 200 Hz and 900 Hz at 100 Hz, reversing their order. Knowing the stipulated original band distinguishes that interpretation from a physical 100-200 Hz input.

This is different information from Offer B's broad permission for content anywhere up to 900 Hz. Merely allowing a higher band does not establish that it is the only band present. A narrow bandwidth figure alone also leaves spectral placement and overlap unresolved. A real bandpass proposal needs its complete acquisition basis.

References: [AD1]

Judge AI input validity against the distinction the task needs

Suppose a proposed industrial AI use requires different conclusions for the two original AD-42 tones. If the only input is the identical sampled sequence, it provides no observation that separates those alternatives. A deterministic feature calculation receives the same values and therefore produces the same features. A model may prefer one interpretation using training assumptions or additional context, but that preference is not a frequency distinction established by this stream alone.

A task concerned only with a pattern in the delivered values might have a different requirement. Aliasing does not by itself establish that every downstream use is invalid. Make the required physical distinction explicit before deciding whether the provider's acquisition evidence supports it. AD-42 supplies no machine-condition label, fault diagnosis or prediction-performance result.

For the proposed feed, retain the channel identity, amplitude meaning, sampling-time basis, admitted-band explanation and any processing that connects acquisition to the supplied values. Identify which evidence excludes competing physical interpretations and which questions remain unresolved. That gives the appointed acquisition and AI owners a concrete input-validity decision to resolve before treating more compute as the answer.

The related OPC UA article addresses a different loss: changes never observed or Notifications not retained. Here every stipulated sample survives and ambiguity remains. The measurement article separates resolution, uncertainty and acceptance rules; the speech article follows already digitized audio into feature arrays. Those explanations complement the acquisition decision without settling it.

Bring a separate physical connection requirement to the cabling scope

If the selected acquisition or compute equipment also needs physical network connections, Leviathan's Structured Cabling & Fiber Installation service provides a way to discuss agreed fiber and copper routing, connections, labels and handoff records. Use the service and project-inquiry links below with a nonconfidential equipment list, approved endpoints and routes, material responsibilities and required installation records.

That physical engagement has its own confirmed scope and availability. Signal-acquisition design, analog treatment, industrial automation, AI software and operational decisions remain with the appointed specialist teams. This article supports an educational comparison of supplied information; it authorizes no field procedure or equipment change.

References: [L1]