A data acquisition (DAQ) system is the hardware and software that turns real-world physical signals, temperature, pressure, vibration, flow, voltage, into digital numbers a computer can store and analyze. In one sentence: it senses a signal, conditions it, converts it from analog to digital, and hands the result to software. Everything a plant later does with data starts here.
DAQ is the unglamorous plumbing beneath every dashboard, every OEE number, and every predictive-maintenance model. If the acquisition is wrong, undersampled, poorly conditioned, badly ranged, no amount of clever analytics downstream can fix it. Garbage in stays garbage. So it is worth understanding what actually happens between a sensor on a machine and a clean number in a database.
What does a DAQ system actually do?
It runs a signal down a short chain, and each link can make or break the measurement. A sensor turns a physical quantity into an electrical signal. Signal conditioning cleans and scales that signal. An analog-to-digital converter (ADC) samples it and turns each sample into a number. Software stores, displays, or analyzes those numbers. Miss any link and the chain fails: a perfect sensor feeding a poorly conditioned input produces noise; a well-conditioned signal sampled too slowly produces a lie.
The important boundary in that chain is the converter. Everything to its left is analog and vulnerable, susceptible to noise, ground loops, temperature, and cable length. Everything to its right is digital and durable, a number is a number, copy it a thousand times and it does not degrade. Good DAQ design pushes the conversion as close to the sensor as practical, so the fragile analog run is short and only clean digital data travels the distance to where it is stored and used.
What is signal conditioning?
Signal conditioning is the electronics between the sensor and the ADC that make a raw signal fit to digitize. Raw sensor output is often too small, too noisy, or too dangerous to feed straight into a converter. Conditioning fixes that with a handful of jobs: amplification scales a tiny signal up to use the converter's full range; filtering removes noise and, critically, blocks high frequencies before sampling (an anti-aliasing filter); isolation protects the system and breaks ground loops so a fault on the machine cannot fry the acquisition hardware; and excitation supplies power to sensors that need it, like strain gauges and RTDs. Skimp on conditioning and you digitize noise with great precision, a common and expensive mistake.
How does sampling work, and what is the Nyquist limit?
An ADC does not capture a signal continuously; it takes snapshots, samples, at a fixed rate. The single most important rule in acquisition governs how fast those snapshots must come. The Nyquist-Shannon sampling theorem states that to faithfully capture a signal you must sample at more than twice its highest frequency. Sample too slowly and you get aliasing: a fast signal masquerades as a slow one that was never there, and the false reading is indistinguishable from real data after the fact. In practice, engineers sample well above the bare minimum, often five to ten times the highest frequency of interest, to leave room for real filters, which never cut off perfectly sharply.
Aliasing matters because it is silent. A clipped signal looks obviously wrong; an aliased one looks perfectly reasonable. Picture a vibration sensor watching a motor spinning at high speed. If the acquisition samples too slowly, the recorded waveform can show a smooth, gentle oscillation that suggests everything is fine, while the real vibration telling you a bearing is failing sat at a frequency the system never sampled fast enough to see. The data looks clean, the trend looks stable, and the machine fails anyway. This is why the anti-aliasing filter belongs before the converter, not after: once a fast component has folded down into the recorded band, no software can separate it from real low-frequency data.
What forms do DAQ systems take?
The same signal chain shows up in very different packages. A benchtop or handheld instrument bundles everything for a single measurement, a vibration meter, a data logger, and is ideal for spot checks. A modular chassis holds swappable input cards so one system can read dozens or hundreds of mixed channels, common in test cells and process skids. A PC-based system puts an acquisition card or USB device next to a computer for heavy analysis and display. And a distributed system spreads acquisition nodes across a plant, digitizing near each machine and sending numbers over the network, which is the pattern most IIoT retrofits follow, because it keeps sensitive analog runs short and moves only clean digital data over distance. The right form depends on channel count, distance, and how much processing happens at the edge versus the center.
What determines the quality of the acquired data?
Four numbers, mostly. Resolution is how finely the ADC divides its input range, set by its bit depth: a 16-bit converter splits the range into roughly 65,000 levels, a 24-bit converter into over 16 million, so more bits means finer detail. Sample rate is how many snapshots per second, governed by Nyquist as above. Range is the span of input the converter expects; set it wrong and you either clip the signal or waste most of your resolution on empty headroom. Channel count is how many separate signals the system reads at once, each with its own wiring, conditioning, and possibly its own converter.
These numbers trade against each other and against cost. Very high sample rates across many channels demand fast converters and fast storage; very high resolution demands quiet, well-conditioned inputs or the extra bits just capture noise. The discipline is to match each specification to the physics you are actually measuring rather than buying the biggest number on the datasheet. A slow-drifting tank temperature needs high resolution and almost no speed; a gearbox vibration signature needs high speed and moderate resolution. Buying the wrong balance is how plants end up with expensive systems that still miss the event that mattered.
How do you specify a DAQ system?
Speccing acquisition is a sequence of decisions that flow from the signal itself, not from a catalog. Work them in order and the hardware chooses itself.
- Characterize the signal. What physical quantity, what amplitude range, and, crucially, what is its highest meaningful frequency? Everything downstream depends on this answer.
- Set the input range. Match the converter's range to the conditioned signal's span so you use most of the resolution without clipping the peaks.
- Choose resolution. Pick bit depth from how small a change you must detect; measuring slow temperature drift needs different resolution than catching a bearing's high-frequency vibration.
- Set the sample rate. Above twice the highest frequency at an absolute minimum, and realistically several times that, with an anti-aliasing filter ahead of the converter.
- Count the channels and place the conditioning. Decide how many signals, whether they sample simultaneously or in turn, and what conditioning each one needs before it reaches the converter.
By the Numbers
The governing facts are settled engineering, not opinion. The Nyquist-Shannon theorem fixes the minimum sample rate at more than twice a signal's highest frequency; a 16-bit converter resolves about 65,536 levels and a 24-bit converter more than 16 million; and every measurement is only as trustworthy as its calibration traceability, which in the United States runs back to the national standards maintained by NIST. The broader business fact is that most plants already generate far more signal than they use, U.S. government adoption surveys consistently show data sitting unexploited, with the barrier being integration rather than sensing (U.S. Census Business Trends and Outlook Survey). Where Harmony fits: Harmony reads the signals your machines and PLCs already acquire, computes true OEE from those source signals rather than estimates, and turns them into searchable, actionable operational data, no new acquisition hardware required. See the connected systems module.
Where does acquired data go next?
Up the stack. Once a DAQ system produces clean numbers, they usually land in a historian and feed a SCADA layer for real-time supervision, then flow upward into machine monitoring OEE, and analytics. High-rate acquisition, vibration in particular, is the raw material for predictive maintenance where catching a bearing's failure signature depends entirely on having sampled fast enough to see it. The pattern that recurs across a modern plant is that the sensors and converters are already there; the missed opportunity is upstream data that stops at a local screen instead of being connected, contextualized, and acted on. For that bigger picture, see IIoT and smart factory technology.