Definition
MEMS (Micro-Electro-Mechanical Systems) integrates mechanical structures and electronic circuits on silicon-based chips, realizing accelerometers, gyroscopes, pressure sensors, etc. Vibration sensors use MEMS or IEPE accelerometers to measure 0.5Hz～10kHz vibrations, supporting edge FFT analysis, suitable for predictive maintenance of rotating machinery.
Sensor Types
MEMS Accelerometer: measures vibration, shock, tilt, outputs analog (voltage) or digital (I2C/SPI); MEMS Gyroscope: measures angular velocity, used for attitude and navigation; MEMS Pressure Sensor: measures air pressure and liquid level; MEMS Microphone: acoustic sensing. Compared with IEPE, MEMS is smaller in size, lower in cost, lower in power consumption, easy to integrate, typically has a frequency response of 0～several kHz, suitable for embedded, high-volume, low-to-medium frequency vibration monitoring.
Use Cases
Rotating machinery vibration monitoring, early warning of bearing faults; structural health monitoring; wearable devices, posture detection; consumer electronics, automotive electronics. Edge gateways integrate MEMS accelerometers and can locally run FFT and envelope analysis. After extracting feature values, the data is reported via MQTT and Modbus, reducing raw data uploads and supporting predictive maintenance platforms.
Range, Bandwidth, and the Mount
A silicon structure turns acceleration into capacitance
A micro-electromechanical sensor builds a moving proof mass and fixed electrodes on a silicon die. The mass displaces with acceleration, the gap changes, and capacitance changes with it. Later circuitry turns that capacitance into a voltage or a digital value. A gyroscope uses a vibrating structure to sense the Coriolis effect and outputs angular rate. A pressure die uses diaphragm deflection to change a cavity capacitance or a piezoresistive element. Semiconductor processing makes these structures in volume, so size, power, and unit cost suit building them into a sensor module. Stiffness, damping, and package stress of the moving part are written directly into the measurement error.
The digital interface is commonly I2C or SPI, and the chip has often already sampled, filtered, and switched range. An analog output is a voltage proportional to acceleration. If the next stage must drive a 4–20 mA transmitter, that driver is extra circuitry, and this stage limits bandwidth again. Noise density, zero-temperature drift, and cross-axis sensitivity decide whether the part can resolve an early, small bearing fault. The "maximum range" on the first page of the data sheet does not.
Suited to mid- and low-frequency rotating machines
It fits machines that do not spin very fast, where the frequencies of interest run from tens of hertz to a few kilohertz, where there are many points, and where you want features computed on the spot: trend monitoring of fans, pumps, and ordinary motors. Gear mesh at high frequency, and faults whose shock pulse is very steep, put their energy higher up. The silicon resonance and the package attenuate those components. Use a piezoelectric measurement chain for that. Attitude sensing cares about the gravity component and slowly changing angular rate. Bandwidth can be kept narrow, but it must not share the same filter settings as vibration monitoring.
The mounting face has to carry mechanical vibration faithfully into the proof mass. Thick paint, soft adhesive on a thin shell, or a magnet stuck to a cover that is already resonating measures the resonance of the mount, not the bearing envelope. Cable pull and temperature cycling also walk the zero slowly. For a spectrum at the edge, the sample rate has to cover more than twice the frequency you care about, and energy outside the chip's analog bandwidth has to be filtered off. Otherwise high frequency folds back and disguises itself as a low-frequency fault.
Mount resonance pulls the spectrum off
The first easy mistake is to look only at the range figure. A ±16 g range can survive a shock, but for a small vibration the useful bits are buried in noise and the trend looks like a flat line. The second is to ignore the axis. A single-axis chip stuck the wrong way never sees horizontal vibration. If a three-axis chip has its axis order reversed in software, the characteristic frequencies of the envelope spectrum are still there, but every amplitude is assigned to the wrong axis.
The third is to upload the raw waveform without limit. The value of a micro-electromechanical design is to extract peak, RMS, and a few characteristic frequencies locally, then report those. Streaming the whole waveform at a high sample rate makes radio power and data volume exceed the sensing itself immediately. Finally, check whether the output is a proportional voltage or an already calibrated engineering unit. If the next stage treats 0–3.3 V as acceleration directly, or reverses the two ends of a 4–20 mA loop, every threshold alarm points at the wrong severity.
Temperature drift and supply noise rewrite the reading as well. The same chip walks its zero as a cabinet warms up. A threshold calibrated cold false-alarms by day and misses by night. If the pull-up and bus capacitance on a digital interface are too large, the sample clock is stretched and the spectrum shows fixed spikes that have nothing to do with vibration. Recording supply, mounting torque, and temperature range in the point file holds a long-term trend more steadily than simply swapping in a higher-range part.