Predictive Maintenance for Rotating Machinery: Vibration and Current-Signature Analysis at the Edge

What predictive maintenance actually is

Predictive maintenance (PdM) is the discipline of detecting the early stages of a mechanical failure in time to plan a repair before the failure produces an unplanned outage. For rotating machinery — motors, pumps, fans, gearboxes, generators — the failure modes are well understood and the underlying physics is observable through vibration and electrical signature analysis decades before catastrophic failure.

The economics are compelling. A planned bearing replacement is a one-hour job at $300 of labor. An unplanned bearing seizure that damages the shaft, the housing, and the connected pump is a multi-day repair at $30,000 of labor and parts, plus the lost throughput of whatever the pump was supporting. For sustainment-focused operations, the latter scenario can also create operational consequences (no power, no water, no fuel) that dwarf the direct repair cost.

The fault modes

Four failure modes dominate rotating-machinery diagnostics in field-deployed systems:

• Unbalance. A rotor whose mass distribution is asymmetric around its axis of rotation. Produces a vibration peak at 1× the rotation frequency, in the radial direction.
• Misalignment. A coupling between two shafts (motor to pump, typically) that is not perfectly collinear. Produces vibration at 1× and 2× rotation frequency with characteristic axial content.
• Bearing defects. A rolling-element bearing developing a flaw on inner race, outer race, ball, or cage. Each defect type produces vibration at a characteristic non-integer multiple of rotation frequency, dependent on bearing geometry.
• Electrical asymmetry in induction motors. Broken rotor bars, stator winding shorts, or air-gap eccentricity produce sidebands in the motor current spectrum around the line frequency at specific slip-related offsets.

The first three are detected primarily through accelerometer-based vibration analysis. The fourth is detected through motor current signature analysis (MCSA) — sampling the motor's input current and looking for spectral features.

Hardware architecture

The edge PdM unit is structurally simple:

• Multi-channel accelerometers, typically 4–8 channels for a machine train, with each axis (radial, axial) on its own channel.
• Current transformers on the motor's three phases, conditioning to feed an ADC.
• A multi-channel synchronous ADC. Sampling rate of 50 kSPS is sufficient for most bearing-defect frequencies; some advanced applications need 100 kSPS.
• A processor with enough horsepower to run FFTs and envelope detection in real time. A modern ARM Cortex-A53-class processor handles this comfortably. For high-channel-count systems (16+), a DSP or FPGA assist is warranted.
• Storage for a baseline reference per machine. A few hundred KB of features per machine, updated periodically as the machine ages naturally.

The architectural decision worth flagging is whether the unit is local-only (decisions made on the machine) or networked (raw data shipped to a central analytics service):

• Local-only for sites with intermittent or no comms — forward bases, remote depots, contested operations.
• Networked for sites with reliable comms — fixed installations, garrison maintenance bays.
• Hybrid is best in practice: local detection runs always, with alerting and routine state reporting over comms when available.

Signal processing chain

Accelerometer ──→ ADC ──→ HP filter ──→ Time-domain features
                                        (RMS, peak, crest factor)
                                ↓
                          Frequency-domain analysis
                                        ↓
                          ┌─── 1×, 2× rotation: unbalance, misalignment
                          │
                          ├─── Bearing fault frequencies (BPFO, BPFI, BSF, FTF)
                          │
                          └─── Envelope demodulation for early-stage bearing faults

The four bearing fault frequencies (Ball Pass Frequency Outer race, Ball Pass Frequency Inner race, Ball Spin Frequency, Fundamental Train Frequency) depend on bearing geometry — number of rolling elements, pitch diameter, contact angle — and rotation speed. These are computed at commissioning and stored as machine-specific configuration.

Envelope demodulation is the technique that catches bearing defects months before they show up in conventional FFT analysis. A nascent bearing defect produces a low-amplitude periodic impact at the bearing fault frequency, which excites the bearing-housing resonance (typically several kHz). The high-frequency band-pass filtering and rectification ("envelope") extracts the impact periodicity from the resonant carrier.

Why MCSA is worth the integration

Motor current signature analysis (MCSA) requires no sensors on the rotating machine — just current transformers in the motor control center, which are often already installed for protection. This makes MCSA particularly valuable for:

• Machines that are remote, sealed, or hazardous to instrument with accelerometers.
• Distributed motor populations where the cost of accelerometer installation is prohibitive.
• Pump and fan applications where the rotating element is inside a wet or pressurized housing.

The diagnostic resolution of MCSA is generally lower than vibration analysis (you find faults later, when they're worse), but the coverage is much broader.

Sustainment relevance

A sustainment platform that ingests PdM signals across a vehicle fleet, a depot's pump population, or a forward base's power-distribution system gets:

• Asset readiness scoring at the population level. The platform can rank assets by predicted time-to-failure and schedule maintenance to minimize operational risk.
• Logistics pre-positioning. When a bearing is detected as degrading, the platform can pre-position replacement parts at the closest supply node, reducing downtime when the planned repair happens.
• Cross-asset correlation. A pattern of bearing failures across one class of equipment that doesn't appear in another can flag a supply-chain issue (bad batch of bearings, contaminated lubricant) that no individual asset would have surfaced.

The on-device PdM is the data source. The cross-asset analytics and the integration into the wider sustainment decision graph is the platform's role.