Predictive maintenance promises to see equipment failure before it happens. On a Penang semiconductor line, that promise lives or dies not on the algorithm but on the calibration of the sensor feeding it. Here is the measurement foundation most PdM programmes skip.
Predictive maintenance in a semiconductor fab is, at its core, a forecasting problem: it watches vacuum pumps, CMP pads and etch chambers and tries to flag an incipient fault before the tool goes down. Critical Manufacturing, which builds MES and manufacturing-analytics software for the industry, frames the hard part plainly — the biggest barrier to fab-wide predictive maintenance is not model sophistication but insufficient data: trustworthy, structured signals from the equipment, the environment and the process. Their read on McKinsey's 'PdM 4.0' is that it is an asset-wide analytics system built to inform trained operators on how to respond to predicted failures, not to replace them.
That is the first uncomfortable truth. A prediction is only as good as the numbers behind it. If a vibration or temperature channel has drifted out of calibration, the model is learning from a corrupted signal — and it will confidently report the wrong answer.
Penang is not an abstract case. SEMI, the global semiconductor industry association, notes the state commanded roughly 5% of global semiconductor exports as of 2019 and has long been described as the 'Silicon Valley of the East'. Invest Penang puts the local electrical-and-electronics ecosystem at more than 6,500 suppliers, with the state recording RM358.1 billion in exports in 2024; Malaysia as a whole is the world's 6th-largest semiconductor exporter and holds about 13% of the global Assembly, Testing and Packaging (ATP) market — the majority of it from Penang.
Density cuts both ways. When a tool stops in a cluster this interconnected, the blast radius runs through multiple suppliers and re-entrant process loops. The natural response is to instrument everything — which multiplies the number of sensors on the floor, which in turn multiplies the calibration burden. More sensors means more drift risk, not less, unless someone owns the measurement discipline.
The international reference for this discipline is ISO 17359:2018, the umbrella standard for condition monitoring and diagnostics of machines. It lays out the general procedure for standing up a monitoring programme — equipment audit, failure-mode analysis, alarm-criteria setting, data acquisition, diagnosis and prognosis — and deliberately points to companion standards for each measurement technique: vibration (ISO 13373-1), thermography (ISO 18434) and tribology (ISO 14830), among others.
The ordering matters. Condition monitoring is a measurement discipline first and an analytics discipline second. A predictive-maintenance programme that bolts machine learning onto uncalibrated, undocumented signals is building on sand; the standard exists precisely to make the measurement step deliberate and auditable.
Sensors drift. Thermocouples, vibration accelerometers and pressure transducers all lose accuracy between calibrations, and the loss is rarely dramatic enough to trip an alarm. An illustrative example: a vibration channel that has drifted 8% still reports numbers that look plausible, but the envelope the model treats as 'normal' has quietly shifted — so a real incipient fault can sit inside the new normal band and be missed entirely. The remedy is not a better algorithm. It is a disciplined calibration and recalibration cadence with traceable records, so the model is reasoning about the real machine and not about a slowly migrating measurement.
This is the part no demo shows, and it is the part that decides whether the dashboard earns trust on the third shift. Without it, predictive maintenance becomes a source of false alarms that operators learn to ignore.
Our habit is to start where value is already being lost — inconsistent measurements, unexplained alarms, slow recovery after a tool event — and unify the operational context around that one workflow: connect OT data, maintenance history, calibration records and shift logs so a question about asset health can actually be answered with numbers you trust. We set the guardrails early: which actions an agent may recommend, which it may execute, and who approves the rest.
That is the whole thesis behind the name. Growing intelligence on the floor means first making the measurement foundation honest; forging it into something that stays behind means the calibration and diagnostic discipline is what makes the prediction usable, shift after shift. Predictive maintenance does not fail because the model is weak. It fails because the measurement underneath it was never owned.
VForge Field Notes are written plainly from real diagnostic engagements and cited public sources. Figures shown here are illustrative of method, not client results. Sources are linked for verification.
