A TLA Innovation Lab product
Lifeline
Lifeline finds the abnormal — and how many hours are left.
Built for plants that cannot staff a reliability lab. Vibration, voltage, temperature, and the signals next to them become an early warning, then a remaining-useful-life estimate — running on a tiny MCU.
The hard part is not another cloud dashboard. The hard part is a model small enough to sit next to the motor, the PSU, or the board, still accurate enough to catch a fault before it becomes downtime.
Why it exists
See the failure coming. Keep the inference on the device.
You do not need a PdM data-science team
Most SMEs already have sensors. They do not have someone to train models on vibration and voltage traces. Lifeline is that loop: detect what is not normal, estimate remaining useful life, and tell the floor in time to act.
The model fits on a tiny MCU
Cloud PdM needs bandwidth and a rack. Lifeline is engineered for the board: compact architecture and aggressive quantization so inference runs on a microcontroller — even when the network is gone.
How it works
From raw signals to remaining life, on the chip.
01
Read the device
Lifeline takes the traces electronics already make — vibration, voltage, temperature, and related signals — and turns them into a picture of health.
02
Catch what is not normal
Models learn the healthy pattern of a unit, then flag the drift that operators would otherwise miss until it is too late.
03
Estimate remaining useful life
Anomaly is the warning. RUL is the plan: how long the device can still run, so maintenance is scheduled instead of improvised.
04
Fit the model on a tiny MCU
Architecture is designed for the microcontroller. Quantization shrinks the weights so the same judgment that lived in a lab model now runs next to the hardware it protects.
05
Act before the stop
The floor gets a signal while there is still time — inspect, swap, or order a spare — not after the line is already down.
What Lifeline does
Signals in. Health out. Inference on the MCU.
Multi-signal health
Vibration, voltage, temperature, and the channels next to them — combined, not read in isolation.
Anomaly detection
The model knows the healthy fingerprint of a device, then raises what does not belong.
Remaining useful life
Hours and cycles left, not only a red light — so you plan the stop instead of surviving it.
MCU deployment
Inference sits on a tiny microcontroller. No GPU rack required for the live decision.
Architecture for the edge
The network is designed to be small from the start — not a lab model squeezed as an afterthought.
Quantization
Weights are compressed so memory and compute fit the chip, without giving up the call that matters.
“The model that predicts the failure has to fit next to the motor, not in a rack.”
Put remaining life on the board.
If the device already shakes, draws current, or runs warm, you already have the raw material. Lifeline is how an SME without a reliability lab still sees the stop coming — and keeps inference on the MCU.