Line Supervisor

Line Supervisor is an autonomous CNC milling cell supervisor that keeps the LLM out of the high-frequency loop until an anomaly requires cognitive reasoning and domain-grounded action. THE PROBLEM Industrial CNC machining cells stream telemetry at 1 Hz (vibration, spindle torque, temperatures, tool wear, scrap rates). Mechanical drift toward catastrophic failure—such as spindle bearing degradation or thermal dissipation failure—manifests as subtle, correlated shifts in signal trends well before hard threshold alarms trigger. Today, detecting this requires an expert operator watching screens or static thresholds that fire too late. Running an LLM directly on 1 Hz telemetry is too slow, too expensive, and unviable for real-time factory operations. WHAT WE BUILT A multi-tier architecture that decouples 1 Hz monitoring from reasoning: 1. Deterministic Baseline Gating: A lightweight statistical monitor samples the OPC UA telemetry stream, establishes a frozen 60-second baseline under healthy conditions, and computes z-scores on sliding windows. Zero LLM tokens are consumed during nominal operation. 2. Dual-Agent Reasoning: When any signal exceeds a 3-sigma drift threshold, the system wakes the Supervisor Agent (powered by Meta Muse Spark 1.3 at reasoning_effort="high"). 3. Domain Grounding: The Supervisor consults a specialized Diagnostician Agent that invokes tool calls to read the cell's maintenance manual, historical telemetry, and live readings to form a quantitative root-cause hypothesis, confidence score, and time-to-failure estimate. 4. Human-in-the-Loop Safety Gate: Mitigations follow an escalation hierarchy. Non-disruptive safe actions (reducing feed rate setpoint to 0.7 to alleviate spindle strain) execute automatically and visibly flatten the telemetry curve in the digital twin. Disruptive actions (operator work orders, machine HOLD) require human approval via interactive, pulsing cards in our real-time dark industrial web HMI. If an operator rejects an action, the Supervisor dynamically recalculates a less disruptive fallback.