On an automotive / two-wheeler parts assembly line, the hard part isn't "getting a clear shot of one cover" — it's making sure every unit is assembled in the same order: load the cover, align, install bolts, tighten, verify torque, paint-mark, then final-check and release. A missing bolt or a skipped torque check must be caught right there.
This article is based on a real assembly-monitoring video and the VisionAgent UI, laid out as field scene → process timeline → workflow / events / state machine / parameter page → runtime dashboard. The UI screenshots are drawn to match the product's real controls and wording, aligned with the actions in the video.
The real assembly station in the video
A top-down camera (labeled "Camera 01" at the lower-right) frames the assembly station: an operator in light workwear and gloves lifts a cast-aluminium engine cover and aligns it onto the case for assembly. The vision system marks "cover bolts" with red detection boxes, "torque check" with a yellow box, and prompts "paint mark" once done; the background shows station rails, pneumatic hoses and part racks — this is exactly the field that auto-parts assembly monitoring must handle.
▲ Field: cover loading & alignment, in progress
▲ Field: AI boxes each "cover bolt" for count check
Six steps: what each one checks
The runtime checklist matches the state-graph event tensor — six ordered steps per the real actions in the video:
A missed assembly step costs more than a surface flaw
One missing cover bolt, a skipped torque check or a forgotten paint mark often only surfaces at end-of-line test or at the customer. Process monitoring must lock "order + presence + torque" on the line side.
Workflow: a five-operator tensor flow
Create a project "Auto-Parts Assembly" and drag operators from the left library, then connect them (matching the real canvas):
▲ Tensor-flow canvas: Camera → Preprocess → Detection → Events → State Machine (real UI style)
Orchestrating events: the GPU Event Orchestrator
Open the "GPU Event Orchestrator"; the header shows "Configuring event: Install cover bolts". Drag "Count = N" and "Object Present" from the primitive library:
- The state-machine list on the right: E0 load cover, E1 align, E2 install bolts, E3 tighten, E4 torque check, E5 paint mark… unfinished steps show "To configure";
- Miss-prone steps can attach "Object Absent" or "Count short" to drive the interception branch.
▲ GPU Event Orchestrator: event "Install cover bolts" = primitives "Object Present + Count" (real UI style)
State machine: the S0–S6 ring
- S0 await cover —E0 load→ S1 await align —E1→ S2 await bolts —E2→ S3 await tighten —E3→ S4 await torque —E4→ S5 await mark —E5→ S6 release;
- after E5 paint mark, it returns to S0 — one loop per unit;
- the event tensor on the right lists EØ (void) and the assembly events E0–E5; the state-action tensor is kept in sync.
▲ State-graph editor: auto-parts assembly ring S0–S6 (real UI style)
Camera & page parameter configuration
Open "Parameter Config" and switch to the Page Params tab. The "Defect Stats" on the left align with the assembly steps, making it easy to count misses by NG type:
- Columns: Defect code (NG0…NG5), Defect name (cover missing, misalignment, bolt missing, not tightened, torque fail, mark missing), Count;
- You can "add / remove defect types"; edit name and description on the right, then "Save parameters" to apply;
- Result display: mask ON, defect name ON, font size 40, line width 2 — matching the on-site overlay;
- On the camera side, configure channel, exposure and ROI at the "Industrial Camera" node so the whole station is framed.
| Defect code | Defect name | Count |
|---|---|---|
| NG0 | Cover missing | 0 |
| NG1 | Misalignment | 1 |
| NG2 | Bolt missing | 3 |
| NG3 | Not tightened | 2 |
| NG4 | Torque fail | 1 |
| NG5 | Mark missing | 0 |
▲ Parameter Config · Page Params: defect types & result-display settings (real UI style)
Go-live: video + six-step checklist
- Left: live view + detection boxes, marking "cover bolts" and "torque check";
- Right: a six-step button checklist — load cover, align, install bolts, tighten, torque check, paint mark; finished steps get a green tick;
- Verdict: all pass → release allowed; any step "object absent / count short" → intercept and log the matching NG.
▲ Runtime monitor: live recognition on the left, six assembly steps advancing on the right
▲ Live footage: the full auto-parts assembly monitoring clip
Comparison table & takeaways
| Aspect | Traditional assembly line | VisionAgent auto-parts assembly |
|---|---|---|
| Recognize parts | Team-lead spot checks | Object detection identifies cover / bolts / torque wrench in real time |
| Manage order | Habit and paper SOP | State graph S0–S6 enforces the order |
| Manage misses | Found at EOL test or by customer | Event "presence + count" step gating + NG stats |
| Manage torque | Relies on tools and diligence | "Torque check" event + paint mark, double confirmation |
| Keep records | Paper ticks filled in later | Step trace and defect counts logged automatically |
- Business requirement: each unit finishes all six steps; cover bolts and torque check must not be missed;
- Decision rule: the current state only accepts its matching event; release is allowed only after everything is complete and torque passes;
- Software config: five-operator tensor flow + GPU event orchestration + state graph + page params (NG dictionary).
Four questions run through it all: what the system looks at, what "complete" means, when to intercept, and how results reach the report.
In closing
The hard part of auto-parts assembly monitoring isn't recognizing one bolt — it's stringing recognize-in-place → report event → advance state → NG stats → release into a chain the floor is willing to use and quality can stand up to review.