Examples

This page provides practical, end-to-end examples showing how to build and use pipelines and action plans in BIION-D.

Example 1: Simple Anomaly Detection Pipeline

A basic pipeline that captures an image and runs anomaly detection.

Scenario

You have one camera (IDS), a light, and one anomaly detection model (MyModel). You want to inspect products for surface defects.

Pipeline Structure

How to Use

Example 2: Multi-Camera pipeline with Lighting

A pipeline that controls lights and captures images from two different cameras.

Scenario

You have two cameras (IDS and 4103822175), a light, and two inference models (MyModel for anomaly detection and MyObjModel for object detection). You want to light the product, capture two images, and run both models.

Pipeline Structure

Example 3: Pipeline with PLC Integration

A pipeline that reads a part ID from the PLC, inspects the product, and writes the result back.

Scenario

Your PLC exposes a PartID node and an InspectionResult node. You want to read the part ID before inspection and write back the compliance result.

Pipeline Structure

Example 4: Custom Script with Webhook Notification

A pipeline that uses a Python script for custom post-processing and a webhook to notify an external system.

Scenario

After anomaly detection, you want to apply custom business logic (combine the anomaly score with a PLC value) and send the final result to an external MES system.

Pipeline Structure

Python Script Code Example

# Read results from previous modules
score = Data.get("anomaly_0.pred_score")
part_id = Data.get("opcua_reader_0.value")

# Custom business logic
if score < 0.3:
    quality_grade = "A"
    compliance = 0
elif score < 0.6:
    quality_grade = "B"
    compliance = 1
else:
    quality_grade = "C"
    compliance = 2

# Export results for downstream modules and history
Data.export("quality_grade", quality_grade)
Data.export("part_id", part_id)
Data.set_compliance(compliance)

print(f"Part {part_id}: Grade {quality_grade} (score: {score:.3f})")

Example 5: Multi-Pipeline Action Plan

An action plan with two separate pipelines that inspect different aspects of the same product.

Scenario

You inspect a product from two perspectives: a surface check and a dimensional check. Each has its own camera and model. The overall result is the worst compliance from both.

Action Plan Structure

Execution Flow

Example 6: Object Detection + Crop + Anomaly Detection

A pipeline that detects an object, crops the image around it using a Plugin, and then runs anomaly detection on the cropped region.

Scenario

You have a camera (IDS) and a light. On the production line, each product can be in a different position within the frame. You want to:

1. Detect the product using an Object Detection model.
2. Crop the original image around the detected bounding box, so that only the product region is isolated.
3. Analyze the cropped image with an anomaly detection model to check for surface defects.

This approach makes anomaly detection more reliable because the model always receives a tightly framed image of the product, regardless of where it appeared in the original capture.

Pipeline Structure

Workflow Description

Python Script Code Example

# Get the original image and object detections
frame = Data.get("Camera_0.frame")
detections = Data.get("ObjectDetection_0.detected")

# Each item in `detected` is an annotation with `class`, `confidence`, and `box`.
if detections:
    annotation = detections[0]
    box = annotation["box"]

    # Use all four vertices so the crop also works with oriented bounding boxes.
    x_values = (box["x1"], box["x2"], box["x3"], box["x4"])
    y_values = (box["y1"], box["y2"], box["y3"], box["y4"])
    height, width = frame.shape[:2]
    x1 = max(0, int(min(x_values)))
    y1 = max(0, int(min(y_values)))
    x2 = min(width, int(max(x_values)))
    y2 = min(height, int(max(y_values)))

    cropped = frame[y1:y2, x1:x2]
    Data.export("cropped_frame", cropped)
    print(f"Cropped region: ({x1},{y1}) -> ({x2},{y2}), shape: {cropped.shape}")
else:
    # No detection: pass the full image as fallback
    Data.export("cropped_frame", frame)
    print("No object detected, using full frame as fallback")

Common Patterns Summary

Tips & Best Practices