Resolving Manufacturing Quality with Machine Vision and MLOps

Think of your manufacturing operation like an orchestra - every instrument needs to play in perfect harmony to create a masterpiece. But instead of violins and cellos, you have machines, sensors, cameras, and control systems all generating their own streams of critical data. For years, manufacturers have struggled to find a conductor who can bring all these instruments together into a cohesive performance.

Today's data and AI technology have changed that dynamic completely. By connecting to any manufacturing data source - from legacy equipment to the newest IoT sensors - companies can finally orchestrate their entire operation in real-time with Crosser and Databricks. This means catching process variations the moment they begin, recommending parameter adjustments to maintain quality, and eliminating the costly delays between detecting issues and fixing them. The result is a manufacturing process that doesn't just collect data but actually puts it to work driving continuous improvement.

MLOps for Event-Driven Manufacturing

Process manufacturers in industries like plastics and papermaking must develop robust real-time monitoring systems to respond immediately to quality issues such as edge cracks and surface defects. By implementing advanced sensor networks connected to automated decision systems, operators can detect defects at their earliest stages and alert operators for quick resolution. This immediate responsiveness prevents minor tears from developing into major defects that result in significant scrap and waste. The most effective solutions combine edge computing for instant analysis with cloud data platforms that continuously improve detection accuracy over time, ultimately transforming reactive quality control into prescriptive maintenance that addresses potential issues before they manifest as physical defects.

Accelerate Defect Detection with Intelligent Edge Simplicity on Crosser

When monitoring for cracks in high-speed production systems, even millisecond delays are significant. Edge simplicity on Crosser fights industrial complexity by enabling immediate defect detection and response where the manufacturing actually happens, eliminating network latency and allowing for actions to be taken before defects cascade into larger problems. Some key aspects of Crosser’s event-driven platform:

• Minimize Latency for Rapid Action - When milliseconds matter, edge inference allows immediate action to minimize the impact of a defect. By processing data onsite, you significantly reduce response times, ensuring minimal downtime and avoiding costly repairs.
• Optimize Bandwidth Usage - High-resolution video streams demand immense bandwidth to transmit data to the cloud. Running models locally keeps data within your environment, reducing reliance on high-speed internet and lowering operational costs.
• Intelligent IT/OT Convergence for AI Model Drift - Smart edge-to-cloud data pipelines selectively transfer only the most informative anomalous data points to cloud systems needed to retrain ML models affected by data drift. This intelligent filtering ensures that model retraining incorporates emerging edge cases and changing production conditions, maintaining detection accuracy without overwhelming central systems with redundant or low-value data.

Databricks + Crosser: End-to-End Machine Vision MLOps

By combining Crosser’s edge analytics platform with Databricks’ Mosaic AI tools, you get a seamless solution for machine vision-based defect detection. Here’s how to implement it:

Step 1: Collect Image Data and Upload to the Cloud with Crosser
Data collection is the foundation of effective AI models. Using Crosser’s FlowApp “Video Capture”, you can easily capture video feeds from local cameras. Each frame is converted to a JPEG image and uploaded to your preferred cloud storage, creating a robust dataset for training your model.

Step 2: Ingest and Govern Image Data
Within the cloud, Databricks Unity Catalog volumes allow users to govern and store various types of content, including images, within a managed or external volume. For machine vision applications, Databricks recommends that you ETL images into a Delta table with the Auto Loader. The Auto Loader helps data management and automatically handles continuously arriving new images.

Step 3: Train and Govern AI Models
Once the images are prepared for model training, the Databricks Runtime for Machine Learning automates the creation of clusters with pre-built machine learning and deep learning infrastructure including the most common machine learning and deep learning libraries. Additionally, with Managed MLflow, Databricks extends the functionality of MLflow, providing model lifecycle management and governance.

For this edge machine vision application, a popular machine vision algorithm, YOLO (You Only Look Once), is considered. YOLO's popularity for edge inference stems from its unique architecture, which processes images in a single pass. This delivers remarkably fast detection speeds and small model footprints while maintaining sufficient accuracy for many industrial applications, making it ideal for resource-constrained edge devices.

The following pseudo-code provides a logical flow of model training and logging as ONNX, which is discussed further in Step 4. The Databricks documentation provides a full machine vision training example using PyTorch.

Step 4: Export Model for Edge Deployment
Once trained in your preferred machine learning framework, the model needs to be exported for edge deployment. ONNX (Open Neural Network Exchange) has emerged as a popular model format for edge deployments due to its exceptional portability across diverse hardware environments. By providing a standardized intermediate representation for neural networks, ONNX allows models trained in frameworks like PyTorch or TensorFlow to be deployed on a wide variety of edge devices without framework-specific dependencies. Additionally, ONNX Runtime's built-in performance optimizations automatically adapt models to the specific hardware characteristics of edge devices, whether they're utilizing CPUs, GPUs, or specialized AI accelerators. This combination of hardware flexibility and optimized inference capabilities makes ONNX particularly valuable for organizations deploying machine learning solutions across heterogeneous edge environments with varying computational constraints.

The mlflow.onnx module provides APIs for logging and loading ONNX models. Within Databricks, a hosted model registry with Unity Catalog provides fully governed APIs used by Crosser to download the model and deploy it to the edge.

Step 5: Download and Perform Edge Inference and Real-Time Alerting with Crosser
Once downloaded, the YOLO ONNX model is prepared for inference with Crosser. Crosser’s FlowApp “Video Crack Detection” demonstrates how to process live video feeds from local cameras, detect crack defects in real-time, and take immediate action.

When cracks are detected:

• Trigger notifications to local HMIs for operator review.
• Send alerts to relevant personnel via notification services.
• Automatically stop the machine to prevent further damage.

Step 6: Re-train and Re-deploy with Databricks and Crosser
Industrial machinery often degrades due to the harsh operating environment and requires continuous maintenance. AI models are no different, and with this architecture pattern Crosser can intelligently capture new image data to send to the cloud and Databricks Lakehouse Monitoring will continuously track data quality and model performance. If drift is detected, Databricks’ orchestration tools can automatically retrain the model and trigger redeployment to Crosser, fulfilling the entire MLOps lifecycle.

Unlock the Full Potential of Machine Vision IT/OT Convergence

The partnership between Databricks and Crosser represents a breakthrough in industrial AI orchestration, creating a seamless bridge between edge processing and AI model training for manufacturing environments. Crosser's edge intelligence platform captures and processes machine vision data in real-time at the production line, while Databricks provides the scalable data lakehouse infrastructure for comprehensive model training and performance monitoring. This integrated approach eliminates the traditional barriers between operational technology and information technology, enabling manufacturers to deploy sophisticated computer vision models that evolve with changing production conditions. By combining Crosser's low-latency edge processing with Databricks' powerful MLflow governance, companies can implement vision AI solutions that not only detect quality issues instantly but continuously improve through automated model retraining cycles. For manufacturers seeking to transform manufacturing quality processes, this collaboration offers a production-ready solution that delivers both immediate operational benefits and long-term AI maturity – turning the manufacturing orchestra from a collection of individual instruments into a harmonious symphony of data-driven excellence.

The Data Intelligence Platform for Manufacturing helps manufacturers deploy Industrial AI at scale with leading ecosystem partners like Crosser. If you’re looking to improve operating margins through AI apps while managing exponential growth in data volumes, contact your Databricks account team to show you how a unified platform brings the power of AI to your data and people, so you can build AI into every process.