-
NPI: A How To Guide for Engineers & Their Leaders
-
Leading from the Front
-
Marcel Tremblay: The Olympic Mindset & Engineering Leadership
-
Anurag Gupta: Framework to Accelerate NPI
-
Kyle Wiens on Why Design Repairability is Good for Business
-
Nathan Ackerman on NPI: Do The Hard Thing First
-
JDM Operational Excellence in NPI
-
Building the Team
-
Quality is Set in Development & Maintained in Production
-
3 Lessons from Tesla’s Former NPI Leader
-
Maik Duwensee: The Future of Hardware Integrity & Reliabilitypopular
-
Reject Fake NPI Schedules to Ship on Time
-
Leadership Guidance for Failure to Meet Exit Criteria
-
-
Screws & Glue: Getting Stuff Done
-
Choosing the best CAD software for product design
-
Screws vs Glues in Design, Assembly, & Repair
-
Design for Instrumental - Simple Design Ideas for Engineers to Get the Most from AI in NPI
-
Best Practices for Glue in Electronics
-
A Practical Guide to Magnets
-
Inspection 101: Measurements
-
A Primer on Color Matching
-
OK2Fly Checklists
-
Developing Your Reliability Test Suite
-
Guide to DOEs (Design of Experiments)
-
Ten Chinese phrases for your next build
-
-
NPI Processes & Workflows
-
-
Production: A Primer for Operations, Quality, & Their Leaders
-
Leading for Scale
-
Greg Reichow’s Manufacturing Process Performance Quadrants
-
8D Problem Solving: Sam Bowen Describes the Power of Stopping
-
Cut Costs by Getting Your Engineers in the Field
-
Garrett Bastable on Building Your Own Factory
-
Oracle Supply Chain Leader Mitigates Risk with Better Relationships
-
Brendan Green on Working with Manufacturers
-
Surviving Disaster: A Lesson in Quality from Marcy Alstott
-
-
Ship It!
-
Production Processes & Workflows
-
Failure Analysis Methods for Product Design Engineers: Tools and Techniques
-
-
Thinking Ahead: How to Evaluate New Technologies
-
How to Buy Software (for Hardware Leaders who Usually Don’t)
-
Adopting AI in the Aerospace and Defense Electronics Space
-
Build vs Buy: A Guide to Implementing Smart Manufacturing Technology
-
Leonel Leal on How Engineers Should Frame a Business Case for Innovation
-
Saw through the Buzzwords
-
Managed Cloud vs Self-Hosted Cloud vs On-Premises for Manufacturing Data
-
AOI, Smart AOI, & Beyond: Keyence vs Cognex vs Instrumentalpopular
-
Visual Inspection AI: AWS Lookout, Landing AI, & Instrumental
-
Manual Inspection vs. AI Inspection with Instrumentalpopular
-
Electronics Assembly Automation Tipping Points
-
CTO of ASUS: Systems Integrators for Manufacturing Automation Don't Scale
-
-
ROI-Driven Business Cases & Realized Value
-
Visual inspection is a critical part of quality control (QC) and quality management systems for electronics manufacturing. It allows engineers to monitor key points during a product’s assembly and can take several forms:
- Manual inspection uses trained human quality inspectors.
- Machine vision inspection uses industrial cameras running conventional computer vision algorithms – applications include automated optical inspection (AOI) and measurement systems like Keyence or Cognex.
- Visual inspection AI, a subset of machine vision, uses industrial cameras running machine-learning algorithms. These algorithms improve with additional data. They can discover novel and identify known defects and support paradigms ill-suited for conventional machine vision.
This article is a primer comparing several visual AI inspection offerings. If you’re interested in other comparisons between the different possibilities above, you can find a comparison of manual inspection and AI inspection here, and machine vision to AI-powered inspection here.
Key Takeaways for Quality Inspection
Visual inspection increases yield and reduces escapes, but it comes with added costs in headcount and equipment. Recent advancements in AI look to address both issues. New offerings like Amazon’s Vision for Lookout, Landing AI, and Instrumental provide novel approaches for automated visual inspection: bespoke AI for your product and process.
Visual inspection AI relies on three fundamental steps:
- A training set of product images are created and uploaded to the AI.
- The AI is trained on those images.
- The AI is deployed to cameras operating on the edge.
Smart engineers hoping to take advantage of this new technology should be aware of its benefits and limitations, especially before purchasing a long-term subscription.
Landing AI: A step up from AOI
Amazon Lookout for Vision: Do It Yourself AOI
Instrumental: For Teams Doing Complex Electronics Manufacturing
Landing AI: A step up from AOI
Landing AI and their LandingLens product are new tools in the AI space with an attractive, low barrier to entry. Landing was founded by famed AI researcher Dr. Andrew Ng, which has added to the product’s popularity. Since its introduction, LandingLens has evolved into a “better training tool” for data science teams, rather than an inspection system for factory personnel and quality engineers.
At its core, LandingLens is a tool with two purposes: classifying images and identifying objects. This may prove sufficient for some types of quality control software (QC): a conveyor belt transporting produce would benefit from visual inspection AI to ensure no contaminants are on the line. However, using LandingLens for QC in a complex electronics factory is laborious.
Similar to conventional computer vision (CV) technology, which relies on rules to identify if something is a pass or fail, LandingLens AI builds are limited to finding the defects they’ve been trained to find via golden units and specific anomalous units in the training set. While this is a big step up for certain inspection use cases where CV and automated optical inspection (AOI) fall short (such as solder inspection or foreign material inspection), it requires significant labeling and training time for each type of defect you want to classify.
Landing AI Limitations in Anomaly Detection
Only finds known defects. LandingLens’s AI models are trained to identify known defects – it does not detect novel defects.
Difficult to Distribute Training Burden. While the tool makes it easy for multiple people to get involved in training the model, Landing’s documentation points out that consistency is very important for performance. Only highly calibrated team members will successfully contribute to algorithm training, thus slowing down deployment and updates while tying up expensive staff.
Retraining is Laborious. Even in production, things change on the line – new suppliers, cost-down initiatives, and process updates. Any visual AI algorithm will need re-training in these scenarios, so if the training process is laborious, it will be felt repeatedly.
Given these limitations, LandingLens is ill-suited for high-variant environments like high-volume electronics NPI or low-volume production applications (where it may be difficult to amass a large enough defect training set). It’s also ill-suited for complex inspection applications, like electronics, where there could be 100s of parameters within the images to review.
Amazon Lookout for Vision: Do It Yourself AOI
Amazon Web Services’ (AWS) signature AI product, Lookout, has been developed for implementation across multiple data-driven applications: Lookout for Metrics analyzes business metrics, Lookout for Equipment brings predictive AI to equipment maintenance, and Lookout for Vision targets industrial applications…even though it wasn’t purpose-built for QC. Lookout has some inherent limitations to be aware of.
Things to Look Out for in Lookout
High Barrier to Entry: Requires Your Own AWS Environment. Interested teams must set up and administrate an AWS instance first - a process requiring software engineers to build and maintain - before they can begin building and training an AI model. Lookout falls in the no-man’s land between IT and OT (Operations Technology) – IT is needed to set it up AWS, but the OT team’s subject matter expertise would be required to get Lookout working for a line. If you already have manufacturing data in an AWS instance, Lookout could be a great fit. This challenge is something AWS acknowledges, which is why Instrumental is an AWS Industrial Software Competency Partner. Instrumental runs on AWS, providing this IT backbone for manufacturing teams so they can focus on the part that drives their business KPIs: quality control.
Only finds known defects. Like Landing, Lookout will only identify defects based on labeled examples in its training set and cannot discover new anomalies. This is an inherent setback in using AI classifiers for QC.
Instrumental: AI Visual Inspection for Complex Electronics Manufacturing
Instrumental technology was designed and built by engineers, for engineers, and focused on building in NPI and mass production. Those teams often don’t have software engineers or data scientists dedicated to building infrastructure or analyzing data for them, so Instrumental built a managed platform and product to skip those steps. If you’re responsible for hitting a first customer shipment date, first pass yield on the line, or field quality and performance – then take a closer look at Instrumental.
Instrumental’s Key Differentiators Make QA Easier
Instrumental’s manufacturing AI and data platform are fundamentally different from AWS Lookout and LandingLens in several key ways:
No IT Needed. Instrumental deploys and maintains your secure cloud instance so you can be up and running in a few days.
Bridging from Defects to KPIs. There are multiple steps between defect detection and impacting the business KPI – like first pass yield. Landing and Lookout only provide Step 1 and leave customers on their own for the rest. Instrumental provides an end-to-end workflow – so those defects can be easily translated into actual yield improvements.
Training Happens Through Usage. Instead of sitting down and labeling a giant dataset, Instrumental labels the dataset as you use the product – taking your feedback and learning from it over multiple sessions. Because it learns through usage, any user can contribute without risking model performance.
New Models in Minutes. Because setting up new algorithms is so easy and works with datasets as small as five units, you can use Instrumental for NPI, low-volume, and high-mix manufacturing.
Discover Novel Defects. Because Instrumental’s technology is built on an anomaly detector rather than a classifier like Landing and Lookout, it can uniquely identify novel defects without any initial training. With a little training, those anomalies can be intercepted as defects in real-time on the line.
Instrumental is the best in the world at what we do: discovering and intercepting novel and known issues and providing a powerful toolset to actually solve them.
Learn more about Instrumental’s product offerings here.