Intelligent automation of industrial assembly line tasks

Introduction

This blog provides updated guidance for manufacturing industry customers using the AWS Cloud to enhance their manufacturing assembly line using intelligent automation. Customers can now leverage Artificial Intelligence (AI) service calls through API operations, or incorporate their own machine learning (ML) models to assist with their assembly line activities.

Customers can build their own solutions with managed services or work with partners their AWS account team to find the right industry package solution available in the AWS Marketplace as Amazon Machine Image (AMI), Containerized, or SaaS solutions. The assembly line tasks we describe here cover some of the options and possibilities.

Amazon has designed and manufactured smart products and distributed billions of products through its globally connected distribution network using cutting-edge automation, machine learning and AI, and robotics, with AWS at its core. Industry customers can take advantage of this in addressing their own specific needs. See the AWS for Industrial quality management page for a broader view.

Industrial assembly line problems and solutions, past to present

Until recently, customers incorporating computer vision driven inferencing for their assembly line have relied on mostly custom-built implementations tied to operations dashboards, compared to assembling inferencing from commercial off the shelf (COTS) or mostly bought components. Now with Amazon Lookout for Vision and Amazon Lookout for Equipment and the AWS Panorama Appliance, manufacturing industry customers are able to tackle the most prevalent industry tasks.

For other tasks, you can continue to use Convolutional Neural Network (CNN) ML models trained on Amazon SageMaker and deployed to the edge using AWS IoT. We observe that the simplified task areas that are benefited include defect identification, deviation traceability, and waste identification.

These industry tasks span ML-specific capabilities that can be addressed through the AWS Cloud and edge capabilities, which include the following:

1. code recognition
2. object recognition
3. completeness check
4. position recognition
5. surface inspection
6. shape and dimension check

We find these tasks are widely requested by our customers for various scenarios, such as monitoring adhesives and abrasive production, auto part kiting, nylon production, paint booth inspection, and steel processing surface defect identification. Now, it’s relatively easy to address missed defects in automotive interiors for example, which can affect the value and overall appeal of a vehicle.

Deviation traceability, and defect detection for strict compliance standards

A reliable inspection solution also enables companies within the aerospace industry to comply with strict industry defect standards and regulations. Deviation traceability within medical and non-woven item manufacturers is used to trace deviations from the source and to audit the results. Quality inspection systems produce real time, reliable, and traceable data to identify defects and separate them from one another. Unsupervised ML solutions are able to deliver up to 40% fewer false accepts than the human standard. Waste identification and reduction, such as used by textiles and composites customers, reduces waste and increases profits for the textile industry by detecting faults at the beginning of the production process. You can read more about broader aspects of quality on AWS here.

The relationships between code, object, and position recognition

An example of machine learning for code recognition is the automation of defect detection on manufacturing lines with ML to automatically capture imperfections with bar codes printed on products. Code recognition denotes the identification of objects using markings on the objects; these are typically standardized bar codes or Data Matrix codes, but can also be custom codes. Typical applications are material flow control and logistics. Code recognition solutions borrow from all areas of image processing, including, for example, edge detection, filtering, and positioning techniques.

Ensuring that the labels are being affixed to the right boxes is an example of object recognition. Object recognition denotes the identification of objects using characteristic features like shape/geometry, dimensions, color, structure/topology, or texture. Object identification includes the distinction of object variants and has applications for other tasks. For example, position recognition or completeness checks may require prior identification of the correct objects in the scene. 3D data is beginning to be used for ML driven object recognition, allowing for completely new ways of matching and evaluating object characteristics, for example, by comparison with computer-aided design (CAD) data. You can read more about engineering design on AWS Cloud here.

Position recognition denotes the determination of position and orientation of an object – or a particular point of an object – in a predefined coordinate system, using feature computation and matching methods. Typical features are center of gravity coordinates and orientation angles. An important distinction is the dimensionality, that is, whether position and orientation have to be determined in two or three dimensions, where the term pose is typically used. Typical applications are assembly-robot guidance, pick-and-place operations, and insertion machines. A variation is the reverse application using images to determine the location and orientation of the camera system itself, used, for example, by autonomous robots to determine their position.

How machine learning improves overall computer vision

Use of computer vision for kiting and scraping processes associated with mechanical assembly of engines is an example of both object recognition and completeness check. Completeness check denotes a categorization of work-pieces as correctly or incorrectly assembled; it checks whether all components are present and in the correct position, often as a precondition for passing the work-piece on to the next assembly step or as a final check before releasing the work-piece to be packed and delivered – or, one step later, the inspection of the package to be completely filled with products of the right type. Adoption of these ML capabilities typically targets 10-12% improvement in output volume as an outcome.

Computer vision for the blade production assembly line is a key use of ML for shape and dimension check. Shape and dimension check denotes the determination of geometrical quantities with a focus on precise and accurate measurements. The importance of this area increases in accordance with rising quality standards as products must meet ever tighter tolerance requirements. Applications can be found wherever work pieces or also tools have to be checked for compliance with nominal dimensions. Because of the required accuracy, these tasks typically impose high demands on sensor equipment in addition to on the mechanical construction of the inspection station.

Visual monitoring with ML for carbon fiber layering on parts is an example of surface inspection.

For example, using ML and image recognition for leather upholstery assembly line computer vision is a combination of ML problems spanning object recognition, completeness check, and surface inspection. Surface inspection can be divided into quantitative surface inspection aiming at the determination of topographical features like roughness, and qualitative surface inspection where the focus is on the recognition of surface defects, such as dents, scratches, pollution, or deviations from desired surface characteristics, like color or texture.

Quantitative measurements of geometrical properties may also be required for judging the surface quality. Qualitative surface inspection tasks are frequently difficult to specify formally, for example, compared to a dimensional check, and are often based on “fault catalogs” used by human inspectors.

Evolving from proof-of-concept experiments to proof-of-value engagements

When you use the pre-trained models and API interfaces of the AWS AIML services noted in this article, you are not repeating the billions of operations resulting in power consumption that contributes to carbon emission. This has positive connotations for sustainability. ML for assembly line vision using AWS Cloud services helps manufacturers introduce robust inferencing with the potential to compensate for changes with minimal down time.

Industry customers who leveraged AIML in the past usually needed increased investment in data science and data platform engineering resources to incorporate assembly line intelligence in their existing solutions. Currently though, achieving cost reduction and disproportionate operational enhancement benefits from AIML on the AWS Cloud no longer requires root-and-branch replacements of existing software given the broad and deep solutions and services available to work and optimize your existing digital investments.

Conclusion

There is an emerging understanding of the “science of the doable” transforming what were traditional proof-of-concept “art of the possible” experiments into proof-of-value engagements that benefit your business now. We invite you to take advantage of these benefits for enhancing your assembly line. Incorporating intelligent automation that provide disproportionate advantages for your assembly line processing is now achievable with AWS support.

Contact us to learn more about this as well as how AWS can help grow your manufacturing or industrial business.