Machine data through sensors and connected devices are generating a wealth of information that can be converted into value-added knowledge for your organization. Traditional approaches to data can allow organizations to monitor for defined conditions but these approaches are naturally backwards looking. The introduction of machine learning to your data can enable predictive analytics and forward-looking analysis to further unlock insights.
Traditionally, it has been difficult to implement and scale the usage of machine learning primarily due to technical complexities and a shortage of AI talent across industries.
This series aims to feature use cases showing how Elipsa enables Approachable AI to allow an organization's existing talent to become AI talent and apply advanced analytics without needing to write code.
Defect Detection
The American Society for Quality (ASQ) suggests that the Cost of Quality is usually around 15–20% of sales, often as high as 40% in some organizations. The cost of poor quality includes both internal and external failure costs. In other words, organizations are spending a considerable amount of money correcting defects found both on the factory floor and in the customer’s hands. This not only leads to a direct monetary loss but also indirectly in the form of reputational risk.
Currently, organizations try to minimize costs through inspections and ensuring specific procedures are followed. However, these are manual, time-consuming processes that become mundane and still lead to errors.
The advent of computer vision opens up the possibility of companies automating inspections both on the line and in the field. Most public computer vision models are able to accurately classify things like faces, dogs, cars, etc. However, those models are not trained to identify the specific defects that are relevant to you and your process.
Data scientists can build custom computer vision models to fit your needs but those resources are expensive and hard to come by.
Certain vendors provide the ability for users to create their own custom models but they require the installation of a new end-to-end hardware solution.
Elipsa’s goal with Approachable AI is to allow users to build custom defect detection models with their existing cameras and software, and without the need for a data scientist.
Problem: Cost of Defects
In our example, we explore examples of a casting manufacturing product found here. Casting is a manufacturing process in which a liquid material is usually poured into a mold, which contains a hollow cavity of the desired shape, and then allowed to solidify. A casting defect is an undesired irregularity in a metal casting process. There are many types of defects in casting like blow holes, pinholes, burr, shrinkage defects, mold material defects, pouring metal defects, metallurgical defects, etc.
Product inspection is a very time-consuming process prone to human error. Defects can be cause to reject an entire order, leading to a large loss for the business.
Our data consists of photos taken from the top view of a submersible pump impeller.
The dataset contains a total 6,633 grey-scaled 300x300 pixels images.
Elipsa’s Approachable AI Applied
In the data used to build the model, we have historical images of defective examples (3,758 images) and images of quality products (2,875 examples). We also set aside a series of images to test against the model to analyze the final accuracy of prediction.
To build the model in Elipsa, we simply create a folder for each label that we are classifying. In our case, there are two labels and thus two folders: defect (def_front) and ok (ok_front). Once our defective training images and ok quality images are uploaded to their respective folders, Elipsa is able to build the computer vision model with a click of a button and no need to write code. The system learns the patterns that are indicative of normal and defective products to be able to classify future examples accurately.
Elipsa automatically splits the images into training (5,307 images) to build the model and testing (1,326) to automatically optimize the model.
Results
Computer vision models can take a considerable amount of time to train, depending on the number of images and the size of these images. Elipsa builds the model on its backend servers and notifies the user when it is complete.
With the model built, we had a total of 713 (261 quality products / 452 defective products) images to run through the model for testing. These are new images that the model has not seen before. In other words, this is the equivalent of deploying this model to production and streaming new product images against it.
Overall, the model was 99.29% accurate at predicting whether the product was defective or not.
Drilling down into those results, the model was 100% accurate at predicting that quality products are OK. For defective products, we ran 452 images against the model. The model was 98.68% accurate at detecting whether they were defects, missing only 5.
Deployment
With high accuracy received, Elipsa users can easily deploy this model to the cloud or to their own edge device with a push of a button. If the accuracy is not to their liking, users can hold off on deploying into production and simply add new images to the respective folders to try and improve the model accuracy.
Summary
We were able to build a highly accurate image classification model for defect detection without the need for a data scientist. In addition, we were able to deploy this model to production without changes to infrastructure.
The use of AI and computer vision for defect detection would allow an organization to cut down on manual processes, enabling more efficient use of the workforce. In addition, by catching defects through AI, companies can prevent defective products from leaving the factory floor; helping to eliminate recalls and buybacks and increase customer satisfaction.
For more information book a demo @ www.elipsa.ai