Landing AI Provides Manufacturers with an Improved Lens on Their Production Lines

Landing AI works with manufacturers to improve quality levels by applying AI and Machine Learning (ML) to inspection processes. The company was founded by Dr. Andrew NG, who previously co-founded Coursera, was founding lead at Google Brain, and also was Chief Scientist at Baidu.

The company’s visual inspection software platform, LandingLens, is currently being used to, among other things, inspect wafers in semiconductor fabrication plants, the soldering quality of components on electronics production lines, vials of pharmaceuticals passing along a production, and to help pharma manufacturers achieve CFR 2.11 (which validates that medical products are fit for use).

Pressure on manufacturers to meet deadlines turns attention to increasing a production line’s yield. An inspection process that involves inspecting a small number of items or relying on staff who can be fatigued and therefore cannot ensure consistent approach to fault identification has its limitations. With AI-based machine vision, those staff members can instead be deployed to train the software rather than spending their days performing inspections.

Landing AI was founded in 2017 and current investors include Samsung, Lenovo, Intel, AIFUND, and Taiwania. The LandingLens platform looks to provide a continuous visual inspection solution helping manufacturing pick up defects (such a scratches, residue, particles damage) by combining their staffs’ expertise and the AI platform. The LandingLens architecture comprises three stages– data collection, modelling, and deployment.

The data collection stage involves the clients’ subject matter experts taking responsibility for defining fault parameters. Faults are labelled and documented in the defect book, forming the data set which the machine learning platform bases its decisions.

In the modeling stage, the software enables users to select a model used previously or one newly created for the task from the library. The subject matter experts run tests to evaluate how well the platform performs—in other words how many faults are identified and whether there are any false positives or worse, missed faults. The model is then refined by looking for anymore inconsistencies.

The model can be deployed at single location or across a manufacturers’ entire plant network. The mantra is to have a single instance of the model which is deployed at scale with results displayed in dashboards. But once deployed, the work is not finished with users monitoring model performance, looking for signs of drift and, if necessary, retraining of data sets.

It’s important to note the user remains in control of LandingLens, with it being the tool for delivering the inspections rather than customers outsourcing their inspection work to Landing AI.

LandingLens is a solution for manufacturers and industrial firms that already appreciate the role technology can play in improving quality levels and want to move away from manual processes. LandingLens helps customers to have consistent inspection processes and also adapt when circumstances change, such as when seasonal changes affect lighting in the facility which can skew results. This flexibility is a critical success factor for LandingAI.

The clear workflows and involvement of customers’ subject matter experts means that adjustments can be easily made. The customer keeps the defect book and models up to date and because the workflows are easy to understand, the models can be tweaked quickly. As a cloud platform, LandingLens can scale quickly across a network of facilities and this also means subject matter experts can collaborate to refine models. Another benefit is that inspectors remain relevant as their role changes from inspecting to designing the inspection process and they don’t necessarily need to have programming expertise. Furthermore, the customer doesn’t need to have a data scientist team to roll out the programs.

The biggest challenge for LandingLens and machine inspection as a whole is when the defect rates are low, resulting in relatively few examples to model. Often a production line is considered to be performing relatively well with hundreds rather than hundreds of thousand defects. The LandingLens platform can classify defects with a relatively limited number of images and learn as the platform is deployed.

Also important is that Landing AI is looking to democratize machine learning by involving frontline staff in the creation and maintenance of the defect book so that automating quality inspection is not perceived a mysterious black box application but rather a tool to make people’s lives easier and improve manufacturers’ quality levels.