Bring AI to your data and improve vision-based product quality inspection
When you can't bring data to AI, see how federated learning (FL) helps everyone build a better, robust model
Advanced applications such as vision-based product quality inspection are making their way into the manufacturing space as part of Industry 4.0. The IoT devices utilized for this are cameras and mobile phones, sometimes mounted onto a collaborative robot arm, monitoring the final product for quality test and defect detection.
Typically, the high-quality image and/or video data captured is sent on to an inference engine where a pre-trained AI model scans it. The inference engine is usually hosted by a public cloud, although large-scale manufacturing organizations can also host an inference engine on a private, local server. Newly observed data (for which the model is not trained) is sent to the cloud or local server for “re-training,” which really means updating the inference engine.
However, due to the pervasive nature of smart vision-based sensors, data is often distributed across different locations and sites. For vision-based product quality inspection use cases, different defects in the same product can be observed across sites.1 It’s important for the inference engine to quickly learn a variety of patterns — which really means “understanding” the defects it finds — from distributed sources of data.
There are a few considerations when bringing distributed data to a single platform:
- Efficiency: Centralized data collection and manual labelling of a large dataset can take many days, which can prove to be inefficient with time-critical manufacturing applications such as product quality inspection.
- Data Privacy: Manufacturing organizations are sensitive about protecting their commercial intelligence, and sending data outside the factory floor is not a popular choice.
- Cost: Centralized, cloud-based solutions can be costly for small- and medium-sized organizations. In addition, uploading high-quality data to a server takes time and network bandwidth.
Bringing AI to the data
When bringing the data to AI becomes unfeasible, the other option is to bring AI to the data. Federated learning (FL) is the key enabler for this.
This iterative process enables different manufacturing sites to train a common model using their own product images and/or video data and to share their model updates with a trusted server. The trusted server aggregates the models sent from the different sites and uses it to build a better, new model that is distributed to all sites for the next round.
The power of working together
A typical FL model occurs when an ecosystem of participatory clients – in this case, manufacturing companies – agree to collaborate and train the federated learning model for the benefit of all.
Take product quality inspection use cases: site-specific model updates capture the patterns (defects) observed in the local data. The FL model then captures all defect data from different companies and sites. This way, not only is the privacy of each site’s data preserved (as the raw data never leaves the premises), but the cost of transmitting thousands of high-quality images and videos is also reduced.
The benefits of a robust FL model are shared by each participant in terms of timely defect detection without even training their individual models on the unseen defects. Small- and mid-sized manufacturers who do not have enough product data to “see” a wide-range of defect patterns truly benefit from federated learning. In addition, some of these organizations can’t afford a cloud infrastructure for centralized data analysis. But because these companies can form a collaborative ecosystem to share their model updates with each other, they are able to bring the AI to their data and get the most out of their resources.
Bringing AI models from experimentation to production involves complex, iterative processes. A significant driver of successful AI investment is access to training data that complies with privacy, governance and locality constraints — especially data moving between different regions, clouds and regulatory environments. Federated learning can boost model training with data collected from complex environments. Moreover, the global push towards collaborative data sharing eco-systems4 is encouraging for manufacturing industry to take a step towards collaborative learning to save costs, time, and network resources.
IBM Resources for manufacturers interested in vision-based product quality inspection
Learn how remote monitoring capabilities enable you to see, predict and prevent issues. IBM Maximo offers advanced AI-powered solutions and computer vision for assets and operations.
To improve overall manufacturing operations, discover why IBM was named a Leader in IDC EAM MarketScape for the Manufacturing industry. Although manufacturers have used EAM solutions for decades, there’s still plenty of opportunities to automate manual tasks, like maintenance execution, work scheduling, spare parts procurement, and asset life-cycle management.
Learn why IDC says IBM Cloud Pak for Data streamlines digital business development and resiliency and helps bring AI to your data – wherever it resides.The Cloud Pak for Data includes a tech preview of federated learning-based solution3 that increases cost savings and efficiencies.
- Mohr, M., Becker, C., Moller, R., Richter, M. (2021). Towards Collaborative Predictive Maintenance Leveraging Private Cross-Company Data. In: Reussner, R. H., Koziolek, A., & Heinrich, R. (Hrsg), INFORMATIK. Gesellschaft fur Informatik, Bonn. (S. 427-432)
- Cloud Pak for Data Footnote
- IBM Federated Learning
- International Data Spaces Association