From Machine Learning to Machine Teaching
Machine Learning offers incredible opportunities to improve manufacturing quality and OEE. IoT platforms can enable enterprises to fully leverage these capabilities with MLOps and Federated Learning, easing deployment from cloud to edge.
For as long as people have produced goods, there has been a need for Quality Assurance. Across industries, a universal form of Quality Assurance has been visual inspection. While many products need to be tested against more quantifiable benchmarks like performance requirements, structural dimensions, or material composition, nearly every manufactured good needs to meet the more difficult to quantify benchmark of “looking right.”
Traditionally, human laborers spend enormous amounts of time inspecting products at multiple stages on a manufacturing line to ensure quality. Often human error is a factor, estimated at 20% to 30%1. Sometimes product changes occur, resulting in frequent reconfiguration of production lines. Another challenge is that manual, visual inspection tends to be inflexible. Product gets marked manually at the point of detection after manufacturing, forcing workers to review and adjust accordingly. Oftentimes it can even cause health issues as human laborers must focus for long periods of time, checking every single item for damage and defects.
Machine Learning can create Machine Sight
With machine learning, devices can analyze unstructured image and video data, automatically verifying product quality and immediately annotating the exact coordinates of all defects in a digital format. This is the same as how manual laborers inspect products on the manufacturing line before reaching the end of cycles – only better. Coupling AI with computer vision technologies enables manufacturers to automate inspection, saving time and money. Quality control can be faster, more accurate, and persistent. Computer vision evaluates 24/7, has no sick leave, and offers a consistent level of quality control, based on its training.
This can even be applied to a wide variety of devices, equipment, and processes, too. On top, the result is digitally available. People have to manually annotate the defect they found (e.g., with markers on textile or through the UI of a quality assessment software). Computer vision immediately produces the coordinates where defects are located in a digital format. Even predictive maintenance, object identification and counting, safety, and security can be monitored, measured, controlled, and improved upon with the power of machine learning-based visual inspection. We also see this particularly with fabric defect localization, classifying tile or raw images appropriately in the sewing industry and digital t-shirt work lines. For these reasons, Allied Market Research estimates the value of AI visual inspection technology will exceed $200 billion by the end of the decade2.
Yet with machine learning, there is one setback – the lack of high-quality data. And that’s exactly where IoT comes in.
The role of IoT in MLOps
IoT operates in a much more advanced atmosphere with devices designed to “learn” and even automate predictive maintenance. Machine learning can then identify more accurate patterns with higher-quality real-time data IoT sensors provide. As IoT connects multiple pieces of equipment and devices into one seamless network, we then have a self-serving, self-learning, streamlined, and evolving production- and process-improvement machine, leading to better quality, healthier workers, and lower cost.
As a result, machine learning operations (MLOps), with the power of IoT interlaced, specifically demand empowerment in the hands of domain experts. That then allows factory-floor technical engineers to create their own AI and ML tools that are hyper-automated and centralized across the cloud and edge. Moreover, PLC products, protocol gateways, and Linux-based operating systems can then easily leverage Cumulocity IoT Thing Edge, even in the cloud. This form of federated learning provides three major advantages.
First, federated learning allows the central model to learn from a diverse and augmented set of learning samples obtained from multiple entities. End user data is obtained from a specific subset population and is therefore unlikely to have been seen by or shared with other end users. This is particularly true of users located in different geographic regions, where data traits likely differ substantially. Because of these differences in end-user populations, the model has not had the opportunity to learn from atypical cases and normally one would advise against using the model calculation at hand. However, a federated learning derived model incorporates data from multiple entities, thereby increasing its external validity. Such a model would be much more likely to generate more accurate results, even for what may be an atypical occurrence at a specific end-user. This type of cooperation therefore allows for the advancement of precision analytics.
Second, the deployment of AI models requires periodic training and updating to remain current. This requirement may place an indue burden on process experts, who must continually label a sufficient volume of data necessary to retrain the model. At certain times when work-load peaks, it may be difficult or even impossible for process experts to produce enough annotated labels. However, these peaks occur, just because interesting analytical events are taking place at the moment. Federated learning mitigates this issue, allowing process experts and data scientists at less busy sites to annotate studies while their counterparts are too busy to do so. In this way, all users can download and use the most up to date model year-round. Your machine models leverage 24/7 machine teaching from IoT.
Third, the federated learning framework brings about auto-scaling at almost no additional cost. When new entities participate, they bring more data and more computational resources. As the loop continues to run, an ever-enlarging dataset is fed to the model, while all computations continue to be made by the end-user. The global model is updated after users have trained their individual models, requiring minimal resources to aggregate models and thus making deployment much more economical.
This is the near future for machine learning training and by offering the IoT, analytics and integrations platforms, Software AG is in the cradle position to make developments in this field, effectively democratizing this challenging process by orchestration of our existing offerings. Note that federated learning exponentially reinforces the value of our “deploy anywhere” strategy. If we manage to eat our own dog food and close the integration loop between our products, this training technique becomes low hanging fruit, since it builds on already existing tools in our product portfolio.