Reasons To Deploy AI At The Edge—And How To Overcome Edge AI Challenges
Artificial intelligence (AI) is pervasive. From online shopping recommenders to weather apps, its adoption is increasing rapidly. The digitization of manufacturing in the realm of Industry 4.0 has fueled the adoption of AI in manufacturing automation with a gamut of applications like defect detection, quality inspection and worker safety, to name a few.
However, the traditional AI that powers shopping recommenders and weather apps tends to reside in the cloud. In such applications, the user request is sent to the cloud, the request is processed by an AI model and the results are sent back to the user. Although widely adopted, this application deployment architecture may not suit all use cases, one such being industrial automation.
AI at the edge is when the data and the AI associated with the data reside closer to the data source or its usage. The requirements governing manufacturing are different from those of a mobile application user, as described below.
Data privacy regulations for the protection of PII (personal identifiable information) and IP (intellectual property) often require that the data be closer to where the data originates. For example, in a factory setting, a machine’s configuration data is the manufacturer’s intellectual property, and they would like to keep it closer to the data source to avoid any data leakage.
The second reason is latency. The time it takes to send data to the cloud and back to the edge may be too long for time-sensitive applications. For example, in the case of weld quality inspection, the AI model that predicts the weld quality has latency requirements in the order of a millisecond. This is especially important when the AI model is part of a fully autonomous application. In the case of welding, configuration changes to a welding controller are needed to rectify the welding fault that was detected as a result of the classification of the weld quality inspection AI model.
The third reason is the bandwidth issue. Streaming large quantities of data to the cloud is expensive and network intensive. In some cases, factories may not even have the network infrastructure to deal with this traffic. For example, dozens of cameras used for defect detection on the factory floor produce thousands of frames per second; it is impractical to send such large amounts of data over the network.
A known issue while training an AI model is imbalanced classes (i.e., skewed class distribution in datasets). This is especially prominent in industrial use cases, as most products manufactured are not defective. The proportion of outliers or defects in such a dataset could be as low as less than 1% to 3%. Traditional ways of developing a model without special techniques on such a dataset would result in a model that performs better on only training samples and does not perform well on other test samples. Hence, the model quality degrades rapidly, decreasing the ROI (return on investment) of an AI model.
Adding context to AI, especially in circumstances with very limited labeled data, can help jumpstart the development of the AI model. This can be done by involving a subject matter expert with knowledge of process heuristics as early as possible in the data exploration phase. Such a data-centric approach yields better accuracy as well as helps with scaling the AI model.
Manufacturing processes, such as building a car, culminate in multiple individual processes. For example, metal sheets go through gluing before a spot weld is applied. Humidity is one factor determining the quality of gluing, which could affect the quality of the weld. Therefore, if the welding process is looked at independently from gluing, this relationship is missed.
Capturing interdependencies and relationships enhances AI model interpretability and dependability instead of viewing AI as a black box. While determining these relationships will likely require a large amount of data and time, it helps AI scale from a simple solution to a more scalable one over time.
AI is still in the early phases of adoption in the Industry 4.0 revolution. Hence, it is not uncommon to see some doubts about its benefits. In such an environment, choosing a high-value problem to apply AI is essential. The impact of solving a rather important problem with high ROI can lead organizations to embrace it. This encourages employees to look for other processes that can be enhanced with AI and, in turn, improve efficiency and quality, thus creating a positive ROI.
