Introduction

Inherent to IoT is the generation of large volumes and varieties of time series telemetry, which makes Generative AI (GenAI) applicable to multiple facets of a typical IoT environment. In this series of articles on GenAI in IoT, we explore several facets where GenAI and IoT intersect. 

This first article provides a base understanding of GenAI capabilities and limitations in the context of IoT. 

Basic differences between AI and GenAI in the context of IoT

AI is software created through machine learning. AI mimics the human brain to accomplish everyday tasks like driving a car, drawing, accounting and summarizing a book. It can provide insights into data (like text or streaming IoT signals) and can support a wide variety of use cases. GenAI is a special kind of AI that generates synthetic data for the use cases it supports. As GenAI is a subset of AI, it also mimics the human brain to accomplish tasks like simulating a machine's behavior, generating a design, translating a language, etc. The main difference between the two is that GenAI uses AI to generate new content to accomplish tasks and mimic human behaviors, whereas AI interprets existing data. 

IoT is about the communication of things over the internet. Effectively, this means that while AI can be used for a wide variety of use cases in IoT, this AI does not have to mimic human interactions or language. Instead, the AI analyzes IoT telemetry and messages from Internet-connected 'things' (aka IoT assets) to identify patterns. The patterns identified are then processed further to generate insights, predictions, etc. 

Having examined the differences between AI and GenAI, along with AI's significance in IoT, let's explore how GenAI enhances the learning capabilities of other AI in two distinct ways:

1. A recent revolution in Large Language Models (LLMs) allows for some non-traditional uses of GenAI in IoT (see reference below). Here, the LLM offers human-understandable interpretations of IoT outcomes and provides guidance on courses of action.
2. GenAI generates synthetic data that can be used to enhance other AI.

Limitations of GenAI in IoT

GenAI in IoT does have limitations, and it is necessary to understand them to make good, informed decisions on when and how to use it. Some of these will be mitigated over time as GenAI LLMs mature. 

• Privacy, security and regulatory/ethical compliance: One should be prudent in exposing any privacy and security data to GenAI; even anonymization and obfuscations should be thoroughly vetted. Furthermore, the additional diversity and inherent nuances of regulatory/ethical compliance further constrains the deployment of GenAI. 
• Data quality, bias and fairness: Since GenAI learns from the data provided to it and then generates synthetic data (that is, it is very susceptible to the garbage-in-garbage-out syndrome) that will include any biases in the training dataset, it must be trained on an abundant volume of very high-quality data for it to yield promising results. 
• Compute resources/energy: In addition to the very high computational resources to train GenAI models, the trained models are large and require comparatively higher computational resources once deployed; this makes it much more challenging to deploy GenAI on resource-constrained devices/equipment often used in IoT and even more so for those powered by batteries. 

Use cases of GenAI in IoT

Now that we understand the potential and limitations of GenAI in IoT, we explore the use cases in a couple of different industries. Of course, this is not an exhaustive list but rather an illustration of how useful GenAI will be in IoT. 

Industrial manufacturing

• When used in conjunction with LLMs like CoPilot, ChatGPT, BARD, etc., GenAI can merge information from the internet to further enhance the equipment telemetry data with suggestions on potential actions and their likely impact, references to additional documents or information that provide additional technical information, success rates of each remediation, etc. 
• GenAI can bridge the data gap that is often encountered in predictive analytics. This topic will be discussed in subsequent blogs in this series. One can assert that GenAI addresses the core of AI in IoT:
  • Equipment failures: Predict failures of equipment/machines before they disrupt operations
  • Optimized maintenance: Predict remaining useful life estimates for timely component replacement and maintenance
  • Digital twins: Accelerate the creation of digital twins for complex operations planning, performance engineering, energy consumption optimization, etc. 
  • Waste reduction: Explore options via what-if simulations and analytics. 

Medical and healthcare

• Privacy protection: Generate representative patient data and protect actual patient data for privacy compliance
• Medical summarization: Generate automated medical summary reports of tests and wearables data
• Prescriptions: Streamline prescription of medications and procedures based on tests and wearables data 
• Image analysis: Perform data augmentation tasks (denoising, reconstruction, registration, etc.) on medical imagery like CTs, MRIs, ultrasounds and xrays 
• Drug research: Accelerate discovery/invention of new drugs

Illustrative scenarios

As this series of blogs explores GenAI in the context of IoT, it always helps to illustrate the concepts with real-world scenarios, and the following scenarios are excellent representations of the use cases in IoT. 

Scenario 1: Issues in manufacturing equipment cause production stoppages

Analysis: Moving components are typical causes of failure, as they often have rotating parts. Rotations cause characteristic vibrations, and anomalies in these vibrations can diagnose and predict failure. 

Challenge: Available vibrational data was insufficient to set a proper baseline for anomaly detection as the equipment was moved and re-mounted; the new location and mounting components caused changes in the vibration patterns and thus created a data gap as many required baseline conditions were absent in the dataset since the move. Effectively, many proper operating conditions were marked as anomalous. 

Remediation: GenAI was leveraged to extend the proper baseline dataset for anomalous vibration detection. 

Scenario 2: Damage assessment and repair estimation for vehicle insurance claims

Analysis: Leverage decades of images of damaged vehicles and their repair costs to perform image analysis of damaged vehicles for repair cost assessment. 

Challenge: Typically, insurance agencies do not have enough examples for all the common types of damage for every vehicle brand, make and model. Damaging vehicles to capture such images is expensive and cumbersome. 

Remediation: GenAI extracts different types of damage information from the existing inventory of damaged vehicle images and then applies the damage information to the desired brand, make and model. These generated images can then be processed using existing AI techniques for automated damage assessment. 

Key takeaways

• GenAI has a limited but important role in IoT
• GenAI can bridge data gaps — we will elaborate on this in subsequent blogs in this series
• GenAI’s LLM can make the predictions of AI more usable 

Conclusion

Having established a sound understanding of AI and GenAI in IoT, our next article will focus on understanding the types and causes of data gaps and their impact on GenAI in IoT. 

References

• What is the Best Generative AI Tool: ChatGPT vs LlaMa vs Google Bard vs Claude, Mariia Yuskevych
• The Rise of Conversational AI: Comparing Bard, Claude, ChatGPT and Bing AI, Mike Onslow
• How 3 healthcare organizations are using generative AI, Aashima Gupta & Greg Corrado
• IBM Consulting Collaborates with Microsoft to Help Companies Accelerate Adoption of Generative AI