6 Identified AI/ML Requirements
Technical Report TR-0068 identified eight representative use cases demonstrating the need for AI/ML integration in oneM2M IoT systems. These use cases span domains such as autonomous driving, delivery logistics, metaverse-based services, video analytics, automated ML, IoT calibration, dataset generation, and AI model lifecycle management.
Through analysis of these use cases, a total of 17 potential AI/ML-related requirements were derived. These requirements highlight the functional gaps in the current oneM2M system with respect to supporting AI data and model management capabilities. The requirements can be categorized as follows:
Table 6.1-1: Collection of potential requirements in TR-0068 [i.2]
| Use case | No. | Potential requirements |
|---|---|---|
| Use case #1. Data augmentation for autonomous driving | 1 | The oneM2M System will be able to handle data augmentation requests for AI/ML purposes. |
| 2 | The oneM2M System will be able to generate augmented data resources from a given source data and data augmentation technique. | |
| 3 | The oneM2M System will be able to manage data for AI/ML purposes such as model training and augmentation of training dataset. | |
| Use case #2. Last mile delivery | 4 | The oneM2M System will be able to manage structured and unstructured data for training, for example, preprocessing data, describing data and inferring meaning. |
| 5 | The oneM2M System will be able to update trained AI/ML model according to continuous measuring data e.g. location, time series and historical data. | |
| 6 | The oneM2M System will be able to provide a classification function (e.g. split data into two parts, training and validating) in supervised Machine Learning. | |
| Use case #3. Smart virtual store using metaverse | 7 | The oneM2M System will be able to synchronize between real and virtual world devices |
| 8 | The oneM2M System will be able to enable Edge/Fog Nodes to run AI/ML models to retrieve information from the real world | |
| Use case #4. Detection of patterns in video streams | 9 | The oneM2M System will be able to support the creation and management of classifiers for AI/ML application as follows:
|
| Use case #5. Autonomous operations using automated machine learning | 10 | The oneM2M System will be able to distinguish the dataset that will be trained and has already been trained. |
| 11 | The oneM2M System will be able to provide automated machine learning under certain conditions, e.g. building a model every week or when the number of datasets reaches 100. | |
| Use case #6. IoT device calibration using ML | 12 | The oneM2M System will be able to manage calibration information and training datasets for ML to eliminate or minimize measurement errors from IoT sensors. |
| 13 | The oneM2M System will be able to perform ML using training datasets from reference IoT devices and notify calibration results to a target sensor that requires calibration. | |
| Use case #7. Dataset creation for AI models | 14 | The oneM2M System will be able to create datasets using the historical data (e.g. IoT sensor) to train AI/ML models. |
| 15 | The oneM2M System will be able to create datasets using the current data (e.g. IoT sensor) to train AI/ML models or make prediction/inference with the trained models. | |
| Use case #8. AI model management | 16 | The oneM2M System will be able to manage AI/ML models with model metadata. |
| 17 | The oneM2M System will be able to support an AI/ML model deployment to IoT devices (e.g. Edge/Fog nodes) and IoT applications. |
The requirements above indicate a broad scope of AI/ML capabilities needed in the oneM2M system, including but not limited to:
- Dataset lifecycle management (creation, augmentation, classification)
- Model lifecycle management (training, deployment, updates)
- Support for edge/fog-based inference
- Synchronization between physical and virtual domains
- Automated machine learning workflows
- Calibration and intelligent data interpretation via ML