Integrating IoT and machine learning: Benefits and use cases

IoT and machine learning

The increasing complexity of real-time process control applications and the untapped value of historical IoT data drive the integration of IoT and ML. That data helps businesses identify trends and better understand how their products and services are created. What sets ML apart from other forms of artificial intelligence (AI) is its ability to learn from real-time or recorded information specific to the IoT process involved. It doesn't train on outside data, but on the job. This provides significant benefits for enterprises:

• IoT systems can generate a large number of events with insufficient links to the processes they control, making practical analysis difficult. ML algorithms address this challenge by learning to identify patterns in data.
• ML systems can convert raw IoT events into meaningful process conditions, which is key to analyzing and automating complex workflows. Whereas a sensor can signal an event, ML can determine what that event means.
• By analyzing signals from multiple sensors, machine learning models develop a comprehensive view of the overall system: how it creates, moves and stores what a company sells. That deeper awareness enables the ML model to provide real-time insights into the system's state -- for example, normal operation or fault -- and recommendations for restoring or optimizing operations.

A recent innovation in the partnership between IoT and ML uses the digital twin, a model of a real-world process that's synchronized with the process using IoT sensors. The digital twin provides ML systems with context, a way of knowing the structure of a real-time process in advance, which improves the ML model's ability to control complex systems.

The role of training

ML systems require training to perform their tasks effectively. In IoT, training aims to transform raw sensor data into meaningful process conditions.

ML training can take four forms:

1. Human-driven cooperation between an ML expert and a process expert. The ML expert writes rules based on guidance from a subject-matter expert on the real-time process being addressed; this forms the basis of the ML application. This was once the dominant approach, and it's still used to create initial ML models.
2. Using application logs for historical data analysis. This gives the ML model a broad view of how the real-time process works. If the logs properly identify events and control actions, they can also teach the ML model the appropriate responses to changing conditions. This is the preferred way to fully train an ML system, whether as the exclusive mechanism for doing so or as a way of building on an existing model.
3. A rule set is created as a foundation model, using generative AI (GenAI) to analyze a broad range of historical data from other IoT applications. This represents the latest advancements in ML training.
4. Real-time learning from the observation of process operations. This requires an application that receives IoT events and passes control commands and depends on a human operator to generate responses that the ML system learns. This can be used to refine existing models, but it often requires considerable time to accumulate experience on the full range of conditions that could be encountered.

Unlike traditional programmatic IoT systems, which recognize only predefined scenarios, ML models can adapt dynamically to new conditions. Once trained, they can analyze sensor data and provide real-time recommendations.

Historical data analysis

Historical IoT data plays a dual role in ML integration:

• Training and pattern recognition. Historical data gives ML models the foundation they need to recognize patterns that represent conditions and, with the support of experts, link those patterns to recommended actions.
• Analysis and forecasting. Analyzing historical data enables ML models to recognize trends, spot developing conditions requiring intervention and compare the value of alternative actions. For example, historical data can define the state of the process, identify it as valid or indicate a problem and recognize steps to restore invalid states to an identified goal state.

Benefits of integrating IoT and machine learning

ML is an alternative to building IoT applications as event processors and using state-event tables or graphs. ML is superior to either of these programming options for several reasons:

• ML is a more complete rule statement. The biggest problem enterprises report with IoT applications is an incomplete rule set, which means that the developer most likely forgot them. ML models learn, meaning that they build experiences into rules and are far less likely to miss important event sequences.
• ML can easily incorporate foundation models or generalized ML models created for a class of real-time processes. These can then be fine-tuned or trained to the specific needs of a target application.
• ML, if it supports real-time training, can automatically accommodate new events, conditions and IoT devices. With traditional programming, these would require development work.
• ML can accommodate more complex conditions in the control loop without significantly increasing latency.
• ML models can be applied hierarchically, integrating with each other or with AI agents.

Machine learning and IoT applications across industries

ML-enabled IoT applications mix real-time and historical data to inform their recommendations and analyses. How that integration is done, and which specific use cases it supports, varies by vertical market segment.

Healthcare

Healthcare is among the fastest-growing verticals for ML-enabled IoT, including both real-time and non-real-time applications -- although the latter has yielded more use cases so far. Consider the following examples:

• ML analysis of broad medical records, combined with patient-specific data and real-time patient vital signs obtained using IoT, can alert the patient care team to trends that require intervention.
• ML is increasingly used to examine medical images of all types. For example, it has proved highly effective in mammography for spotting cancer early and in difficult situations. Because ML can learn from a radiologist's review of the results, its accuracy improves over time.
• The same information, when collected for historical analysis, can drive reviews and changes to overall patient care protocols, treatment plans, medications and medical equipment. Many of these use cases dovetail with other uses of ML in healthcare, such as analyzing medical imaging, electrocardiogram data and blood test results.

Industrial and manufacturing

Industrial settings are a significant area for IoT and ML use cases and a fast-growing domain for real-time applications. ML-driven process control is more flexible than traditional fixed programming of automated systems.

Manufacturing use cases for IoT and ML include the following:

• Flexible or highly adaptable manufacturing. Because ML algorithms can adapt over time, they help improve process control systems by learning from and adjusting to new conditions.
• Parts and assembly inspection. ML can inspect images of parts or assemblies to spot defects before they're further integrated into products, reducing waste.
• Assembly lines and product transportation. ML has been used in assembly lines and to move supplies and products with autonomous vehicles to improve safety. One manufacturer reported a 22% reduction in insurance claims and rates through this application.
• Production processes and logistics. IoT sensor data collected over time can be combined with real-time data from production processes and logistics, such as manufacturing steps, transportation and storage of parts and finished goods. This can improve efficiency in parts delivery and traditional just-in-time manufacturing practices, thereby reducing costs, vehicle miles traveled and carbon footprint.
• Equipment and vehicle predictive maintenance. ML models can use IoT-reported process activity data to learn long-term relationships and anticipate failures. When paired with historical cost information on past maintenance from core business applications, ML can also help determine when to retire vehicles or equipment.

Utilities management

The utility industry stands out as a committed adopter of IoT, ML and both in combination. This is due to the sector's multifaceted nature, which combines elements from areas including transportation, customer support, regulatory compliance and BPM.

Consequently, utilities involved in distributing or handling power, gas, water and wastewater are a fast-growing area for ML and IoT. These applications combine real-time missions, such as early problem detection, with historical analysis to inform capacity planning, resource allocation and environmental impact management.

Business process management

In BPM, ML and IoT applications typically rely on non-real-time analysis. These systems often repurpose IoT data collected for other applications to track business processes in detail.

BPM has the unique and explicit goal of assembling real-time IoT data that tracks business processes and combining it with other historical company data. Traditional business analytics focuses on accumulated transactional data about a company's operations, which often lacks any link with how it creates, distributes and manages its products and resources.

By combining IoT-derived insights with transactional data analysis, ML algorithms can give decision-makers a more complete picture of operations. This application is also being increasingly augmented by AI agents built on small or large language models.

Tom Nolle is founder and principal analyst at Andover Intel, a consulting and analysis firm that looks at evolving technologies and applications first from the perspective of the buyer and the buyer's needs. By background, Nolle is a programmer, software architect, and manager of software and network products. He has provided consulting services and technology analysis for decades.