Artificial intelligence (AI) could turn NB-IoT from a simple data pipe into a network that predicts, adapts, and acts on its own, a new review suggests.
According to a paper published in Springer’s Journal on Wireless Communications and Networking, AI lets these low-power networks handle spectrum, energy, and data management without constant human control. The studies they analysed report gains in accuracy, speed, and energy efficiency – from farms that use less water to wearables that flag heart rhythm problems at home and factories that catch faults before machines break down.
The paper was authored by seven experts from Indian institutions, including Ramdeobaba University and the Manipal Institute of Technology. It analyses recent advances in applying AI to NB-IoT across agriculture, healthcare, industry, and smart cities.
One point to note, though, is that most of the AI-powered NB-IoT systems surveyed are still prototypes. And, as the authors themselves admit, the industry still lacks a common yardstick to measure how much AI actually helps.
From Tomato Fields to Heartbeats
The review rounds up dozens of studies across agriculture, healthcare, industry, and smart cities. Some of the numbers are eye-catching.
In farming, a machine learning-guided irrigation system for processing tomatoes used 32.66% less water than full irrigation with no yield loss. A soil-monitoring system sent 84 days of field data over NB-IoT with a 95.87% communication success rate. In simulations, a single drone could serve more than 2,000 underground sensors across 20 hectares, with sensors lasting over six years.
Healthcare offers another telling example. A wearable NB-IoT electrocardiogram sensor detected atrial fibrillation with a specificity of 0.99 and a sensitivity of 0.71. In other words, it rarely raises a false alarm, but misses roughly three in ten cases.
Still, the authors see promise. The result suggests “low-bitrate NB-IoT communication is sufficient for arrhythmia monitoring in a home setting,” they observed.
On the industrial side, a TinyML model combined with federated learning read old mechanical dial meters with around 97% accuracy. It consumed only 87.5 µWh per day – a cheap way to turn legacy meters into smart ones. In smart cities, deep reinforcement learning was used to tune NB-IoT access control and keep congestion close to optimal levels in dense deployments.
Why NB-IoT Needs a Brain
NB-IoT was built to last. The technology standardised by the 3rd Generation Partnership Project (3GPP) squeezes into just 180 kHz of bandwidth, reaches deep indoors with a link budget of up to 164 dB, and can keep a sensor running on the same battery for more than ten years.
Flexibility, however, was never its strong suit. “It also has a poor reaction time and is not adaptable to changing situations,” the authors noted. Static resource allocation and rigid scheduling make traffic spikes hard to handle, while saving energy usually comes at the expense of latency.
The proposed fix is to push intelligence as close to the device as possible. Edge AI processes data on the device or at a nearby gateway, so only alerts or summaries cross the network. TinyML goes further, running compact models on microcontrollers with less than 256 KB of RAM.
Federated learning, in turn, lets devices train a shared model without handing over raw data. That is good news for privacy, but not necessarily for the network. “Model updates do not necessarily consume less bandwidth than sensor data uploads,” the authors warned.
Stuck Between the Lab and the Field
So, is AI-powered NB-IoT ready for prime time? Not quite, by the authors’ own assessment.
The review grades each domain using technology readiness levels (TRL), a nine-step scale running from basic research (level 1) to proven commercial deployment (level 9). Agriculture and smart city projects sit between levels 3 and 6. Industrial solutions range from 4 to 6, and healthcare from 4 to 7.
Most healthcare systems “have been tested on datasets, prototypes, or small field trials but are not ready for full-scale deployment yet,” the experts pointed out.
Reaching levels 8 and 9 will require more than better algorithms. According to the review, isolated solutions will need to come together into digital twin systems with shared data models, security protocols, and lifecycle tools.
Show Me the Benchmarks
The list of hurdles is long. Devices remain computationally limited, and over-the-air model updates add to bandwidth and energy bills. Data are another headache: failures, attacks, and leaks are rare, so training sets are dominated by “normal” behaviour.
The result, according to the authors, is models that look good on paper but “fail miserably when it comes to detecting rare, but critical events.” Security is also a concern, with threats ranging from poisoned model updates to inputs crafted to trick AI into wrong decisions.
Another big gap may be measurement. With no standardised performance metrics, “it is rather difficult to compare solutions of vendors,” the review acknowledged. The authors call for benchmarks that report median and tail latencies, data sent, and packet delivery rates under defined hardware and network settings.
“Without such comparisons, the architectural advantages must be considered qualitative instead of actual performance gains,” they admitted.
None of this dampens the long-term vision. On the path to AI-native 6G, the authors expect NB-IoT devices to “not only sense and connect, but also learn, adjust, and take actions at the edge of the network.”
