📄 DeepFeatIoT: Unifying Deep Learned, Randomized, and LLM Features for Enhanced IoT Time Series Sensor Data Classification in Smart Industries
Official Implementation of "DeepFeatIoT" accepted at IJCAI 2025 under the special track of AI4Tech: AI Enabling Critical Technologies as a full-length original research article after a double blind peer-review process. This year the acceptance rate of full-length research articles were less than 18%. IJCAI is a CORE A* conference and a flagship international research conference in Artificial Intelligence.
Internet of Things (IoT) sensors are ubiquitous technologies deployed across smart cities, industrial sites, and healthcare systems. They continuously generate time series data that enable advanced analytics and automation in industries. However, challenges such as the loss or ambiguity of sensor metadata, heterogeneity in data sources, varying sampling frequencies, inconsistent units of measurement, and irregular timestamps make raw IoT time series data difficult to interpret, undermining the effectiveness of smart systems. To address these challenges, we propose a novel deep learning model, DeepFeatIoT, which integrates learned local and global features with non-learned randomized convolutional kernel-based features and features from large language models (LLMs). This straightforward yet unique fusion of diverse learned and non-learned features significantly enhances IoT time series sensor data classification, even in scenarios with limited labeled data. Our model's effectiveness is demonstrated through its consistent and generalized performance across multiple real-world IoT sensor datasets from diverse critical application domains, outperforming state-of-the-art benchmark models. These results highlight DeepFeatIoT's potential to drive significant advancements in IoT analytics and support the development of next-generation smart systems.
We introduce DeepFeatIoT, which uniquely integrates: learnable local (multi-scale learnable convolutions) and global feature (bi-directional gated recurrent units) with non-learned randomized convolutional kernel features, and pretrained LLM features to enhance classification performance across various IoT sensor datasets, even in scenarios with limited labeled data. Our approach achieves state-of-the-art performance on heterogenous IoT Time Series Sensor data classification task.
.
├── datasets/ # Contains datasets
├── scripts/ # Source code
├── paper/ # paper related other resources
├── requirements.txt # Python dependencies
└── README.md
Clone this repository and install dependencies:
git clone https://github.com/skinan/DeepFeatIoT-IJCAI-2025.git
cd DeepFeatIoT-IJCAI-2025Conda environment setup
conda create -n deepfeatiot python=3.9.0
conda activate deepfeatiot
pip install -r requirements.txt
We evaluate our method on the following datasets (see paper/appendix.pdf for more details):
- Swiss Experiment (abbvr. as Swiss) – This dataset [Montori et al., 2023] contains highly noisy time series sensor readings from sensors located within the Swiss Alps mountain range [Calbimonte et al., 2012]. The sensors in this dataset come from heterogeneous environments, exhibiting variability in timestamp ranges, frequency, and sampling ratios, along with significant phase shifts in the data. This is one of the few highly heterogeneous IoT sensor time series datasets available in the domain. It serves as a practical example of the heterogeneity present in ubiquitous real-world sensor data, which is used for various sensor applications in different industries (e.g. traffic, media etc.) [Calbimonte et al., 2012]. The class-labels available in this dataset are: CO2(Carbon Dioxide), Humidity, Lysimeter, Moisture, Pressure, Radiation, Snow Height, Temperature, Voltage, Wind Speed, and Wind Direction.
- Urban Observatory (abbvr. as Urban) - This dataset [Montori et al., 2023] contains one day of data from a city-wide urban sensor network, initiated by the University of Newcastle, United Kingdom. The dataset includes real-time sensor time series from multiple domains, such as smart buildings, traffic control, weather stations, remote sensing, and more. It consists of a set of highly correlated data. The sensor types availables on this dataset are: NO2 (Nitrogen Dioxide), Wind Direction, Humidity, Wind Speed, Temperature, Pressure, Wind Gust, Rainfall, Soil Moisture, Average Speed, Congestion, Traffic Flow, Journey time, Sound, CO (Carbon Monoxide) and NO (Nitrogen Monoxide).
- Iowa ASOS (abbrv. as Iowa)- This dataset [Inan et al., 2023] comprises time series sensor readings with an hourly frequency and a total of one week of data for each sample, spanning a six-month period overall. These sensors are deployed in airports and can produce observations every minute or every hour, depending on the requirements of the airport authority, to support aviation opera- tions and facilitate smart airport or aviation management using Automated Surface Observing Systems (ASOS). The sensor types or class-label available in this dataset are: Air Temperature, Dew Point Temperature, Relative Humidity, Wind Direction, Pressure Altimeter, Visibility, Wind Gust, and Apparent Temperature (Heat Index).
- Smart Building Automation System (abbrv. as SBAS) - This dataset [Hong et al., 2017] contains time series sensor data (resampled at a frequency of 1 hour) derived from sensors installed in 51 rooms of the Sutardja Dai Hall (SDH) at UC Berkeley. These sensor data can support investigations of the physical properties of rooms within the building. This dataset contain 5 unique type of sensors including CO2 concentration, room air humidity, room temperature, luminosity, and PIR motion sensor.
Main results from the paper:
If you use this code or any resources of the paper (including datasets), please kindly cite our paper:
@inproceedings{ijcai2025p1025,
title = {DeepFeatIoT: Unifying Deep Learned, Randomized, and LLM Features for Enhanced IoT Time Series Sensor Data Classification in Smart Industries},
author = {Inan, Muhammad Sakib Khan and Liao, Kewen},
booktitle = {Proceedings of the Thirty-Fourth International Joint Conference on
Artificial Intelligence, {IJCAI-25}},
publisher = {International Joint Conferences on Artificial Intelligence Organization},
editor = {James Kwok},
pages = {9223--9231},
year = {2025},
month = {8},
note = {AI4Tech: AI Enabling Technologies},
doi = {10.24963/ijcai.2025/1025},
url = {https://doi.org/10.24963/ijcai.2025/1025},
}This repository also contains the DeepHeteroIoT model, published in MobiQuituous 2023. The DeepHeteroIot model code could found in the "scripts/models" directory. Please, kindly cite the DeepHeteroIoT paper, if you use Iowa dataset or DeepHeteroIoT model:
@inproceedings{inan2023deepheteroiot,
title={DeepHeteroIoT: Deep Local and Global Learning over Heterogeneous IoT Sensor Data},
author={Inan, Muhammad Sakib Khan and Liao, Kewen and Shen, Haifeng and Jayaraman, Prem Prakash and Georgakopoulos, Dimitrios and Tang, Ming Jian},
booktitle={International Conference on Mobile and Ubiquitous Systems: Computing, Networking, and Services},
pages={119--135},
year={2023},
organization={Springer Nature Switzerland}
}