SenDaL: An Effective and Efficient Calibration Framework of Low-Cost Sensors for Daily Life

TL;DR

SenDaL addresses the challenge of calibrating low-cost IoT sensors within tight resource limits by introducing a novel bottom-up training regime and a top-down inference mechanism. The framework refines data, embeds time-series into a three-layer structure (high-accuracy component, linear, and embedding classifier), and alternates between models to balance accuracy with latency and energy use on embedded hardware. Key contributions include the three-part embedding design, the soft-label classifier training, and the unified fine-tuning that enables efficient top-down inference without sacrificing accuracy. Empirical results on real-world fine-dust sensors across multiple embedded platforms show SenDaL outperforms baselines in accuracy while matching or exceeding the efficiency of linear models, underscoring its practical impact for daily-life IoT calibration.

Abstract

The collection of accurate and noise-free data is a crucial part of Internet of Things (IoT)-controlled environments. However, the data collected from various sensors in daily life often suffer from inaccuracies. Additionally, IoT-controlled devices with low-cost sensors lack sufficient hardware resources to employ conventional deep-learning models. To overcome this limitation, we propose sensors for daily life (SenDaL), the first framework that utilizes neural networks for calibrating low cost sensors. SenDaL introduces novel training and inference processes that enable it to achieve accuracy comparable to deep learning models while simultaneously preserving latency and energy consumption similar to linear models. SenDaL is first trained in a bottom-up manner, making decisions based on calibration results from both linear and deep learning models. Once both models are trained, SenDaL makes independent decisions through a top-down inference process, ensuring accuracy and inference speed. Furthermore, SenDaL can select the optimal deep learning model according to the resources of the IoT devices because it is compatible with various deep learning models, such as long short-term memory-based and Transformer-based models. We have verified that SenDaL outperforms existing deep learning models in terms of accuracy, latency, and energy efficiency through experiments conducted in different IoT environments and real-life scenarios.

Figures (5)

• Figure 1: Overview of SenDaL. Both high-cost and low-cost sensors are used to train SenDaL on a machine through a bottom-up training process. Then, a top-down inference process enables SenDaL calibrate low-cost sensors in IoT-controlled devices with high accuracy and fast inference speed.
• Figure 2: Overall data refinement process.
• Figure 3: Overall structure of SenDaL framework. SenDaL can perform direct inference through either the linear layer (marked in a blue box) or the component layer (marked in a red box).
• Figure 4: Two types of decision-making concept.(a) Embedding layer (Classifier) learns a model determination method using the results of several calibration models, in a bottom-up manner. (b) Once the embedding layer is trained, the embedding layer employs a top-down strategy during the inference process, independently formulating decisions without assistance from any calibration model. (c) Through different training and inference processes, fast and accurate inference results can be made in an actual hardware environment.
• Figure 5: The three-step bottom-up training process of SenDaL. Freeze notation in unified model training denotes a layer in which weight does not change during backpropagation.