Scholarlink

<p>Low-cost environmental sensors are increasingly used for real-time monitoring, but they often suffer from limitations such as accuracy, calibration drift, short lifespan, and sensitivity to environmental factors. These issues are particularly evident in airborne pollen monitoring devices like Beenose, which provide live high-frequency data yet frequently deviate from manual reference methods such as Hirst traps. To address some of these challenges, we propose a deep learning framework for anomaly correction in multivariate time series collected from Beenose sensors, using Hirst trap measurements as a free anomaly reference. Our approach extends LSTM-based autoencoders with a penalized loss function that explicitly integrates the reference data. The main contributions are: (i) the introduction of data augmentation strategies to address limited training data, and (ii) the adaptation of an LSTM-based autoencoder trained with a regularized loss function. Ablation experiments show that the main performance gain comes from the reference-based regularization term, while the architectural components play a complementary role.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Regularized LSTM models for cleaning multivariate time series from low-cost sensors for pollen detection in the air

  • 期刊论文

  • 发表日期：

    2026年6月30日

查看全文
  • M’Hammed Oudrane,
  • Pierre Houedry,
  • Valérie Monbet,
  • Johann Lauthier,
  • Houssam Al Azari

摘要

Low-cost environmental sensors are increasingly used for real-time monitoring, but they often suffer from limitations such as accuracy, calibration drift, short lifespan, and sensitivity to environmental factors. These issues are particularly evident in airborne pollen monitoring devices like Beenose, which provide live high-frequency data yet frequently deviate from manual reference methods such as Hirst traps. To address some of these challenges, we propose a deep learning framework for anomaly correction in multivariate time series collected from Beenose sensors, using Hirst trap measurements as a free anomaly reference. Our approach extends LSTM-based autoencoders with a penalized loss function that explicitly integrates the reference data. The main contributions are: (i) the introduction of data augmentation strategies to address limited training data, and (ii) the adaptation of an LSTM-based autoencoder trained with a regularized loss function. Ablation experiments show that the main performance gain comes from the reference-based regularization term, while the architectural components play a complementary role.