Skip to main content

LUP Student Papers

LUND UNIVERSITY LIBRARIES

Unsupervised Anomaly Detection for Humidity Sensor Data

Manasboonpermpool, Pinyapat LU and Baars, Dirk Willeboord LU (2022) DABN01 20221
Department of Statistics
Department of Economics
Abstract
Detecting anomalies in loT sensor devices for humidity is crucial for the maintenance and safety of households and buildings. An Auto-Encoder (AE) can detect anomalies in sensor data. However, the reality is that most sensor data is unlabelled, while it is recommended to use labelled data when training the model for anomaly detection with AE. Therefore together with Sensative AB we developed a Long-Short-Term Memory (LSTM)-AE model that is tested with labelled data and later applied to unlabelled humidity series. We found a detection threshold of 99.5% in the experiment with labelled data is optimal when using LSTM-AE on sensor data. Based on the experiment, a 99.5% threshold is applied to unlabelled humidity series, detecting four... (More)
Detecting anomalies in loT sensor devices for humidity is crucial for the maintenance and safety of households and buildings. An Auto-Encoder (AE) can detect anomalies in sensor data. However, the reality is that most sensor data is unlabelled, while it is recommended to use labelled data when training the model for anomaly detection with AE. Therefore together with Sensative AB we developed a Long-Short-Term Memory (LSTM)-AE model that is tested with labelled data and later applied to unlabelled humidity series. We found a detection threshold of 99.5% in the experiment with labelled data is optimal when using LSTM-AE on sensor data. Based on the experiment, a 99.5% threshold is applied to unlabelled humidity series, detecting four collective anomalies in humidity series 1217 and two collective anomalies in humidity series 2595 with LSTM-AE. (Less)
Please use this url to cite or link to this publication:
author
Manasboonpermpool, Pinyapat LU and Baars, Dirk Willeboord LU
supervisor
organization
course
DABN01 20221
year
type
H1 - Master's Degree (One Year)
subject
keywords
LSTM, Auto-Encoder, Anomaly Detection, Time series, Deep Learning
language
English
id
9086768
date added to LUP
2022-10-10 08:42:56
date last changed
2022-10-10 16:03:26
@misc{9086768,
  abstract     = {{Detecting anomalies in loT sensor devices for humidity is crucial for the maintenance and safety of households and buildings. An Auto-Encoder (AE) can detect anomalies in sensor data. However, the reality is that most sensor data is unlabelled, while it is recommended to use labelled data when training the model for anomaly detection with AE. Therefore together with Sensative AB we developed a Long-Short-Term Memory (LSTM)-AE model that is tested with labelled data and later applied to unlabelled humidity series. We found a detection threshold of 99.5% in the experiment with labelled data is optimal when using LSTM-AE on sensor data. Based on the experiment, a 99.5% threshold is applied to unlabelled humidity series, detecting four collective anomalies in humidity series 1217 and two collective anomalies in humidity series 2595 with LSTM-AE.}},
  author       = {{Manasboonpermpool, Pinyapat and Baars, Dirk Willeboord}},
  language     = {{eng}},
  note         = {{Student Paper}},
  title        = {{Unsupervised Anomaly Detection for Humidity Sensor Data}},
  year         = {{2022}},
}