A Multi-Working States Sensor Anomaly Detection Method Using Deep Learning Algorithms

Di Wu ¹, Kari Koskinen ¹, Eric Coatanea ¹

⁰Faculty of Engineering and Natural Sciences, Tampere University, 33720 Tampere, Finland.

Sensors (basel, Switzerland) +

|

September 27, 2025

View abstract on PubMed

Summary

This study introduces a novel method for sensor anomaly detection and isolation using Long Short-Term Memory (LSTM) networks, effectively handling unlabeled data from varying machine operational states to improve data quality.

Area Of Science

Data Science
Machine Learning
Sensor Technology

Background

Sensor data often contains anomalies due to malfunctions or communication issues.
Existing data-driven methods struggle with inadequately labeled sensor data and changing operational states.

Purpose Of The Study

To propose a sensor anomaly detection and isolation method robust to unlabeled data and varying machine operational states.
To enhance the accuracy and reliability of sensor data processing.

Main Methods

Utilized Long Short-Term Memory (LSTM) networks to predict sensor measurements based on historical data.
Implemented an input-selection strategy based on prediction errors from a small dataset to optimize LSTM performance.
Calculated the residual between predicted and actual measurements to identify anomalies.

Main Results

The proposed LSTM-based method accurately detected drift and stall anomalies in real-world truck sensor data.
The input-selection method improved prediction accuracy and reduced the impact of redundant sensors.
The approach effectively addressed challenges posed by unlabeled data and dynamic operational states.

Conclusions

The developed sensor anomaly detection and isolation method is effective for handling unlabeled data in dynamic environments.
LSTM networks offer a promising approach for improving the integrity of sensor data in industrial applications.
The method demonstrates practical applicability in real-world scenarios, such as mining operations.

Keywords:

LSTM data-driven deep learning sensor anomaly detection