Anomaly Detection in Smart City Sensor Data Using Isolation Forest and Autoencoder Models

Komara, Indra and Sulaksono, Danang Haryo and Pambudi, Wahyu Setyo and Noche, Elmar B. (2026) Anomaly Detection in Smart City Sensor Data Using Isolation Forest and Autoencoder Models. Journal of Data Science, 2026 (20). pp. 344-367. ISSN 2805-5160

[img] Text
jods2026_20.pdf - Published Version
Available under License Creative Commons Attribution.

Download (232kB)
[img] Text
967 - Published Version
Available under License Creative Commons Attribution.

Download (34kB)
Official URL: http://ipublishing.intimal.edu.my/jods.html

Abstract

Smart city operations depend on continuous sensing across transport, energy, and environmental systems. Although these data streams enable data-driven urban management, they also expose city operators to abnormal patterns caused by sensor malfunction, communication dropout, sudden load changes, and cybersecurity-relevant manipulation of telemetry values. This study develops and evaluates a hybrid anomaly detection framework that combines Isolation Forest and Autoencoder models for multivariate smart city sensor data. Isolation Forest is used to perform scalable initial screening of sparse and high-deviation observations, while an Autoencoder learns nonlinear normal operating patterns and refines alerts through reconstruction error. The framework was assessed on CitySense-AD, a controlled multivariate smart city benchmark containing traffic, energy, environmental, and device-health measurements across multiple urban zones. Model performance was evaluated using precision, recall, F1-score, receiver operating characteristic area under the curve, precision-recall area under the curve, false-positive rate, detection latency, noise robustness, and explanation stability. The hybrid model achieved an F1-score of 0.837 and ROC-AUC of 0.944, outperforming standalone Isolation Forest, standalone Autoencoder, one-class support vector machine, and local outlier factor baselines. Robustness analysis showed that the hybrid model retained a higher F1-score under injected sensor noise and partial missingness, while feature-level reconstruction and isolation contributions provided operationally meaningful explanations for alert triage. The findings indicate that combining tree-based isolation with neural reconstruction can improve anomaly detection reliability in heterogeneous smart city sensor environments without requiring extensive labeled attack data.

Item Type: Article
Uncontrolled Keywords: Anomaly detection; Smart city; Isolation Forest; Autoencoder; Sensor analytics; Sustainable cities and communities
Subjects: Q Science > QA Mathematics > QA75 Electronic computers. Computer science
Q Science > QA Mathematics > QA76 Computer software
T Technology > TD Environmental technology. Sanitary engineering
Depositing User: Unnamed user with email masilah.mansor@newinti.edu.my
Date Deposited: 18 Nov 2026 09:11
Last Modified: 18 Nov 2026 09:11
URI: http://eprints.intimal.edu.my/id/eprint/2391

Actions (login required)

View Item View Item