Zhang, Shaobo and Guo, Xiao
2026.
A voting-based unsupervised framework for multivariate anomaly detection in vertical farming.
Presented at: 3rd International Conference on Big Data Science and Engineering,
Kunming, China,
12-14 June 2026.
Procedings of ICBDSE 2026.
IEEE,
10.1109/icbdse70225.2026.11635673
|
Preview |
PDF
- Accepted Post-Print Version
Available under License Creative Commons Attribution. Download (416kB) | Preview |
Abstract
In vertical farming systems, anomaly detection is crucial for preventing resource waste, mechanical failure, and safety accidents. However, anomalies are often hidden in high-dimensional multivariate series data, making them difficult to isolate. The lack of labelled data complicates both anomaly detection and the evaluation of detection performance.This paper proposes an unsupervised anomaly detection framework with a pseudo-label-based surrogate classifier for efficient deployment in vertical farming systems. After missing values are addressed, four complementary unsupervised detectors—Anomaly Transformer (AT), Deep Isolation Forest (Deep iForest), One-Class Support Vector Machine (OCSVM), and Clustering-Based Local Outlier Factor (CBLOF)—are combined through majority voting to identify contextual anomalous time steps. The voting results are first converted into pseudo-labels. The pseudo-labels are then used to train a lightweight eXtreme Gradient Boosting (XGBoost) surrogate classifier. This surrogate model approximates the behavior of the voting ensemble and is designed for efficient real-time anomaly screening.Experiments on a real tomato controlled-environment agriculture dataset show that only 2.06%–3.87% of samples are identified as contextual anomalous time steps. After removing the identified time steps, the LSTM shows a clear improvement in forecasting performance, with R2 increasing from 0.129 to 0.675 and RMSE decreasing from 245.57 to 147.63. This change suggests that the detected time steps are associated with reduced downstream forecasting utility. The lightweight XGBoost surrogate reaches a ROC-AUC of 0.95 and a PR-AUC of 0.41 when evaluated against the voting-generated pseudo-labels. These results reflect a high level of agreement with the ensemble decisions, while also maintaining low computational cost.
| Item Type: | Conference or Workshop Item - published (Paper) |
|---|---|
| Date Type: | Published Online |
| Status: | In Press |
| Schools: | Schools > Engineering |
| Additional Information: | RRS policy applied |
| Publisher: | IEEE |
| ISBN: | 9798331552442 |
| Date of First Compliant Deposit: | 2 September 2026 |
| Last Modified: | 02 Sep 2026 13:45 |
| URI: | https://orca.cardiff.ac.uk/id/eprint/189031 |
Actions (repository staff only)
 |
Edit Item |
Dimensions
Dimensions