2026, Vol.31, No.3

Journal of The Korea Society of Computer and Information 2024 KCI Impact Factor : 0.81

pISSN : 1598-849X / eISSN : 2383-9945
https://journal.kci.go.kr/jksci

This study proposes a time-series model that simultaneously predicts the battery cell average temperature (cTavg) and state of charge (SOC) based on high-frequency (5-second) operational logs from a single residential PV-ESS system operated in a low-temperature environment during winter (2024-12-29 ~ 2025-02-13, total 812,859 records). It compares LSTM encoder-decoder and Transformer-based architectures and suggests a method for early detection of low-SOC/low-temperature events by linking the multi-output predictions to a preventive maintenance logic. The model's performance is evaluated against low-SOC days and low-temperature periods observed in actual data, and the operational benefits provided by the prediction-based alerts are quantitatively discussed.

Cloud-based speech recognition services provide high convenience and accessibility, leading to their widespread adoption across various applications. However, the processing of speech data on remote servers has raised growing concerns about potential privacy breaches stemming from the exposure of speaker information. Although numerous speaker de-identification techniques have been proposed to address this issue, the preservation of semantic information in speech has received relatively little attention. To achieve a balance between speaker de-identification and speech recognition performance, this study proposes an adversarial noise-based approach. The proposed approach is designed based on the differences in input processing between speaker and speech recognition systems. Experimental results show that the proposed method can achieve a meaningful trade-off between speaker de-identification performance and speech recognition performance, with a balance between the two observed particularly in the 10%–20% noise intensity range. These findings suggest that the proposed method can serve as a practical alternative for privacy-preserving speech security in cloud-based speech service environments.

With the proliferation of IoT emphasizing the need for high-performance Intrusion Detection Systems (IDS) in edge environments, deep learning-based IDS research is actively pursued; However, existing approaches that directly utilize tabular data as input for deep learning models are limited in their ability to capture the complex inherent relationships of network traffic. To address this, we propose a 3-stage lightweight IDS framework that converts tabular data into images to leverage the CNN. (1) First, feature selection based on Shapley Additive exPlanations (SHAP) is performed to compress data by retaining only critical features. (2) The selected data is transformed into images using the LLM-categorized Vortex Feature Positioning (LVFP) technique, which reconstructs tabular data into CNN-optimized spatial patterns by assigning semantically categorized feature groups to RGB channels and rearranging them through vortex feature allocation. (3) Finally, we construct a lightweight CNN encoder and pre-train it on the converted images via contrastive learning to establish generalizable feature representations. As downstream tasks, evaluations on 6 IDS & IoT benchmark datasets demonstrate that the proposed model outperforms existing models while using a minimal number of parameters.