Hyun Joon Park

Mr. Hyun Joon Park

School of Industrial and Management Engineering

Research Focus & Metrics

Academic Career Length

6Years (2021-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
25 April 2023 in IEEE Transactions on Aerospace and Electronic Systems
Verified
Journal Article
Closed access
JKJin Sob Kim
Jin Sob Kim
HPHyun Joon Park
Hyun Joon Park
WSWooseok Shin
Wooseok Shin
DPDongil Park
Dongil Park
SHSung Won Han
Sung Won Han
The Automatic Identification System (AIS) has recorded near-real-time vessel monitoring data over the years, paving the way for data-driven maritime surveillance methods; concurrently, the data suffer from unrefined, reliability issues and irregular intervals. In this paper, we address the problem of vessel destination estimation by exploiting the global-scope AIS data. We propose a differentiated data-driven approach recasting a long sequence of port-to-port international vessel trajectories as a nested sequence structure. Based on spatial grids, this approach mitigates the spatio-temporal bias of AIS data while preserving the detailed resolution of the original. Further, we propose a novel deep learning architecture (WAY) that is designed to effectively process the reformulated trajectory and perform the long-term estimation of the vessel destination ahead of arrival with a horizon of days to weeks. WAY comprises a trajectory representation layer and channel-aggregative sequential processing (CASP) blocks. The representation layer produces the multi-channel vector sequence output based on each kinematic and non-kinematic feature collected from AIS data. Then CASP blocks include multi-headed channel- and self-attention architectures, where each processes aggregation and sequential information delivery respectively. Then, a task-specialized learning technique, Gradient Dropout (GD), is also suggested for adopting many-to-many training along the trajectory progression on single labels. The technique prevents a surge of biased feedback by blocking the gradient flow stochastically using the condition depending on the length of training samples. Experimental results on 5-year accumulated AIS data demonstrated the superiority of WAY with recasting AIS trajectory compared to conventional spatial grid-based approaches, regardless of the trajectory progression steps. Moreover, the data proved that adopting GD in a spatial grid-based approach leads to the performance gain. In addition, the possibilities of improvement and real-world application with WAY's expandability in multitask learning for the estimation of ETA was explored.

4 December 2021 in Applied Sciences
Verified
Journal Article
Open access
PDF file
HPHyun Joon Park
Hyun Joon Park
MLMin Seok Lee
Min Seok Lee
DPDong Il Park
Dong Il Park
SHSung Won Han
Sung Won Han
An accurate vessel fuel consumption prediction is essential for constructing a ship route network and vessel management, leading to efficient sailings. Besides, ship data from monitoring and sensing systems accelerate fuel consumption prediction research. However, the ship data consist of three properties: sequential, irregular time interval, and feature importance, making the predicting problem challenging. In this paper, we propose Time-aware Attention (TA) and Feature-similarity Attention (FA) applied to bi-directional Long Short-Term Memory (LSTM). TA acquires time importance by nonlinear function from irregular time intervals in each sequence and emphasizes data depending on the importance. FA emphasizes data based on similarities of features in the sequence by estimating feature importance with learnable parameters. Finally, we propose the ensemble model of TA and FA-based BiLSTM. The ensemble model, which consists of fully connected layers, is capable of simultaneously capturing different properties of ship data. The experimental results on ship data showed that the proposed model improves the performance in predicting fuel consumption. In addition to model performance, visualization results of attention maps and feature importance help to understand data properties and model characteristics.

16 April 2024 in IEEE Signal Processing Letters
Verified
Journal Article
Closed access
HPHyun Joon Park
Hyun Joon Park
WSWooseok Shin
Wooseok Shin
JKJin Sob Kim
Jin Sob Kim
SHSung Won Han
Sung Won Han
To improve real-time speech enhancement (SE) while maintaining efficiency, researchers have adopted knowledge distillation (KD). However, when the same network type as the real-time SE student model is used as a teacher model, the performance of the teacher model can be unsatisfactory, thereby limiting the effectiveness of KD. To overcome this limitation, we propose cross-network non-causal knowledge distillation (CNNC-Distill). CNNC-Distill enables knowledge transfer between networks of different types, allowing the use of a teacher model with a different network type compared to the real-time SE student model. To maximize the KD effect, a non-real-time SE model unconstrained by causality conditions is adopted as the teacher model. CNNC-Distill transfers the non-causal knowledge of the non-real-time SE teacher model to a real-time SE student model using feature and output distillation. We also introduce a time-domain network, RT-SENet, used as the real-time SE student model. Results on the Valentini dataset show the efficiency of RT-SENet and the significant performance improvement achieved by CNNC-Distill.

19 August 2022 in IEEE Access
Verified
Journal Article
Open access
HPHyun Joon Park
Hyun Joon Park
Taehyeong KimTaehyeong Kim
Taehyeong Kim
YKYoung Seok Kim
Young Seok Kim
JMJinhong Min
Jinhong Min
KSKi Woo Sung
Ki Woo Sung
SHSung Won Han
Sung Won Han
Reliability prediction has been studied in many industries for managing stocks and reducing quality assurance costs and production costs. Particularly, in the automotive industry, reliability prediction is performed based on two automobile reliability perspectives, time and mileage. To maximize cost savings, researchers attempted reliability prediction with short-term inputs. However, limited information on short-term inputs resulted in unsatisfactory prediction results for the long warranty periods. Additionally, the overall evaluation metrics could not reflect the pattern-wise performance, such as the increasing failure patterns. This study proposes Complementary Reliability perspective Transformer (CRFormer) based on Transformer encoder to achieve enriched representations from a short-term input sequence. CRFormer fuses different automobile reliability perspective information and automobile features to compensate for the limited information on short-term input. The performance of CRFormer is evaluated based on automobile claim data accumulated over 16 years. Results showed that compared to previous methods in terms of overall, pattern-wise, and pattern similarity evaluation metrics, CRFormer achieved outstanding performance in time and mileage reliability prediction. Lastly, visualization results and survival analysis based on accurate model prediction can be used to support decision-making to reduce quality assurance costs and production costs.

20 January 2026 in IEEE Access
Verified
Journal Article
Open access
JYJeong Yeol Yang
Jeong Yeol Yang
HPHyun Joon Park
Hyun Joon Park
SLSoyoung Lee
Soyoung Lee
SHSung Won Han
Sung Won Han
The extraction of particular sounds from noisy, overlapping mixtures of audio signals remains a significant practical challenge. Universal separation is computationally intractable in modern real-world situations, with query-conditioned extraction emerging as an indispensable targeted approach. Among conditioning signals, natural language exhibits the greatest expressivity, enabling the specification of objects, attributes, and relations without auxiliary references. Although existing contrastive language-audio pre-training (CLAP)-based models enhance language-audio alignment, they typically process hierarchical audio features in a layer-wise manner without modeling their semantic flow explicitly. Thus, their computational cost is identical to that of Transformer blocks. In this paper, we introduce CLAPMamba, a hybrid architecture that enhances conditioning performance while reducing the computational overhead simultaneously. In QueryNet, text embeddings enable conditioning for CLAP audio features, and the Mamba integration stage treats layer outputs sequentially to produce a single, structured conditioning vector that preserves hierarchical semantics. In SeparationNet, a Mamba Forward-Attention Block (MFAB) is used to combine the global context of self-attention with linear-time sequence modeling based on a bidirectional Mamba, capturing long-range dependencies more efficiently. Trained on AudioCaps and evaluated on multiple datasets, including AudioSet, ESC-50, FSDKaggle2018, and MUSIC21, CLAPMamba matches or outperforms Transformer baselines while incurring lower resource costs. Its SDRi and SI-SDRi are higher than those of CLAPSep by 0.25–0.46 and 0.09–0.85 dB, respectively, while using fewer parameters and MACs and exhibiting lower inference latency. Moreover, it increases semantic consistency, as reflected by text–audio similarity and layerwise alignment analyses. Ablation studies indicate complementary improvements based on Mamba-based conditioning in QueryNet and the forward-attention module in SeparationNet. Overall, the method exhibits high generalizability across datasets and sequence lengths, and achieves consistent separation quality with an efficiency suitable for on-device scenarios and latency-sensitive applications.

31 October 2022 in arXiv
Verified
Preprint
Open access
PDF file
WSWooseok Shin
Wooseok Shin
HPHyun Joon Park
Hyun Joon Park
JKJin Sob Kim
Jin Sob Kim
LHLee Byung Hoon
Lee Byung Hoon
SHSung Won Han
Sung Won Han
Recent deep learning models have achieved high performance in speech enhancement; however, it is still challenging to obtain a fast and low-complexity model without significant performance degradation. Previous knowledge distillation studies on speech enhancement could not solve this problem because their output distillation methods do not fit the speech enhancement task in some aspects. In this study, we propose multi-view attention transfer (MV-AT), a feature-based distillation, to obtain efficient speech enhancement models in the time domain. Based on the multi-view features extraction model, MV-AT transfers multi-view knowledge of the teacher network to the student network without additional parameters. The experimental results show that the proposed method consistently improved the performance of student models of various sizes on the Valentini and deep noise suppression (DNS) datasets. MANNER-S-8.1GF with our proposed method, a lightweight model for efficient deployment, achieved 15.4x and 4.71x fewer parameters and floating-point operations (FLOPs), respectively, compared to the baseline model with similar performance.