20 June 2024 in IEEE Access
Time-series clustering remains a challenge in data mining. Although novel deep-learning-based representation learning integrated with deep clustering methods have considerably enhanced the performance of time-series clustering, efficiently capturing the various temporal patterns inherent in the data is difficult in representation learning for time-series data. In this study, we proposed a novel representation learning method called temporal multi-features representation learning (TMRL) to capture various temporal patterns embedded in time-series data. Based on TMRL, we introduce the temporal multi-features representation clustering (TMRC) framework for performing time-series clustering. The proposed framework decomposes the input time-series data into k temporal patterns and uses k LSTM autoencoders to extract specialized features for each decomposed diverse temporal pattern through TMRL. Variational-mode decomposition is used to extract temporal multi-features. Finally, temporal multi-features derived from TMRL are ensembled for time-series clustering. To evaluate the superiority of the proposed method, comparative experiments were conducted with 36 publicly available time-series datasets against 16 baseline models. In the comparative experiments, we achieved the highest RI and normalized mutual information values in 12 time-series datasets. Particularly, on datasets consisting of eight types of activities and spectral types, the proposed method attained the highest RI and NMI values in six datasets. Furthermore, visualization results of the learned features through TMRL demonstrated superior representation learning compared with existing methods. These results indicated that the proposed TMRC framework is highly suitable for the learning representations of time-series data and can be effectively used for time-series clustering.
6 December 2025 in Transportation Research Part D: Transport and Environment
10 June 2026 in SSRN Electronic Journal
Self-supervised learning for time-series representation learning is dominated by two paradigms: contrastive methods, which face challenges in constructing positive or negative pairs, and masking-based methods, which disrupt the temporal continuity of time-series signals. Joint-Embedding Predictive Architecture (JEPA) offers a promising alternative by predicting in representation space rather than reconstructing raw inputs. However, existing time-series JEPA variants still rely on masking and therefore inherit its continuity problem. Crop-based Forward JEPA (CF-JEPA) is proposed as an innovative mask-free framework that replaces masking with multi-horizon forward prediction: random crops serve as context views, and short-, mid-, and long-horizon future representations are predicted in the forward temporal direction, directly leveraging the inherent temporal ordering of time-series data as a learning signal. A strong asymmetry is also identified between the online encoder and the exponential moving average (EMA) target encoder, both produced from a single training run: the online encoder develops higher-rank discriminative features, while the EMA target encoder develops smoother, lower-rank temporal features. Exploiting this asymmetry, classification is routed to the online encoder and forecasting or anomaly detection to the EMA target encoder, achieving a 27% reduction in multivariate forecasting mean squared error (MSE) at no additional training cost. Across 126 University of California, Riverside (UCR) and 26 University of East Anglia (UEA) classification datasets, eight electricity transformer temperature forecasting benchmarks, and Key Performance Indicator /Yahoo anomaly detection, CF-JEPA achieves the highest average accuracy and rank on UCR and UEA among self-supervised baselines and ranks second on univariate forecasting and k-nearest neighbors-scored anomaly detection.
20 September 2026 in SSRN Electronic Journal
Container terminals identify containers by reading the eleven-character code specified in ISO 6346 from camera images at gates and cranes. On site, this reading must run on edge hardware with limited computation and memory, yet existing pipelines inherit deep backbones, multiscale branches, and sequence-modeling modules from general-purpose detectors and scene text recognizers, capacity the task does not need because the code has a restricted character set and a standardized layout. The contribution to artificial intelligence is an architecture-tailoring strategy that converts such format constraints into removed capacity at design time rather than compressing a large model afterwards; the engineering application is automated container identification in ports. The resulting pipeline pairs a detector that concentrates multiscale structure on the high-resolution features that locate small code regions with a convolution-only recognizer trained by hierarchical knowledge distillation, followed by format-based postprocessing that uses the check digit to stabilize readings. On imagery from an operating terminal, it read 90.2 percent of codes correctly with 0.82 million parameters, 2.12 billion multiply-accumulate operations, and 14.0 milliseconds of processor latency, the lowest cost among 18 compared pipelines and within 1.9 percentage points of the most accurate baseline. Both learned components stayed effective when paired with models they were not designed alongside, and the distilled task-specific model beat post-hoc compression of large models. Detection failed on 1.6 percent of images, recognition errors concentrated at the owner-code prefix and the check digit, and the reported confidence was well calibrated (expected calibration error 0.028), giving terminals an acceptance criterion.
5 November 2026 in Knowledge-Based Systems