Search Results (574)

We set out to fix a “contamination” problem in reconstruction-based Temporal Convolutional Network VAEs (TCN-VAEs) for cross-dataset IoT flow anomaly detection: when attack flows share an encoder window with benign flows, the shared latent code is allegedly distorted, inflating benign reconstruction error and producing false positive rates (FPRs) of 22–65% despite an ROC-AUC above 0.93. Our proposed fix, TCN-Pred, excludes the target flow from the encoder and scores it by next-flow prediction error, reducing FPR to 0.65–13%. We subjected this causal explanation to a battery of controlled ablations, holding architecture, decoder, loss, and thresholding fixed while varying one factor at a time. Each one falsified the original hypothesis: target inclusion/masking changes FPR by at most 0.001; context shuffling/reversing/zeroing changes it by at most 0.003; a context-blind constant-output predictor matches TCN-Pred’s FPR and F1 to three decimal places on all three datasets. The actual cause, confirmed on the original trained models with no retraining, is a scoring-granularity mismatch: the TCN-VAE threshold is calibrated from per-window errors averaged over 20 flows but applied to per-flow errors at evaluation (standard deviation $\approx \sqrt{20}\times$ higher, measured ratio 4.46 against a predicted 4.47). Recalibrating the identical model at matching granularity drops FPR from 22.7/47.6/64.6% to 0.65/5.0/12.5% on BoT-IoT, IoT-23 and ToN-IoT, closing 89–97% of the reported FPR gap without changing a single model weight. We report this diagnostic chain, together with an attack-prevalence sensitivity analysis, sample-disjoint calibration, normality diagnostics, and label-free and redundancy-aware (mRMR) feature-selection benchmarks, as a methodology other work should apply before attributing fixed-threshold performance to architecture. The pipeline is supervised source-domain feature selection followed by benign-only detector training, not fully unsupervised, a distinction we quantify later in the paper. Investigating dataset representativeness, we found that all three provided files reduce to only ≈6000 genuinely distinct flows via an undocumented row-duplication procedure, causing 97.8% BoT-IoT train/test near-duplicate overlap; a leakage-free re-evaluation changes FPR by only 0.23 percentage points. We also found that the TLS-metadata columns are already transformed upstream of every available artefact, so the proportion of genuinely TLS-encrypted flows cannot be recovered, and we soften the paper’s encrypted-traffic framing accordingly. Full article

Pathological gait classification plays an important role in objective motor function assessment, early clinical screening, and rehabilitation evaluation. However, traditional clinical gait assessment methods are highly dependent on expert experience and may fail to detect subtle gait abnormalities. Moreover, existing inertial measurement unit (IMU)-based gait recognition methods often rely on single-sensor configurations or single-scale temporal models, limiting their ability to capture complex pathological gait patterns. In this study, a convolutional neural network–bidirectional long short-term memory–temporal convolutional network (CNN-BiLSTM-TCN) multi-branch feature fusion framework was proposed for pathological gait classification using a publicly available clinical multi-inertial measurement unit dataset containing 260 subjects. The proposed model employs three parallel branches to extract local instantaneous motion variations, continuous temporal dynamics, and relatively broader temporal dependencies within the 2 s input window, respectively, followed by feature-level fusion and end-to-end joint optimization. Experimental results show that the proposed model achieves a test accuracy of 0.9818 and an F1-score of 0.9700, outperforming conventional machine learning methods, single-branch models, voting-based fusion methods, and other temporal models, including Support Vector Machine (SVM), Temporal Convolutional Network (TCN), and Convolutional Neural Network-long short-term memory (CNN-LSTM). Five repeated experiments with stratified random splits demonstrate minimal performance variation, indicating good robustness and stability. The proposed framework provides a potential approach for pathological gait screening and quantitative rehabilitation assessment. Full article

Multi-site water quality sequences are jointly affected by seasonal periodicity, meteorological disturbances, and inter-site differences in the Jianghuai Watershed region. Conventional quality prediction models struggle to simultaneously achieve multi-scale feature extraction, spatial heterogeneity characterization, and prediction uncertainty expression. This study used daily-scale monitoring data on dissolved oxygen (DO), pH, and ammonia nitrogen (NH3N) from 32 monitoring stations within the region in 2025 and proposed the FT-TransONet (Fourier-enhanced Temporal Transformer Operator Network) multi-site probabilistic water quality prediction model. Within a Transformer framework, the model employed a FourierTime learnable frequency-domain filtering module, a GeoBias (Geographic Bias) attention bias mechanism, and a multi-residual ensemble strategy composed of a multilayer perceptron (MLP), a gated recurrent unit (GRU), and a temporal convolutional network (TCN) combined with a mass conservation constraint, thereby achieving both point and interval prediction of key water quality indicators. The results showed that FT-TransONet achieved the lowest Macro_RMSE among all compared methods on the multi-site water quality prediction task. At a prediction horizon of three days, its Macro_RMSE reached 0.2255, which was 21.89% lower than that of the long short-term memory network and 5.57% lower than that of the strongest baseline MC-Dropout. For the three individual indicators, the model attained coefficients of determination of 0.9135, 0.9338, and 0.8660 for dissolved oxygen, pH, and ammonia nitrogen, with corresponding root-mean-square errors of 0.5145, 0.1098, and 0.0523, confirming its potential to characterize the temporal variation in the main water quality indicators. Under multi-step prediction, the error grew gently, with the Macro_RMSE rising only from 0.2255 to 0.2384 as the horizon extended from three to seven days, and the ablation experiments, together with the probabilistic prediction results, further supported the effectiveness of the proposed structural design. Validated on 32 water quality monitoring stations in the Jianghuai Watershed, the method improved multi-site prediction accuracy while accounting for stability and uncertainty quantification, providing a preliminary reference for regional water quality early warning and management. Full article

Falls are a major cause of severe injury and mortality among older adults, requiring rapid state recognition and alerts in healthcare and caregiving environments. However, many existing fall-recognition approaches primarily focus on post-fall detection or assume access to cloud/GPU computing, leaving limited evidence for short-term pre-fall recognition on low-end CPU-based edge devices. This study proposes a lightweight CNN–TCN–Attention framework for recognizing fall and pre-fall states on resource-constrained edge devices. Using MediaPipe, three-dimensional coordinates and visibility scores of 33 human body landmarks are extracted from input videos. Raw RGB frames are not directly used by the classifier, thereby reducing the processing of visually identifiable information. A total of 2300 fall-related videos from the AI-Hub dataset were reorganized into three classes: normal, pre-fall, and fall. The pre-fall onset was defined as the point at which the downward vertical velocity of the nose landmark exceeded a dataset-specific threshold. The model uses a CNN to extract frame-level skeletal patterns, a temporal convolutional network (TCN) to learn temporal changes in posture, and an attention module to aggregate temporal features. It was deployed on a Raspberry Pi Zero 2 W using ONNX Runtime. The proposed model achieved 94.71% accuracy and a macro F1-score of 93.63%, with an average state-decision latency of 325.88 ms/decision. It improved accuracy by 2.83 percentage points over the lightweight CNN–LSTM–Attention baseline while maintaining comparable latency and achieved approximately 1.75× faster processing than the MobileNetV3–TCN–Attention model. Full article

Full-field water saturation is central to waterflood surveillance but cannot be observed continuously across a reservoir, whereas well histories provide sparse dynamic evidence. We formulate target-time saturation reconstruction as a mapping from static geology, well locations, scheduled controls, simulated multi-well production responses, and development time to a two-dimensional saturation field. The water-balance-guided saturation network (WB-SatNet) combines a U-Net spatial pathway, a fixed-order gated recurrent unit (GRU) history encoder, explicit time conditioning, and a closed-boundary water-storage consistency term. Experiments used 400 geological realizations, 800 simulation runs, six target times, realization-wise train/validation/test splitting, and three random seeds. On the held-out test set, WB-SatNet achieved a mean absolute error (MAE) of 0.01175, a coefficient of determination (${\mathrm{R}}^2$) of 0.98145, a structural similarity index measure (SSIM) of 0.99177, a flooded-area intersection over union (IoU) of 0.95523, and a global storage-consistency error of 0.00944. Its mean MAE was 6.31% lower than that of TCN-U-Net, the strongest temporal convolutional network baseline. Target-time, component-ablation, flooded-area, storage-consistency, history-window, noise, and operating-regime analyses support a monitoring-oriented interpretation. These results indicate that WB-SatNet provides an effective framework for production-history-conditioned water-saturation reconstruction and waterflood state monitoring in the investigated setting. Full article

Wind power interval forecasting remains challenging due to the uncertainty and strong variability of wind generation. To capture temporal dependency and predictive uncertainty, this paper proposes a hybrid interval forecasting framework that integrates an Informer-based point prediction model with a temporal convolutional network (TCN) conditional quantile model. The Informer is used to generate deterministic forecasts, while the TCN models the temporal dependency of prediction residuals and estimates conditional quantiles for interval construction. To further improve interval quality, an improved osprey optimization algorithm (IOOA) is introduced to optimize key TCN hyperparameters. The Coverage–Width Criterion (CWC) on the validation set is adopted as the optimization objective for hyperparameter tuning and adaptive quantile-pair selection. To maintain the nominal 90% confidence level, candidate quantile pairs are constrained to have a fixed quantile span of 0.90. Experiments on real-world wind power datasets demonstrate that, when averaged across the two wind farms, the proposed framework achieves a prediction interval coverage probability (PICP) of 0.910, satisfying the nominal coverage level of 90%, and a mean prediction interval width (MPIW) of 7.48, the lowest among all compared methods. Specifically, it reduces the mean interval width by 8.89–28.35% relative to the benchmark models, indicating that the proposed framework generates sharper prediction intervals without compromising coverage reliability and achieves a better trade-off between reliability and sharpness. Full article

Feature engineering has become a common step in AI-based HVAC forecasting, often involving variables calculated from raw building management system (BMS) measurements, such as temperature differences, setpoint tracking deviations, airflow balance indicators, rolling statistics, and temporal or seasonal descriptors. Accurate short-term forecasting can provide a baseline of expected operation for anomaly and fault detection and can support control optimization and operator decision making. However, real-world deployment is complicated due to differences in BMS sensor availability and data quality, as well as the preprocessing and maintenance burden associated with complex feature sets. The actual contribution of these features to the performance of AI forecasting remains underexplored, particularly for short-term prediction of air handling unit (AHU) operation. This study evaluates the impact of features on short-term AHU forecasting using three deep learning (DL) architectures: Temporal Convolutional Networks (TCNs), Long Short-Term Memory (LSTM) networks, and a hybrid CNN–LSTM model. An actual operational AHU dataset from a BMS was used to predict key operational variables, including supply and extract air temperatures, supply and extract fan operating signals, and supply air temperature setpoint-tracking error. Fan signal balance was additionally evaluated as a derived indicator calculated from the two predicted fan signals. Four input configurations were evaluated: (i) full (74 inputs), containing raw BMS measurements, short-cycle temporal variables, engineered and dynamic features, and annual-calendar information; (ii) no annual calendar (68 inputs), identical to full but excluding annual-calendar variables; (iii) raw + short-cycle temporal (20 inputs); and (iv) raw-only (12 inputs). The models used a 60-min input history to forecast the following 30-min at one-minute resolution. Persistence and Ridge models were included as reference baselines. All models were trained and tested on identical data splits and forecasting horizons to ensure a fair comparison. Each DL experiment was repeated across five independent runs, and performance was evaluated using MAE, RMSE, and R². The TCN showed the strongest overall DL performance. Raw-only achieved the highest mean R² in 11 of 15 architecture–target comparisons using just 12 inputs. The best mean DL R² ranged from 0.916 for the fan signals to 0.993 for extract air temperature. Annual-calendar features improved the TCN results but provided no consistent benefit for the LSTM or CNN–LSTM. Ridge slightly outperformed the best DL configurations for temperature-related targets, reflecting the strong short-term continuity of these signals. These findings show that recent raw BMS measurements contain most of the information needed for accurate 30-min AHU forecasting, while explicit seasonal and engineered features provide limited additional value. The resulting simpler models may in the future be used as forecasting components in predictive control and fault detection systems. However, their control and energy-saving benefits must be tested separately. Full article

Ocean sound speed profile (SSP) is a key parameter for underwater acoustic detection, remote sensing, and seafloor geodetic positioning, and its temporal prediction is essential for improving acoustic positioning accuracy. Conventional direct measurements are inefficient and spatially sparse, while statistical and acoustic inversion methods fail to capture the strong nonlinear evolution of the sound speed field. Among existing time series models, LSTM, a recurrent network for time series forecasting, lacks an explicit receptive field. In contrast, the original TCN, a temporal convolutional network with dilated convolutions, poorly captures local fine structures and relies heavily on empirical tuning. To overcome these limitations, we propose an improved TCN-based SSP prediction method and apply it to seafloor geodetic positioning. The approach first constructs a sound speed increment field via first-order time differencing to remove global trends and highlight local variations. It then employs Optuna (version 4.9.0), a Bayesian sampling-based automatic optimization framework, to automatically tune key TCN parameters within a predefined search space, reducing reliance on manual tuning. The predicted high-resolution sound speed time series is finally used for ray tracing positioning to enhance seafloor geodetic accuracy. Experiments on the GLORYS12V1 reanalysis dataset show that LSTM and the original TCN achieve root mean square error (RMSE) and mean absolute error (MAE) values of 0.414 and 0.299 m/s, as well as 0.360 and 0.258 m/s, respectively, whereas our improved TCN reduces these to 0.205 and 0.131 m/s, substantially outperforming both baselines. In simulated Global Navigation Satellite System–Acoustics (GNSS-A) seafloor positioning, the 3D positioning RMSE drops to about 0.075 m, with improved stability. The proposed method offers an effective solution for accurate SSP time series forecasting and high-precision seafloor geodesy. Full article

Power cable tunnels are typical enclosed cable-routing spaces in which cable overheating, insulation aging, and partial discharge may gradually develop into fire hazards. During the early stage of cable fires, abnormal sensor responses are often weak, localized, and continuously evolving, which increases the difficulty of early warning based on a single sensor or a single temporal feature. Motivated by cable fire early warning in power cable tunnels, this study uses cable-fire records from a publicly available indoor EN 54 fire-test-room dataset with distributed multi-sensor nodes to evaluate the proposed model under controlled laboratory conditions. In monitoring scenarios with fixed sensor nodes, temporal-only modeling methods often struggle to simultaneously characterize short-term variations, temporal evolution, and spatial differences in node responses. To address this limitation, this study proposes a multi-sensor spatiotemporal feature fusion model that integrates a gated recurrent unit (GRU), a Modern Temporal Convolutional Network (ModernTCN), and an enhanced graph convolutional network (GCN+). The proposed model adopts the ModernTCN as the temporal modeling backbone. A GRU module is introduced at the front end to encode local fluctuations and short-term continuous changes between consecutive time steps, while GCN+ is embedded at the intermediate feature stage of the backbone to model spatial correlations and cross-node coordinated responses among fixed sensor nodes. Experimental results show that the proposed model achieves strong classification performance in the cable fire early warning discrimination task, with a test accuracy of 0.9884 and a false negative rate (FNR) reduced to 0.0150. The comparative experimental results indicate that the proposed model achieves better overall performance than typical temporal baseline models. The ablation study further shows that, under the experimental settings of this study, the introduction of GRU and GCN+ leads to overall improvements in the main evaluation metrics, suggesting that both modules provide a certain enhancement to the cable fire early warning discrimination performance of the ModernTCN backbone. Full article

Autonomous docking is a core capability enabling full autonomy of unmanned surface vehicles (USVs), whose practical deployment demands visual pose estimation with high efficiency, temporal stability, and closed-loop control compatibility. This paper proposes a lightweight monocular visual docking perception framework based on MobileNetV2 and a temporal convolutional network (TCN). In this framework, a MobileNetV2 backbone is adopted to perform end-to-end regression of the USV’s relative pose with respect to the dock from a single monocular image, while a feature-level TCN module fuses sequential visual features across consecutive frames to enhance the short-term stability of pose estimation. To validate the performance and reliability of the proposed method, a high-fidelity simulation environment is established to conduct closed-loop USV docking tests. Comparative results demonstrate that the MobileNetV2 backbone reduces inference latency compared with the VGG19 architecture, and the embedded TCN module effectively suppresses inter-frame pose fluctuations and abnormal estimation jumps. The proposed method provides an efficient and temporally consistent visual perception solution for simulation-validated USV autonomous docking systems. Full article

Bridge settlement prediction is essential for understanding structural deformation evolution and supporting bridge condition assessment. However, settlement monitoring sequences usually exhibit nonlinear characteristics, non-stationarity, and multi-scale fluctuations, which increase the difficulty of accurately capturing deformation evolution trends. This study develops an integrated settlement prediction and dynamic early warning framework for concrete bridges by combining Variational Mode Decomposition (VMD), Temporal Convolutional Network (TCN), and Bidirectional Long Short-Term Memory Network (BiLSTM). In the proposed framework, VMD is employed to decompose settlement monitoring sequences into multiple components with different frequency characteristics, TCN is used to extract local temporal variation features, and BiLSTM is applied to capture long-term temporal dependencies for settlement prediction. Based on the predicted settlement responses, settlement increment, settlement rate, and differential settlement indicators are further incorporated to establish a dynamic early warning method for deformation risk identification. Continuous monitoring data from an in-service reinforced concrete bridge were used for validation. The results show that the proposed framework achieves better prediction performance than benchmark models, with an R² of 0.956, MAE of 0.38 mm, RMSE of 0.51 mm, and MAPE of 1.52%. Furthermore, simulated abnormal settlement scenarios were constructed to evaluate the response capability of the warning framework under different deformation conditions. The results demonstrate that the proposed framework can effectively integrate settlement prediction and deformation risk assessment, providing a potential approach for bridge settlement monitoring and early warning analysis. Full article

This study addresses the challenge of incomplete indoor environmental monitoring data under alternate-floor sensor deployment in multi-story residential buildings. To enable cost-effective environmental sensing, a Building Environmental Data Reconstruction Framework (EDRF) is proposed for estimating unmonitored floor conditions. The EDRF integrates a convolutional neural network (CNN) for spatial feature extraction, a temporal convolutional network (TCN) for temporal dependency modeling, residual connections for stable feature propagation, and a multi-task learning (MTL) strategy for simultaneous reconstruction of multiple environmental variables. The model is trained and validated using real-world data collected from Floors 2–10 of a residential building, focusing on illuminance, temperature, and relative humidity. Experimental results demonstrate that the proposed framework achieves high reconstruction accuracy, with average R² values of 0.992 for temperature and 0.988 for relative humidity. Even under a reduced sensor deployment rate of 33.3%, the model maintains robust performance with an overall R² of 0.987. Ablation studies further confirm the effectiveness of each component in improving reconstruction accuracy. The proposed method provides a practical and scalable solution for reconstructing missing indoor environmental data and supports low-density sensor deployment in building monitoring systems. Full article

Large integrated station–bridge buildings combine track-bearing members, station floors, transfer structures, columns, and urban-rail facilities within a single coupled structural system. For such buildings, refined train–track–station dynamic simulations can reproduce train-induced vibration, but repeated time-history analysis remains costly when many operating conditions must be screened. This study develops a simulation-based response-database framework for multi-horizon forecasting of front-end carbody vertical acceleration (FCVA), defined here as the vertical acceleration at the front-end floor evaluation point of the leading carbody, in the integrated station–bridge building of Yichang North Railway Station. The project-specific database contains 700 operating cases constructed from 100 Latin-hypercube-sampled combinations of a dimensionless track-spectrum amplitude multiplier (TSA), structural damping ratio (DR), and track-spectrum initial moving position (TSIP), each evaluated at seven train speeds. With a sampling interval of 0.002 s, supervised samples were constructed using a 200-point historical window, and prediction horizons from 20 to 300 steps (0.04–0.60 s) were evaluated under a case-level split. Classical regression, tree ensembles, a multilayer perceptron, recurrent networks, a temporal convolutional network, and a Transformer were compared after automated hyperparameter selection. For the 20-step task, Extra Trees achieved the best performance, with a root mean squared error (RMSE) of 0.00336 m/s² and $R^2$ = 0.9958. In the independently refitted reference-fixed horizon experiment, Extra Trees retained $R^2$ = 0.9526 at the 300-step horizon, while the temporal convolutional network (TCN) RMSE increased from 0.00394 to 0.01502 m/s². The results show that the response database preserves exploitable short- to medium-range dynamic continuity, although phase drift and peak-timing uncertainty increase as the forecast horizon becomes longer. Parameter analysis indicates that train speed dominates both response energy and forecast error, whereas TSA mainly affects amplitude-related response metrics. On a common central processing unit (CPU) platform, the saved Extra Trees model processed 10,000 held-out windows in $0.1404\pm 0.0008$ s. The proposed method provides a computationally efficient response-screening and post-processing layer for design-stage assessment and operating-scenario comparison within the modeled parameter domain, complementing rather than replacing refined dynamic simulation and field validation. Full article

Profiling side-channel attacks based on deep learning can recover cryptographic information from power-consumption traces, but their effectiveness may be reduced when desynchronisation displaces informative leakage across temporal positions. This study investigates whether Temporal Convolutional Networks (TCNs) can model the local and long-range dependencies present in desynchronised traces. Three complementary TCN variants are designed to isolate different temporal modelling strategies: a single-kernel dilated architecture (TCN1), a multi-scale architecture using parallel kernel sizes (TCN2), and a residual architecture intended to support stable hierarchical feature learning (TCN3). The models are evaluated on the ASCAD v1 fixed-key benchmark under synchronised conditions and maximum temporal shifts of 25, 50, and 75 samples, using validation loss and complementary key-ranking metrics. Under the most challenging setting, TCN1 achieves the highest Rank Success Rate and reaches its best rank substantially earlier than the conventional CNN baselines, although the lowest Final Rank is obtained by a CNN. These results indicate that TCN1 provides the most favourable trade-off among convergence speed, ranking consistency, and architectural complexity, without establishing uniform TCN superiority across all metrics. Gradient saliency, LIME, and occlusion analyses further identify temporal regions that influence the model predictions and are consistent with expected leakage patterns. The main contribution is a controlled comparison of complementary TCN design strategies, combined with an explainability analysis, for profiling side-channel attacks under trace desynchronisation. Full article

Photovoltaic (PV) power exhibits high variability and non-stationarity due to irradiance fluctuations, cloud shading, and seasonal changes, which complicate ultra-short-term multi-step forecasting. This study proposes a Solar-Historical Prior Network (SHPNet) for forecasting at 5 min resolution. SHPNet integrates a solar-geometry clear-sky prior-guided temporal convolutional network (SGCP-TCN), a similar historical window (SHW) branch, and horizon-wise adaptive fusion (HA). SGCP-TCN estimates clear-sky power potential from site coordinates and timestamps and reformulates direct power prediction as clear-sky power ratio forecasting. SHW retrieves training windows from the same intra-day time slot that exhibit similar power–irradiance evolution, thereby constructing a non-parametric historical prior, while HA determines horizon-specific fusion weights based on validation errors. Unlike purely data-driven predictors and conventional similar-day methods, SHPNet combines a physically interpretable power scale with input-window-level historical evolution patterns and adaptively balances the two priors across forecasting horizons. Across the two sites, SHPNet reduced the mean MAE and RMSE by 14.07% and 10.26%, respectively, compared with the original TCN, while increasing the mean R² from 0.8279 to 0.8613. Evaluations under different weather conditions and across seasons demonstrate consistent forecasting performance, while convergence analysis confirms stable training behavior. Full article