Sliding Window Research Articles

Adaptive soft sensor design using a regression neural network and bias update strategy for non-linear industrial processes

Soft sensing of quality parameters in process industries has been an active area of research for the past two decades. To improve the performance of soft sensors in the scenario of time varying process states, adaptation capability is incorporated into the soft sensor model. In this work, recursive (R), sliding window (SW) and just-in-time learning (JITL) frameworks are used for adaptive soft sensor design. A rarely explored modeling technique in the adaptation framework, the generalized regression neural network (GRNN) is used as a local modeling strategy. A bias update procedure is applied during the model adaptation activity to improve the prediction accuracy. Further, the performances of the developed models are tested against input–output data dimension mismatch along with various concept drift phenomena by considering a different number of labeled samples for inputs and outputs. The proposed adaptation strategy is applied on two benchmark industrial processes. Simulation results show that the GRNN local modeling approach combined with the bias update strategy gives higher prediction accuracy than other adaptive soft sensors proposed in the literature. Moreover, GRNN local modeling strategy using SW adaptation mechanism has the least computation time among the three adaptation methods due to the use of a low number of samples for model development.

Improving BERT with local context comprehension for multi-turn response selection in retrieval-based dialogue systems

Multi-turn response selection is an important branch in the field of natural language processing, which aims to select the most appropriate response based on multi-turn dialogue. Most state-of-the-art models adopt pre-trained language models (PrLMs) and multiple auxiliary tasks to enhance their ability to understand the semantics in multi-turn dialogue. However, some critical challenges still remain to be addressed. Optimizing multiple auxiliary tasks simultaneously may significantly increase the training cost. Meanwhile, the semantic gap between the optimization objectives of the main and auxiliary tasks may bring noise to pre-trained language models. To address these challenges, we propose an efficient BERT-based neural network model with local context comprehension (BERT-LCC) for multi-turn response selection. First, we propose a self-supervised learning strategy, which introduces an auxiliary task named Response Prediction in Random Sliding Windows (RPRSW). In a multi-turn dialogue, the RPRSW task takes utterances falling within a random sliding window as input and predicts whether the last utterance within the sliding window is the appropriate response for the local dialogue context. This auxiliary task can enhance BERT’s understanding of local semantic information. Second, we propose a local information fusion (LIF) mechanism that collects multi-granularity local features at different dialogue stages and employs a gating function to fuse global features with local features. Third, we introduce a simple but effective domain learning strategy to learn rich semantic information at different dialogue stages during pre-training. Experimental results on two public benchmark datasets show that BERT-LCC outperforms other state-of-the-art models.

Dosimetric comparison of VMAT standard optimization (SO) and multi-criteria optimization (MCO) treatment plans with standard mode delivery (STD) or sliding window (SW) for head and neck cancer.

A new development on the RayStation treatment planning system (TPS) allows a plan to be planned by imposing a constraint on the leaf sequencing: all leaves move in the same direction before moving again in the opposite direction to create a succession of sliding windows (SWs). The study aims to investigate this new leaf sequencing, coupled with standard optimization (SO) and multi-criteria optimization (MCO) and to compare it with the standard sequencing (STD). Sixty plans were replanned for 10 head and neck cancer patients (two dose levels simultaneously SIB, 56 and 70Gy in 35 fractions). All plans were compared, and a Wilcoxon signed-rank test was performed. Pre-processing QA and metrics of multileaf collimator (MLC) complexity were studied. All methodologies met the dose requirements for the planning target volumes (PTVs) and organs at risk (OARs). SO demonstrates significantly best results for homogeneity index (HI), conformity index (CI), and target coverage (TC). SO-SW gives best results for PTVs (D98% and D2% ) but the differences between techniques are less than 1%. Only the D2%,PTV-56Gy is higher with both MCO methods. MCO-STD offer the best sparing OARs (parotids, spinal cord, larynx, oral cavity). The gamma passing rates (GPRs) with 3%/3mm criteria between the measured and calculated dose distributions are higher than 95%, slightly lowest with SW. The number of monitor units (MUs) and MLC metrics are higher in SW show a higher modulation. All plans are feasible for the treatment. A clear advantage of SO-SW is that the treatment plan is more straightforward to planning by the user due to the more advanced modulation. MCO stands out for its ease of use and will allow a less experienced user to offer a better plan than in SO. In addition, MCO-STD will reduce the dose to the OARs while maintaining good TC.

Unsupervised video segmentation for multi-view daily action recognition

Multi-layer representations have achieved outstanding performances in the complex daily action recognition. However, the fixed length of the sliding window (SW) leads to the segmented motion atoms incomplete or non-unique. To deal with this problem, we propose unsupervised video segmentation (UVS) for multi-view daily action recognition. Firstly, the average cosine similarity is designed to ensure the integrity of motion atoms. Secondly, we utilize the ordered combination of motion atoms to construct the table of multi-scale motion phrases in a top-down manner, instead of the fixed-scale traditional motion phrases. Finally, the experimental results based on the WVU dataset, the NTU RGB-D 120 dataset, and the N-UCLA dataset show that the proposed UVS method has state-of-the-art performance, compared with the classic methods such as IDT, MoFAP, JLMF, FGCN, and MVMLR.

A novel hybrid model for building heat load forecasting based on multivariate Empirical modal decomposition

Accurate heat load forecasting is crucial for the high precise real-time operational control of buildings in winter. The inconsistency of frequencies between features and heat load, however, constrains the improvement of the ultra-short-term forecasting accuracy of heat load. This study proposed a novel hybrid model built upon Multivariate Empirical Mode Decomposition (MEMD) and Support Vector Regression (SVR) with hyper-parameters optimized by Particle Swarm Optimization (PSO), which is able to improve forecasting accuracy significantly. Meanwhile, Sliding Window (SW) is employed to overcome the limitations of MEMD in forecasting, and feature selection is carried out using eXtreme Gradient Boosting (XGBoost) before modeling to minimize errors and reduce the workload. The principle of the proposed SW-MEMD-PSO-SVR hybrid model is to decompose the associated features and heat load into several groups of components by MEMD, maintaining a constant frequency of feature components and heat load components for each forecasting. A real building in Inner Mongolia, China, have been considered as an example to verify the superiority of the proposed hybrid model. The proposed SW-MEMD-PSO-SVR hybrid model has the best values of MAPE, NMBE, CVRMSE and R2, being 2.68%, 0.09%, 3.52%, and 84.90%, respectively. The results demonstrated that the proposed hybrid model is a promising alternative for improving accuracy of ultra-short-term building heat load forecasting.

A deep learning based four-fold approach to classify brain MRI: BTSCNet

Incorrect diagnosis of brain tumor types prevent appropriate response to medical assistance and reduces patients' chances of survival. Examining MRI images of the patient's brain manually is one traditional method for distinguishing brain tumors, but it is time consuming and prone to human errors. Thus, an automated and new deep learning based four-fold method, Brain Tumor Segmentation and Classification Network (BTSCNet), is proposed in this article which will be helpful for physicians to classify three types of brain tumors (meningioma, glioma, and pituitary tumor) from T1-weighted contrast-enhanced MRI (CE-MRI) images properly. The proposed four folds includes: segmentation of brain tumor region using Brain Tumor Segmentation Network (BTSNet), ROI selection using morphological operation, feature extraction using multi-region gray level co-occurrence matrix (MR-GLCM), andclassification using Sliding Window Euclidean distance (SWED) measure. The lack of annotated training samples is the main challenge in deep-learning-based brain MRIimage classification. Thus, while training the proposed system, the scale, orientation and flip of the input image is randomly changed in the first fold of BTSCNetso that the network can be trained with varied or augmentedimages with respect toscale, orientation and flip.Four performance measures are used to measure the efficiency of the proposed method. The output of the proposed method is tested on a public database where the correct classification rates obtained are96.6% (meningioma), 98.1% (glioma), and 95.3% (pituitary tumor), when considering MR-Contrast feature, MR-Correlation and MR-Homogeneity feature respectively, which indicates that the efficiency of the proposed method.

Multiplication free Fast-Adaptive Binary Range Coder using ISW

Data compression is defined as the process of encoding, converting and modifying the bits-structures of data in such a way that reduces less-spaces on the disk. Fast-ABRC, a new context ABRC for compressing the image and video. This paper introduces novel hardware F-ABRC (Fast-adaptive binary range coder) and architecture of VLSI, as it doesn’t have requirement of LUTs (Look-up-Tables) and also it is completely multiplication free. To get the result, we will combine the utilization of simple operation to compute the approximation after encoding every single symbol and the PE (probability estimation) on the basis of ISW (Imaginary Sliding Window) with approximation of the multiplication. We have represented our introduced algorithm, which is faster and in comparison, to the existing model it gives superior compression efficiency and the comparison takes place on the basis of two parameters such as power dissipation (Dynamic and Static) and device utilization.

Facial Expression Recognition Using Local Sliding Window Attention.

There are problems associated with facial expression recognition (FER), such as facial occlusion and head pose variations. These two problems lead to incomplete facial information in images, making feature extraction extremely difficult. Most current methods use prior knowledge or fixed-size patches to perform local cropping, thereby enhancing the ability to acquire fine-grained features. However, the former requires extra data processing work and is prone to errors; the latter destroys the integrity of local features. In this paper, we propose a local Sliding Window Attention Network (SWA-Net) for FER. Specifically, we propose a sliding window strategy for feature-level cropping, which preserves the integrity of local features and does not require complex preprocessing. Moreover, the local feature enhancement module mines fine-grained features with intraclass semantics through a multiscale depth network. The adaptive local feature selection module is introduced to prompt the model to find more essential local features. Extensive experiments demonstrate that our SWA-Net model achieves a comparable performance to that of state-of-the-art methods with scores of 90.03% on RAF-DB, 89.22% on FERPlus, 63.97% on AffectNet.

How Secure is Exponent-blinded RSA–CRT with Sliding Window Exponentiation?

This paper presents the first security evaluation of exponent-blinded RSA–CRT implementation with sliding window exponentiation against cache attacks. Our main contributions are threefold. (1) We demonstrate an improved cache attack using Flush+Reload on RSA–CRT to estimate the squaring–multiplication operational sequence. The proposed method can estimate a correct squaring–multiplication sequence from one Flush+Reload trace, while the existing Flush+Reload attacks always contain errors in the sequence estimation. This is mandatory for the subsequent steps in the proposed attack. (2) We present a new and first partial key exposure attack on exponent-blinded RSA–CRT with a random-bit leak. The proposed attack first estimates a random mask for blinding exponent using a modification of the Schindler–Wiemers continued fraction attack, and then recovers the secret key using an extension of the Heninger–Shacham branch-and-prune attack. We experimentally show that the proposed attack on RSA–CRT using a practical window size of 5 with 16-, 32-, and 64-bit masks is carried out with complexity of 225.6, 267.7, and 2161, respectively. (3) We then investigate the tradeoffs between mask bit length and implementation performance. The computational cost of exponent-blinded RSA–CRT using a sliding window with a 32- and 64-bit mask are 15% and 10% faster than that with a 128-bit mask, respectively, as we confirmed that 32- and 64-bit masks are sufficient to defeat the proposed attack. Our source code used in the experiment is publicly available.

Improving Multi-Class Motor Imagery EEG Classification Using Overlapping Sliding Window and Deep Learning Model

Motor imagery (MI) electroencephalography (EEG) signals are widely used in BCI systems. MI tasks are performed by imagining doing a specific task and classifying MI through EEG signal processing. However, it is a challenging task to classify EEG signals accurately. In this study, we propose a LSTM-based classification framework to enhance classification accuracy of four-class MI signals. To obtain time-varying data of EEG signals, a sliding window technique is used, and an overlapping-band-based FBCSP is applied to extract the subject-specific spatial features. Experimental results on BCI competition IV dataset 2a showed an average accuracy of 97% and kappa value of 0.95 in all subjects. It is demonstrated that the proposed method outperforms the existing algorithms for classifying the four-class MI EEG, and it also illustrates the robustness on the variability of inter-trial and inter-session of MI data. Furthermore, the extended experimental results for channel selection showed the best performance of classification accuracy when using all twenty-two channels by the proposed method, but an average kappa value of 0.93 was achieved with only seven channels.

Improving Biomedical Question Answering by Data Augmentation and Model Weighting.

Biomedical Question Answering aims to extract an answer to the given question from a biomedical context. Due to the strong professionalism of specific domain, it's more difficult to build large-scale datasets for specific domain question answering. Existing methods are limited by the lack of training data, and the performance is not as good as in open-domain settings, especially degrading when facing to the adversarial sample. We try to resolve the above issues. First, effective data augmentation strategies are adopted to improve the model training, including slide window, summarization and round-trip translation. Second, we propose a model weighting strategy for the final answer prediction in biomedical domain, which combines the advantage of two models, open-domain model QANet and BioBERT pre-trained in biomedical domain data. Finally, we give adversarial training to reinforce the robustness of the model. The public biomedical dataset collected from PubMed provided by BioASQ challenge is used to evaluate our approach. The results show that the model performance has been improved significantly compared to the single model and other models participated in BioASQ challenge. It can learn richer semantic expression from data augmentation and adversarial samples, which is beneficial to solve more complex question answering problems in biomedical domain.

FAST: A Forecasting Model With Adaptive Sliding Window and Time Locality Integration for Dynamic Cloud Workloads

The workload predictor has attracted attention as a key component of the proactive service operation management framework. However, the request and resource workloads of cloud applications are highly dynamic. Existing approaches decompose the original workload into trend, seasonal, and random components, establish models accordingly, and then combine all outputs to generate results. Indeed, the random component usually has significant heteroscedasticity and noise, having little or even a negative effect on model accuracy improvement. In our model, trend and seasonal components are seen as macro workload changes, and the micro workload changes are obtained by an adaptive sliding window algorithm. Therefore, we propose an ensembling model named FAST for Forecasting workloads with Adaptive Sliding window and Time locality integration. Notably, we propose an adaptive sliding window algorithm that considers trend correlation, time correlation, and random fluctuations of workload for online regression to achieve higher accuracy with lower overhead; and for the error-based integration strategy, we propose a time locality concept for local-predictor behavior and develop a multi-class regression algorithm for model integration. Finally, we conduct experiments on Google cluster trace datasets which show FAST has better accuracy than all other state-of-the-art models for dynamic workloads.

A Novel Short-Term Ship Motion Prediction Algorithm Based on EMD and Adaptive PSO–LSTM with the Sliding Window Approach

Under the influence of variable sea conditions, a ship will have an oscillating motion comprising six degrees of freedom, all of which are connected to each other. Among these degrees of freedom, rolling and pitching motions have a severe impact on a ship’s maritime operations. An accurate and effective ship motion attitude prediction method that makes the prediction in a short period of time is required to guarantee the safety and stability of the ship’s maritime operations. Traditional methods are based on time domain analysis, such as the autoregressive moving average (ARMA) models. However, these models have limitations when it comes to predicting the nonlinear and nonstationary characteristics of real ship motion attitude data. Many intelligent algorithms continue to be applied in nonlinear and nonstationary ship attitude prediction, such as extreme learning machines (ELMs) and the long short-term memory (LSTM) neural network, as well as other deep learning methods, showing promising results. By using the sliding window approach, the time-varying dynamic characteristics of the ship’s motion attitude can be preserved better. The simulation results demonstrate that the proposed model performs well in terms of predicting the nonlinear and nonstationary ship motion attitude.

Internet Financial Credit Risk Assessment with Sliding Window and Attention Mechanism LSTM Model

Internet Financial Credit Risk Assessment with Sliding Window and Attention Mechanism LSTM Model

Ship behavior during encounters in ports and waterways based on AIS data: From theoretical definitions to empirical findings

Currently, the research on ship behavior during encounters focuses on evasive behavior during specific situations with existing risks of collision. However, the preliminary selection of encounters to refine the presented ship behavior is biased. To obtain a full understanding of all ship behavior during different encounters in ports and waterways, the encounter is defined from the viewpoint of the spatial-temporal co-existence of ships in the same waterway segments during the same period. Based on this definition, this paper investigates ship behavior through the encounter process with other ships. The proposed approach starts from the moment when the distance in between is minimum as the critical moment to recognize ship behavior change (course alteration and speed change) based on the Sliding Window algorithm. Thus, the encounter process is identified by the key behavior feature point into phases, being before decision-making, before the critical moment, after the critical moment, and after being past and clear. The relative movement factors are calculated according to the behavior status of both ships to describe the conditions, timing, and objective of behavior change during the dynamic process of encounters. The empirical findings based on one-year Automatic Identification System data in the port of Rotterdam are presented. In the overtaking encounters, as the give-way ship, about 14% of the overtaking ships do not take any evasive actions. Among the ships with behavior changes, the preference for course alteration and speed change is equal. As the stand-on ship, about 87% of the overtaken ships take cooperative maneuvers to facilitate the encounter, in which deceleration seems the primary choice. The timing of overtaken ship's behavior change is later than overtaking ship. For overtaking ships, the objective of course alteration is a clear passing distance of about 5 times her beam, 100m for overtaken ships irrespective of her own size. Regarding speed, the overtaking ship aims to reach a relative speed of 0.3 times her own SOG, while the objective for the overtaken ship is fixed at around 2–3 m/s. In the encounters of ships sailing in the opposite direction, most of the ships take maneuvers to change their course or speed. However, within the influence distance of 2 km, over 76% of the ships do not take any evasive behavior, which implies a passing-by situation. Based on the recognized key feature points of behavior change, statistical tests show the objective of clear passing distance has been reached beforehand. The behavior change during head-on situations could be due to the precautionary behavior of officers onboard in case of interaction between ships. The findings enrich the knowledge of ship behavior during different types of encounters in real-life navigation, which can be further applied to simulation models for ship behavior in ports and waterways.

Disease identification method based on graph features between pulse cycles

The pulse signal at the wrist communicates a range of cardiovascular information and reflects a person’s state of health. In recent years pulse signals have been used extensively in the field of disease identification. As a typical quasi-cyclic physiological signal, many researchers have simply used single-cycle pulse signals for disease diagnosis, ignoring the influence of inter-cycle differences on the diagnosis. In this paper, we propose a disease identification method based on graph features between pulse cycles, to focus on the differences in pulse cycles brought about by the disease. The pulse was pre-processing using a Sliding Window, followed by the construction of the correlation coefficient matrix, the setting of a threshold and the construction of the connection graph to obtain the graph features; the Multiscale Permutation Entropy method was used to extract the pulse’s entropy features. The ReliefF algorithm is then used to select a total of 18 features and a classification algorithm is used to distinguish between patients and healthy people. Both the self-constructed dataset and the open dataset from the PLA 211 Hospital had high recognition rates for the experiments. The accuracy of identifying hypertension was 94.16% with a precision of 90.41%; the accuracy of identifying chronic diseases was 97.23% with a precision of 97.81%. The experimental results from the two datasets demonstrate both the validity of the difference between cycles for disease classification and the applicability of the method proposed in this paper.

Comprehensive analysis on error-robustness of FeFET computation-in-memory for hyperdimensional computing

This work comprehensively analyzes the error robustness of hyperdimensional computing (HDC) by using FeFET-based local multiply and global accumulate computation-in-memory. HDC trains and infers with hypervectors (HVs). Symmetric or asymmetric errors, which simulate read-disturb and data-retention errors of FeFET, are injected into Item memory and/or Associative memory before/after or during training in various cases when solving European language classification task. The detailed error injection reveals that HDC is acceptable for both symmetric and asymmetric error rate up to 10−1. Based on the detailed analysis of error robustness, training window slide (TWS) improves the error robustness against memory errors by removing data which contain different amount of errors. TWS shows 10 times higher error robustness. In addition, parallelization of HV encoding in training achieves fast training with up to 10 000 parallelism while maintaining the inference accuracy.

Predicting Spacecraft Telemetry Data by Using Grey–Markov Model with Sliding Window and Particle Swarm Optimization

Predicting telemetry data is vital for the proper operation of orbiting spacecraft. The Grey–Markov model with sliding window (GMSW) combines Grey model (GM (1, 1)) and Markov chain forecast model, which allows it to describe the fluctuation of telemetry data. However, the Grey–Markov model with sliding window does not provide better predictions of telemetry series with the pseudo-periodic phenomenon. To overcome this drawback, we improved the GMSW model by applying particle swarm optimization (PSO) algorithm a sliding window for better prediction of spacecraft telemetry data (denoted as PGMSW model). In order to produce more accurate predictions, background-value optimization is specially carried out using the particle swarm optimization technique in conventional GM (1, 1). For verifying PGMSW, it is utilized in the prediction of the cyclic fluctuation of telemetry series data and exponential variations therein. The simulation results indicate that the PGMSW model provides accurate solutions for prediction problems similar to the pseudo-periodic telemetry series.

Adaptive Decoding Algorithm With Variable Sliding Window for Double SC-LDPC Coding System

In this letter, an adaptive decoding algorithm with a variable sliding window for joint source-channel coding (JSCC) scheme based on double spatially coupled low-density parity-check (DSC-LDPC) codes is investigated. For the traditional sliding window decoding (SWD) algorithm, one or more errors occurring in a decoding window may affect the decoding of the next window, which will trigger error propagation. To mitigate the effects, a window extension algorithm based on average log-likelihood ratios (LLRs) will be proposed, but the implement of extension implies higher decoding complexity. In order to trade-off the error rate performance and complexity, the source and channel threshold that determines the window extension are analyzed. In the simulation, the proposed SWD algorithm with acceptable complexity has better error rate performance in the water fall region and the error floor region.

Aberrant dynamic functional network connectivity in vestibular migraine patients without peripheral vestibular lesion.

This study aimed to investigate changes in dynamic functional network connectivity (FNC) in patients with vestibular migraine (VM) and explore their relationship with clinical manifestations. Resting-state functional magnetic resonance imaging (rs-fMRI) data were scanned from 35 VM patients without peripheral vestibular lesion and 40 age-, sex- and education-matched healthy controls (HC). Independent component analysis (ICA), sliding window (SW) and k-means clustering analysis were performed to explore the difference in FNC and temporal characteristics between two groups. Additionally, Pearson's partial correlation analysis was adopted to investigate the relationship between clinical manifestations and rs-fMRI results in patients with VM. Compared with HC, patients with VM showed increased FNC in pairs of extrastriate visual network (eVN)-ventral attention network (VAN), eVN-default mode network (DMN) and eVN-left frontoparietal network (lFPN), and exhibited decreased FNC in pairs of VAN-auditory network (AuN). The altered FNC was correlated with clinical manifestations of patients with VM. Additionally, we found increased mean dwell time and fractional windows in state 2 in VM patients compared with HC. Mean dwell time was positively correlated with headache impact test-6 (HIT-6) scores, fractional windows was positively associated with dizziness handicap inventory (DHI) scores. Our results indicated that patients with VM showed altered FNC primarily between sensory networks and networks related to cognitive, emotional and attention implementation, with more time spent in a state characterized by positive FNC between sensor cortex system and dorsal attention network (DAN). These findings could help reinforce the understanding on the neural mechanisms of VM.