Online Stage Research Articles

Physics-informed two-tier neural network for non-linear model order reduction

AbstractIn recent years, machine learning (ML) has had a great impact in the area of non-intrusive, non-linear model order reduction (MOR). However, the offline training phase often still entails high computational costs since it requires numerous, expensive, full-order solutions as the training data. Furthermore, in state-of-the-art methods, neural networks trained by a small amount of training data cannot be expected to generalize sufficiently well, and the training phase often ignores the underlying physical information when it is applied with MOR. Moreover, state-of-the-art MOR techniques that ensure an efficient online stage, such as hyper reduction techniques, are either intrusive or entail high offline computational costs. To resolve these challenges, inspired by recent developments in physics-informed and physics-reinforced neural networks, we propose a non-intrusive, physics-informed, two-tier deep network (TTDN) method. The proposed network, in which the first tier achieves the regression of the unknown quantity of interest and the second tier rebuilds the physical constitutive law between the unknown quantities of interest and derived quantities, is trained using pretraining and semi-supervised learning strategies. To illustrate the efficiency of the proposed approach, we perform numerical experiments on challenging non-linear and non-affine problems, including multi-scale mechanics problems.

A non-intrusive reduced-order model for finite element analysis of implant positioning in total hip replacements.

Sophisticated high-fidelity simulations can predict bone mass density (BMD) changes around a hip implant after implantation. However, these models currently have high computational demands, rendering them impractical for clinical settings. Model order reduction techniques offer a remedy by enabling fast evaluations. In this work, a non-intrusive reduced-order model, combining proper orthogonal decomposition with radial basis function interpolation (POD-RBF), is established to predict BMD distributions for varying implant positions. A parameterised finite element mesh is morphed using Laplace's equation, which eliminates tedious remeshing and projection of the BMD results on a common mesh in the offline stage. In the online stage, the surrogate model can predict BMD distributions for new implant positions and the results are visualised on the parameterised reference mesh. The computational time for evaluating the final BMD distribution around a new implant position is reduced from minutes to milliseconds by the surrogate model compared to the high-fidelity model. The snapshot data, the surrogate model parameters and the accuracy of the surrogate model are analysed. The presented non-intrusive surrogate model paves the way for on-the-fly evaluations in clinical practice, offering a promising tool for planning and monitoring of total hip replacements.

Distributed model predictive control for consensus of nonlinear systems via parametric sensitivity

To handle the nonlinear consensus problem, a distributed model predictive control (DMPC) scheme is developed via parametric sensitivity. A two-stage input computation strategy is adopted for enhancing optimization efficiency. In the background stage, each agent first establishes its next-step optimization problem based on communication topology, and then performs distributed optimization to calculate the future inputs. In the online stage, all the agents build their sensitivity equations based on new information. Three variants of sensitivity equation are developed based on the level of communication load capacity, and the corresponding computation strategies are proposed. After solution, the background inputs are corrected and implemented. The optimality and robustness of the proposed algorithm are rigorously derived. Finally, the superiority of this DMPC scheme is demonstrated in the multi-vehicle system with two different topologies.

An adversarial diverse deep ensemble approach for surrogate‐based traffic signal optimization

AbstractSurrogate‐based traffic signal optimization (TSO) is a computationally efficient alternative to simulation‐based TSO. By replacing the simulation‐based objective function, a surrogate model can quickly identify solutions by searching for extreme points on its response surface. As a popular surrogate model, the ensemble of multiple diverse deep learning models can approximate complicated systems with a strong generalizability. However, existing ensemble methods barely focus on strengthening the prediction of extreme points, which we found can be realized by further diversifying base learners in the ensemble. The study proposes an adversarial diverse ensemble (ADE) method for online TSO with limited computational resources, comprising two stages: In the offline stage, base extractors are diversified with unlabeled data by a designed adversarial diversity training algorithm; in the online stage, base predictors are trained in parallel with limited labeled data, and the ensemble then serves as the surrogate model to search for solutions iteratively for TSO. First, it is demonstrated that the prediction accuracy on extreme points, and associated solution quality, can be constantly improved with base learners’ diversity enhanced by ADE. Case studies of TSO conducted on a four‐intersection arterial further demonstrate the superior solution quality and computational efficiency of the ADE surrogate model in a wide range of traffic scenarios. Moreover, a large‐scale online TSO experiment under dynamic traffic demand proves ADE's effectiveness in practical applications.

Non-Intrusive Reduced Basis two-grid method for flow and transport problems in heterogeneous porous media

Due to its non-intrusive nature and ease of implementation, the Non-Intrusive Reduced Basis (NIRB) two-grid method has gained significant popularity in numerical computational fluid dynamics simulations. The efficiency of the NIRB method hinges on separating the procedure into offline and online stages. In the offline stage, a set of high-fidelity computations is performed to construct the reduced basis functions, which is time-consuming but is only executed once. In contrast, the online stage adapts a coarse-grid model to retrieve the expansion coefficients of the reduced basis functions. Thus it is much less costly than directly solving a high-fidelity model. However, coarse grids in heterogeneous porous media of flow models are often accompanied by upscaled hydraulic parameters (e.g. hydraulic conductivity), thus introducing upscaling errors. In this work, we introduce the two-scale idea to the existing NIRB two-grid method: when dealing with coarse-grid models, we also employ upscaled model parameters. Both the discretization and upscaling errors are compensated by the rectification post-processing. The numerical examples involve flow and heat transport problems in heterogeneous hydraulic conductivity fields, which are generated by self-affine random fields. Our research findings indicate that the modified NIRB method can effectively capture the large-scale features of numerical solutions, including pressure, velocity, and temperature. However, accurately retrieving velocity fields with small-scale features remains highly challenging.

An adaptive learning strategy for surrogate modeling of high-dimensional functions - Application to unsteady hypersonic flows in chemical nonequilibrium

Many engineering applications rely on the evaluation of expensive, non-linear high-dimensional functions. In this paper, we propose the RONAALP algorithm (Reduced Order Nonlinear Approximation with Adaptive Learning Procedure) to incrementally learn, on-the-fly as the application progresses, a fast and accurate reduced-order surrogate model of a target function. First, a combination of nonlinear auto-encoder, community clustering, and radial basis function networks allows us to learn an efficient and compact surrogate model with limited training data. Secondly, an active learning procedure overcomes any extrapolation issues during the online stage by adapting the surrogate model with high-fidelity evaluations that fall outside its current validity range. This approach results in generalizable, fast, and accurate reduced-order models of high-dimensional functions. The method is demonstrated on three direct numerical simulations of hypersonic flows in chemical nonequilibrium. Accurate simulations of these flows rely on detailed thermochemical gas models that dramatically increase the cost of such calculations. Using RONAALP to learn a reduced-order thermodynamic model surrogate on-the-fly, the cost of such simulations was reduced by up to 75% while maintaining an error of less than 10% on relevant quantities of interest.

A Real-Time Global Re-Localization Framework for a 3D LiDAR-Based Navigation System

Place recognition is widely used to re-localize robots in pre-built point cloud maps for navigation. However, current place recognition methods can only be used to recognize previously visited places. Moreover, these methods are limited by the requirement of using the same types of sensors in the re-localization process and the process is time consuming. In this paper, a template-matching-based global re-localization framework is proposed to address these challenges. The proposed framework includes an offline building stage and an online matching stage. In the offline stage, virtual LiDAR scans are densely resampled in the map and rotation-invariant descriptors can be extracted as templates. These templates are hierarchically clustered to build a template library. The map used to collect virtual LiDAR scans can be built either by the robot itself previously, or by other heterogeneous sensors. So, an important feature of the proposed framework is that it can be used in environments that have never been visited by the robot before. In the online stage, a cascade coarse-to-fine template matching method is proposed for efficient matching, considering both computational efficiency and accuracy. In the simulation with 100 K templates, the proposed framework achieves a 99% success rate and around 11 Hz matching speed when the re-localization error threshold is 1.0 m. In the validation on The Newer College Dataset with 40 K templates, it achieves a 94.67% success rate and around 7 Hz matching speed when the re-localization error threshold is 1.0 m. All the results show that the proposed framework has high accuracy, excellent efficiency, and the capability to achieve global re-localization in heterogeneous maps.

Strengthening public health system resilience to disasters in Türkiye: Insights from a scorecard methodology

BackgroundTürkiye is a country that faces many disasters, especially earthquakes and floods, which have serious short- and long-term consequences for public health. The importance of disaster risk reduction activities in building resilience before a disaster strikes is increasingly recognized. Furthermore, resilience assessment is considered to be the starting point for these activities. However, there is a substantial gap in the scientific evidence on systematic assessments of the resilience of the public health system and also a serious lack of activities to strengthen the system against disasters in Türkiye. AimUsing a Scorecard methodology, the study aimed to systematically assess the resilience of the public health system in Türkiye by examining key indicators related to disaster risk management and preparedness, and subsequently to recommend priority actions. A systematic assessment of resilience can provide the scientific evidence needed to identify weaknesses in the system. Furthermore, identifying priority actions based on this evidence allows progress to be made towards strengthening the system. MethodsThis mixed-methods study was conducted in two separate regions of Türkiye with the highest earthquake (Esenler/Istanbul) and flood risk (Ortahisar/Trabzon). Based on the Scorecard methodology, two-stage workshops (online and face-to-face) were held. During the online stage, qualitative data were collected by interviewing participants about their evaluation of the situation in the counties regarding the Scorecard indicators, and quantitative data were collected through scoring. In the face-to-face stage, strategies were developed to improve the resilience of the public health system. Quantitative data were expressed with numbers and percentages, and thematic analysis was utilized for qualitative data analysis. ResultsThe lowest scoring indicators were as follows (Esenler and Ortahisar, respectively): addressing mental health needs associated with disasters (2.17 ± 1.29; 2.27 ± 1.16), resilience of key healthcare facilities (2.22 ± 0.55; 2.91 ± 1.08), and protection and identification of ecosystem services (2.39 ± 1.04; 2.13 ± 1.49). In this study, participants recommended increasing the resilience of the public health infrastructure and improving human resources and ecosystem services to ensure the resilience of the public health system. ConclusionsThe study found that the weakest areas in terms of public health system resilience were the seismic safety and infrastructure of health facilities, the capacity of disaster-related mental health services, and disaster-related ecosystem services. To improve the resilience of public health systems, it is essential to strengthen these areas of weakness and ensure collaboration between all stakeholders in the system. The identification of weaknesses in the study has guided the determination of priority actions to build a baseline of resilience. In addition, the results of the study have highlighted priority areas for investment to improve disaster response and the overall health of the population in the context of disasters.

Bi-directional Long Short-Term Memory with Bird Mating Optimizer based Spectrum Sensing Technique for Cognitive Radio Networks

Cognitive radio networks (CRN) enable the wireless devices to sense the radio spectrum, determine the frequency state channels, and reconfigure the communication variables for satisfying the QoS needs by reducing the energy utilization. In cognitive radio, the detection of principal user signals is crucial for secondary users in order to make the best use of available spectrum (CR). The problem with conventional spectrum sensing approaches is that they have a high rate of missed detections and false alarms, which makes it difficult to make effective use of the spectrum. In instruction to recover the correctness of the detection of free spectrum, deep learning-based spectrum sensing is employed. For resolving the drawbacks of traditional energy detection models, this paper presents a new spectral sensing technique for cognitive radio networks (SST-CRN). Recently published research in spectrum sensing has placed a high value on deep learning that is model-agnostic as a result of this. In particular, long-short term memory (LSTM) networks perform exceptionally well at extracting spatial and temporal information from input data in deep learning. The proposed model, Bidirectional Long Short-Term Memory (Bi-LSTM) with Bird Mating Optimization (BMO), makes it possible to create nonlinear threshold-based systems more quickly and easily than previously possible. The proposed Bi-LSTM with BMO technique involves two stages of operations namely offline and online. The offline stage creates the non-linear threshold value to detect energy. In addition, the online stage automated selects a decision function saved in offline stage for determining the existence of primary user. The experiments were carried out and the results were analyzed using the help of the RadioML2016.10b dataset. A greater spectrum detection performance over previous sensing models has been obtained using a combination of the B-LSTM model and the BMO model, it has been discovered.

Multi-Fingerprints Indoor Localization for Variable Spatial Environments: A Naive Bayesian Approach.

Fingerprint-based indoor localization has been a hot research topic. However, the current fingerprint-based indoor localization approaches still rely on a single fingerprint database, where the average level of data at reference points is used as the fingerprint representation. In variable environmental conditions, the variations in signals caused by changes in the environmental states introduce significant deviations between the average level and the actual fingerprint characteristics. This deviation leads to a mismatch between the constructed fingerprint database and the real-world conditions, thereby affecting the effectiveness of fingerprint matching. Meanwhile, the sharp noise interference caused by uncertainties such as personnel movement has a significant interference on the creation of the fingerprint database and fingerprint matching in online stage. Examination of the sampling data after denoising with Robust Principal Component Analysis (RPCA) revealed distinct multi-fingerprint characteristics with clear boundaries at certain access points. Based on these observations, the concept of constructing a fingerprint database using multiple fingerprints is introduced and its feasibility is explored. Additionally, a multi-fingerprint solution based on naive Bayes classification is proposed to accurately represent fingerprint characteristics under different environmental conditions. This method is based on the online stage fingerprints. The corresponding state space is selected using the naive Bayes classifier, enabling the selection of an appropriate fingerprint database for matching. Through simulations and empirical evaluations, the proposed multi-fingerprints construction scheme consistently outperforms the traditional single-fingerprint database in terms of positioning accuracy across all tested localization algorithms.

PLS-based hellinger distance method for fault detection in chemical engineering systems

Fault detection is vital in chemical engineering systems to maintain operational efficiency, product quality, and safety through timely identification and correction of deviations from expected behavior. Although partial least squares (PLS) has proven effective in monitoring due to its ability to handle highly correlated variables, traditional detection metrics of PLS may fail to identify small abnormal changes as they rely solely on recent observations. This paper integrates PLS modeling framework with Hellinger Distance (HD)-based fault detection index to overcome the limitations of conventional detection metrics. The utilization of HD is motivated by its sensitivity to quantifying any dissimilarity between distributions, which makes it well-suited for detecting small deviations in process behavior. The HD-based index will be computed between the residuals obtained from the model in the offline stage and the online stage. The HD metric involves careful inspection and comparison of the residuals, which enables it to capture the sensitive details in the data, thus, enhancing the detection of faults. For increased flexibility, kernel density estimation is employed to establish the reference threshold of the PLS-HD approach. The performance of this approach will be evaluated using data from simulated Continuous Stirred-Tank Heater (CSTH) and Continuous Stirred-Tank Reactor (CSTR) processes, by considering various fault types such as bias, freezing, and sensor drift faults. The results demonstrate the superior performance of the proposed PLS-HD approach compared to conventional PLS monitoring methods.

Inverse distance weight-assisted particle swarm optimized indoor localization

Indoor localization using wireless fidelity (WiFi) fingerprint has attracted considerable attention due to its extensive deployment and low cost. Most indoor fingerprint localization methods could be computationally inefficient and behave with unsatisfactory positioning accuracy. Therefore, an inverse distance weight (IDW) assisted particle swarm optimized (PSO) indoor localization (IDWPSOInLoc) algorithm is proposed. It is an optimized positioning process in the online stage rather than an offline optimization approach. The main idea is to find the optimal particle with the most similar generated fingerprint to the real-time testing one in the online stage. K nearest neighbors and boundary constraints are combined to initialize particles’ positions. The generated fingerprint of each particle is interpolated based on the IDW algorithm and nearby reference points within a specified range. A new fitness function, the sum of standard deviations between the generated and real-time testing fingerprints, is proposed to find the optimal particle. The position of the optimal particle with the minimal fitness value is taken as the estimated coordinate through iterations updating the particles’ velocities, positions, and generated fingerprints. IDWPSOInLoc achieves a mean positioning accuracy of 2.477 meters in the local environment experimental test and that of about 6 meters using selected floor in the UJIIndoorLoc dataset. Compared with state-of-the-art algorithms, positioning error is decreased by at least 11.9%. In addition, IDWPSOInLoc has the characteristic of low computational complexity. Experimental results show that the proposed IDWPSOInLoc outperforms the considered WiFi indoor positioning algorithms.

FCA-based efficient prediction of effective properties for 3D braided composites considering the effect of strain rate

The existing fast concurrency algorithms are difficult to predict the dynamic mechanical performance of three-dimensional (3D) braided composites from different scales. A modified Finite Element Method Cluster Analysis (FCA) method is employed to address this challenge. The modified FCA method is consisted of an offline and an online stage. In the offline stage, the high-fidelity Representative Unit Cell (RUC) is compressed to form a cluster-based RUC via the k-means method. Additionally, the cluster interaction matrix is calculated. In the online stage, the cluster interaction matrix is updated according to the strain rate levels for predicting the stress–strain curves for the RUC. Based upon the comparison with the data obtained from the Split Hopkinson Pressure Bar (SHPB) experiments, it is found that the modified FCA method can accurately predict the stress–strain curves under different strain rates for RUCs. Additionally, the modified FCA method provides a great improvement on computing efficiency when predicting the dynamic behavior of 3D braided composites.

Advancing passive BCIs: a feasibility study of two temporal derivative features and effect size-based feature selection in continuous online EEG-based machine error detection.

The emerging integration of Brain-Computer Interfaces (BCIs) in human-robot collaboration holds promise for dynamic adaptive interaction. The use of electroencephalogram (EEG)-measured error-related potentials (ErrPs) for online error detection in assistive devices offers a practical method for improving the reliability of such devices. However, continuous online error detection faces challenges such as developing efficient and lightweight classification techniques for quick predictions, reducing false alarms from artifacts, and dealing with the non-stationarity of EEG signals. Further research is essential to address the complexities of continuous classification in online sessions. With this study, we demonstrated a comprehensive approach for continuous online EEG-based machine error detection, which emerged as the winner of a competition at the 32nd International Joint Conference on Artificial Intelligence. The competition consisted of two stages: an offline stage for model development using pre-recorded, labeled EEG data, and an online stage 3 months after the offline stage, where these models were tested live on continuously streamed EEG data to detect errors in orthosis movements in real time. Our approach incorporates two temporal-derivative features with an effect size-based feature selection technique for model training, together with a lightweight noise filtering method for online sessions without recalibration of the model. The model trained in the offline stage not only resulted in a high average cross-validation accuracy of 89.9% across all participants, but also demonstrated remarkable performance during the online session 3 months after the initial data collection without further calibration, maintaining a low overall false alarm rate of 1.7% and swift response capabilities. Our research makes two significant contributions to the field. Firstly, it demonstrates the feasibility of integrating two temporal derivative features with an effect size-based feature selection strategy, particularly in online EEG-based BCIs. Secondly, our work introduces an innovative approach designed for continuous online error prediction, which includes a straightforward noise rejection technique to reduce false alarms. This study serves as a feasibility investigation into a methodology for seamless error detection that promises to transform practical applications in the domain of neuroadaptive technology and human-robot interaction.

A system-level interface sampling and reduction method for component mode synthesis with varying parameters

This paper presents a new interface sampling and reduction method for varying parameters within the framework of the component mode synthesis (CMS) method. In the conventional interpolation-based parametric reduced-order model (IB-PROM) based on CMS, reduced subsystems are synthesized, and interface degrees-of-freedom are reduced via the secondary eigenvalue analysis in the online stage, which is essential to obtain highly reduced system matrices. However, it inhibits rapid parameter updating and requires extra computational resources in the online stage. In the present study, the pre-reduced interface sampling method is proposed by synthesizing substructures and reducing the interface before beginning the online stage. The interface solution is obtained through offline and online adaptation, likewise the procedure of computing the interior solution of previous studies. Additionally, congruence transformations are applied to each reduced interface, which is mandatory for interpolating symmetric positive definite matrices with respect to a new input parameter. As a result, the online performance of IB-PROM is greatly enhanced, showing rapid and near real-time adaptations to the variations in system parameters.

DLLog: An Online Log Parsing Approach for Large-Scale System

Syslog is a critical data source for analyzing system problems. Converting unstructured log entries into structured log data is necessary for effective log analysis. However, existing log parsing methods demonstrate promising accuracy on limited datasets, but their generalizability and precision are uncertain when applied to diverse log data. Enhancements in these areas are necessary. This paper proposes an online log parsing method called DLLog, which is based on deep learning and has the longest common subsequence. DLLog utilizes the GRU neural network to mine template words and applies the longest common subsequence to parse log entries in real-time. In the offline stage, DLLog combines multiple log features to accurately extract the template words, creating a log template set to assist online log parsing. In the online stage, DLLog parses log entries by calculating the matching degree between the real-time log entry and the log template in the log template set. This method also supports the incremental update of the log template set to handle new log entries generated by systems. We summarized the previous works and validated DLLog using real log data collected from 16 systems. The results demonstrate that DLLog achieves high parsing accuracy, universality, and adaptability.

Online prediction of sustained muscle force from individual motor unit activities using adaptive surface EMG decomposition

Decoding movement intentions from motor unit (MU) activities to represent neural drive information plays a central role in establishing neural interfaces, but there remains a great challenge for obtaining precise MU activities during sustained muscle contractions. In this paper, we presented an online muscle force prediction method driven by individual MU activities that were decomposed from prolonged surface electromyogram (SEMG) signals in real time. In the training stage of the proposed method, a set of separation vectors was initialized for decomposing MU activities. After transferring each decomposed MU activity into a twitch force train according to its action potential waveform, a neural network was designed and trained for predicting muscle force. In the subsequent online stage, a practical double-thread-parallel algorithm was developed. One frontend thread predicted the muscle force in real time utilizing the trained network and the other backend thread simultaneously updated the separation vectors. To assess the performance of the proposed method, SEMG signals were recorded from the abductor pollicis brevis muscles of eight subjects and the contraction force was simultaneously collected. With the update procedure in the backend thread, the force prediction performance of the proposed method was significantly improved in terms of lower root mean square deviation (RMSD) of around 10% and higher fitness (R2) of around 0.90, outperforming two conventional methods. This study provides a promising technique for real-time myoelectric applications in movement control and health.

A local ROM for Rayleigh–Bénard bifurcation problems

This work presents a local reduced-order method for computing bifurcation diagrams in 2D Rayleigh–Bénard convection problems. The proposed method is based on Proper Orthogonal Decomposition, and employs a reduced-order study of the regularity of solutions to detect new solution branches within the bifurcation diagram. The locality of the method is achieved through k-means clustering, and the selection of local problems in the online stage is done in the solution space. All hyperparameters of the reduced method, such as the number of clusters or the number of POD modes, are estimated by deterministic criteria. Furthermore, the offline–online splitting strategy for online calculations is explicitly outlined. The method is applied to a single-parameter problem and a two-parameter problem, showing its ability to rapidly compute bifurcation diagrams with small errors. Compared to a standard approach that samples each branch separately, the local approach produces more accurate results in less computational time.

Solver-free reduced order homogenization for nonlinear periodic heterogeneous media

Reduced-order homogenization (ROH) and related methods are important computational tools for simulating the material behavior of composites. These methods generally sacrifice accuracy in exchange for superior computational efficiency, relative to methods such as classical computational homogenization (CCH). In this study, building on the recently developed solver-free CCH, we propose a fine-scale solver-free reduced order homogenization approach that avoids solving the fine-scale equilibrium equations and approximates the phase-average eigenstrains by sampling the fine-scale eigenstrain at a small number of points. The proposed method, which we call solver-free ROH, works by pre-computing history-dependent eigenstrain influence function tensors and sampling point contribution factors based on training data from a small set of CCH simulations. Then, during the online stage of the computation, phase-average eigenstrains are computed from sampling point eigenstrains and used to compute homogenized stresses and strains. Focusing on small-deformation problems, this paper formulates and verifies the solver-free ROH approach for nonlinear periodic heterogeneous media. First, in the formulation, we delineate a sampling point approximation of the phase eigenstrains and describe the use of this approximation within the coarse-scale stress update. Next, we verify the solver-free ROH using loading cases outside the training data set. Finally, we use the proposed method to simulate a multilayer composite plate in three point bending (3pt-bend) and open hole tension (OHT), demonstrating the method’s efficiency and accuracy relative to the CCH.

Offline and Online Optical Flow Enhancement for Deep Video Compression

Video compression relies heavily on exploiting the temporal redundancy between video frames, which is usually achieved by estimating and using the motion information. The motion information is represented as optical flows in most of the existing deep video compression networks. Indeed, these networks often adopt pre-trained optical flow estimation networks for motion estimation. The optical flows, however, may be less suitable for video compression due to the following two factors. First, the optical flow estimation networks were trained to perform inter-frame prediction as accurately as possible, but the optical flows themselves may cost too many bits to encode. Second, the optical flow estimation networks were trained on synthetic data, and may not generalize well enough to real-world videos. We address the twofold limitations by enhancing the optical flows in two stages: offline and online. In the offline stage, we fine-tune a trained optical flow estimation network with the motion information provided by a traditional (non-deep) video compression scheme, e.g. H.266/VVC, as we believe the motion information of H.266/VVC achieves a better rate-distortion trade-off. In the online stage, we further optimize the latent features of the optical flows with a gradient descent-based algorithm for the video to be compressed, so as to enhance the adaptivity of the optical flows. We conduct experiments on two state-of-the-art deep video compression schemes, DCVC and DCVC-DC. Experimental results demonstrate that the proposed offline and online enhancement together achieves on average 13.4% bitrate saving for DCVC and 4.1% bitrate saving for DCVC-DC on the tested videos, without increasing the model or computational complexity of the decoder side.