Update Model Parameters Research Articles

Sequential selection for accelerated life testing via approximate Bayesian inference

AbstractAccelerated life testing (ALT) is typically used to assess the reliability of material's lifetime under desired stress levels. Recent advances in material engineering have made a variety of material settings readily available. A critical question is how to efficiently conduct ALT to optimize reliability performance over different material settings. We propose a sequential selection approach to solve this problem. The proposed approach contains (1) a model updating mechanism to incorporate new experimental data in each step, and (2) a design criterion to guide new experiments that maximizes the potential to find the optimal material setting. To guarantee a tractable statistical mechanism for information collection, we develop explicit model parameter update formulas via approximate Bayesian inference. Theories show that our explicit update formulas give consistent parameter estimates. To guarantee that the design criterion in each step can make improvement on the identification of optimal material setting, this paper adopts an expected improvement‐based design criterion for optimizing the material setting under target stress factor levels. We also give a heuristic on this design criterion to justify the statistical consistency of approximate Bayesian estimates. Simulation studies and a case study show that the proposed sequential selection approach can significantly improve the probability of identifying the material alternative with best reliability performance compared to other design approaches.

Probabilistic history matching of multi-scale fractured reservoirs: Integration of a novel localization scheme based on rate transient analysis

In this paper, a new assisted history-matching workflow is presented, where rate transient analysis (RTA) results are used to constrain not only the initial discrete fracture network (DFN) models, the interpreted flow regimes are also used to formulate a localization scheme for more efficient updating of the pertinent DFN model parameters. The outcome is an ensemble of DFN realizations that are calibrated to both geologic and dynamic production data.RTA interpretations and other pertinent geological data are used to infer the prior probability distributions of the unknown fracture parameters, from which an ensemble of initial DFN models is sampled. The DFN models are subjected to numerical multiphase flow simulation. The fracture parameters are adjusted following an indicator-based probability perturbation method, which is capable of minimizing the objective function and reducing the uncertainties in the unknown fracture parameters simultaneously. A key feature is that the flow regimes identified from RTA are used to formulate a localization strategy, where individual segments of the production data are used to tune only a specified subset of the unknown model parameters.In a case study, the method is applied to characterize the probability distributions of four parameters in a multifractured shale gas well: primary fracture transmissivity, aperture of the secondary fracture, transmissivity of the secondary induced fracture, and global fracture intensity. Their probability distributions are updated following the proposed approach to match the production history. Multiple realizations of the DFN model are sampled.A key novelty is that the proposed probabilistic approach facilitates the representation of uncertainties in fracture parameters via multiple equally-probable DFN models and their corresponding upscaled flow-simulation models. A more comprehensive and robust approach is presented for integrating specific RTA interpretations and estimations into various steps of the history-matching process.

A Sub-Model Detachable Convolutional Neural Network

In this research, we propose a Convolutional Network with sub-Networks (CNSN), i.e., a Convolutional Neural Network (CNN) or base-model that can be divided into sub-models on demand. The CNN architecture, entails that feature map shapes are varied throughout the model, therefore, the hidden layer within CNN may not directly process an input image without modification. To address this problem, we propose a step-down convolutional layer, which is a convolutional layer acting as an input layer for the sub-model. This step-down convolutional layer reshapes and processes an input image to a preferred representation to the submodel. To train CNSN, we treat the base-model and sub-models as distinct models. Each model is forward- and back-propagated separately. Using multi-model loss, i.e., a linear combination of losses from base-model and sub-models, we thus update model parameters that can be utilized in both base-model and sub-models.

A data-driven model for building energy normalization to enable eco-feedback in multi-family residential buildings with smart and connected technology

In this paper, we present a new unit-level data-driven modelling approach to normalize heating and cooling (HC) energy usage in multi-family residential buildings based on easily accessible data from smart thermostats and WiFi-enabled power metres. Our physics-informed approach starts from a heat balance equation to derive a linear regression model and uses a Bayesian mixture model to identify groups of units that have similar regression coefficients. Our model captures the effect of behaviour on HC energy consumption by normalizing the effect of building characteristics and accounting for the inter-unit heat transfer and unobserved variables. Our probabilistic approach incorporates unit- and season-specific prior information and sequential Bayesian updating of model parameters when new data become available. Using yearly data collected in a multi-family building, our model identifies distinct normalized HC energy use groups in different seasons and provides more accurate rankings compared to the case without normalization.

VoI-Based Optimization of Structural Assessment for Spatially Degrading RC Structures

Before implementing a bridge monitoring strategy, a bridge manager would like to know the return on investment. Moreover, in order to spend the available budget as efficiently as possible, the monitoring strategy should be optimized, i.e., the type of measurements but also the time and locations at which these are performed. For this purpose, the Value of Information (VoI) can be used. The VoI represents an estimate of the benefit that can be gained from a monitoring strategy before it is actually implemented. By comparing the VoI of different alternative strategies, the one with the highest VoI can be selected. As such, the VoI is a tool for objective decision-making. The calculation of the VoI is based on pre-posterior analyses, including Bayesian updating of model parameters based on yet unknown monitoring outcomes. When calculating the VoI for an actual case, some challenges arise. First, the calculation of the VoI requires a number of assumptions on different input parameters. Second, the VoI is computed by evaluating life-cycle costs for different possible outcomes of the monitoring strategy, leading to a high computational cost. However, for practical implementations, results are preferably available within an acceptable time span and are robust with respect to the chosen input parameters. In this work, the implementation of the VoI approach for optimization of monitoring strategies is investigated by a problem statement in a case study where a reinforced concrete girder bridge is considered. To perform this optimization, the VoI for different monitoring strategies is compared. The calculation time required for the Bayesian updating of the model parameters based on the available data is limited by using Maximum A Posteriori (MAP) estimates to approximate the posterior distribution. The VoI can be used both to optimize a monitoring strategy or for comparison of different strategies. To limit the number of required (computationally expensive) evaluations of the VoI, optimization of the monitoring strategy itself can be simplified by determining the optimal sensor locations beforehand, based on a different metric than the VoI. For this purpose, the information entropy is used, which expresses the difference between the prior and posterior uncertainty of the model parameters. Finally, the sensitivity of the VoI to different input parameters is investigated.

Sequent extended Kalman filter capacity estimation method for lithium-ion batteries based on discrete battery aging model and support vector machine

Precise battery capacity estimation plays an important role in the future intelligent battery management system. In this paper, a fusion estimation method based on support vector machine and discrete battery aging model is put forward to enhance the online capacity estimation accuracy of lithium-ion batteries under variable temperature conditions. During the constant current charging process, the support vector machine is developed to estimate the battery capacity, which first trains a single 18650 battery offline and then tests the accuracy of the model using two other batteries of the same type intermittently. Subsequently, the discrete aging model of the battery is proposed to continuously estimate the capacity of the battery. However, unmodelled dynamics between battery aging model and real physical battery is easily occur in the process of modeling, which affects the accuracy and robustness of the model. Therefore, a sequent extended Kalman filter algorithm is deployed for solving the problem. The first Kalman filter takes the identified value of support vector machine as observation value to update the model parameters of discrete battery aging model. The second Kalman filter fuses the identified value of support vector machine and the discrete battery aging model after updating model parameters to improve the precision of online battery capacity estimation. The experimental results indicate that the proposed discrete battery aging model and support vector machine have good applicability, and the algorithm used can online modify the parameters of the model. When the model parameters are modified four times, the fusion estimation error is less than 2%.

Integrated Workflow of Geomechanics, Hydraulic Fracturing, and Reservoir Simulation for Production Estimation of a Shale Gas Reservoir

In this research, an integrated workflow from geomechanics to reservoir simulation is suggested to accurately estimate performances of a shale gas reservoir. Rather than manipulating values of hydraulic fracturing such as fracture geometry and transmissibility, the workflow tries to update model parameters to derive reliable hydraulic fracturing results. A mechanical earth model (MEM) is built from seismic attribute and drilling and diagnostic fracture injection test results. Then, the MEM is calibrated with microseismic measurements obtained in a field. Leakoff coefficient and horizontal stress anisotropy are sensitive parameters of the MEM that influence the propagation of the fracture network and gas productions. Various combinations of calibration parameters from a single-well simulation are evaluated. Then, an appropriate combination is chosen from the whole simulation results of a pad to reduce the uncertainty. Finally, production estimations of the four wells which have slightly different fracture design are compared with seven-year production history. Their results are reasonably matched with actual data having 8% of global error due to successful development of the reservoir model with geomechanical parameters.

Sequential Data Assimilation for Streamflow Forecasting: Assessing the Sensitivity to Uncertainties and Updated Variables of a Conceptual Hydrological Model at Basin Scale

AbstractSkillful streamflow forecasts provide key support to several water‐related applications. Because of the critical impact of initial conditions (ICs) on forecast accuracy, ever‐growing interest is focused on improving their estimates via data assimilation (DA). This study aims to assess the sensitivity of the DA‐based estimation of forecast ICs to several sources of uncertainty and the update of different model states and parameters of a lumped conceptual rainfall–runoff model over 232 watersheds in France. The performance of two sequential ensemble‐based techniques, namely, the ensemble Kalman filter (EnKF) and the particle filter (PF), is compared in terms of efficiency and temporal persistence (up to 10 days) of the updating effect through the assimilation of observed discharges. Several experiments specifically address the impact of the meteorological, state, and parameter uncertainties. Results show that an accurate estimate of the initial level of the routing store of the conceptual model ensures the most benefit to the DA‐based estimation of forecast ICs. While EnKF‐based forecasts outperform PF‐based ones when accounting for meteorological uncertainty, the more comprehensive representation of the state uncertainty makes it possible to greatly improve the accuracy of PF‐based predictions, with a longer‐lasting updating effect. Conversely, forecasting skill is undermined when accounting for parameter uncertainty, owing to the change in hydrological responsiveness. This study extensively addresses several sensitivity analyses in order to provide useful recommendations for designing DA‐based streamflow forecasting systems and for diagnosing possible deficiencies in existing systems.

Sounding out the hidden data: A concise review of deep learning in photoacoustic imaging.

The rapidly evolving field of photoacoustic tomography utilizes endogenous chromophores to extract both functional and structural information from deep within tissues. It is this power to perform precise quantitative measurements in vivo-with endogenous or exogenous contrast-that makes photoacoustic tomography highly promising for clinical translation in functional brain imaging, early cancer detection, real-time surgical guidance, and the visualization of dynamic drug responses. Considering photoacoustic tomography has benefited from numerous engineering innovations, it is of no surprise that many of photoacoustic tomography's current cutting-edge developments incorporate advances from the equally novel field of artificial intelligence. More specifically, alongside the growth and prevalence of graphical processing unit capabilities within recent years has emerged an offshoot of artificial intelligence known as deep learning. Rooted in the solid foundation of signal processing, deep learning typically utilizes a method of optimization known as gradient descent to minimize a loss function and update model parameters. There are already a number of innovative efforts in photoacoustic tomography utilizing deep learning techniques for a variety of purposes, including resolution enhancement, reconstruction artifact removal, undersampling correction, and improved quantification. Most of these efforts have proven to be highly promising in addressing long-standing technical obstacles where traditional solutions either completely fail or make only incremental progress. This concise review focuses on the history of applied artificial intelligence in photoacoustic tomography, presents recent advances at this multifaceted intersection of fields, and outlines the most exciting advances that will likely propagate into promising future innovations.

Fast just-in-time-learning recursive multi-output LSSVR for quality prediction and control of multivariable dynamic systems

Aiming at quality prediction and control of blast furnace (BF) ironmaking process characterized by complicated nonlinear time-varying dynamics, this paper proposes a just-in-time-learning (JITL) recursive multi-output least squares support vector regression (JITL-R-M-LSSVR) algorithm with fast nonlinear local learning capability for multivariable dynamic systems. The proposed fast JITL-R-M-LSSVR effectively combines the online local learning of JITL with the multi-output LSSVR (M-LSSVR) based on multi-task transfer learning, and focuses on how to ensure the rapid verification of the local model during online learning of M-LSSVR, and how to perform model pruning while recursively updating the model parameters to improve the calculation efficiency. To this end, the proposed algorithm uses a derived multi-output incremental learning algorithm to recursively update model parameters online in a gentle way, which has better modeling stability and smoothness than the traditional way that discards old models. At the same time, when the model is pruned, a novel multi-output reverse decremental learning algorithm is proposed to adaptively delete the modeling data, so as to effectively control the sample size and reduces the calculation cost. In particular, the model verification of the proposed algorithm only needs to construct the M-LSSVR modeling matrix and the matrix inverse operation once, and the matrix after deleting each modeling sample can be easily obtained by reverse decremental learning of the original modeling matrix, which can achieve fast and efficient model verification. Finally, the effectiveness and practicability of the proposed method are verified by applying it to prediction modeling and predictive control of the molten iron quality in BF ironmaking process.

On-line Update Method for the Sensitivity Consistency Model of Interconnected Power Grid Based on Electric Power Big Data

Real-time fast calculation of the power flow of the interconnected power grid is an important guarantee for the reliable operation of the interconnected power grid. The topology of the interconnected power grid is complex, and the calculation of the power flow of the whole network is large and timeconsuming. The sensitivity equivalent model can effectively simplify the interconnected power grid and shorten the time of the power flow calculation of the whole network. The operating state of the power grid is constantly changing. In order to ensure the accuracy of the power flow calculation results, it is necessary to update the uniform sensitivity equivalent model in real time. Due to factors such as the vertical management system between the interconnected power grids and the principle of commercial confidentiality, it is difficult to share information between interconnected power grids in real time, and the sensitivity equivalent model cannot be updated in real time, resulting in too much error in the calculation results and no reference value. To solve this problem, this paper proposes an online update method for the sensitivity equivalent model of the interconnected power grid based on power big data to solve the problem of excessive power flow calculation errors caused by the untimely update of the equivalent model parameters, and to ensure the operational reliability of the interconnected power grid.

Joint-level force sensing for indirect hybrid force/position control of continuum robots with friction

Continuum robots offer the dexterity and obstacle circumvention capabilities necessary to enable surgery in deep surgical sites. They also can enable joint-level ex situ force sensing (JEFS), which provides an estimate of end-effector wrenches given joint-level forces. Prior works on JEFS relied on a restrictive embodiment with minimal actuation line friction and captured model and frictional actuation transmission uncertainties using a configuration space formulation. In this work, we overcome these limitations. First, frictional losses are canceled using a feed-forward term based on support vector regression in joint space. Then, regression maps and their interpolation are used to account for actuation hysteresis. The residual joint-force error is then further minimized using a least-squares model parameter update. An indirect hybrid force/position controller using JEFS is presented with evaluation carried out on a realistic pre-clinically deployable insertable robotic effectors platform (IREP) for single-port access surgery. Automated mock force-controlled ablation, exploration, and knot tightening are evaluated. A user study involving the daVinci Research Kit surgeon console and the IREP as a surgical slave was carried out to compare the performance of users with and without force feedback based on JEFS for force-controlled ablation and knot tightening. Results in automated experiments and a user study of telemanipulated experiments suggest that intrinsic force-sensing can achieve levels of force uncertainty and force regulation errors of the order of 0.2 N. Using JEFS and automated task execution, repeatability, and force regulation accuracy is shown to be comparable to using a commercial force sensor for human-in-the-loop feedback.

Probabilistic Load Forecasting Based on Adaptive Online Learning

Load forecasting is crucial for multiple energy management tasks such as scheduling generation capacity, planning supply and demand, and minimizing energy trade costs. Such relevance has increased even more in recent years due to the integration of renewable energies, electric cars, and microgrids. Conventional load forecasting techniques obtain single-value load forecasts by exploiting consumption patterns of past load demand. However, such techniques cannot assess intrinsic uncertainties in load demand, and cannot capture dynamic changes in consumption patterns. To address these problems, this paper presents a method for probabilistic load forecasting based on the adaptive online learning of hidden Markov models. We propose learning and forecasting techniques with theoretical guarantees, and experimentally assess their performance in multiple scenarios. In particular, we develop adaptive online learning techniques that update model parameters recursively, and sequential prediction techniques that obtain probabilistic forecasts using the most recent parameters. The performance of the method is evaluated using multiple datasets corresponding with regions that have different sizes and display assorted time-varying consumption patterns. The results show that the proposed method can significantly improve the performance of existing techniques for a wide range of scenarios.

Mechanics‐based model updating for identification and virtual sensing of an offshore wind turbine using sparse measurements

Offshore wind turbines are complex systems operating in harsh environment. The dynamic demands in these systems often differ from values used in design, leading to unexpected mechanical and structural failures. This signifies the importance of remote monitoring technologies for damage diagnosis and prognosis in offshore wind turbines. This study is focused on developing mechanics-based digital twins for offshore wind turbine monitoring through a model-updating process using sparse measurement data. Digital twins can be used to estimate the system unmeasured response (i.e., virtual sensing) and to predict the remaining useful fatigue life and failure point of different structural components. A time-domain sequential Bayesian finite element model updating is proposed for mechanics-based digital twinning. This approach is formulated for application to offshore wind turbine and jointly estimates the updating model parameters and the time history of unknown input forces. A classical modal-based model updating followed by modal expansion method is also implemented for comparison. In this approach, updating model parameters are estimated to minimize the discrepancies between the identified and model-predicted modal parameters of the turbine. The performance of these two approaches are studied on a 2-MW offshore wind turbine at the Blyth wind farm in the United Kingdom. Strain response time history at mudline is estimated through both approaches and compared with actual measurements for validation. It is observed that both approaches are capable of accurate response prediction while the Bayesian approach leads to slightly better results. Furthermore, the Bayesian approach allows for identification of input loads and uncertainty quantification.

Model prediction-based battery-powered heating method for series-connected lithium-ion battery pack working at extremely cold temperatures

The degraded performance of lithium-ion batteries at low temperatures is a key obstacle to the development of battery energy storage system applied in extremely cold environment. Therefore, this paper proposes a heating method based on model prediction to support the low-temperature operation of battery pack without additional power sources. Battery pack model is developed based on Thevenin equivalent circuit model. A co-estimator is established to update model parameters and state-of-charge online using adaptive recursive least squares and extended Kalman filter. The permissible discharging current of pack is predicted based on multiple constraints to prevent over-discharge. Then, the battery-powered heating structure, control circuit, and heating strategy are designed. The strategy contains a preheating process for cold-start and a holding process for stabilizing cell temperature. The method is verified experimentally through systematic battery-in-the-loop tests at the environmental temperature of – 40 °C. Results show that the method can uniformly preheat all in-pack cells from − 40 °C to − 20 °C in 330 s consuming 4.7% of nominal capacity. In holding process, it is energy-efficient to raise cell temperature continuously and then maintain at 5 °C, which makes 68.3% of nominal capacity available when loading a modified federal urban driving schedule.

Soliciting Human-in-the-Loop User Feedback for Interactive Machine Learning Reduces User Trust and Impressions of Model Accuracy

Mixed-initiative systems allow users to interactively provide feedback to potentially improve system performance. Human feedback can correct model errors and update model parameters to dynamically adapt to changing data. Additionally, many users desire the ability to have a greater level of control and fix perceived flaws in systems they rely on. However, how the ability to provide feedback to autonomous systems influences user trust is a largely unexplored area of research. Our research investigates how the act of providing feedback can affect user understanding of an intelligent system and its accuracy. We present a controlled experiment using a simulated object detection system with image data to study the effects of interactive feedback collection on user impressions. The results show that providing human-in-the-loop feedback lowered both participants’ trust in the system and their perception of system accuracy, regardless of whether the system accuracy improved in response to their feedback. These results highlight the importance of considering the effects of allowing end-user feedback on user trust when designing intelligent systems.

PIRATE: A Blockchain-Based Secure Framework of Distributed Machine Learning in 5G Networks

In fifth-generation (5G) networks and beyond, communication latency and network bandwidth will be no longer be bottlenecks to mobile users. Thus, almost every mobile device can participate in distributed learning. That is, the availability issue of distributed learning can be eliminated. However, model safety will become a challenge. This is because the distributed learning system is prone to suffering from byzantine attacks during the stages of updating model parameters and aggregating gradients among multiple learning participants. Therefore, to provide the byzantine-resilience for distributed learning in the 5G era, this article proposes a secure computing framework based on the sharding technique of blockchain, namely PiRATE. To prove the feasibility of the proposed PiRATE, we implemented a prototype. A case study shows how the proposed PiRATE contributes to distributed learning. Finally, we also envision some open issues and challenges based on the proposed byzantine- resilient learning framework.

Multi-state joint estimation for a lithium-ion hybrid capacitor over a wide temperature range

Lithium-ion hybrid capacitors (LIHCs) have the advantages of high energy and power densities and long cycle life. This paper proposes a multi-state joint estimation method for LIHCs. The effects of temperature on the key LIHC parameters are analyzed. A first-order resistor-capacitor (RC) equivalent circuit model is used to characterize the LIHC dynamics. The coupling relationships among multiple states and LIHC parameters are derived. A recursive least squares algorithm with a variable forgetting factor (VFF-RLS) is employed to adaptively update model parameters at different temperatures. Subsequently, the online identified open-circuit voltage (OCV) is corrected by using a long short-term memory (LSTM) artificial neural network model and an ampere-hour integration method. With the corrected OCV curves and the other model parameters, the estimations of the state of charge (SOC), the remaining useful energy (RUE) and the state of energy (SOE) under dynamic operating conditions over a wide temperature range are achieved. The accuracy of the multi-state joint estimation is verified via the Federal Urban Driving Schedule (FUDS) test and the dynamic stress test (DST) at −10 °C, 25 °C and 50 °C. The root-mean-square errors (RMSEs) of the SOC, RUE and SOE estimations are 2.1%, 0.9 Wh, and 2.3%, respectively.

Deep Learning for Portfolio Optimization

We adopt deep learning models to directly optimise the portfolio Sharpe ratio. The framework we present circumvents the requirements for forecasting expected returns and allows us to directly optimise portfolio weights by updating model parameters. Instead of selecting individual assets, we trade Exchange-Traded Funds (ETFs) of market indices to form a portfolio. Indices of different asset classes show robust correlations and trading them substantially reduces the spectrum of available assets to choose from. We compare our method with a wide range of algorithms with results showing that our model obtains the best performance over the testing period, from 2011 to the end of April 2020, including the financial instabilities of the first quarter of 2020. A sensitivity analysis is included to understand the relevance of input features and we further study the performance of our approach under different cost rates and different risk levels via volatility scaling.

Dynamic Scheduling for Stochastic Edge-Cloud Computing Environments Using A3C Learning and Residual Recurrent Neural Networks

The ubiquitous adoption of Internet-of-Things (IoT) based applications has resulted in the emergence of the Fog computing paradigm, which allows seamlessly harnessing both mobile-edge and cloud resources. Efficient scheduling of application tasks in such environments is challenging due to constrained resource capabilities, mobility factors in IoT, resource heterogeneity, network hierarchy, and stochastic behaviors. Existing heuristics and Reinforcement Learning based approaches lack generalizability and quick adaptability, thus failing to tackle this problem optimally. They are also unable to utilize the temporal workload patterns and are suitable only for centralized setups. However, asynchronous-advantage-actor-critic (A3C) learning is known to quickly adapt to dynamic scenarios with less data and residual recurrent neural network (R2N2) to quickly update model parameters. Thus, we propose an A3C based real-time scheduler for stochastic Edge-Cloud environments allowing decentralized learning, concurrently across multiple agents. We use the R2N2 architecture to capture a large number of host and task parameters together with temporal patterns to provide efficient scheduling decisions. The proposed model is adaptive and able to tune different hyper-parameters based on the application requirements. We explicate our choice of hyper-parameters through sensitivity analysis. The experiments conducted on real-world data set show a significant improvement in terms of energy consumption, response time, Service-Level-Agreement and running cost by 14.4, 7.74, 31.9, and 4.64 percent, respectively when compared to the state-of-the-art algorithms.