Black-box System Identification Research Articles

An online reinforcement learning approach for HVAC control

Heating, Ventilation and Air Conditioning (HVAC) optimization for energy consumption reduction is becoming ever more a topic of the utmost environmental and energetic concerns. The two most employed methodologies for optimizing HVAC systems are Model Predictive Control (MPC) and Reinforcement Learning (RL). This paper compares three different RL approaches to HVAC optimization: one based on a black-box system identification model trained on historical data, one based on a white-box model of a building and one online method based on an imitation learning pretraining phase on historical data. The three approaches are compared with a literature baseline and an EnergyPlus baseline. Results show that the overall best method in terms of energy consumption reduction (65% decrease) and thermal comfort increase (25% increase) is the approach based on the white-box model. However, the proposed methodology, based on online and imitation learning, demonstrates remarkable efficiency, achieving comparable improvements in energy consumption after just a few months of online training, while maintaining thermal comfort at around the same level as the baseline. These results prove a direct online RL approach, which avoid the use of costly simulations, can provide a reliable and inexpensive solution to the problem of HVAC optimization.

Flydeling: Streamlined Performance Models for Hardware Acceleration of CNNs through System Identification

The introduction of deep learning algorithms, such as Convolutional Neural Networks (CNNs) in many near-sensor embedded systems, opens new challenges in terms of energy efficiency and hardware performance. An emerging solution to address these challenges is to use tailored heterogeneous hardware accelerators combining processing elements of different architectural natures such as Central Processing Unit (CPU), Graphics Processing Unit (GPU), Field Programmable Gate Array (FPGA), or Application Specific Integrated Circuit (ASIC). To progress towards heterogeneity, a great asset would be an automated design space exploration tool that chooses, for each accelerated partition of a CNN, the most appropriate architecture considering available resources. To feed such a design space exploration process, models are required that provide very fast yet precise evaluations of alternative architectures or alternative forms of CNNs. Quick configuration estimation could be achieved with few parameters from representative input sequences. This article studies a solution called flydeling (as a contraction of fly weight mo deling ) for obtaining these models by inspiring from the black-box System Identification (SI) domain. We refer to models derived using the proposed approach as fly weight mo dels ( flydels ). A methodology is proposed to generate these flydels , using CNN properties as predictor features together with SI techniques with a stochastic excitation input at a feature map dimensions level. For an embedded CPU-FPGA-GPU heterogeneous platform, it is demonstrated that it is possible to learn these Key Performance Indicators (KPIs) flydels at an early design stage and from high-level application features. For latency, energy, and resource utilization, flydels obtain estimation errors varying between 5% and 10% with less model parameters compared to state-of-the-art solutions and are built automatically from platform measurements.

Ultrasound shear wave phase velocity imaging using black-box system identification (BSI): a data-driven approach

Objective. Shear wave elasticity imaging (SWEI) is an established approach for diagnosing lesions in human tissue. However, the shear wavelengths used by traditional SWEI are usually not short enough, leading to inferior accuracy in small-target (<10 mm) reconstruction. To exploit short shear wavelengths (high-frequency content), this study introduces a new phase velocity (PV) estimation technique as an alternative to the conventional group velocity (GV) modality. Approach. We propose using a black-box model instead of a fully physics-based model to describe the transition process of two arbitrary shear wave signals. With this representation, local PV can be obtained via black-box system identification (BSI). For validation, two PV estimation scenarios were established: (numerical) dispersion measurements in viscoelastic media, and (real) imaging targets in a CIRS elasticity phantom. BSI was compared with a state-of-the-art PV imaging method that uses local wavenumber estimation (LWE). Main results. BSI showed excellent accuracy in the dispersion estimation for all three viscoelastic media in the simulations. In the phantom study, the two PV methods exhibited good agreement in the frequency dependence of target quantification, and could both generate a higher target reconstruction accuracy than GV. LWE images were strongly affected by noise-induced faulty estimates, whereas BSI showed no notable artifacts. Significance. This study demonstrates the advantage of the PV modality over the GV modality as the former can achieve better target visualization by increasing imaging frequency. It also implies the feasibility of data-driven modeling for soft tissue characterization.

Nonlinear ensemble gray and black-box system identification of friction induced vibrations in slender rotating structures

Due to the nonlinear bit-rock and drill–string–borehole interactions, the modeling and analysis of drill string dynamics are a challenging work. These slender structures are subjected to torsional, lateral, and axial vibration modes, which are generally prejudicial to the drilling process. For this reason, proper modeling of the system dynamics is necessary to optimize drilling performance. In the present paper, we focus on the torsional dynamics and model of an experimental system subjected to friction torque. An ensemble approach, combining gray and black-box modeling techniques, is used to calibrate some of the dynamical system parameters, particularly those related to friction. This combination is chosen with the aim of improving the prediction accuracy of a pure gray-box approach while retaining interpretability in the dynamic responses. The present paper compares four well-known friction models with increasing levels of complexity using experimental data. It is shown that the ensemble model proposed can improve the gray-box by up to 92.68% in terms of prediction errors, by adding a non-physical layer to it. The present analysis demonstrates that better results can be obtained by using an ensemble of gray and black-box identification techniques for friction modeling and simulation, aiming at improving data-driven model construction of complex torsional dynamics in slender structures.

Impedance Parameters Estimation of An RLCM Ladder Network Using Subspace and Similarity Transformation Approach

ABSTRACT Numerous physical systems such as transformers, sensors, and biomaterials are analysed using their equivalent ladder network model. Impedance parameters of a ladder network consist of combinations of resistances, inductances, and capacitances. Accurately determining any device impedance parameters directly from the measured time-domain input-output data is challenging. The use of non-linear optimization algorithms does not always work due to the problem being ill-conditioned. Also, the obtained parameters may not be physically realizable. In this work, the problem is approached from a black-box system identification point of view, followed by a physical parameter extraction step. The necessary current and voltage equations obtained from first principles are utilized to build a physical model in the state-space domain. With the use of a subspace system identification algorithm, the black-box parameter matrices are determined from input-output data, which do not correspond to the physical parameter matrices. However, both sets of parameters can be related through a similarity transformation matrix, leading to the extraction of the physical parameters. The novelty of the approach lies in solving the non-linear parameter estimation problem linearly without initialization. The presence of mutual inductances between various transformer winding sections makes its ladder network the most complex one. Therefore, in this paper, the ladder network model of a transformer winding is considered. As an example of the methodology, a six-section transformer ladder network is simulated and the parameter estimation results are used to validate the proposed algorithm. The same approach can be applied to estimate any other device impedance parameters.

A new reference model-based fixed-structure controller for industrial applications

This paper introduces a new approach to design a fixed-structure controller for those industrial applications, where mathematical model of the plant is not known. The design methodology is composed of two steps. First, black-box system identification is used to obtain a mathematical model of plant and second, a fixed-structure controller is designed based on reference model defined by closed-loop performance. For controller identification, prediction error method (PEM) is considered, which uses the off-line generated input-output inherently noise free data using the reference model. Due to fixed design structure, the controller is independent of plant's order. The proposed approach is tested on a lab-based real-time magnetic levitation system (MLS). The results show effectiveness of the proposed approach with good reference tracking and robustness properties. The proposed approach emerged a lower order controller which is computationally efficient and easy to implement. Compare to other data-driven techniques, inverse mapping of the reference model is not required and the performance versus design/tuning time comparison with conventional proportional integral derivative (PID) controller is also improved.

Estimation of Humidity Sensor Equivalent Circuit Parameters Utilizing Frequency Response Measurements Combined With Subspace Identification and Similarity Transformations

This article presents an algorithm to estimate the impedance parameters of a humidity sensor from the measured frequency response data. The sensor is modeled as a multielement two-terminal network. The procedure accurately identifies the impedance parameters using a noniterative algorithm. Traditional nonlinear least-squares procedures rely on having initial parameter values close to the actual values for converging. We approach the problem in two steps: a black-box system identification and a physical parameter extraction step. Based on the first principles, a physical state-space model is developed, where the system matrices contain the impedance parameters in very complicated ways. By strategically introducing extra variables that are a function of these problematic parameters, the state-space model is converted into an observable canonical form. Using frequency response data, a subspace algorithm identifies a model of the same order as the physical model. An unknown similarity transformation matrix is then used to establish the relationship between the identified model and the canonical model. In the first step, we determine the additional variables and the transforming matrix. The impedance parameters are obtained from the relationship between the additional variables and the physical parameters. Simulations and experiments validate the proposed algorithm.

Non-Autoregressive vs Autoregressive Neural Networks for System Identification

The application of neural networks to non-linear dynamic system identification tasks has a long history, which consists mostly of autoregressive approaches. Autoregression, the usage of the model outputs of previous time steps, is a method of transferring a system state between time steps, which is not necessary for modeling dynamic systems with modern neural network structures, such as gated recurrent units (GRUs) and Temporal Convolutional Networks (TCNs). We compare the accuracy and execution performance of autoregressive and non-autoregressive implementations of a GRU and TCN on the simulation task of three publicly available system identification benchmarks. Our results show, that the non-autoregressive neural networks are significantly faster and at least as accurate as their autoregressive counterparts. Comparisons with other state-of-the-art black-box system identification methods show, that our implementation of the non-autoregressive GRU is the best performing neural network-based system identification method, and in the benchmarks without extrapolation, the best performing black-box method.

Condition Monitoring of a Paper Feeding Mechanism Using Model-of-Signals as Machine Learning Features

Prognostics and Health Management of machine devices and parts is a hot topic in the Industry 4.0 era. In this fashion, automated procedures to evaluate machinery working conditions are essential to minimize downtime and maintenance costs. In this work, we study how to monitor the decrease in performance of a paper sheet feeder for the packaging industry under heavy-duty cycle operations. The main measurable outcome of such degradation is the increase in backlash among the device moving components. A wide variety of methods and procedures is available to tackle this monitoring problem. In this paper, we analyze the use of a simple yet efficient diagnosis methodology that can exploit machinery controllers (i.e., Programmable Logic Controllers) edge-computing capabilities. Vibration measurements are known in the literature to retain information about the system's mechanics. Model-of Signals, a data-driven approach based on black box system identification, allows to extract that information reliably during machinery working cycle. The refinement of those data using machine learning allows the retrieval of knowledge about the health state of the machine. In this study, the feeder mechanism is run to failure with its parts backlash measured at given time intervals. Accelerometer signals are modelled as AutoRegressive processes whose coefficients are then considered as features to feed to machine learning algorithms, which are employed to perform severity evaluation of the ongoing degradation. Estimation and prediction are both implementable on-board the controller, while the learning task can be carried out remotely, in a cloud computing perspective. The exploitation of AutoRegressive modelling gives a simple and inherent methodology for feature selection, serving as a foundation of the machine learning stage. We make use of a Support Vector Machine algorithm to analyze how obtained models represent the various levels of backlash in the device and develop a suitable predictor of the degradation severity. Finally, the results of the application of the methodology to the case study are shown.

Nonlinear black-box system identification through coevolutionary algorithms and radial basis function artificial neural networks

The present work deals with the application of coevolutionary algorithms and artificial neural networks to perform input selection and related parameter estimation for nonlinear black-box models in system identification. In order to decouple the resolution of the input selection and parameter estimation, we propose a problem decomposition formulation and solve it by a coevolutionary algorithm strategy. The novel methodology is successfully applied to identify a magnetorheological damper, a continuous polymerization reactor and a piezoelectric robotic micromanipulator. The results show that the method provides valid models in terms of accuracy and statistical properties. The main advantage of the method is the joint input and parameter estimation, towards automating a tedious and error prone procedure with global optimization algorithms.

A Novel Method and Implementation of Digital Algorithm for First Order Non Linear Spherical Tank Level Control Process with an Integral Windup

The primary goal of the process control is to control, monitor and to maintain different kinds of process stations at the desired stable conditions for improving the productivity and product quality. In industries most of the process are non - linear, controlling such type of system is difficult. So, spherical tanks are one among the non-linear systems which can be used to store maximum quantity of product in minimum area. These types of non-linear systems violate the superposition principle and are very difficult to control and monitor. Generally, conventional PID Controller is used for controlling a process in a plant. In Level control process, there exists the sudden change in final control element which leads to maximum overshoot and larger settling time, this stated phenomenon requires tuning of gain values to control the process variable to the set point. In order to overcome the above-mentioned problem, Anti reset windup concept is preferred. The mathematical modelling of a given process is obtained using black box system identification methodology. The accuracy in controlling the process variable is complicated due to its dynamic behaviour and characteristics. So in order to maintain the proper characteristics of non-linear system the mathematical modelling has been obtained by conducting the open loop test. In ancient stage the Ziegler Nicholas method is implemented to tune the PID parameters such as Kp, τi and τd which takes larger settling time, maximum Integral Square error and maximum peak overshoot. After tuning the PID gain values with the help of anti reset windup, the overshoot and settling time will be reduced. In this proposed work, Deadbeat and Dahlin’s Algorithms are used to convert the transfer function of the level process from ‘S’ domain to ‘Z’ domain (digital controller). The transfer function is used in MATLAB (Matrix Laboratory) to analyse the output of level process and the simulations are taken for the plant with above mentioned two different algorithms. For real time interface the LabVIEW platform is used and the results are discussed. This proposed digital controller with anti reset windup provides lower steady state error, minimum peak overshoot and minimum error criteria indices.

Black Box Attack of Simulation Secret Key using Neural-Identifier

The issue of the cryptanalysis is defined as the unknown issue or problem related to the identification of the system where the major goal of the cryptanalysis is for the design of the system for various steps involved. Neural networks will be ideal tool for Black-Box system identification. The black-Box attacks against secret key cryptosystems (stream cipher system) would be presented by considering a Black-Box Neuro-Identifier model to retain two different objectives: first is for finding out the key from the provided plaintext-cipher to pair, while the second objective is to emulator construction a neuro-model for the target cipher system. There are many researches going on considering the various models of encryption where ANN is being used as single layered or multi layered perceptron. The above defined cryptographic techniques are sometimes also termed as the Neural Cryptography. As the ANN model relies on the feedforward working criteria means it can be used for the generation of some effective and efficient encryption methodologies. Cryptanalysis is considered as significant footstep for evaluating and checking quality of any cryptosystem. A portion of these cryptosystem guarantees secrecy and security of huge data trade from source to goal utilizing symmetric key cryptography. The cryptanalyst researches the quality and distinguishes the shortcoming of the key just as enciphering calculation. With the expansion in key size, the time and exertion required anticipating the right key increments. These systems for cryptanalysis are changing radically to decrease cryptographic multifaceted nature. In this paper a point by point study has been directed. Much cryptography strategies are accessible which depend on number hypothesis however it has the hindrance of necessity a substantial computational power, unpredictability and time utilization. To defeat these disadvantages, artificial neural networks (ANNs) have been connected to take care of numerous issues. The ANNs have numerous qualities, for example, learning, speculation, less information necessity, quick calculation, simplicity of usage, and programming and equipment accessibility, which make it exceptionally alluring for some applications. This paper gives a cutting-edge survey on the utilization of counterfeit neural systems in cryptography and concentrates their execution on estimation issues identified with cryptography.

Transformer frequency response: a new technique to analyze and distinguish the low-frequency band in the frequency response analysis spectrum

Smart monitoring applications in all aspects of power systems are becoming increasingly popular because of the indirect effect on energy management and the smart grid. In this regard, high voltage equipment monitoring and real-time assessment are critical. Among these evaluations, transformer life and condition assessment has been considered for many years [1], [2]. Currently, advanced methods such as transfer function measurement are employed to determine the transformer condition [3]. Transfer function measurement is generally used for black box system identification [4]. This measurement might be implemented in the time domain as an impulse response or in the frequency domain using a swept signal; the latter approach is commonly called frequency response analysis (FRA) [5].

Information Theoretic Causality Measures for System Identification of Mechanical Systems

Parameter estimation and model order reduction (MOR) are important system identification techniques used in the development of models for mechanical systems. A variety of classical parameter estimation and MOR methods are available for nonlinear systems but performance generally suffers when little is known about the system model a priori. Recent advancements in information theory have yielded a quantity called causation entropy (CSE), which is a measure of influence between elements in a multivariate time series. In parameter estimation problems involving dynamic systems, CSE can be used to identify which state transition functions in a discrete-time model are important in driving the system dynamics, leading to reductions in the dimensionality of the parameter space. This method can likewise be used in black box system identification problems to reduce model order and limit issues with overfitting. Building on the previous work, this paper illustrates the use of CSE-enabled parameter estimation for nonlinear mechanical systems of varying complexity. Furthermore, an extension to black-box system identification is proposed wherein CSE is used to identify the proper model order of parameterized black-box models. This technique is illustrated using nonlinear differential equation (NDE) models of physical devices, including a nonlinear spring–mass–damper, a pendulum, and a nonlinear model of a car suspension. Overall, the results show that CSE is a promising new tool for both gray-box and black-box system identification that can speed convergence toward a parameter solution and mitigate problems with model overfitting.

Frequency Domain Subspace Identification of Multivariable Dynamical Systems for Robust Control Design

Black-box system identification is subjected to the modelling uncertainties that are propagated from the non-parametric model of the system in time/frequency-domain. Unlike classical H1/H2 spectral analysis, in the recent robust Local Polynomial Method (LPM), the modelling variances are separated to noise contribution and nonlinear contribution while suppressing the transient noise. On the other hand, without an appropriate weighting on the objective function in the system identification methods, the acquired model is subjected to bias. Consequently, in this paper the weighted regression problem in subspace frequency-domain system identification is revisited in order to have an unbiased estimate of the frequency response matrix of a flexible manipulator as a multi-input multi-output lightly-damped system. Although the unbiased parametric model representing the best linear approximation (BLA) of the system in this combination is a reliable framework for the control design, it is limited for a specific signal-to-noise (SNR) ratio and standard deviation (STD) of the involved input excitations. As a result, in this paper, an additional step is carried out to investigate the sensitivity of the identified model w.r.t. SNR/STD in order to provide an uncertainty interval for robust control design.

Real-time Audio Processing with a Cascade of Discrete-Time Delay Line-Based Reservoir Computers

Background: Real-time processing of audio or audio-like signals is a promising research topic for the field of machine learning, with many potential applications in music and communications. We present a cascaded delay line reservoir computer capable of real-time audio processing on standard computing equipment, aimed at black-box system identification of nonlinear audio systems. The cascaded reservoir blocks use two-pole filtered virtual neurons to match their timescales to that of the target signals. The reservoir blocks receive both the global input signal and the target estimate from the previous block (local input). The units in the cascade are trained in a successive manner on a single input output training pair, such that a successively better approximation of the target is reached. A cascade of 5 dual-input reservoir blocks of 100 neurons each is trained to mimic the distortion of a measured guitar amplifier. This cascade outperforms both a single delay reservoir having the same total number of neurons as well as a cascade with only single-input blocks. We show that the presented structure is a viable platform for real-time audio applications on present-day computing hardware. A benefit of this structure is that it works directly from the audio samples as input, avoiding computationally intensive preprocessing.

A new structure for non-linear black-box system identification using the extended Kalman filter

A new tool for black-box non-linear system identification of multi-input multi-output systems is presented in this paper. The new structure extends the conventional linear state-space model into a non-linear framework, where each parameter is a non-linear function of the inputs or the states. The method works iteratively in the time domain using an extended Kalman filter. The model retains a state-space structure in modal canonical form, which ensures that a minimal number of parameters need to be identified and also produces additional information in terms of the system eigenvalues and the dominant modes. This structure is a completely black-box system, which requires no physical understanding of the process for successful identification, and it is possible to expand easily the order and the complexity of non-linearities, while ensuring good parameter conditioning. A simple non-linear example illustrates the method, and identification of a highly non-linear brake model is also presented. These examples show that the method can be applied as a mechanism for model order reduction; it is equally very suitable as a tool for non-linear plant system identification. In both capacities this new method is valuable, particularly as the generation of simplified models for the whole vehicle and its subsystems is an increasingly important aspect of modern vehicle design.

Robust Estimation of Balanced Simplicity-Accuracy Neural Networks-Based Models

Neural networks are powerful tools for black box system identification. However, their main drawback is the large number of parameters usually required to deal with complex systems. Classically, the model's parameters minimize a L2-norm-based criterion. However, when using strongly corrupted data, namely, outliers, the L2-norm-based estimation algorithms become ineffective. In order to deal with outliers and the model's complexity, the main contribution of this paper is to propose a robust system identification methodology providing neuromodels with a convenient balance between simplicity and accuracy. The estimation robustness is ensured by means of the Huberian function. Simplicity and accuracy are achieved by a dedicated neural network design based on a recurrent three-layer architecture and an efficient model order reduction procedure proposed in a previous work (Romero-Ugalde et al., 2013, “Neural Network Design and Model Reduction Approach for Black Box Nonlinear System Identification With Reduced Number of Parameters,” Neurocomputing, 101, pp. 170–180). Validation is done using real data, measured on a piezoelectric actuator, containing strong natural outliers in the output data due to its microdisplacements. Comparisons with others black box system identification methods, including a previous work (Corbier and Carmona, 2015, “Extension of the Tuning Constant in the Huber's Function for Robust Modeling of Piezoelectric Systems,” Int. J. Adapt. Control Signal Process., 29(8), pp. 1008–1023) where a pseudolinear model was used to identify the same piezoelectric system, show the relevance of the proposed robust estimation method leading balanced simplicity-accuracy neuromodels.

An Empirical Human Controller Model for Preview Tracking Tasks.

Real-life tracking tasks often show preview information to the human controller about the future track to follow. The effect of preview on manual control behavior is still relatively unknown. This paper proposes a generic operator model for preview tracking, empirically derived from experimental measurements. Conditions included pursuit tracking, i.e., without preview information, and tracking with 1 s of preview. Controlled element dynamics varied between gain, single integrator, and double integrator. The model is derived in the frequency domain, after application of a black-box system identification method based on Fourier coefficients. Parameter estimates are obtained to assess the validity of the model in both the time domain and frequency domain. Measured behavior in all evaluated conditions can be captured with the commonly used quasi-linear operator model for compensatory tracking, extended with two viewpoints of the previewed target. The derived model provides new insights into how human operators use preview information in tracking tasks.

A MISO model for power consumption in virtualized servers

Energy efficiency is always a concern in hosting servers. When any new development is added to a host server, the power consumption of the host server must be theoretically and empirically re-evaluated. Because of the ongoing development trends in computing systems at the hardware, software and middleware levels, deriving a direct mathematical model for quantifying the power consumption of a host server is difficult. Therefore, a system identification is used to construct the power consumption model for virtualized hosting servers. To date, three types of system identifications have been used in the literature for defining the power consumption model: the first-principles, the black-box and the gray-box identification approaches. To the best of our knowledge, the majority of these approaches are apparently used to model the power consumption in a single-input single-output (SISO) system model, in which a hardware component is reconfigured to meet the power budget target. In this paper, to accommodate the ongoing development trends in computing systems, we propose a multi-input single-output (MISO) model for modeling the power consumption of virtualized hosting servers. We use the black-box system identification method, and we utilize the Auto-Regressive eXogenous (ARX) mathematical model to construct the MISO power model. We compare our MISO power model with the SISO power model that is used in existing state-of-the-art works. Empirically, our MISO power model exhibits higher accuracy than the existing SISO power model in predicting power consumption. Using our model, we can achieve approximately 98 % accuracy in predicting the power consumption of virtualized hosting servers.