Debarshi Patanjali Ghoshal

Selected Publications

• Debarshi Patanjali Ghoshal, Hannah Michalska. Finite interval estimation of LTI systems using differential invariance, instrumental variables, and covariance weighting. The 2020 American Control Conference (ACC 2020)
  See publication
  Abstract: It is shown how the kernel approach to joint parameter and state estimation can be improved to handle large measurement noise. High accuracy of estimation results from combining the powers of the kernel representation of the differential invariance in the system, a feasible recursive version of the generalized least squares with covariance weighting to eliminate regression dilution and suitable choices of instrumental variables to compensate for the error-in-the-variable in the stochastic regression formulation.

• Debarshi Patanjali Ghoshal, Hannah Michalska. Finite-interval kernel-based identification and state estimation for LTI systems with noisy output data. The 2019 American Control Conference (ACC 2019)
  See publication
  Abstract: This note extends previous results pertaining to algebraic state and parameter estimation of linear systems based on a special construction of kernel system representations that incorporate system differential invariants. Main results include explicit expressions for the kernel functions for single-input, single-output LTI systems of arbitrary order. A recursive regression type algorithm is also proposed for the purpose of joint system identification and finite interval filtering. As compared with previous results the proposed non-asymptotic estimation method proves remarkably robust to Gaussian noise in output measurements. The approach has been shown to extend to linear time-varying and linear parameter-varying systems in a multivariate setting. The idea of system-related kernels can further be employed to enhance convergence properties of moving-window and minimum energy nonlinear filtering methods.

• Debarshi Patanjali Ghoshal, Shaunak Sinha, Hannah Michalska. Algebraic nonlinear identification and output tracking control of synchronous generator using differential flatness. The 23rd International Conference on System Theory, Control and Computing (ICSTCC 2019)
  See publication
  Abstract: A kernel-based approach is explored to enhance robustness of flatness-based nonlinear tracking control design for a synchronous generator machine. The design involves full system identification and nonlinear filtering of the system state, to permit effective implementation of a nonlinear controller based on differential flatness of the model. The difficulty associated with robust implementations of flatness-based controllers resides in the necessity of fast and accurate estimation of higher order derivatives of the noisy, observed flat output. The recently developed forward-backward kernel estimation methods, lend themselves powerfully for this task. Two LTI surrogate models are used with the nonlinear model of the machine to serve identification and filtering of the state, and are switched seamlessly to generate persistent excitation for the purpose of a complex nonlinear identification of all the system parameters. The approach does not require a separate start-up phase for identification purposes. The need for on-line adaptive identification and associated re-tuning of the controller is detected and implemented during full operation of the machine. Neither the identification nor the state estimation procedures need any re-initialization while rendering improved accuracy of derivative estimates due to the forward-backward smoothing feature of the kernels involved.

• Abhishek Pandey, Debarshi Patanjali Ghoshal, Hannah Michalska. Variational approach to joint linear model and state estimation. The 2018 American Control Conference (ACC 2018)
  See publication
  Abstract: An approach similar to variational data assimilation is presented for simultaneous model and state estimation of linear systems from output measurement perturbed by large noise of unknown characteristic. Estimation can be carried out over time windows of arbitrary length and is based on a special kernel representation of a hypothetical linear model. A functional criterion is employed to assess the distance of a model to a data cloud. The reproducing kernel representation of the model is particularly advantageous as it permits a Hilbert subspace representation of the model sought. The approach is similar to variational data assimilation as it relies on the adjoint method for calculation of the gradient of the cost functional involved.

• Deepak Sridhar, Debarshi Patanjali Ghoshal, Hannah Michalska. B-splines in joint parameter and state estimation in linear time-varying systems. The 2018 American Control Conference (ACC 2018)
  See publication
  Abstract: A kernel functional representation of linear time-varying systems is employed in conjunction with B-spline functional approximation techniques to construct non-asymptotic state and parameter estimators for LTV systems. Total observability of the estimated system must be assumed. Practical identifiability conditions for parametric estimation are also stated. In the absence of output measurement noise the observer provides almost exact reconstruction of the system state and delivers high fidelity functional estimates of the time varying system parameters. It also shares the usual superior features of algebraic observers such as independence of the initial conditions of the system and good noise attenuation properties. Other advantages of the kernel and B-spline based identification of linear time-varying systems are elucidated.

• Debarshi Patanjali Ghoshal, Hannah Michalska. Double-sided kernel observer for linear time-varying systems. IEEE Conference on Control Technology and Applications (CCTA 2017)
  See publication
  Abstract: A double-sided input-output kernel functional representation is developed for the class of totally observable linear time-varying systems with inputs. The double-sided kernel representation is immediately applicable as part of a non-asymptotic state observer for observable LTV systems. In the absence of output measurement noise the observer provides exact state values of the system state in arbitrarily short time. It also shares the usual superior features of algebraic observers such as independence of the initial conditions of the system and good noise attenuation properties. Other advantages of the double-sided input-output kernel functional representation of linear systems are elucidated as the concept can be employed to construct state and parameter estimators for flat nonlinear systems.

• Debarshi Patanjali Ghoshal, Kumar Gopalakrishnan, Hannah Michalska. Kernel-based adaptive multiple model target tracking. IEEE Conference on Control Technology and Applications (CCTA 2017)
  See publication
  Abstract: The novel adaptive multiple-model target tracking algorithm presented here employs a non-asymptotic state and parameter estimator whose design hinges on a non-standard integral system representation. The same estimator can be used for target maneuver detection and isolation and hence constitutes the principal ingredient of the tracking algorithm. The algorithm does not maintain a model bank, but creates and identifies new models in an attempt to best track the measurement data. Such an approach is rendered uniquely possible by the fact that the state and parameter estimator is essentially dead-beat. Practical model identifiability, persistent excitation condition for the measured signal are discussed. Although this first version of the algorithm is deterministic and employs threshold-based maneuver detection, it exhibits good robustness with respect to Gaussian measurement noise.

• Debarshi Patanjali Ghoshal, Kumar Gopalakrishnan, Hannah Michalska. Algebraic parameter estimation using kernel representation of linear systems. The 20th World Congress of the International Federation of Automatic Control (IFAC 2017)
  See publication
  Abstract: This work makes a contribution to algebraic parameter estimation as it proposes a simple alternative to the derivation of the algebraic estimation equations. The idea is based on a system representation in the form of an evaluation functional which does not exhibit any singularities in the neighbourhood of zero. Implied is the fact that algebraic estimation of parameters as well as system states can then truly be performed in arbitrary time and with uniform accuracy over the entire estimation interval. Additionally, the result offers a geometric representation of a linear system as a finite dimensional subspace of a Hilbert space, that readily suggests powerful noise rejection methods in which invariance plays a central role.

• Debarshi Patanjali Ghoshal, Kumar Gopalakrishnan, Hannah Michalska. Using invariance to extract signal from noise. The 2017 American Control Conference (ACC 2017)
  See publication
  Abstract: It is shown how differential invariance can be used to extract an underlying signal from its noisy measurement towards constructing a non-asymptotic state estimator for linear systems. While the model of the system is assumed known, the noise can have arbitrary characteristics. The differential invariance is rendered by the Cayley-Hamilton theorem and the system is represented in terms of a output reproducing functional on a Hilbert subspace. High accuracy, full state estimation of the system is achieved over arbitrary time intervals by way of orthogonal projection onto the subspace that represents the system invariance. Although the results are presented here primarily with reference to SISO LTI systems they readily extend to LTV systems with multiple outputs.

• Debarshi Patanjali Ghoshal, Niladri Das, Samrat Dutta, Laxmidhar Behera. Robot learns from human teacher through modified kinesthetic teaching. International conference on Advances in Control and Optimization of Dynamic Systems (ACODS 2014)
  See publication
  Abstract: Teaching new motor tasks to robots through physical interactions is an important goal for both robotics and machine learning. Most monolithic machine learning approaches fail to scale when going beyond basic skills. In this paper we present a simple framework for teaching the robot (to play tennis) through direct physical interaction with a human teacher (i.e. Kinesthetic Teaching). Current popular established method of kinesthetic teaching generally uses a two-stage approach: First, a library of motor primitives is generated through direct physical manipulation of the robot. In second stage, a reinforced learning (“reward” stage) is implemented to dynamically adjust the policy of choosing from motor primitive library. In this paper, we show that by proper modification of the first stage of Kinesthetic Teaching and incorporating the domain experience of the human teacher, we can remove the necessity of the second stage. This approach has multiple advantages: (i) We can make the whole training process much simpler. This would go a long way in making our training algorithm scalable (for much increased number of basic moves, etc). (ii) One potential problem with the “reward” learning phase is that there may be subjective difference of what is a “good” shot from the kinesthetic teaching and from the bystander viewpoint (later in “reward” stage). Even a little difference in this regard will result in a confusing feedback (“reward”), and hence it would be difficult to correctly figure out which library training samples should be reassigned what weight values. Our approach eliminates this problem altogether.

• Ankush Roy, Debarshi Patanjali Ghoshal. Number plate recognition for use in different countries using an improved segmentation. IEEE Technically Sponsored National Conference on Emerging Trends and Applications in Computer Science (NCETACS 2011)
  [Awarded best paper of the conference]
  See publication
  Abstract: Automatic Number Plate Recognition (ANPR) is a real time embedded system which identifies the characters directly from the image of the license plate. It is an active area of research. ANPR systems are very useful to the law enforcement agencies as the need for Radio Frequency Identification tags and similar equipments are minimized. Since number plate guidelines are not strictly practiced everywhere, it often becomes difficult to correctly identify the non-standard number plate characters. In this paper we try to address this problem of ANPR by using a pixel based segmentation algorithm of the alphanumeric characters in the license plate. The non-adherence of the system to any particular country-specific standard & fonts effectively means that this system can be used in many different countries — a feature which can be especially useful for trans-border traffic e.g. use in country borders etc. Additionally, there is an option available to the end-user for retraining the Artificial Neural Network (ANN) by building a new sample font database. This can improve the system performance and make the system more efficient by taking relevant samples. The system was tested on 150 different number plates from various countries and an accuracy of 91.59% has been reached.

Back