Fraud and money laundering detection in financial transaction data is a highly regulated field where explanations and human decisions are paramount to deploying any automatic detection models. Artificial intelligence (AI) methods are already routinely used for the detection of suspicious transactions, but their application in the final stage of analysis is underexplored. To assess the prospects for AI in this task, we analyze the work of a human analyst as a data processing agent. Specifically, we analyze textual feedback from real human analysts to determine the focus of their work and identify how AI can improve it. We conclude that analysts do not spend their time reanalyzing or explaining the raw transaction data but actively seek additional information. This paradigm is formalized in machine learning as learning with costly features. We applied state-of-the-art methods to studied data and concluded that using the costly features paradigm leads to an optimal trade-off between the cost of the feature acquisition and model performance. We believe that this approach is ready to be applied in the production stage. We provide a full code of our approach as well as a modification of publicly available data to reproduce our results.

Many popular machine learning methods for classification, anomaly detection, clustering, and dimensionality reduction rely on distance functions. However, these methods' theoretical foundations and practical performance critically depend on well-defined and meaningful distance functions on the input space. While distances are well-defined in Euclidean spaces, extending them to popular structured data stored in formats such as JSON, XML, ProtoBuffer, or MessagePack remains challenging. To fill this gap, this work proposes the Hierarchically-Structured Tree Distance (HTD), a fully differentiable distance tailored to these heterogeneous data formats. HTD is flexible, parameterized by weights, and supports automatic recursive construction, enabling it to be used on various datasets and tasks. This effectiveness and generality is demonstrated on a wide range of tasks, such as classification, anomaly detection, fast retrieval (indexing), and clustering, and on diverse datasets. The experimental comparison shows that classical distance-based methods with the HTD often rival or outperform state-of-the-art neural network models with orders of magnitude more parameters. HTD thus opens the door to scalable, interpretable, and efficient modeling of hierarchically structured data.

This paper investigates the utilization of the zero-sequence component (ZSC) to increase the voltage and enhance the operation area of five-phase electric drives with permanent magnet synchronous machines (PMSM). Although the ZSC allows reaching higher voltage, its application in multiphase drives is constrained by the usage of x-y planes activated by harmonic components. Since some odd harmonics allow torque enhancement, these planes are typically involved in voltage and current control as part of field-oriented control strategies (FOC) to maximize the torque while minimizing the losses (e.g., Joule losses in the Maximum Torque per Ampere strategy). In five-phase systems, the third harmonic is used for torque enhancement, which affects the maximum value and shape of the resulting voltage waveform across the torque-speed map. Consequently, the space available for ZSC injection is limited and depends on the actual operation point. This paper presents an analysis of how ZSC can be leveraged within these constraints to improve the torque-speed operating map of a five-phase PMSM drive, especially in the field-weakening region. Experimental results obtained on a laboratory five-phase PMSM setup validate the theoretical findings and demonstrate the practical benefits of the proposed approach.

Deep generative models (DGMs) have recently demonstrated remarkable success in capturing complex probability distributions over graphs. Although their excellent performance is attributed to powerful and scalable deep neural networks, it is, at the same time, exactly the presence of these highly non-linear transformations that makes DGMs intractable. Indeed, despite representing probability distributions, intractable DGMs deny probabilistic foundations by their inability to answer even the most basic inference queries without approximations or design choices specific to a very narrow range of queries. To address this limitation, we propose probabilistic graph circuits (PGCs), a framework of tractable DGMs that provide exact and efficient probabilistic inference over (arbitrary parts of) graphs. Nonetheless, achieving both exactness and efficiency is challenging in the permutation-invariant setting of graphs. We design PGCs that are inherently invariant and satisfy these two requirements, yet at the cost of low expressive power. Therefore, we investigate two alternative strategies to achieve the invariance: the first sacrifices the efficiency, and the second sacrifices the exactness. We demonstrate that ignoring the permutation invariance can have severe consequences in anomaly detection, and that the latter approach is competitive with, and sometimes better than, existing intractable DGMs in the context of molecular graph generation.

Flux linkage maps (FLMs) are routinely used in high-precision control and modeling of synchronous machines. Common methods often consider only the dependence of the FLM on the stator currents, allowing for convenient representation in lookup tables or neural networks. However, the flux linkage also depends on speed, position, and other state variables. Although this is formally simple to add as an additional input to neural models of FLM, the estimation with additional inputs becomes more demanding. We demonstrate that the conventional approach of FLM training using the assumption of a steady-state regime is insufficient to learn the dependency on the rotor position. It is necessary to use the complete ordinary differential equation of the current to learn the FLM model. Even for a shallow neural model of the FLM, the estimation procedure yields a deep learning task known as neural ODE. This procedure essentially generates multistep ahead prediction of differential equations and minimizes the mismatch between the mathematical model and data. The efficiency of this approach is demonstrated on the FLM of a synchronous machine considering flux saturation, speed dependence, and slot harmonics. The proposed approach significantly improves current prediction, yielding improved deadbeat current control. The results are experimentally verified on a 4.5 kW laboratory prototype.

Daily internet communication relies heavily on tree-structured graphs, embodied by popular data formats such as XML and JSON. However, many recent generative (probabilistic) models utilize neural networks to learn a probability distribution over undirected cyclic graphs. This assumption of a generic graph structure brings various computational challenges, and, more importantly, the presence of non-linearities in neural networks does not permit tractable probabilistic inference. We address these problems by proposing sum-product-set networks, an extension of probabilistic circuits from unstructured tensor data to tree-structured graph data. To this end, we use random finite sets to reflect a variable number of nodes and edges in the graph and to allow for exact and efficient inference. We demonstrate that our tractable model performs comparably to various intractable models based on neural networks.

Neural networks are universal function approximations used in modern methods of artificial intelligence (AI) technology. With the increasing adoption of this technology, the need to integrate neural networks into embedded systems grows. However, neural networks are not the only methods with universal approximation properties. In this contribution, we compare the performance of two neural network architectures and one alternative for approximating benchmark functions. Specifically, Multi-Layer Perceptron (MLP), temporal Convolutional Neural Network (1D-CNN), and Piece-Wise Affine function (PWA) are compared. The alternatives are trained and implemented using publicly available tools. The execution time of the resulting code is measured on a system-on-chip (SoC) ARM processor and field programmable gate array (FPGA). The CNNs were the best option for accurate approximation of high dimensional functions, while PWA is the fastest method with good accuracy for low dimensional functions.

Detection of malicious behavior in a large network is a challenging problem for machine learning in computer security, since it requires a model with high expressive power and scalable inference. Existing solutions struggle to achieve this feat—current cybersec-tailored approaches are still limited in expressivity, and methods successful in other domains do not scale well for large volumes of data, rendering frequent retraining impossible. This work proposes a new perspective for learning from graph data that is modeling network entity interactions as a large heterogeneous graph. High expressivity of the method is achieved with neural network architecture HMILnet that naturally models this type of data and provides theoretical guarantees. The scalability is achieved by pursuing local graph inference, i.e., classifying individual vertices and their neighborhood as independent samples. Our experiments exhibit improvement over the state-of-the-art Probabilistic Threat Propagation (PTP) algorithm, show a further threefold accuracy improvement when additional data is used, which is not possible with the PTP algorithm, and demonstrate the generalization capabilities of the method to new, previously unseen entities.

Maximum torque per current (MTPA) and other optimal curves are well-known for three-phase machines. Since multiphase machines dispose by more degrees of freedom than their three-phase counterparts, the solution is much more complicated and a variety of regimes can be found. In this paper, optimal current setpoints for five-phase synchronous machines are analyzed. Besides the first harmonic, the optimization includes also the third harmonic, used for both torque enhancement and waveform shaping. It is shown that the MTPA curves, field weakening (FW) and maximum current (MC) regions have an inner structure with multiple regimes. This variety comes from the phase voltage and phase current limits that yield optimizing waveforms containing multiple peaks (peak shaving). A taxonomy of these regimes is provided, and their mutual relations are demonstrated on a case study machine. The optimal current setpoints in their waveforms are analyzed in selected operating points.

An accurate estimation of flux linkage maps is essential for the proper control and modeling of synchronous machines. We propose to use neural networks as the flux linkage model with training procedure respecting the differential equation of the stator current. Moreover, the neural network allows straightforward extension of the number of input variables. We demonstrate this ability to estimate the flux linkage as a function of rotor speed and position modulated by slot harmonics. The proposed approach is demonstrated on real interior permanent magnet synchronous machine data. The results demonstrate a significant improvement in current prediction compared to commonly used methods.

Deep generative models are challenging the classical methods in the field of anomaly detection nowadays. Every newly published method provides evidence of outperforming its predecessors, sometimes with contradictory results. The objective of this article is twofold: to compare anomaly detection methods of various paradigms with a focus on deep generative models and identification of sources of variability that can yield different results. The methods were compared on popular tabular and image datasets. We identified that the main sources of variability are the experimental conditions: 1) the type of dataset (tabular or image) and the nature of anomalies (statistical or semantic) and 2) strategy of selection of hyperparameters, especially the number of available anomalies in the validation set. Methods perform differently in different contexts, i.e., under a different combination of experimental conditions together with computational time. This explains the variability of the previous results and highlights the importance of careful specification of the context in the publication of a new method. All our code and results are available for download.

The field of blind image deblurring was for a long time dominated by Maximum-A-Posteriori methods seeking the optimal pair of sharp image-blur of a suitable functional. Recently, learning-based methods, especially those based on deep convolutional neural networks, are proving effective and are receiving increasing attention by the research community. In 2020, Ren et al. proposed a deblurring method called SelfDeblur which combines the model-driven approach of traditional MAP methods and the generative power of neural nets. The method is capable of producing very high-quality results, yet it inherits some problems of MAP methods, especially possible convergence to a wrong local optimum. In this paper we propose several easy-to-implement modifications of SelfDeblur, namely suitable initialization, multiscale processing, and regularization, that improve the average performance of the original method and decrease the probability of failure.

Efficiency of an IPMSM motor is influenced by the operating point (OP) of the machine. The optimal operating point can be found by using either direct search methods or model-based methods. Model-based methods are sensitive to parameter uncertainties of the equivalent circuit of the drive. In this paper, the impact of various parameter estimation methods on the motor efficiency is compared. Studied methods are Recursive Least Squares (RLS), frequency domain identification at standstill, and the flux linkage map method. Results are compared to direct search methods, where the efficiency is evaluated using the power consumption measurement and direct measurement of the torque. Comparison is performed on a grid of various setpoints (currents and speed). The RLS is tested in two versions of linearization of the flux: one using only the inductances, the second estimating also their offset. Each method is able to obtain good results for some OPs and bad results for other OPs. Overall, good performance was obtained for direct search, offline identified parameters and flux linkage map and RLS with flux offset. RLS without the flux offset does not yield consistent results which implies significantly lower efficiency.

Conventional Neural Networks can approximate simple arithmetic operations, but fail to generalize beyond the range of numbers that were seen during training. Neural Arithmetic Units aim to overcome this difficulty, but current arithmetic units are either limited to operate on positive numbers or can only represent a subset of arithmetic operations. We introduce the Neural Power Unit (NPU) that operates on the full domain of real numbers and is capable of learning arbitrary power functions in a single layer. The NPU thus fixes the shortcomings of existing arithmetic units and extends their expressivity. We achieve this by using complex arithmetic without requiring a conversion of the network to complex numbers. A simplification of the unit to the RealNPU yields a highly transparent model. We show that the NPUs outperform their competitors in terms of accuracy and sparsity on artificial arithmetic datasets, and that the RealNPU can discover the governing equations of a dynamical system only from data.

Two main aspects of a LQ control of a single-phase converter with a LCL filter are studied in this paper. First, the choice of a penalization structure that prevents instability of the closed loop caused by input saturation due to limited voltage of the DC link. We show that all tested variants of the penalization can achieve stable performance of the system. Second, adaptation of the controller gains to a change of parameter of the LCL filter. We compare adaptation rules based on the classical Riccati equation and on the reinforcement learning method of the temporal difference gradient descent. The latter method has lower computational cost and the potential to be more extensible to non-linear systems. We show that the standard approach with a constant learning rate is inefficient and better results can be achieved using adaptive learning rate. Performance of the method with adaptive learning rate is tested in simulation of current tracking after step change of the output inductance.

Prediction of junction temperatures in power semiconductor modules is essential to improve reliability of the device and prevent module failures due to thermal stress. Lumped parameter network is a popular approach for temperature modeling. Calibration of the thermal model is based on thermal measurements of the junction temperatures that are difficult to obtain. We aim to combine the knowledge of internal model structure and as little measurements as possible. Specifically, we use a state space thermal model with structure determined by the module layout, and propose to use the Expectation-Maximization algorithm from that can utilize data from different incomplete experiments. The identification procedure is introduced in detail in this paper and the applicability of the proposed approach is demonstrated on simulated and experimental data.

Non-negative matrix factorization (NMF) is generally an ill-posed problem which requires further regularization. Regularization of NMF using the assumption of sparsity is common as well as regularization using smoothness. In many applications it is natural to assume that both of these assumptions hold together. To avoidad hoccombination of these assumptions using weighting coefficient, we formulate the problem using a probabilistic model and estimate it in a Bayesian way. Specifically, we use the fact that the assumptions of sparsity and smoothness are different forms of prior covariance matrix modeling. We use a generalized model that includes both sparsity and smoothness as special cases and estimate all its parameters using the variational Bayes method. The resulting matrix factorization algorithm is compared with state-of-the-art algorithms on large clinical dataset of 196 image sequences from dynamic renal scintigraphy. The proposed algorithm outperforms other algorithms in statistical evaluation.

We propose a modification of the recently proposed optimal torque control for time varying parameters. The previous method allows to compute optimal trajectories for synchronous machine considering stator resistance and mutual inductances, however the strategies of transitions between these curves are changed at precomputed speeds. In this contribution, we introduce an algorithm that computes the transition strategies online at the cost of slightly higher computational burden. We also study numerical and DSP implementation issues and show that the computational cost is affordable in conventional DSP. The results are experimentally validated on IPMSM drive of rated power of 4. SkW.

Dual-motor single inverter drives reduce installation space, weight and overall costs. On the other hand, complexity of the control increases significantly, since the condition on each electric motor can vary in a wide range. Today, majority of control algorithms are based on field oriented control. The consequences of parallel connection of induction motors are addressed by prioritizing model of the motors with higher relevance to the control. The control problem becomes even more complicated when the drive is equipped with common sensors measuring phase currents at the inverter output only. Performance of the field oriented control strongly depends on the model with full state, which is not available trough measurement, due to the lack of current sensor for particular motors. We propose to simplify the model of dual motor drive and use it with finite control set model predictive control (FCS-MPC). The steady state solution of the control problem is used to improve the control performance. This is achieved as an additional term in the cost function. The algorithm is designed with respect to maximum simplicity and its properties are tested in simulations as well as in a real experiment on the laboratory prototype of a dual induction motor drive of rated power 2×4.5 kW.

Image restoration is typically defined as an ill-posed problem which has to be regularized to obtain an acceptable solution. In Bayesian interpretation, regularization is equivalent to prior model of the image. An added value of Bayesian point of view is the ability to form a hierarchical model and estimate the hyperparameters of the prior from the data. Many prior models are available, usually based on automatic relevance determination principle applied to the transformed image. However, the transformation (the most common is a differential operator) is assumed to be known. In this letter, we propose to relax this assumption and estimate the image transformation from the data. The resulting algorithm is analytically tractable using the variational Bayes method. Properties of the new prior are demonstrated on the problem of image superresolution.

Finite control set model predictive control (FCS-MPC) has become very popular form of predictive control among the technical society. Thank to its flexibility and well understood and simple structure, it is well suited for control of ac drives. It represents very interesting alternative for conventional control schemes based on cascade structures with PI controllers. The substantial advantage of MPC based approaches, is the ability of control so called MIMO (Multiple Input Multiple Output) systems in a simple way. This paper deals with the design of FCS-MPC for control of multiple induction machine fed from single inverter. Such drive configuration can offer significant cost and installation space reduction. However the complexity of a control design significantly increases. In multi - motor drives, the condition on each electric machine can vary. Typically, this is a case of traction drives, where the load on each electric machine can vary due to the inhomogeneous friction conditions on the railway track. We propose to use FCS-MPC algorithm as a simple alternative to conventional cascade control approaches. In order to verify the performance of proposed algorithm, we have tested the control technique in simulations and experiments on laboratory prototype of rated power 4.5 kW.

Blind deconvolution is a strongly ill-posed problem comprising of simultaneous blur and image estimation. Recent advances in prior modeling and/or inference methodology led to methods that started to perform reasonably well in real cases. However, as we show here, they tend to fail if the convolution model is violated even in a small part of the image. Methods based on variational Bayesian inference play a prominent role. In this paper, we use this inference in combination with the same prior for noise, image, and blur that belongs to the family of independent non-identical Gaussian distributions, known as the automatic relevance determination prior. We identify several important properties of this prior useful in blind deconvolution, namely, enforcing non-negativity of the blur kernel, favoring sharp images over blurred ones, and most importantly, handling non-Gaussian noise, which, as we demonstrate, is common in real scenarios. The presented method handles discrepancies in the convolution model, and thus extends applicability of blind deconvolution to real scenarios, such as photos blurred by camera motion and incorrect focus.

Temperature measurement on IGBT modules using voltage-based, direct and infrared methods is presented. The measured temperatures from all three methods are used in determining thermal impedances of IGBT and diodes for air cooled IGBT modules at different air flowrates. Thermal impedances for both active and passive switching elements are measured and the corresponding Foster network is established. The measured temperature and corresponding thermal impedances calculated from measured temperatures using the direct and voltage-based temperature measurement approach are validated using infrared temperature measurements and corresponding thermal impedances calculated from infrared measured temperatures. Finally, thermal impedances calculated from maximum measured IGBT and diode temperatures are compared to those obtained from measured average IGBT and diode temperatures.

Finite control set model predictive control (FCS-MPC) is very effective and simple control approach which has been found suitable for control of various types of power converters. It is a flexible tool because it evaluates an arbitrary cost function which directly influences converter behavior according to particular demand. The limitation of this approach is that efficient calculation is available only for very short prediction horizons. We will show that the input AC part of the CSR is a linear system which can be solved on a much longer horizon using approximation by a LQR solution. In effect, the LQR solution provides a reference for the converter input current which is then added to the cost function of the FCS-MPC algorithm. Comparison of the proposed approach with the conventional FCS-MPC cost is provided in terms of output current tracking, input current THD and average switching frequency. Final results will be introduced by simulation study in MATLAB/Plecs.

A simple adaptive linear current derating strategy for safe operation at low coolant flow rates and optimum operation at temperatures close to the allowed maximum operating junction temperature for electric drivetrain inverters is presented. An inverter junction temperature observer is implemented to calculate the maximum junction temperature which is used as an input to the derating strategy. Results of the proposed strategy are compared with those of the instantaneous derating strategy for different cooling conditions. It is observed that the proposed strategy is robust and eliminates unwanted drivetrain oscillations resulting from abrupt current reduction at low coolant flow rates using instantaneous derating.

Finite control set model predictive control (FCS-MPC) is very effective and simple control approach which has been suitable for control of various types of power converters. It is a very flexible tool because it evaluates an arbitrary cost function which directly influences converter behavior according to particular demand. For an accurate numerical calculation of the cost function, discretization of state space model is the most accurate approach among the single step calculation methods. Common definition of the current-source rectifier state space model covers only the input ac part of the rectifier whereas the dc part is calculated separately by Euler method. In this paper, definition of full state space model enabling an accurate discretization of all converter variables at once will be introduced, described and compared with a common solution. A behavior of both models will be compared in terms of their performance, robustness and simplicity. Final results will be introduced by simulation study in MATLAB/Plecs.

A common problem of imaging 3-D objects into image plane is superposition of the projected structures. In dynamic imaging, projection overlaps of organs and tissues complicate extraction of signals specific to individual structures with different dynamics. The problem manifests itself also in dynamic tomography as tissue mixtures are present in voxels. Separation of signals specific to dynamic structures belongs to the category of blind source separation. It is an underdetermined problem with many possible solutions. Existing separation methods select the solution that best matches their additional assumptions on the source model. We propose a novel blind source separation method based on probabilistic model of dynamic image sequences assuming each source dynamics as convolution of an input function and a source specific kernel (modeling organ impulse response or retention function). These assumptions are formalized as a Bayesian model with hierarchical prior and solved by the Variational Bayes method. The proposed prior distribution assigns higher probability to sparse source images and sparse convolution kernels. We show that the results of separation are relevant to selected tasks of dynamic renal scintigraphy. Accuracy of tissue separation with simulated and clinical data provided by the proposed method outperformed accuracy of previously developed methods measured by the mean square and mean absolute errors of estimation of simulated sources and the sources separated by an expert physician. MATLAB implementation of the algorithm is available for download.

Removing blurs from a single degraded image without any knowledge of the blur kernel is an ill-posed blind deconvolution problem. Proper estimators together with correct image priors play a fundamental role in accurate blind de-convolution. We demonstrate a superior performance of the variational Bayesian estimator and discuss suitability of automatic relevance determination distributions as image priors. Restoration of real photos blurred by out-of-focus and motion blur, and comparison with a state-of-the-art method is provided.

Start-stop system of an internal combustion system of a hybrid drive is designed to reduce energy consumption of the drive and reduce the produced emissions. Effectiveness of the start-stop system heavily depends on its ability to correctly predict when the combustion engine will be idle. This makes the problem suitable for model predictive control. In this contribution, we study the case of a hybrid city bus for which we have available recorded profiles of power consumption from repeated drives of the same route. We propose to use these historical data as predictions of the power consumption in the predictive control approach. We show in simulation with recorded data, that the resulting model predictive start-stop system outperforms the rule based approach and yield fuel savings of 5% and operational cost over 2%.

A new sensorless control algorithm for surface mounted permanent magnet synchronous machines with direct torque control (DTC) is proposed. The DTC usually operates with a short sampling period, therefore estimation algorithms need to be computationally cheap. Moreover, injection of high-frequency (hf) signals into a given axis can be a challenging task in comparison with a standard vector control with a PWM modulator. In this paper, a modification of the DTC algorithm is proposed to enable hf injections in a separate d-axis in a rotating coordinate system. Demodulation of this signal and sensorless rotor position estimation based on both phase-locked loop and modified extended Kaiman filter are described and evaluated. Verification was performed by both simulations and experiments on a laboratory prototype of a PMSM drive with rated power of 11 kW.

Finite control set model predictive control (FS-MPC) has been shown to be a very effective approach to control of PMSM drives. FS-MPC is a very flexible tool since it can evaluate an arbitrary loss function. However, design of the appropriate loss function for the problem can be a challenge especially when the design input is visible only on the long horizon. An example where this problem becomes apparent is the main propulsion drive of a traction vehicle fed from a dc catenary. Specifically, the catenary voltage is subject to short circuits, fast changes, harmonics and other disturbances which can vary in very wide range. Therefore, the drive is equipped with the trolley-wire input LC filter. The filter is almost undamped by design in order to achieve maximum efficiency and the control strategy needs to secure active damping of the filter to guarantee the drive stability. While it is possible to introduce active damping terms to the loss function, it is hard to predict its properties. In this paper, we consider decomposition of the problem to control of the LC filter and PMSM drive. We show that the resulting controllers can be elegantly combined using Bellman's principle of optimality. The resulting controller is easy to design and its performance is demonstrated in simulation and experiments on 10.7 kW drive.

The aim of this paper is an investigation of an extended Kaiman filer (EKF) based on a model in phase coordinates for sensorless control of an interior permanent magnet synchronous motor (IPMSM) with higher harmonics in the motor flux. The challenge for the EKF is to minimize its execution time especially for model in phase coordinates. This is achieved by using state space model of the drive with only two state variables, the rotor speed and the rotor position, which is known as the reduced-order model. This paper describes proposed EKF based estimator, verifies in simulations and experiments, that the EKF with this reduced model has low computational cost and high estimation accuracy. All experiments were carried out on a laboratory prototype of the IPMSM drive with rated power of 4 kW.

The aim of this paper is to study the use of the Extended Kalman filer (EKF) for sensorless control of a permanent magnet synchronous motor (PMSM) drive controlled by direct torque control (DTC). In contrast to the vector control, the DTC allows for better compensation of the dead-time effects, however, it also needs to run with much shorter sampling period. The challenge for the EKF is to minimize its execution time. This is achieved by using state space model of the drive with only two state variables, the rotor speed and the rotor position, which will be known as the reduced model. We show in simulations and experiments, that the EKF with this reduced model has the same performance as the four-dimensional full order model at much lower computational cost. This allows to use the EKF even in DTC with very short sampling time and take full advantage of the dead-time compensation. Due to better reconstruction of the voltage vector, the sensorless DTC is able to operate the drive at lower speed than the vector control with the same estimator. All experiments were carried out on a laboratory prototype of the drive with rated power of 11 kW and 44 poles. Control algorithms using DTC with sampling periods of 30 or 50 μs and vector control with sampling period of 125 μs were compared. The sensorless DTC with the reduced model EKF was found to be equal to DTC with the full model EKF in terms of performace and superior in terms of computational cost. Both sensorless DTC algorithms outperformed the vector control algorithm in accuracy of the estimation especially at low speed.

Marginalized particle filter is a stochastic filter combining Kalman filters with particle filters. It decomposes the model into linear and nonlinear part and applies the Kalman filter for the former and the particle filter for the latter. In effect, this allows to represent accurately the inherent non-Gaussianity and nonlinearity of the model. This allows estimation of the rotor position of the PMSM drive in the full speed range, including the standstill. The main disadvantage is its high computational cost. In this paper, we present an implementation of the marginalized particle filter in the field programmable logic array (FPGA). The parallel nature of the MPF algorithm allows to use pipelining which yields speedup in the order of magnitude in comparison to the DSP implementation. The sensorless control of the drive is implemented on a board with both DSP and FPGA, where the drive control runs on the DSP and the MPF estimator in the FPGA. Execution time of the estimator is thus negligible in the execution time of the sensorless control. Performance of the resulting sensorless control algorithm is evaluated on a developed drive prototype of rated power of 10.7kW.

Methods of sensorless control of PMSM drives are commonly divided into model-based and high-frequency injection based approaches. Each of these approaches uses a different algorithm for estimation of the rotor position and speed. Typically a Kalman filter is used for the model-based approach and phase-locked loop (PLL) for the hf injection based approach. In this paper, we show that the PLL is a steady state solution of the Kalman filter for a special state space model. Since this model has a commonly used state equations, we can easily combine the observation equations from the model-based approach with those from the hf injections. Several possibilities of combination are described and tested in the paper.We illustrate properties of these algorithms on experimental data in sensored mode. Sensorless control strategy based on the presented models is demonstrated on a laboratory prototype of surface mounted permanent magnet synchronous motor (PMSM) drive of rated power of 10.7kW.

This paper presents a novel estimation approach to sensorless control of wound rotor synchronous motor drives. The proposed rotor position estimation strategy is motivated by principle function of a resolver and utilizes design similarity between the controlled motor and the resolver based rotor position sensor. The rotor circuit is fed by controlled three-phase bridge rectifier which produces an ac component of frequency of 300Hz in the rotor excitation (flux) current. This ac component of rotor excitation current can be understood like an injection signal. Its response on the stator is evaluated by simplified Kalman filter which estimates the rotor position. This paper describes physical principle and functionality of the proposed sensorless rotor position estimation technique and verifies the theoretical foundations by simulation and experimental results made on developed drive prototype of rated power of 10kW.

Marginalized particle filter is a stochastic filter combining Kalman filters with particle filters. It decomposes the model into linear and nonlinear part and applies the Kalman filter for the former and the particle filter for the latter. Its application in sensorless control of permanent magnet synchronous motor (PMSM) drives is based on separate treatment of the state variables: the rotor position is represented by a set of samples (particles), and the rotor speed is estimated by the Kalman filters associated with each sample. In effect, this allows to represent accurately the inherent non-Gaussianity and nonlinearity of the model. We show that the resulting filter is capable to estimate the rotor position in the full speed range, including the standstill. Analysis of the filter performance is presented on open-loop off-line analysis of data recorded on a drive prototype. Execution time of optimized implementation of the algorithm for six particles in DSP is comparable to that of the Extended Kalman filter for full state-space model. Closed-loop performance of the filter (a sensorless drive control) is evaluated on developed drive prototype of rated power of 10.7kW.

In this study, we investigate online Bayesian estimation of the measurement noise density of a given state space model using particle filters and Dirichlet process mixtures. Dirichlet processes are widely used in statistics for nonparametric density estimation. In the proposed method, the unknown noise is modeled as a Gaussian mixture with unknown number of components. The joint estimation of the state and the noise density is done via particle filters. Furthermore, the number of components and the noise statistics are allowed to vary in time. An extension of the method for the estimation of time varying noise characteristics is also introduced.

In Bayesian filtering, the model may allow analytical marginalization over a subset, θ1,t, of the parameters. The marginalized (Rao-Blackwellized) particle filter (MPF) exploits this, by requiring stochastic sampling only in the remaining parameters, θ2,t, with the potential for major computational and convergence speed-ups. The marginalized filtering distribution in θ1,tis expressed as a mixture of n analytical components, each conditioned on one of the n particle trajectories in θ2,t; i.e. sufficient statistics must be stored and updated for each particle trajectory. In this paper, the variational Bayes (VB) approximation is used as a one-step approximation to extract necessary moments from the n particles in a principled manner, yielding a single-component marginalized filtering distribution. This formalizes and extends a recently reported certainty equivalence approach to accelerating MPFs. The comparative performance of the full and accelerated MPFs is explored via a scalar nonlinear filtering example.

The Variational Bayes (VB) approximation is applied in the context of Bayesian filtering, yielding a tractable on-line scheme for a wide range of non-stationary parametric models. This VB-filtering scheme is used to identify a Hidden Markov Model with an unknown non-stationary transition matrix. In a simulation study involving soft-bit data, reliable inference of the underlying binary sequence is achieved in tandem with estimation of the transition probabilities. The performance compares favourably with a proposed particle filtering approach, and at lower computational cost.

The autoregressive (AR) model is extended to cope with a wide class of possible transformations and degradations. The variational Bayes (VB) procedure is used to restore conjugacy. The resulting Bayesian recursive identification procedure has many of the desirable computational properties of the classical RLS procedure. During each time-step, an iterative variational EM (VEM) procedure is required to obtain the necessary moments. The procedure is used to reconstruct an outlier-corrupted AR process and a noisy speech segment. The VB scheme appears to offer improved performance over the related quasi-Bayes (QB) scheme in the case of time-variant component weights.