With the increasing use of advanced control methods such as model predictive control for nonlinear systems, the demand for real-time computational power continues to increase. Meeting this demand can be especially challenging when multiple nonlinear systems need to be controlled with the potentially limited embedded hardware of the individual systems. Utilizing a central computing unit, such as a cloud, can provide the required additional computing power at the expense of increased economic and energy costs.In this work, we consider a scenario where embedded controllers based on neural networks that imitate a model predictive controller need to be supported by a central computing unit in case of errors or model changes. To minimize the necessary amount of central computing used to guarantee a safe simultaneous operation of multiple systems, we propose a reinforcement learning-based scheduling of the different necessary computing tasks. Furthermore, we show that the proposed control structure is safe for the closed-loop control of nonlinear systems. A case study illustrates the benefits of the proposed control structure.

Efficient approximations of predictive controllers using neural networks can enable the deployment of highperformance controllers virtually everywhere. Such approximations can run on simple embedded hardware but have significant drawbacks. The approximate controllers need a computationally expensive training in advance, and they do not inherit the feasibility and stability guarantees of the original predictive controllers. This paper considers possible future control architectures in an industrial setting where a cloud, or a centralized computing unit, can be used to simultaneously mitigate these drawbacks by supervising the performance of the embedded controllers based on the computation of safe sets and predicted constraint violations, which can trigger a direct control by the cloud as well as a retraining of the embedded approximate controllers. We demonstrate the potential of the proposed approach with a simulation study.

Based on the principle of onion routing, the Tor network achieves anonymity for its users by relaying user data over a series of intermediate relays. This approach makes congestion control in the network a challenging task. As of this writing, this results in higher latencies due to considerable backlog as well as unfair data rate allocation. In this article, we present a concept study of PredicTor, a novel approach to congestion control that tackles clogged overlay networks. Unlike traditional approaches, it is built upon the idea of distributed model predictive control, a recent advancement from the area of control theory. PredicTor is tailored to minimizing latency in the network and achieving max-min fairness. We contribute a thorough evaluation of its behavior in both toy scenarios to assess the optimizer and complex networks to assess its potential. For this, we conduct large-scale simulation studies and compare PredicTor to existing congestion control mechanisms in Tor. We show that PredicTor is highly effective in reducing latency and realizing fair rate allocations. In addition, we strive to bring the ideas of modern control theory to the networking community, enabling the development of improved, future congestion control. Thus, we demonstrate benefits and issues alike with this novel research direction.

AI tasks based on deep neural networks have been widely applied in industrial applications, such as process control, quality inspection or predictive maintenance. Deep neural network classifiers are particularly successful, as they provide powerful and reliable algorithms for many applications such as object recognition and fault diagnosis. However, most deep classifier applications are not able to recognize class samples that are beyond the scope of their training data. Samples of unknown classes (denoted as open set data) lead to significant drops in performance, as the output of deep classifiers is limited to the known classes of the training data (denoted as closed set data). This paper presents a method to recognize open set samples without changing the neural network architecture, the training process, nor the trained models. In our method, we firstly train a neural network for normal closed set fault diagnosis. Then we compare the feature maps of testing samples and known class samples during inference using local outlier factor to recognize open set samples. We evaluate our method with two public datasets and show that our method can increase the overall accuracy by 40% when classifying open set data. Besides, we also compared our method to the state-of-the-art open set recognition approach for fault diagnosis applications and the results show that our method leads to better F1-scores.

Resonant power converters offer improved levels of efficiency and power density. In order to implement such systems, advanced control techniques are required to take the most of the power converter. In this context, model predictive control arises as a powerful tool that is able to consider nonlinearities and constraints, but it requires the solution of complex optimization problems or strong simplifying assumptions that hinder its application in real situations. Motivated by recent theoretical advances in the field of deep learning, this article proposes to learn, offline, the optimal control policy defined by a complex model predictive formulation using deep neural networks so that the online use of the learned controller requires only the evaluation of a neural network. The obtained learned controller can be executed very rapidly on embedded hardware. We show the potential of the presented approach on a hardware-in-the-loop setup of an field-programmable gate array-controlled resonant power converter.

We show that artificial neural networks with rectifier units as activation functions can exactly represent the piecewise affine function that results from the formulation of model predictive control (MPC) of linear time-invariant systems. The choice of deep neural networks is particularly interesting as they can represent exponentially many more affine regions compared to networks with only one hidden layer. We provide theoretical bounds on the minimum number of hidden layers and neurons per layer that a neural network should have to exactly represent a given MPC law. The proposed approach has a strong potential as an approximation method of predictive control laws, leading to a better approximation quality and significantly smaller memory requirements than previous approaches, as we illustrate via simulation examples. We also suggest different alternatives to correct or quantify the approximation error. Since the online evaluation of neural networks is extremely simple, the approximated controllers can be deployed on low-power embedded devices with small storage capacity, enabling the implementation of advanced decision-making strategies for complex cyber-physical systems with limited computing capabilities.

Resonant power converters have enabled the development of highly efficient systems in a wide range of applications. In particular, induction heating has taken advantage of this technology to provide high-performance solutions to millions of users. Modern induction heating systems are based on multi-coil systems, requiring advanced power conversion techniques. While in the past many different power converter topologies have studied, little or no attention has been paid to the control of such topologies considering high performance and complex restrictions. This paper proposes the use of model predictive control to address these challenges and provides an example of application for a multi-coil induction heating system.

Model predictive control (MPC) is an optimization-based strategy for high-performance control that is attracting increasing interest. While MPC requires the online solution of an optimization problem, its ability to handle multivariable systems and constraints makes it a very powerful control strategy specially for MPC of embedded systems, which have an ever increasing amount of sensing and computation capabilities. We argue that the implementation of MPC on field programmable gate arrays (FPGAs) using automatic tools is nowadays possible, achieving cost-effective successful applications on fast or resource-constrained systems. The main burden for the implementation of MPC on FPGAs is the challenging design of the necessary algorithms. We outline an approach to achieve a software-supported optimized implementation of MPC on FPGAs using high-level synthesis tools and automatic code generation. The proposed strategy exploits the arithmetic operations necessaries to solve optimization problems to tailor an FPGA design, which allows a tradeoff between energy, memory requirements, cost, and achievable speed. We show the capabilities and the simplicity of use of the proposed methodology on two different examples and illustrate its advantages over a microcontroller implementation.