مجله کنترل
سال هفدهم شماره 2 (تابستان 1402)

Deep learning-based models appropriately perform in modeling complex problems in computer vision and natural language processing (NLP). With this in mind, nonlinear system identification methods can benefit from tools developed in deep learning, leading to enriched frameworks to choose from. For this purpose, we are going to review some potential structures and methods of deep learning that can be used in nonlinear system identification. Although we comprehensively review the applicable tools of deep learning to system identification, this paper mainly focuses on using latent variable models (LVM) for identifying nonlinear state space models. LVMs are powerful tools for extending generative models primarily developed for only generating static data, yet by a combination of recurrent neural network (RNN) and variational auto-encoders (VAE), they can also generate sequential data. A structured version of introduced models compatible with control systems will also be given. The study shows that the deep learning models have a comparative performance to traditional and classic ones.

Due to recent advancements in artificial intelligence, the utilization of image and video data along with image processing technologies has been significantly used in automating surgical procedures and their training. Today, one or more cameras or imaging devices are commonly used in most operating rooms to capture crucial medical information for further analysis. These visual data can be leveraged to design and develop automated image-guided systems aimed at assisting specialized surgeons during procedures and facilitating their training. Furthermore, these systems can serve as the cognitive knowledge for assisting surgical robots. An image-guided surgical system comprises of various components, among which the identification, segmentation, and tracking of surgical instruments and critical anatomical regions, phase recognition, motion analysis, and gesture recognition, and more importantly as surgical skill assessment, are the most crucial components. Automating these components by the use of image processing and machine vision will certainly contribute to the system's deep and independent understanding of the surgical scene. This article introduces the important datasets relevant to image processing and machine vision in surgical applications. Subsequently, several research endeavors in the realm of image processing, machine vision, on the above-mentioned applications are presented, and some prospect research areas are introduced.

In recent years, with the advancement of industry and technology, we see that systems are becoming more and more complex. This complexity in the industry has required a parallel progress in control systems (controllers), which has multiplied the need for advanced and intelligen t controllers. One of the most important criteria in the design of any control system is the accurate knowledge of the system or, more precisely, the modeling of the system. Considering these two challenges, this paper examines intelligent systems and specifically neural networks from the perspective of classical theory and their application in modeling and control of complex systems. At first, the basic element of neural networks, i.e. neuron, is introduced and its types of models (collective and radial) are also presented. Then the types of neural networks such as multilayer perceptron neural networks (MLP-NN), radial basis neural networks (RBF-NN) and recurrent neural networks (RNN) are described and the appropriate number of layers and also the appropriate number of neurons in the hidden layers in neural networks for different applications and especially for the approximation of nonlinear functions (modeling) are discussed. Then, it is tried to build a bridge between the concepts of neural networks (intelligent world) and the concepts of classical world and analyze neural networks from the perspective of classical theory. It is shown that from the point of view of classical theory, a neural network can be considered as a model structure and its weights and biases as unknown parameters of this structure. In neural networks, learning algorithms are used to determine the unknown parameters of the network (weights and biases). Therefore, learning algorithms of neural networks are studied from the point of view of classical theory and specifically in relation to numerical optimization and a bridge will be built between learning algorithms and numerical optimization methods. Also, the most important neural network learning algorithms are introduced along with their advantages and disadvantages. In the end, two important applications of neural networks, modeling and control, are discussed and for each of these applications, several illustrative examples are presented and simulated to show the effectiveness of neural networks.

Over the past two decades, an increasing number of engineers and researchers in the field of control engineering have shifted their focus towards data-driven methods for system analysis and design. The defining characteristic of these data-driven approaches is their departure from conventional models and their associated assumptions. Instead, these methods harness the wealth of readily available, cost-effective, and reliable data derived from real, complex, and complex adaptive systems. Their primary objective is to facilitate system analysis and control solely through measured data, without relying on explicit or implicit model utilization. In this article, we commence by presenting an overview of the design principles embedded in model-based control systems, with a specific emphasis on adaptive and robust control system design methods. Subsequently, we delve into the fundamental principles of data-driven control system design. To comprehensively examine these data-driven methods, we categorize them into two groups: those grounded in machine learning and soft computing, and those based on traditional control systems analysis and design methods, often referred to as classical methods. Within this article, we initiate with a concise review of machine learning and soft computing-based methods before delving into a more comprehensive exploration of classical methods in this field.

In this paper, recently proposed methods of pattern recognition and similarity analysis in alarm management systems are reviewed, and the impact of these methods on improving the efficiency of these systems, with emphasis on alarm floods, is discussed. Various types of alarms and their patterns are analyzed, highlighting the potential for more precise predictions, timely response strategies, and improvements in system maintenance. The technological advancements that have strengthened these methods are explained. More specifically, the role of recent advancements in artificial intelligence and machine learning in enhancing pattern recognition and similarity analysis, especially during alarm floods, is studied. The combination of these modern technologies with traditional alarm management systems paves the way for significant improvements in industrial safety. Examples from various industries are provided to demonstrate the practical application and impact of these methods.

Networked Control Systems (NCS) is a new field in control systems that has emerged by the use of communication networks in control systems. NCS refers to a control system in which sensors, actuators, and controllers are connected via a communication network and enable remote monitoring and control. The use of communication networks in control systems have many advantages, like reducing the cost of wiring, increasing flexibility, and improving scalability. However, the integration of communication networks in control systems will also carries new challenges such as data transmission delay, data packet loss, and network congestion, that require new methods for modeling and designing control systems. Research in the networked control systems is focused on the development of new modeling, estimating, identifying, and designing methods, which the effects of communication networks on the system’s behavior will be taken into account. In general, the field of NCS is an important research area because it allows the control systems to be applied in a wide range of applications and develop it in more efficient and flexible way. NCS can be found in various applications such as industrial automation, smart energy grids, transportation systems, and house automation. This article describes the emergence of control systems from the field of continuous time to networked control and compares it with traditional control systems. Furthermore, the most important challenges in NCS and related approaches will be discussed. Important applications of networked control systems by reviewing some featured case studies in Iran and other countries are also discussed. In the end, possible future directions in the development of networked control systems will be presented.

Financial time series are fundamentally complex, dynamic, noisy, non-linear, non-parametric and chaotic, so forecasting financial markets is one of the most challenging fields in engineering and economics. With the increasing progress of artificial intelligence and the emergence of deep learning methods, the problem of stock market forecasting has been faced with significant developments, especially in the field of prediction models and big data. Four important steps to create a systematic intelligent forecasting model include model inputs, selection of forecasting algorithms and design of the general structure of the forecasting model, using appropriate loss functions to train the algorithm and finally suitable evaluation of the results, according to the desired criteria. In this paper, a comprehensive review of recent approaches to stock market forecasting is provided, focusing on the above steps. The most important achievements of this paper are: 1- A comprehensive review of the problem, including: reviewing the types of model inputs, different prediction structures, training the model and types of loss functions used, and the evaluation metric of the results, in a fully classified and structured way to easily provide the road map and existing challenges for the enthusiasts and also an important research field of each section for the researchers. 2- Complete analysis of each part, specifying the application of each method and discussing their advantages and disadvantages based on the latest developments and providing perspectives on the research boundaries. 3- Determining the ongoing research path, future approaches and open issues for researchers interested in this field.

One of the most important and basic components in every industrial plant such as refinery, power plant, cement factory, automotive industry, etc. is its control system, Which is responsible for ensuring safety, performance, optimal operation with high efficiency, meeting environmental requirements and responding appropriately to user requests. With the advancement of technology, the design and implementation of control systems and related equipment has been completely changed and upgraded from pneumatic and semi-automatic control systems to electronic and intelligent systems so that the evolution of control equipment technology is several times the main mechanical equipment of industrial plants. In this article, some new technologies with related practical points for the implementation of control systems in industrial plants such as: Smart Transmitter, Soft Sensor, Autotuning PID Controller, Intelligent Actuators, Hardware in the loop Simulator, Digital twin, Application of artificial intelligence and Machine Learning are briefly reviewed. Today, when working with industrial control systems, familiarity with these topics is very necessary

Utilizing cloud computing and distributed computing has led to various advantages like enhanced performance, enabling outsourcing of complex computations, and higher scalability in a vast range of network control systems such as smart energy networks, smart buildings, and intelligent transportation. However, confidentiality breaches and manipulation of sensitive and private information, as well as lack of public trust in cloud-based decentralized and distributed approaches where agents are reluctant to share their information due to privacy concerns are among the emerging challenges in control of cyber-physical systems. This paper reviews the privacy-preserving encrypted control methods that address some of these challenges. In encrypted control methods, all the required computations are performed directly on the encrypted data, and thus, no intermediate decryption of private data is needed. In this way, the access of adversaries to the crucial information of the control system will be very restricted. Since implementing a complex cyberattack usually requires an in-depth knowledge of the system’s data, protecting the privacy of the system’s signals in the entire control loop considerably reduces the possibility of more complex cyberattacks. Therefore, in this paper, homomorphic encryption and secure multi-party computation methods are introduced as the basis for preserving the privacy of data and designing secure control approaches. Then, various control and optimization methods are reviewed. Shortcomings and challenges of existing results are discussed and the roadmap to further research in this emerging topic in control engineering is drawn.