Smart cities: the role of Internet of Things and machine learning in realizing a data-centric smart environment

We elaborate on cases from different smart city application fields with a general discussion on the topics, implementations, and algorithms.

We emphasize the importance of data in driving innovation and improvement. The paper explains how a data-centric smart environment can be created by integrating IoT devices and machine learning algorithms to collect, analyze, and act upon data in real-time.

We evaluate an assessment of data management, containing the acquisition of data, processing of data, and distribution of data from an algorithmic viewpoint.

We present the significant challenges that must be addressed in terms of data privacy, security, and ethical considerations to realize the full potential of smart cities.

We consider the given framework from knowledge discovery in the machine and deep learning point of view.

Lastly, we demonstrate the future recommendations and scope of this study.

Smart Street Light Components: There are three main components for enacting a smart street light system in real life. These basic components are [38]:

Skin diseases with parasitic and intestinal problems in waste collecting workers.

Plague and flea-borne fever diseases exist in cats, dogs, and rats.

Dust from this waste pollutes the air causing breathing problems.

Flies, wasps, and mosquitoes breed on these waste piles, spreading diseases such as hepatitis, cholera, diarrhea, etc.

Mosquitoes spread malaria and yellow fever.

Smart buildings: The term can be defined as a building capable of responding according to the needs of residents, businesses, and society. Smart building is sustainable, conserve energy and water utilization, and are less polluted with emissions and waste products. It should have health, safety, and security for occupants and be capable of functioning to the requirement [77]. The term “smart buildings” is not very new and has been used for more than two decades to propose the idea of intelligent equipment, energy efficiency, and networking devices in the building. Back in the 1970s, the idea was to design buildings with energy-efficient capability, and in the 1980s, it was managed by a house PC. Nowadays, smart buildings use additional components for controlling and managing renewable energy sources, house appliances, and energy consumption using, most often, wireless communication technology. The appliances and gadgets within the smart buildings can communicate with each other and the outside environment or other buildings [78]. Generally, the smart building comprises of following components:

All these components create an automated building with some additional units in some cases, like power storage capacity and small-scale renewable power resources, [77]. These buildings are created and designed for the benefit of and constructed by people for individuals and the community. Efficient design management and integration guarantee that the systems and processes work effectively. Today, smart buildings can manage temperature control, airflow, and humidity in areas with harsh environmental conditions. The advances in air conditioning systems, electric lighting, and smart coating systems in buildings have eased the way of living and revolutionized civil engineering architecture.

Intelligent buildings management systems: These are the most recent and novel aspects of smart monitoring systems, structures, and infrastructures. During the past few years, many buildings have been categorized as smart/intelligent, but the real applications of this term are yet to be explored with its actual capability. Intelligent buildings ensure the enhancement of business opportunities and revenue as they consider social requirements and environmental conditions for residents’ well-being, leading to better work efficiency [78]. The latest intelligent building systems can connect the devices, the residents, and the systems within the buildings to provide full personal control and management. In the past few years, 5G technology has completely revolutionized the buildings, architecture, and construction business due to researchers’ and academia’s recent advances in 5G applications. Various enterprises, like IBM and Intel, have already conceived and advertised intelligent buildings to illustrate upcoming 5G features [79]. There are only a few studies on implementing 5G-based smart buildings, and many areas still require deep research. It is critical to realize the technical and scientific needs of the building industry as the recent 5G sector mainly concentrates on expertise and business benefits, and the end users are not considered the most substantial shareholders. 5G development has significantly improved the use of smart mobile devices with the growth of communication technologies resulting in new ideas of economic opportunities for business and industry [80]. By the year 2020 and onwards, a considerable percentage of people will be shifted to 5G facilities which approve the requirement of developing 5G-based smart cities. Smart buildings utilize the recent 5G communication and management technologies to provide individuals with a comfortable and healthy environment [79]. The recent earthquake-related issues and studies are very important to improve the literature review for global aspects concerning future smart cities [81]. The seismic issues are aspects of critical importance concerning the overall safety level in highly-populated cities, therefore the innovative smart cities of the future bestowed with artificial intelligence-based interconnected tools may provide innovative solutions for attempting to preserve human life against natural disasters.

Challenges in integrating 5G-based smartness in existing buildings: The 5G network has been established to help and encourage the population that is becoming more mobile and connected. Moreover, it will facilitate the flourishing of innovative business prototypes. However, the major challenge is to allow 5G functions in pre-existing building structural design, as the architecture cannot be altered once it has been built. Thus, renovation is a massive task in creating a smart city and incorporating an already established building into it. The volume of information to be transmitted and mobile broadband traffic will develop considerably in the upcoming years as mobile broadband is used in different situations and locations. With the help of 5G, new, scalable customer services will be created to encourage the growth of existing mobile broadband services [80]. This implies that 5G will sustain a wide range of uses, from those that require small amounts of data to those that need massive amounts of data quickly. Various network appliances, techniques, controllers, and sensors are required to make smart building residents have a relaxed and comfortable life. Therefore, in smart buildings, there is a network of remotely controlled communications [79].

Overview of prior efforts: This section provides detail of the most widely used datasets for smart cities and overviews state-of-the-art papers which used these datasets to get formal results. Barely any significant datasets portray their attributes.

CDnet dataset: The Cnet dataset exhibited at the IEEE Change Detection Symposium contained 31 records, including ship, automobile, truck, and scenes inside and outside of people walking in various challenging situations. These records are taken by several low-resolution IP cameras, medium-resolution cameras, and PTZ cameras that heat the camera. As a result, the spatial resolution of the Cnet record changes from \(320 \times 240\) to \(720 \times 576\). Besides, different illumination conditions and pressure parameters change the level of disturbance and pressure disturbance from one video to the next. Recording time ranged from 1000 to 8000 h. Recordings taken with a low-end IP camera were incompletely affected by long-term adverse effects. We believe that the way records fall within a specific range will help bias this data set to a particular group’s progress positioning technology. According to the history of each test, papers are divided into six categories. The purpose of recording selection is that one type of testing is one of the categories. For example, only records in the “shadow” category contain solid shadows, and documents in the “dynamic background” category contain the parasitic motion of the concrete on the ground. Such a conference is fundamental to identifying the quality and disadvantages of various change discovery strategies. The creator captured each video in addition to the video in the “gauge” category of the PETS 2006 dataset [91].

Internet of things: IoT is a network of physical devices, automobile devices, and additional components integrated with electronic components, software, sensors, actuators, and connectors that allow these objects to join and interchange data [92]. IoT applications are launching innovative city programs globally. It provides remote monitoring, management, and equipment control and creates new knowledge and actions from massive data stream real-time information. The central characteristic of a smart city is the degree of integration of information technology and the general application of information assets.

Sensor networks: The sensor network contains a small set of devices driven by batteries and wireless Infrastructure that monitors and records every environmental condition from the factory to the data center, hospital laboratory, and even the field. The sensor network is connected to the Internet, the enterprise’s WAN or LAN, or a dedicated industrial network and can send the collected data to the analysis and application program backend system. Extensive sensor networks (environmental checking, waste administration, health observing, smart grid, etc.) have become essential for gathering intelligent city data. Several sensors (motion sensors, cameras, sensors to manage environmental constraints) have been introduced [19]. The main challenge of the sensor network is to deliver appropriate semantics for varied data used in diverse smart city applications rather than placing a vast number of sensors in cities. Sensor networks are one of the critical technologies for data acquisition [93].

Mobile ad-hoc networks (MANETs): Mobile Ad-hoc Networks (MANETs) is a network that does not contain the Infrastructure for mobile devices, and they interconnect with one another without a physical medium [102]. The critical trial is to preserve the information needed for the correct route [103]. Evaluate MANET’s different routing protocols for VoIP, HTTP, and FTP applications in a smart city environment. Vehicle Ad-hoc Networks (VANETs) deliberate in the situation of intelligent transportation and intelligent transportation systems. Meanwhile, VANET is a particular type of MANET, and difficulties affecting MANET also affect VANET’s data collection process. VANET uses various apps to gather data and offer to clients to serve them [104]. These applications are divided into security and information entertainment applications. Safety applications used to advance road safety, i.e., frontal collision warnings and information and entertainment applications, are designed to provide comfort and entertainment for vehicle occupants, such as traffic information and weather information [105].

Device-to-device (D2D): Device-to-Device (D2D) communication in a cellular network is defined as a serial connection and communication between two mobile users without going through a base station (BS) or core network [106]. In traditional cellular networks, all communications must pass through the BS even though both parties of the conversation are in the D2D communication range [107]. This architecture is suitable for traditional mobile services at low data rates, such as voice calls and text messages, which are often not enough for users to communicate directly with each other. However, users of mobile devices in current cellular networks use high-speed data services (e.g., video sharing, and games). D2D refers to a wireless communication technology in which the device can interchange data openly without traversing base stations [106]. Non-infrastructure facilities can profit significantly from D2D communication. One possibility is to provide a secure connection if the Infrastructure is damaged. With the increasing number of mobile devices and IoT devices in the city, 4G plays an essential role in smart city applications. 4G is helping in the data acquisition in the smart city and the deployed intelligent towns. The work in [108] studied and used a hybrid of 5G/4G/3G, IoTs, and D2D for data gathering in smart cities.

Machine learning: Machine learning is generally divided into two types: supervised and unsupervised learning and reinforcement. In supervised learning, the computer gives an example input provided by the “teacher” and its expected result to learn the universal instructions for mapping input to output. The input signal may be merely partly accessible to precise comments in a particular case. In unsupervised learning, tags are not given in a learning algorithm, leaving only the label to discover the structure in their input. Unsupervised learning can itself be a goal (a hidden pattern of finding data) or a means of achieving it (learning function) [110, 111]. This pre-processing step can meaningfully advance the results of the following machine learning algorithms and is known as feature extraction [112]. There is no result (i.e., classification) of training data in intensive learning, which is true in many smart city applications, but the right choice of behavior is rewarded [109]. The utmost significant contests and limitations in machine learning are determining applicable data sets from the enormous amount of smart city data collected by data acquisition techniques (discuss above). The gathered data is now trained to fulfill the essential needs of the application. For example, measuring (the movement of people, object features, and the changes in the city with the development) requires covering these dynamic features. These data sets generate significant encounters for the administration, collection, and withdrawal of public data.

Machine learning algorithms: Machine learning is additionally divided into various algorithms based on varying problem perspectives. It includes clustering, instance base, and decision trees, which help solve complex problems [110]. These algorithms are supervised and unsupervised learning. For example, public data can be grouped using mutual properties k-means and k-medians algorithms. Instance-Based is the type of supervised machine learning which trains data. With this metric, the algorithms treasure the best and most accurate match for the novel data. k-Nearest Neighbor and Self-Organizing Map are examples of instance-based supervised machine learning algorithms. Indecision tree methods form a tree shape structure for decision-making based on fundamental values. With new data, the tree traverses the whole tree for the accurate result until a final decision is reached. Speed and accuracy are the advantages of these algorithms [112]. Examples of tree methods are classification and regression trees. There are many more machine learning algorithms; some of them are summarized in Table 6.

Machine learning and smart cities: Machine learning is the only field used widely in smart city deployment. Machine learning extracts meaningful information from the massive amount of smart city data [100]. Analysis of big data in smart cities needs techniques using machine learning. Machine learning helps clean the data, and various techniques help train the data. Oriented data work as a learned system. These oriented systems help develop better smart cities, capable of predicting the need for resources to manage the growing demand of the population the machine learning technique used in the health care system. A better healthcare system is a vital requirement of smart cities. Some of the deployed system in the different towns for real-world data processing is mentioned below in Table 7.

Deep learning: Deep learning is a particular machine learning type that expresses the world as a hierarchy of nested concepts, achieving high power and flexibility. Each idea is associated with a more straightforward concept and a more abstract representation. The promise of deep learning is not that the computer starts thinking like a human being. It is like trying to change an apple to an orange. Instead, suppose a sufficiently large dataset, a fast enough processor, and a reasonably sophisticated algorithm. In that case, the computer can complete the tasks that once remained in the human perception world. Deep Learning, such as Artificial Neural Networks (ANN), is successfully applied in applications such as information retrieval [127], image recognition, object tracking, and language processing [128, 129]. The whole objective of Deep Learning is to solve problems characterized by High dimensionality, which has no rules. In recent years, advances in hardware, software, and integrated systems enabled billions of smart devices to connect to the Internet. This ecosystem is jointly called the Internet of things. Many people are actively moving to cities, and primary resources become increasingly short. If everyone wants to support it, cities must manage Infrastructure such as water, electricity, and transportation very effectively. But how do you do this: the quality and format of the collected data will change, and its practical use will be hindered. For this, we will discuss several structural attacks.

Convolutional neural network: Convolution Neural Networks (CNN) are like standard Neural Networks. It comprises neurons with learnable weights and biases, which help classify the input into different layers. An artificial neural network may be best known among all bio-heuristic computing technologies, as most people have at least a question of whether they understand their work roughly. Unfortunately, most programmers are afraid to find a simple neural network to sit and work on. CNN combines artificial neural networks and discrete convolution for image processing, automatically extracting features. So, it is specially planned for the recognition of 2D data, such as pictures and videos [135]. The neural network (NN) entails a set of neurons linked by edges. The central function of the neural network is to gradually execute complex operations on the gathered data and use the output to resolve the issue. NN is used for altered applications such as character recognition, image compression video processing, and the like. CNNs are preferred for the reasons such as the simplicity of their training process, fast running in a test phase, and ease of application among Deep Learning Algorithms. CNN consists of 5 different stages.

Recurrent neural networks Humans never think from the beginning every second. Read this article and understand each word based on the understanding of the previous word. You will not lose everything; you will rethink from scratch. Your thoughts will last. The primary strategy of RNNs is to process the Sequential data. Like CNN, the RNN is also the traditional feed-forward neural network model. First, data is used as an input and flows through the hidden layer, and then the last layer is the output layer. It is dissimilar from other feed-forward networks in its capability to direct information over time steps. The neuron is not connected to this layer. The drawback of this architecture is not solve all the problems related to the neural network. The main reason is that when we use many data as input if this data has any relationship, it cannot solve the scenario [138]. For example, if we write a paragraph, all the words depend on one another, and no word can be independent of the article. As we look back on history, these networks are challenging to train. Still, research, network architecture, optimization, graphics processing unit (GPU), and parallelism have become more accessible for experts [139]. Circular neural networks are trained such that the output of each time step produces a sequence based on the input of the current information and all previous time steps. Frequently, recursive neural networks use an algorithm called back-propagation through time (BPTT) to calculate gradients.

Gated recurrent unit (GRU): The GRU has two doors: a resume door r and an update door z. Automatically, the reset gate regulates how the new entry is combined with the previous memory, and the update gate defines the quantity of the last memory. Setting all the resets to 1 and updating the doors to all zeros returns you to the standard RNN model. The basic idea of learning long-term dependency using the gate mechanism is the same as the basic idea of LSTM, but there are some crucial differences. Closed repeat unit (GRU) [140] is a controlled RNN with characteristics analogous to LSTM. Nonetheless, there are some variances. There are no separate memory units [141] in the GRU, rather than three gate layers. There are only two gate layers, restart and update. The closed-loop unit (GRU) introduced in 2014 is the activation mechanism of the recurrent neural network [141]. Its polyphony modeling and voice signal modeling performance is similar to long-term, short-term memory. A closed repeat unit is a somewhat simplified variant of LSTM. Combine the oblivion and entrance door with an “update door” with an additional “resume door.” The final model is simpler and more popular than the standard LSTM model [140].