Introduction
-
There is a need to understand how a cognitive computing-based system will be implemented on a large-scale smart city environment where data is growing, and scalability is a concern.
-
There is a lack of research addressing the flexibility of cognitive data to provide not only single cognitive computing AI-based solutions but multiple solutions using the same data.
-
Current cognitive systems are not flexible enough to provide solutions for multiple smart city-based applications. They lack scalability and flexibility which makes them unable to offer multiple real-time solutions.
-
This paper presents the related work done to identify the necessary foundations of an IoT, cognitive IoT and smart city architecture in the context of IoT.
-
The proposed CIoT-Net architecture is the first architecture which demonstrates different domains of a smart city network such as smart home, buildings, energy and transportation share similar sets of cognitive data used to build multiple cognitive computing-based applications.
-
The technologies that enable the proposed CIoT-Net architecture to function, which include cognitive enabled artificial intelligence and big data in cognitive computing.
-
The opportunities and challenges that arise in implementing the CIoT-Net architecture. We discuss the requirement of scalability in big data analytics, the concerns of data security and privacy and the role of artificial intelligence in addressing security concerns.
Related work
IoT architecture
-
Application layer: One of the major objectives in this layer of IoT is the creation of an intelligent environment such as smart buildings, smart home, smart health, smart, and smart industry and also this layer of application guarantees data integrity, authenticity, and confidentiality.
-
Network layer: Network layer consists of a set of interconnected devices sharing resources, information, and services, according to Badra, research [7] and Wu M, Ling F-Y research [8] proposed this layer which responsible to connect the IoT infrastructure, it collects data from the lower layer known as the perception layer also helps to transmit the communication to the upper layer IoT architecture. Cui A, Stolfo SJ research [9] and Mattern, Floerkemeier research [10] proposed that the communication medium in which is considered as wireless or wired, their different technologies used which can be Zigbee, wi-fi, wi-fi and LTE, Bluetooth low energy (BLE).
-
Perception layer: The perception layer considered one of the closest levels to the environment, known as the sensor layer in the IoT in which it is responsible for collecting packets and converting this information into a digital signal and identifying objects.
Cognitive IoT architecture
-
Sensor: In sensor components there are several types of sensors and basically, it is a device that has the function of detecting and efficiently responding to some stimulus and are used to collect the environment data in the IoT, acquire all critical information regarding the context of physical systems and allow the elaboration of the semantic goal model of the physical world. Networks of wireless sensors play an essential role in directing collected data to a central server. Sensors are used to allow data collection from the Internet environment of Things. Sensing share decision requires spectrum usage characterization and prediction to achieve channel selection [22].
-
Machine learning: Known as advanced algorithm optimization, has the ability to improve performance based on existing semantic models to provide the system with self-learning capabilities. Based on Bhattacharya et al.’s [23] showed that in machine learning Various Analytics applications in building.
-
Semantic modeling: The cognitive processes involved in the modeling, transformation, It uses the perceived information to construct semantic models, that later facilitates and automates for the elaboration of the physical and semantic reasoning.
Smart city architecture in the context of IoT
CIoT based smart city network architecture
Smart city platform
-
Smart buildings: Smart buildings contain many sensors which enable to collect data from different sources to help optimizing services in light, elevators and HVAC systems, etc. Buildings collect data such as emotions, environment, etc. which collectively provide multiple solutions such as optimizing of energy consumption using thermal analysis and control over lighting, air quality control, elevators, etc. based on user preferences for maximizing their comfort.
-
Smart home: IoT sensors provide data concerning the occupant’s movement patterns, the schedule for managing home activities, preferences set for indoor temperatures. Application of cognitive computing can provide humans with better control to enhance their safety and well-being.
-
Smart energy: A cognitive visualization tool is formed by combining and using data collected from both perceptual and logical part of the brain. Connecting the cognitive tool with imagery and weather analysis, energy companies provide advanced safety solutions to power plants such as in detecting operational risks in machinery while maximizing uptime.
-
Smart transportation: Using data collected from human brain activity, emotions and environment, a cognitive computing-based tool is used with voice analysis to provide real-time data on optimum speed and directions to follow for transportation to reach their destinations on time. Railway services and automobiles have managed to provide solutions for more reliable and on-time services to their users.
-
Smart agriculture: Sensors in smart agriculture can collect sensor data which includes perceptual and logical activity from the brain sensors and combine it with imagery analysis to efficiently detect where the harvested fruits and vegetables are located and pick them. Labor gaps are filled with robotic technology-based harvesting equipment which learns from data collected from these sensors.
-
Smart industry: Combining data collected from environment sensors and brain sensors to include human information such as workflow process and contextual knowledge, industry performance can be optimized, productive decisions in real-time can be made and raw resources in industries can be better managed.
IoT layer
Data layer
-
Brain activity: The human brain consists of two parts. The left portion and the right portion of the brain. The left side is responsible for the logical reasoning which helps humans in presenting a scientific approach to a problem. The right side of the brain is concerned with the perceptual view of the brain which aids in providing intuition, visualization and a creative approach to issues. Combined with these two parts of the brain, cognitive computing offers a human brain-like approach and analysis towards any challenge.
-
Emotions: Emotions are used to detect a range of feelings human beings express such as surprise, doubt, anger, distress, stress, etc. These range of emotions are learned by detecting facial impressions such as tensing of facial expressions, muscle stress, pupil dilation, etc. These range of emotions can be taught and used in domains such as healthcare and robotic technology.
-
Environment: Environment data is observed based on human interaction with the surrounding environment. Our decisions vary based on the different contextual environment with their associated situation such as position, time, goals, tasks and conditions. Learning rational thoughts and feelings is insufficient for the cognitive computing algorithm without the proper contextual knowledge of the environment.
-
Social media: Human beings have a natural understanding of detecting sentiments being expressed. Whether it is a happy, sad or even a sarcastic sentiment, we can quickly identify the difference. However, machines are not as good at this. Training human analysis of opinions posted on social media with cognitive-based artificial intelligence, we can derive more in-depth analytics of emotions expressed towards organizations.
-
Voice recognition: Extracting and training cognitive based AI based on voice has many applications especially in customer service. Human voice-tones, expression of urgency and stress levels can be recorded and used in training AI to help provide personalized support to customers.
-
Gesture recognition: Gesture recognition relies on learning the user’s hand movement coupled with brain activity. Combining this with AI, there are multiple uses such as gesture-controlled systems and aid to physically challenged people, patients with stroke, etc.
-
Spatial–temporal data: Spatial–temporal data is applied to train robots for navigation purposes and to develop awareness of time and space. This skill in humans is often used for puzzle solving and organizational skills by first visualizing the problem in front of them and then approaching it with the solution.
Cognitive computing layer
-
Data preprocessing: Raw data collected from sources may contain noisy data or incomplete data. Including this data will result in inefficient cognitive computing based artificial intelligence-based models. The missing data is removed if it is inconsequential to the outcome, or it can be replaced to avoid losing any vital information.
-
Data analysis: Data analysis refers to the selection of cognitive traits that are essential to the training of our model. Multiple traits are selected which serve various use cases instead of building a separate cognitive model for each different use case in the smart city. For example, brain activity involving both rational and perceptual understanding can be used for solutions for both smart energy and agriculture systems. User contextual environment data and emotions can be used to serve optimization purposes in both smart home and transportation.
-
Cognitive traits extraction: We extract the required data using sensors placed in different facets of the smart city allowing the collection of user data in real-time to train the machine learning model. Psychological data is gathered by collecting user emotions, voice, spatial–temporal data and brain activity data. Activity traits are collected from the user context-based environment and gestures. Social traits are extracted from brain activity, emotions and social media.
-
Machine learning: Machine learning allows training systems to provide analytics or prediction based on the training received. Machine learning has evolved to think intelligently with which it can develop its learning patterns on its own and provide better analysis. This evolution is called artificial intelligence. Cognitive computing has taken it one more step forward by merging the way how humans think and trained machines to think with a much more natural approach. Deep learning and Reinforcement learning are commonly used to develop algorithms from large data sources such as IoT devices which actively learn and improve themselves with no intervention needed. We discuss this in detail in Sect. 4.
Service layer
-
Law: Most law firms require their legal team to keep up with the latest legal rulings, law changes and all historical data on their clients and their legal entanglements. Cognitive computing leverages past and future legal documents with voice recognition and spatial–temporal information to assist legal teams in arriving on more effective solutions. It aids legal organizations with suggestions on how to proceed with a case and is seen as a supportive member of the legal team rather than as a replacement.
-
Firefighting: Firefighting is a dangerous job and requires the firefighters to be always alert at all times. Cognitive-based applications are now able to provide support to firefighters based on system trained with image analysis, historical missions and cognitive-based contextual environment, gesture recognition and brain activity to provide real-time analysis of the emergency. The system can detect and alert the level of risk involved, the scenario they will face once they enter the building and provide real-time analysis should the risk factor increase. Cognitive-based applications can help save lives of both firefighters and all else involved.
-
Police department: Cognitive computing applications can retain all past criminal records and the city database of every individual’s address and the registration of each building and home. Combining this data with brain data, environment, emotions and social media, police officers can locate where the perpetrator may be located. If the person is in a building, it will provide a more context-aware situational awareness of what they may face upon entering the building. Police officers can be assisted with historical and contextual information of a crime scene they are actively investigating.
-
Medical care: Healthcare industry leverages cognitive applications for enhanced patient care, identifying symptoms and providing an effective patient diagnosis. Combining patient medical history, illness symptoms with cognitive data such as environment, brain activity, emotions, gesture and voice recognition, the system will suggest diagnosis based on a study of symptoms which may have been overlooked by human error. Cognitive systems excel at pattern recognition which can be ignored by human error. Patient needs can be better understood, and medicine directly administered without any doctor/nurse delay. While the discussed use case involves hospital care, it can be used in smart homes to regulate medical care, especially for children and seniors.
-
Autonomous vehicles: Humans driving automobile require logical thinking and perceptual thinking as part of brain activity to detect and avoid any physical obstacle in their path. Humans can detect and respond immediately to such obstructions, and such with the growth of autonomous vehicles, cognitive computing plays a significant role in improving driver safety. Cognitive data such as brain activity, environment and spatial–temporal data is used to provide real-time analysis of the road and enhance human safety.
-
Retail industry: Human emotions and environment offer an active role in identifying individual shopping behavior. Retail outlets are now able to identify customer shopping patterns which help in better management of store inventory during their highest footfall. It suggests which items are more likely to be in demand based on a user preference based on the environment. Environment data includes weather conditions and geo data.
-
Media industry: Media advertisements have grown more customer preference aware base on enhanced customer insights enriched from cognitive computing applications. Sources such as social media, emotions, and user tone obtained from brain activity and voice provide businesses to present advertisements which provide higher customer engagement. With the addition of environment data, advertisers can provide more personalized content which exhibits higher user interest.
Enabling technologies in cognitive computing
Cognitive enabled artificial intelligence
-
Deep Learning: Systems can be trained to behave with human brain-like thinking using Deep Learning algorithms. A human brain is divided into two parts, the left and the right portion which serve two different purposes. The left portion of the brain is responsible for logical and rational thinking whereas the right portion of the brain is responsible for visual thinking and emotion recognition and expression [31, 32]. This individual application of both the logical and perceptive side of their minds is simulated via data analysis in cognitive computing systems. Logical and rational thinking such as measurements of an object can be defined. Perceptual reasoning will require the mapping of features to help correlate between the input and output. A cognitive system is capable of perceiving objects the way humans do where it identifies unknown objects which it has not been previously trained to recognize as done in traditional machine learning methods [33].
-
Perceptual thinking or feature mapping in case of cognitive computing is essential to train systems to think as humans do. Rational thought or analytic thinking is insufficient on its own. A system may be trained to recognize a human face by learning features such as the shape of ears, nose and mouth. However, with the addition of factors such as facial hair, distinctive expressions, spectacles and photographic angle, it becomes difficult for the system to recognize the face [34]. The mapping relation between the image of the person and the result is obtained using intuition as humans do. In deep learning, image classification obtained using deep learning mimics human perceptual thinking. A large amount of image features is required for training the deep learning model, and the mapping data is used in cognitive systems to provide personalized solutions.
-
Reinforcement learning: Reinforcement learning behaves in the same way often how a human learns. Based on an incentive or reward-based learning, reinforcement learning learns from the environment and improves its behavior. The system is rewarded if from among multiple paths the chosen action is an optimal solution towards reaching the desired objective [35]. The initial approach opted might be good for the system, but it may not be the optimal one. Constant learning and improving the approach results in refining the approach.
Role of big data in cognitive computing
New opportunities and challenges
Scalable big data analytics
-
Big data management: Is one of the biggest challenges, considered the management and storage current technologies are not suitable for Big Data; Algorithms are not efficient to work with data heterogeneity; the storage capacity grows more slowly than the amount of data; a big amount of data that cannot be analyzed will take the Internet of Things with current manual approaches, was presented evaluate some big data platforms and technically advanced analyzes to analyze data are available [44].
-
Big data aggregation: Big data aggregation is another great challenge that allows you to synchronize distributed Big Data forms and external data sources such as repositories, applications, sensor networks, etc. with the internal infrastructures of an organization.
-
Big data analytics: It brings potential transformer to multiple sectors and great opportunities; but otherwise, it also presents unprecedented challenges for tapping large volumes of data. Based on the Wang et al. research [45] suggested, for many reasons the big data analysis is still challenging, since the complex nature of big data is based on 5Vs, the performance to analyze and the need for scalability such heterogeneous data sets with real-time response.
Large scale, diversified instrumentation
-
Scalable data processing: Data-centered CIoT services are expected to provide noisy data cleansing, support for reduced data redundancy, and real-time data processing and delivery. In particular, an efficient service design model is required to scale with the large number of IoT services shared between multiple applications with different processing requirements.
-
CIoT Challenge: In environment, various IoT sensors and devices such as cellular, multi-sensors and wearable would have led to a substantial increase and expected to exceed largest connected devices. For different practical applications of large-scale CIoT, it is much more challenging to process massive sensing data and which may be mixed characteristics, which includes high dimensionality, heterogeneity, and others. Bhattacharya and D. Culler evaluated that, it would lead to great challenges related to the integration of devices into infrastructures as well as analyzes of backend systems [46]. Approaches are needed to address these challenges, allowing heterogeneous devices to be automatically integrated into analytical infrastructures.
Preserving security and privacy
-
Unsecured IoT devices: Many services such as medical care, autonomous vehicle, and firefighting require a continuous stream of data from sensors to provide critical services which directly affect human health. Sensors implemented in medical care, transportation, building, and household are computationally low powered IoT devices. Cyber-attacks are frequent on IoT devices which allow an attacker to seize control of them to steal user data and disable the machine. Disabling of the sensors will adversely affect the security and performance of the proposed CIoT-Net architecture.
-
Impact of data loss: An example of the impact of disabling the sensors on the security of the CIoT-Net architecture is in the case of firefighting. Data such as gesture recognition and brain activity in the context of the surrounding environment directly affects a firefighter’s ability to approach the incident with full awareness. Each firefighting incident requires new data accumulated from the sensors capturing the surrounding environment and user’s health and stress level to approach the disaster incident tactfully. Without the flow of data from sensors, a firefighter’s job and the victim’s health are both at an increased risk.
-
Data exploitation: Data theft from sensors will affect privacy concerns of the user and compromise the architecture. An example of privacy concern is with data comprising of human emotions linked with gesture and voice recognition patterns which are collectively used in medical care services. A malicious attacker can use this collected data to expose private medical information to the public or place ransom requests to the user for financial compensation in exchange for the data. The data can be further exploited by selling it off to other organizations who may use it for advertising fraudulent medical services to users.
-
Artificial Intelligence: We require a solution that caters to the growing need of big data security. AI can help resolve both privacy and security concerns in the CIoT-Net architecture. AI-based models are not only trained using data but perform better as the size of the data grows, i.e. the model’s accuracy improves as it learns from a larger volume of data. AI-based models operate on two phases. Firstly, they require a collection of data to train the model, which is its initial training, the foundation upon which the model is built. Secondly, they need a continuous stream of data to improve the model and provide better security solutions. AI-based solutions do not cease developing and improving their models, and they continuously require new information or data to ensure security against unknown or ground-zero attacks. AI and Big data go hand in hand whether it is to provide human-like personalized services to users or provide security solutions against cyber-attacks.
-
Blockchain technology: Considering that there is a vast stream of data collected from sensors which power cognitive computing, solutions such as blockchain technology are not ideal. Though a blockchain based network ensures sensor data security by storing data as transactions in blocks, data in blockchain based systems are immutable, i.e. it cannot be altered. However, as the size of data grows, there are concerns about scalability issues that emerge in blockchain based systems.
Context-awareness
-
Scalability: Data gathered from sensors is not utilized to its full potential. Existing AI-powered systems use structured data to build solutions. However, cognitive computing-based AI systems benefits from their ability to learn from both unstructured and structured data.
-
Flexibility: Various existing data are used to train a single cognitive computing application for a specific industry. The CIoT-Net architecture implements the same data used to power one application to be used for several other applications. A common dataset can be used to build and train multiple cognitive computing powered AI applications.
-
Data type: Collecting data to benefit multiple applications requires to identify the possible different types of data and their sources from where to be obtained. In the data layer of the CIoT-Net architecture, we list the various sources from where data can be collected such as brain activity, environment and emotions and the data type such as human voice-tones, facial and vocal expressions which are essential for building cognitive computing-based AI solutions.
-
Efficient algorithms: We identify that traditional machine learning methods are unable to improve themselves continuously from a flowing stream of data. Deep learning and deep reinforcement learning are the most efficient AI algorithms that can provide optimum solutions for cognitive computing applications.
-
Applications: For a stream of data to build different applications, it is essential to know which type of applications can benefit from which type of shared source of data. In the service layer of the CIoT-Net architecture, we present multiple applications which utilize common data, for example, healthcare and autonomous vehicles both require brain activity and environment data to train their applications. Human emotions and environment data both help training and building applications for retail and media industries.
-
Big data: Cognitive computing greatly benefits from Big data as it is a valuable and abundant source of data to train and continually improve existing AI models. Another significant benefit using cognitive computing is in its ability to use natural language processing to learn quickly from unstructured data collected from multiple sources, including social media.