In a few years, the world will be populated by billions of connected devices that will be placed in our homes, cities, vehicles, and industries. Devices with limited resources will interact with the surrounding environment and users. Many of these devices will be based on machine learning models to decode meaning and behavior behind sensors’ data, to implement accurate predictions and make decisions. The bottleneck will be the high level of connected things that could congest the network. Hence, the need to incorporate intelligence on end devices using machine learning algorithms. Deploying machine learning on such edge devices improves the network congestion by allowing computations to be performed close to the data sources. The aim of this work is to provide a review of the main techniques that guarantee the execution of machine learning models on hardware with low performances in the Internet of Things paradigm, paving the way to the Internet of Conscious Things. In this work, a detailed review on models, architecture, and requirements on solutions that implement edge machine learning on Internet of Things devices is presented, with the main goal to define the state of the art and envisioning development requirements. Furthermore, an example of edge machine learning implementation on a microcontroller will be provided, commonly regarded as the machine learning “Hello World”.
16 June 2022 in IEEE Journal of Radio Frequency Identification
Indoor positioning of objects and people is becoming of great importance in the Internet of Things (IoT), in-home automation, and navigation in malls, airports, or very large buildings. Positioning is determined by multiple distance measurements between reference points and sensors. Distance measurement uses the time of flight of an ultrasonic signal traveling from an emitter to receiving sensors. This requires close synchronization between the emitter and the sensors and a sharp time resolution of the time of arrival (TOA) of the ultrasonic signal. Usually, TOA is detected using cross-correlation processing requiring significant computational resources at the sensors level. In this work, the synchronization is done using the RFID standard protocol features. The TOA detection is performed firstly by training off-line a Machine Learning model using as input the peaks indexes of the ultrasonic signal received and the output of a cross-correlation based positioning system, as ground-truth. In a second phase, the positioning is evaluated and tested on-board using the previously trained model on a microcontroller. The system architecture is presented and experimental results on the positioning accuracy are shown accordingly. Results show a mean positioning error below 25 cm in 95% of the positionings in a typical room.
The growing interest for indoor position-based applications and services, as well as ubiquitous computing and location aware information, have led to increasing efforts toward the development of positioning techniques. Many applications require accurate positioning or tracking of people and assets inside buildings, and some market sectors are waiting for such technologies for starting a fast growth. Ultrasonic systems have already been shown to possess the desired positioning accuracy and refresh rate. However, they still require accurate synchronization between ultrasound emitters and receivers to work properly. Usually, synchronization is carried out through radio frequency (RF) signals, adding system complexity and raising the cost. In this work, this limit is overcome by introducing a novel self-synchronizing indoor positioning technique. Ultrasonic signals travel from emitters placed at fixed reference positions to any number of mobile devices (MD). The travelled distance is computed from the time of flight (TOF), which requires in turn synchronism between emitter and receiver. It is shown that this synchronism can be indirectly estimated from the time difference of arrival (TDOA) of the ultrasonic signals. The obtained positioning information is private, in the sense that the positioning infrastructure is not aware of the number or identity of the MDs that use it. Computer simulations and experimental results obtained in a typical office room are provided.
The design and testing phase of photovoltaic (PV) power systems requires time-consuming and expensive field-testing activities for the proper operational evaluation of maximum power point trackers (MPPT), battery chargers, DC/AC inverters. Instead, the use of a PV source emulator that accurately reproduces the electrical characteristic of a PV panel or array is highly desirable for in-lab testing and rapid prototyping. In this paper, we present the development of a low-cost microcontroller-based PV source emulator, which allows testing the static and dynamic performance of PV systems considering different PV module types and variable operating and environmental conditions. The novelty of the simple design adopted resides in using a low-cost current generator and a single MOSFET converter to reproduce, from a fixed current source, the exact amount of current predicted by the PV model for the actual load conditions. The I–V characteristic is calculated in real-time using a single diode exponential model under variable and user-selectable operating conditions. The proposed method has the advantage of reducing noise from high-frequency switching, reducing or eliminating ripple and the demand for output filters, and it does not require expensive DC Power source, providing high accuracy results. The fast response of the system allows the testing of very fast MPPTs algorithms, thus overcoming the main limitations of state-of-art PV source emulators that are unable to respond to the quick variation of the load. Experimental results carried on a hardware prototype of the proposed PV source emulator are reported to validate the concept. As a whole result, an average error of ±1% in the reproduction of PV module I–V characteristics have been obtained and reported.
This paper presents the design and hardware implementation of open-source hardware dedicated to smart converter systems development. Smart converters are simple or interleaved converters. They are equipped with controllers that are able to online impedance match for the maximum power transfer. These conversion systems are particularly feasible for photovoltaic and all renewable energies systems working in continuous changing operating conditions. Smart converters represent promising solutions in recent energetic scenarios, in fact their application is deepening and widening. In this context, the availability of a hardware platform could represent a useful tool. The platform was conceived and released as an open hardware instrument for academy and industry to benefit from the improvements brought by the researchers’ community. The usage of a novel, open-source platform would allow many developers to design smart converters, focusing on algorithms instead of electronics, which could result in a better overall development ecosystem and rapid growth in the number of smart converter applications. The platform itself is proposed as a benchmark in the development and testing of different maximum power point tracking algorithms. The designed system is capable of accurate code implementations, allowing the testing of different current and voltage-controlled algorithms for different renewable energies systems. The circuit features a bi-directional radio frequency communication channel that enables real-time reading of measurements and parameters, and remote modification of both algorithm types and settings. The proposed system was developed and successfully tested in laboratory with a solar module simulator and with real photovoltaic generators. Experimental results indicate state-of-art performances as a converter, while enhanced smart features pave the way to system-level management, real-time diagnostics, and on-the-flight parameters change. Furthermore, the deployment feasibility allows different combinations and arrangements of several energy sources, converters (both single and multi-converters), and modulation strategies. To our knowledge, this project remains the only open-source hardware smart converter platform used for educational, research, and industrial purposes so far.
Increasing efforts toward the development of positioning techniques testify the growing interest for indoor position-based applications and services. Many applications require accurate indoor positioning or tracking of people and assets, and some market sectors are starting a rapid growth of products based on these technologies. Ultrasonic systems have already been demonstrating their effectiveness and to possess the desired positioning accuracy and refresh rates. In this work, it is shown that a typical signal used in ultrasonic positioning systems to estimate the range between the target and reference points—namely, the linear chirp—due to the effects of acoustic diffraction, in some cases, undergoes a shape aberration, depending on the shape and size of the transducer and on the angle under which the transducer is seen by the receiver. In the presence of such signal shape aberrations, even one of the most robust ranging techniques, which is based on cross-correlation, provides results affected by a much greater error than expected. Numerical simulations are carried out for a typical ultrasonic chirp, ultrasonic emitter, and range technique based on cross-correlation and for a typical office room, obtained using the academic acoustic simulation software Field II. Spatial distributions of the ranging error are provided, clearly showing the favorable low error regions. The work demonstrates that particular attention must be paid to the design of the acoustic section of the ultrasonic positioning systems, considering both the shape and size of the ultrasonic emitters and the shape of the acoustic signal used.