In this paper, we tackle the challenging task of reconstructing Received Signal Strength (RSS) maps by harnessing location-dependent radio measurements and augmenting them with supplementary data related to the local environment. This side information includes city plans, terrain elevations, and the locations of gateways. The quantity of available supplementary data varies, necessitating the utilization of Neural Architecture Search (NAS) to tailor the neural network architecture to the specific characteristics of each setting. Our approach takes advantage of NAS’s adaptability, allowing it to automatically explore and pinpoint the optimal neural network architecture for each unique scenario. This adaptability ensures that the model is finely tuned to extract the most relevant features from the input data, thereby maximizing its ability to accurately reconstruct RSS maps. We demonstrate the effectiveness of our approach using three distinct datasets, each corresponding to a major city. Notably, we observe significant enhancements in areas near the gateways, where fluctuations in the mean received signal power are typically more pronounced. This underscores the importance of NAS-driven architectures in capturing subtle spatial variations. We also illustrate how NAS efficiently identifies the architecture of a Neural Network using both labeled and unlabeled data for Radio Map reconstruction. Our findings emphasize the potential of NAS as a potent tool for improving the precision and applicability of RSS map reconstruction techniques in urban environments.

Short-term residential load forecasting plays a crucial role in smart grids, ensuring an optimal match between energy demands and generation. With the inherent volatility of residential load patterns, deep learning has gained attention due to its ability to capture complex nonlinear relationships within hidden layers. However, most existing studies have relied on default loss functions such as mean squared error (MSE) or mean absolute error (MAE) for neural networks. These loss functions, while effective in overall prediction accuracy, lack specialized focus on accurately predicting load peaks. This article presents a comparative analysis of soft-DTW loss function, a smoothed formulation of Dynamic Time Wrapping (DTW), compared to other commonly used loss functions, in order to assess its effectiveness in improving peak prediction accuracy. To evaluate peak performance, we introduce a novel evaluation methodology using confusion matrix and propose new errors for peak position and peak load, tailored specifically for assessing peak performance in short-term load forecasting. Our results demonstrate the superiority of soft-DTW in capturing and predicting load peaks, surpassing other commonly used loss functions. Furthermore, the combination of soft-DTW with other loss functions, such as soft-DTW + MSE, soft-DTW + MAE, and soft-DTW + TDI (Time Distortion Index), also enhances peak prediction. However, the differences between these combined soft-DTW loss functions are not substantial. These findings highlight the significance of utilizing specialized loss functions, like soft-DTW, to improve peak prediction accuracy in short-term load forecasting.

Nowadays, set of cooperative drones are commonly used as aerial sensors, in order to monitor areas and track objects of interest (think, e.g., of border and coastal security and surveillance, crime control, disaster management, emergency first responder, forest and wildlife, traffic monitoring). The drones generate a quite large and continuous in time multimodal (audio, video and telemetry) data stream towards a ground control station with enough computing power and resources to store and process it. Hence, due to the distributed nature of this setting, further complicated by the movement and varying distance among drones, and to possible interferences and obstacles compromising communications, a common clock between the nodes is of utmost importance to make feasible a correct reconstruction of the multimodal data stream from the single datagrams, which may be received out of order or with different delays. A framework architecture, using sub-GHz broadcasting communications, is proposed to ensure time synchronization for a set of drones, allowing one to recover even in difficult situations where the usual time sources, e.g. GPS, NTP etc., are not available for all the devices. Such architecture is then implemented and tested using LoRa radios and Raspberry Pi computers. However, other sub-GHz technologies can be used in the place of LoRa, and other kinds of single-board computers can substitute the Raspberry Pis, making the proposed solution easily customizable, according to specific needs. Moreover, the proposal is low cost, since it does not require expensive hardware like, e.g., onboard Rubidium based atomic clocks. Our experiments indicate a worst case skew of about 16 ms between drones clocks, using cheap components commonly available in the market. This is sufficient to deal with audio/video footage at 30 fps. Hence, it can be viewed as a useful and easy to implement architecture helping to maintain a decent synchronization even when traditional solutions are not available.

Multi-robot path planning has evolved from research to real applications in warehouses and other domains; the knowledge on this topic is reflected in the large amount of related research published in recent years on international journals. The main focus of existing research relates to the generation of efficient routes, relying the collision detection to the local sensory system and creating a solution based on local search methods. This approach implies the robots having a good sensory system and also the computation capabilities to take decisions on the fly. In some controlled environments, such as virtual labs or industrial plants, these restrictions overtake the actual needs as simpler robots are sufficient. Therefore, the multi-robot path planning must solve the collisions beforehand. This study focuses on the generation of efficient collision-free multi-robot path planning solutions for such controlled environments, extending our previous research. The proposal combines the optimization capabilities of the A* algorithm with the search capabilities of co-evolutionary algorithms. The outcome is a set of routes, either from A* or from the co-evolutionary process, that are collision-free; this set is generated in real-time and makes its implementation on edge-computing devices feasible. Although further research is needed to reduce the computational time, the computational experiments performed in this study confirm a good performance of the proposed approach in solving complex cases where well-known alternatives, such as M* or WHCA, fail in finding suitable solutions.

Motor intention decoding is one of the most challenging issues in brain machine interface (BMI). Despite several important studies on accurate algorithms, the decoding stage is still processed on a computer, which makes the solution impractical for implantable applications due to its size and power consumption. This study aimed to provide an appropriate real-time decoding approach for implantable BMIs by proposing an agile decoding algorithm with a new input model and implementing efficient hardware. This method, unlike common ones employed firing rate as input, used a new input space based on spike train temporal information. The proposed approach was evaluated based on a real dataset recorded from frontal eye field (FEF) of two male rhesus monkeys with eight possible angles as the output space and presented a decoding accuracy of 62%. Furthermore, a hardware architecture was designed as an application-specific integrated circuit (ASIC) chip for real-time neural decoding based on the proposed algorithm. The designed chip was implemented in the standard complementary metal-oxide-semiconductor (CMOS) 180 nm technology, occupied an area of 4.15 mm2, and consumed 28.58 μW @1.8 V power supply.

Consumer-centric energy management approaches are emerging as a major solution for future power systems. In this context, intelligent home management systems should control different kinds of devices existing in the houses assuring convenient comfort levels and understanding the users’ behaviour. At the same time, the home management systems should be able to interact with other actors such as energy communities, aggregators, and system operators. The main contribution of this work is a new methodology allowing intelligent management, in near real-time (1 minute), of different types of energy resources existing in a smart home. The energy resources include appliances and other loads, micro-generation, and electric vehicles. The proposed system includes a permanent evaluation of the operation state of each energy resource considering their functional model and the behaviour and comfort level defined by the users. Participation in demand response programs reducing the power consumption limits is also considered showing the advantage of the proposed approach. The case study contains two scenarios considering a demand response program of power limitation with 120 minutes duration. To guarantee the participation in these demand response events, the system should evaluate the priority of each device according to its model. A domestic consumer with 45 energy resources (appliances, generation, and electric vehicles) is used for demonstration purposes.

Image binarization is one of the fundamental methods in image processing and it is mainly used as a preprocessing for other methods in image processing. We present an image binarization method with the primary purpose to find markers such as those used in mobile 3D scanning systems. Handling a mobile 3D scanning system often includes bad conditions such as light reflection and non-uniform illumination. As the basic part of the scanning process, the proposed binarization method successfully overcomes the above problems and does it successfully. Due to the trend of increasing image size and real-time image processing we were able to achieve the required small algorithmic complexity. The paper outlines a comparison with several other methods with a focus on objects with markers including the calibration system plane of the 3D scanning system. Although it is obvious that no binarization algorithm is best for all types of images, we also give the results of the proposed method applied to historical documents.

Activity recognition technologies only present a good performance in controlled conditions, where a limited number of actions are allowed. On the contrary, industrial applications are scenarios with real and uncontrolled conditions where thousands of different activities (such as transporting or manufacturing craft products), with an incredible variability, may be developed. In this context, new and enhanced human activity recognition technologies are needed. Therefore, in this paper, a new activity recognition technology, focused on Industry 4.0 scenarios, is proposed. The proposed mechanism consists of different steps, including a first analysis phase where physical signals are processed using moving averages, filters and signal processing techniques, and an atomic recognition step where Dynamic Time Warping technologies and k-nearest neighbors solutions are integrated; a second phase where activities are modeled using generalized Markov models and context labels are recognized using a multi-layer perceptron; and a third step where activities are recognized using the previously created Markov models and context information, formatted as labels. The proposed solution achieves the best recognition rate of 87% which demonstrates the efficacy of the described method. Compared to the state-of-the-art solutions, an improvement up to 10% is reported.

Machine learning models work better when curated features are provided to them. Feature engineering methods have been usually used as a preprocessing step to obtain or build a proper feature set. In late years, autoencoders (a specific type of symmetrical neural network) have been widely used to perform representation learning, proving their competitiveness against classical feature engineering algorithms. The main obstacle in the use of autoencoders is finding a good architecture, a process that most experts confront manually. An automated autoencoder architecture search procedure, based on evolutionary methods, is proposed in this paper. The methodology is tested against nine heterogeneous data sets. The obtained results show the ability of this approach to find better architectures, able to concentrate most of the useful information in a minimized coding, in a reduced time.

In recent years, incremental sampling-based motion planning algorithms have been widely used to solve robot motion planning problems in high-dimensional configuration spaces. In particular, the Rapidly-exploring Random Tree (RRT) algorithm and its asymptotically-optimal counterpart called RRT <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" alttext="{}^{*}" display="inline" overflow="scroll"> <mml:msup> <mml:mi/> <mml:mo>*</mml:mo> </mml:msup> </mml:math> are popular algorithms used in real-life applications due to its desirable properties. Such algorithms are inherently iterative, but certain modules such as the collision-checking procedure can be parallelized providing significant speedup with respect to sequential implementations. In this paper, the RRT and RRT <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" alttext="{}^{*}" display="inline" overflow="scroll"> <mml:msup> <mml:mi/> <mml:mo>*</mml:mo> </mml:msup> </mml:math> algorithms have been adapted to a bioinspired computational framework called Membrane Computing whose models of computation, a.k.a. P systems, run in a non-deterministic and massively parallel way. A large number of robotic applications are currently using a variant of P systems called Enzymatic Numerical P systems (ENPS) for reactive controlling, but there is a lack of solutions for motion planning in the framework. The novel models in this work have been designed using the ENPS framework. In order to test and validate the ENPS models for RRT and RRT <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" alttext="{}^{*}" display="inline" overflow="scroll"> <mml:msup> <mml:mi/> <mml:mo>*</mml:mo> </mml:msup> </mml:math> , we present two ad-hoc implementations able to emulate the computation of the models using OpenMP and CUDA. Finally, we show the speedup of our solutions with respect to sequential baseline implementations. The results show a speedup up to 6x using OpenMP with 8 cores against the sequential implementation and up to 24x using CUDA against the best multi-threading configuration.