Jevgenijs Telicko

Sound event detection models commonly rely on spectrogram representations of audio signals and recent approaches have adapted image-based object detection architectures to acoustic domains. This paradigm is suitable for respiratory monitoring, where breathing events are visually distinguishable even under noisy conditions. In this study, we propose a Representation Enhancement for Neural Imaging (RENI) framework that combines a modified You Only Look Once (YOLO) object detection head with a trainable spectrogram front-end implemented using nnAudio. The front-end enables GPU-accelerated waveform-to-spectrogram conversion while allowing adaptive learning of Short-Time Fourier Transform (STFT) and Melody (Mel) basis functions. The model was trained for breathing-phase localization and respiratory rate estimation from 44.1 kHz audio recordings acquired during exercise. The results show that the trainable Mel representation improves respiratory-rate accuracy compared with static and trainable STFT configurations, achieving a mean absolute error of 1.15 breaths per minute. Bootstrap 95% confidence intervals and one-sided permutation tests show statistically significant gains for selected trainable STFT and Mel configurations under the min-MAE confidence thresholding protocol, while pooled effects remain directionally favorable for the trainable Mel front-end. These findings demonstrate improved exhale-based respiratory rate estimation under the studied conditions, while broader external validation is still required.

The operational and indoor environmental quality of buildings has a significant impact on global energy consumption and human quality of life. One of the key directions for improving building performance is the optimization of building control systems. In modern buildings, the presence of numerous actuators and monitoring points makes manually designed control algorithms potentially suboptimal due to the complexity and human factors. To address this challenge, model predictive control based on artificial neural networks can be employed. The advantage of this approach lies in the model’s ability to learn and understand the dynamic behavior of the building from monitoring datasets. It should be noted that the effectiveness of such control models is directly dependent on the forecasting accuracy of the neural networks. In this study, we adapt neural network architectures such as GRU and TCN for use in the context of building model predictive control. Furthermore, we propose a novel hybrid architecture that combines the strengths of recurrent and convolutional neural networks. These architectures were compared using real monitoring data collected with a custom-developed device introduced in this work. The results indicate that, under the given experimental conditions, the proposed hybrid architecture outperforms both GRU and TCN models, particularly when processing large sequential input vectors.

As more and more people are using wearable devices for tracking their sports activities, the amount of data generated by these devices is growing. Such wearables are worn day to day, tracking the sleep quality, stress level, and other measurements about the wellness of the user. In addition, some wearables are optimized for sports activities, allowing the user to track his activity sessions, and measure how his performance and results change with time. As the data about the training sessions is saved, the user could potentially use it not only to track his progress, but also to personalize his training program to train effectively. As of now different machine learning and neural network models are used for forecasting and predicting values in many fields, for example, in energetics, finance. In our paper, we propose how data collected from a wearable can be used to forecast heart rate during a running session using a sequence to sequence approach. The data from a future hypothetical training can be used to forecast the heart rate curve during the activity, allowing the user to plan the intensity of the training and the route taken, therefore personalizing his training process for better results. Furthermore, the proposed approach takes into account historical well-being data collected from the wearable, that could potentially increase the prediction quality. Our experiments with several forecasting models demonstrate how this approach could be carried out, while also showcasing how different neural network models perform on such a task. In addition, we also prove, that wellness data collected during the previous day, namely, hours slept, steps taken, calories burned and others can reduce the prediction error, therefore showing that it is important to take into account such features when creating forecasts for heart rate during the running session.

Buildings consume about 40 % of all energy. Ventilation plays a significant role in both the energy consumption of buildings and the comfort of occupants. To achieve energy efficiency and comfort, smarter ventilation control algorithms can be employed, such as those with feedback based on CO₂ levels. Furthermore, by knowing the current number of people in a space, ventilation can theoretically be adjusted to maintain a constant CO₂ level without wasting energy when people are not present. An additional benefit of such control could arise due to occupants’ habits. For example, if a person senses elevated CO₂ levels, even if the ventilation system has started operating more intense, they might choose to open a window, potentially compromising energy efficiency. Therefore, if the control algorithm were to maintain a constant CO₂ level, occupants may be less likely to open windows. In our work, we explore a model in combination with a custom monitoring system based on computer vision to implement such control. The monitoring system combines outside and inside CO₂ sensors with precise people counting based on computer vision to provide data to the model. The model relies on the mass balance equation for CO₂ and considers the historical data of the number of occupants and their activities to estimate the overall CO₂ generation in indoor spaces. The results suggest that the model can effectively forecast CO₂ dynamics with an absolute deviation of 40 ppm. However, it was observed that the analysis of the actual air exchange level could be compromised by several factors.

In the present day, monitoring and automated control stand as pivotal factors for the energy-efficient and comfortable operation of buildings. As the demand for indoor climate control grows, building management systems have become more intricate, making their control challenging due to the increasing number of controllable elements. Replacing manual human analysis of complex systems can be achieved through the utilization of algorithms like model-based control. It is important to note that performance of this method usually relies on the accuracy of neural network-based building state forecasts. Studying the internal dynamics of climate as influenced by temperature changes necessitates a brief record of measurements, whereas evaluating structural modifications through moisture transfer demands data covering a more extended period. Neural networks such as Long Short-Term Memory have the potential to lose information within lengthy time-series data, and the intricate nature of moisture transfer further adds complexity to the task of approximating functions, ultimately leading to a reduction in energy efficiency. In order to improve the precision of indoor climate predictions, our suggestion involves not only assessing changes in temperature but also considering alterations in U-values triggered by temperature variations and moisture transfer. Our preliminary assessment of the influence of U-value, conducted through numerical simulations using WUFI6, exposes variations of up to 10 % of U-value in certain scenarios. Dealing with these computations in real time using physical models proves to be demanding due to computational requirements and limited data availability. To tackle this issue, we present an innovative preprocessing approach for on-the-fly evaluation of U-values. Empirical trials involving three years of monitoring data indicate that the suggested technique led to an approximate 8 % reduction in the average mean squared error of climate predictions based on neural network models, in specific instances.

For an efficient indoor air purification, it is important to know the detailed airflow distribution in the room. A series of numerical simulations are carried out for five heating regimes using an air–air heat pump, capillary mat on the ceiling, capillary mat on the walls, heated floor, and radiator. The most homogeneous temperature field is obtained for the case with the heated floor. The highest velocity is obtained for the air–air heat pump, while the lowest is obtained for the capillary mat on the ceiling. A portable air cleaner based on the prototype device is introduced into the model and its influence on the velocity and temperature distributions is calculated. Our simulations additionally consider is the transport of an infectious aerosol and its purification inside the air cleaner. The time dependency of the concentration is exponential, and the purification rate depends on the air cleaner’s orientation and heating regime. The efficiency is higher for a purifier with flow in the upwards direction compared to in a horizontal one. In the experimental part, an NaCl solution is dispersed into the air, and the efficiency of purification in the case of the air–air heat pump is evaluated by measuring the time-dependent particle concentrations. These experimental results corroborate the numerical model.

Nowadays, monitoring and automated control are key for the power-efficient and comfortable use of buildings. As the demand for the indoor climate grows, the complexity of building control increases due to the increased number of controlled systems within the building. Therefore, control methods capable of assessing the dynamics of a task and controlling a building independently, are gaining popularity. One of these methods is a model-based predictive control, the energy efficiency of which is directly dependent on the accuracy of predictions of the state of the building. To analyse changes in the dynamics of the internal climate of a building caused by temperature changes, a relatively short history of measurements is needed. However, the analysis of changes in structural properties caused by moisture transfer requires much longer-term historical data. It is known that even neural network architectures with large core memories such as Long Short-Term memory could lose information on large time-series data. In addition, from mathematical point of view, moisture transfer is more complex than heat transfer. As a result, the approximating function for a neural network becomes more complicated, which leads to a decrease in energy performance. To solve this problem, a pre-processing technique to obtain the U value in real time is introduced. The results of a numerical simulation in WUFI6, verified by measurements, showed that the U value of an experimental construction could change by ~10 % due to moisture transfer. The experiments with 3-year monitoring data showed that usage of the proposed method in some cases reduces the average mean squared error of neural network in indoor climate forecast by ~8 %.

Building performance has significant impact on humans’ life and ecology as buildings account for 39% of total greenhouse gas emissions and consume about 40% of total global energy consumption. Smart building control is one of the key points to archive high energy efficiency. Each year, the complexity of building state control grows due to the increase in the number of controlled elements that are used to achieve better indoor climate. Therefore, in the manual analysis and implementation of the building control program, an error can easily appear due to the human factor. Artificial intelligence (AI) algorithms could be used as an alternative solution as they could evaluate building dynamics independently. One of strategies for automatic building control adaptation to its dynamic is model based predictive control where neural network is used for different control strategies evaluation. Performance of such control technique is highly dependent on control strategies evaluation accuracy. To achieve top accuracy, several hyperparameters of neural network could be tuned as well as data set for specific construction must be prepared. Preparation of data set could be a problem as random control of building for generation of dataset could be not unacceptable for building users as well as it could damage construction.In this paper authors process optimization of experimental building heating system control algorithm to achieve smaller fluctuations of temperature indoors. For dataset generation were used several data from weather station as well as heating system parameters and temperature indoors while building was controlled by thermostat with build in PID regulation. For evaluation of building dynamics were used temporal convolutional neural network. To achieve high accuracy results of control strategies evaluation, several hyperparameters of neural network were tested. Finally resulting model were tested on physical building. Results indicate that in some cases developed control model could prevent temperature fluctuations which could be caused by limits of heating system power.

With solar panels increasingly used in nearly zero energy building solutions it is important to clarify if the panels can achieve the efficiency indicators specified by their producers in real operation conditions. To determine the efficiency of poly- and monocrystalline panels depending on their spatial orientation and the seasons a set of test panels was installed in 2018 in Riga, Latvia for long-term monitoring of the amount of generated power and produced energy. Here we summarize the results of the first year of monitoring. Data indicate that orientation has high significance, very small amounts of energy are generated in autumn/winter and that there is a good correspondence of the actual efficiency to the standardized estimated values. The energy production is significantly reduced by the cloudiness that is characteristic for Latvia’s climate from October till January. To make a generalization of the results possible it is planned to continue the monitoring for several years.