Introduction

Smart Home Care (SHC) with assistive technologies have been a major topic in both the research and manufacturing industries in recent decades, with a focus on improving social interaction, supporting healthcare, commerce, education, and everyday activities1. The number of elderly people with unmet care and support needs is increasing substantially. Emerging technological developments have the potential to address some of the challenges of caring and supporting older people. There are various types of assistive technologies such as smart mobile applications, automated home appliances, mobile and assistive autonomous robots, self-driving vehicles, AI-enabled wearable health devices, new types of mechanical medicine dispensing boxes, wearable diagnostics, voice-activated devices, virtual, augmented and mixed reality, and Smart Home (SH)2. The World Health Organization (WHO) suggests that the development of smart, physical, social and age-friendly environments will improve the quality of life of older adults. Social Companion Robots (SCR) integrated with various sensing technologies such as vision, voice, and haptic can interact with other smart devices in the home environment and can enable the development of advanced AI solutions to an age-friendly smart space3. Another example is an autonomous mobile robot that can safely deliver packages to the desired location using the Global Positioning System (GPS)4 within Smart Cities (SC). For autonomous mobile robots (AMR) in SH and for the development of new context-aware applications, knowledge of the spatial distribution of physical variables such as radio frequency (RF) interference, pollution, magnitude of geomagnetic field, temperature, humidity, sound and light intensity is important5. Another important skill in AMR is the ability to resolve movement and mobility in an environment with multiple navigation points6. Barber et al. present the design of a heterogeneous multi-robot system consisting of a small mobile robot that monitors the well-being of elderly people living alone and suggests activities to keep them positive and active, and a mobile home handling robot that helps to perform household tasks7. Using major advances in computer vision and machine learning, robots can be used as hotel concierges, museum guides, waiters in cafes and restaurants, home assistants, and automated delivery drones within the SC platform8. Robotics9, artificial intelligence (AI), and the Internet of Things (IoT) support various processes in many scenarios of modern life10,11,12,13,14,15.

Novelty and gaps in knowledge

Assistive technologies have become an important area of research, innovation, and production in recent decades, with a major focus on improving social interactions and providing support for medical, business, educational, and everyday activities. The following assistive technologies can be used to support independent living and home life for the elderly and disabled in Smart Home Care (SHC) environments: wearable headband and electronic fabrics for long term forehead EEG signal sensing16, multirobotic (MR) system in the assisted home environment (AHE) to support seniors in their daily living (DL)7, personalized home-care (PHC) robot support for the elderly17, artificial intelligence - based smart comrade robot for elders healthcare (HC) with strait rescue system18, healthcare live-in prognostic robot (or HLPR)19, automatic pathological gait recognition (PGR) by a mobile robot using ultra wideband-based localization and a depth camera20, bridging gaps in the design and implementation of socially assistive technologies (SAT) for dementia care: the role of occupational therapy21, visualizing domiciliary human activity (HA) perception and comprehension22, The AI devices utilized in elderly healthcare were summarized as robots23, virtual reality technologies (VRT), smart wearables, and robots were used to provide telerehabilitation services (TRS)24 etc.

In order to provide building control technological support for independent living of elderly and disabled people in the SHC home environment, information on occupancy of monitored areas without the use of cameras (to ensure privacy) is needed to optimize the management of operational and technical functions. A detailed description of scientific articles on the topic of monitoring and prediction of building occupancy (measured values used, focus of the study) is described in Table 1.

Table 1 A study investigating the monitoring and prediction of building occupancy based on measured values.
Full size table

In addition, we note that suitable neural network (NN) algorithms can also be used to provide energy management for building control optimalization.34,35,36,37,38,39

What is missing from the above studies that describe the use of robots and assistive technologies to support independent living in the home environment in SHCs and in studies describing occupancy monitoring and prediction in SHCs is a description of a robust technological end-to-end solution for integrating AMR with building automation technologies (e.g., KNX technology) in SHCs in an IoT framework. There is a lack of studies on the use of a globally deployed secure standard designed for building automation to ensure that they can communicate securely within IoT using 5G networks for a transformational change towards more efficient, responsive, and adaptive healthcare systems and the provision of appropriate services to support independent living of older people and people with disabilities at home.

In the present article, the authors describe a newly proposed implementation of interoperability and secure data transfer for communication between autonomous mobile robot (AMR) and building automation KNX technology in SHC using 5G networks within IoT. In addition, a novel design and application of a time-ahead \(\textrm{CO}_2\,\)concentration prediction method are proposed to send presence and occupancy information to monitored Smart Home Care spaces. The novelty lies in the adoption of a mechanism that avoids cameras in autonomous mobile robots to detect deviations from the daily routine. The newly proposed approach is based on nonlinear input-output neural network models and nonlinear autoregressive neural network models with exogenous inputs. The Levenberg-Marquardt Algorithm (LMA), the Bayesian Regularization Algorithm (BRA), and the Scaled Conjugate Gradient Algorithm (SCGA) were used as the learning algorithm. The operational and technical variables measured by the KNX technology were used to indirectly locate people, eliminating the need for cameras by using secure communication of autonomous mobile robots in the IoT through 5G networks (Fig. 3). The experiment described below advances the state-of-the-art in this specific application of humanoid robots in healthcare. Benefits and novelty of the proposed method is using information from standard sensors used in building automation to monitor the current state of operational technical variables (T, \(\textrm{CO}_2\,\), rH, E, W1-W3, D1-D4) without cameras (Fig. 5); ensuring secure data connectivity and interoperability between building automation technology (KNX and AMR technology to optimize human-AMR interaction (Fig. 6); ensuring a secure standard using KNX technology, which is standardized according to EN 50090-3-4; ensuring data connectivity of KNX technology and AMR to IoT in a secure 5G network design (Fig. 7); SHC occupancy monitoring and transfer of occupancy information of monitored premises from KNX technology (SHC automation) to AMR using time forward prediction for faster determination of AMR mode (I. Mode-Robot Invisible, II. Mode-Cleaning Mode, III. Mode-Be Nearby (Fig. 4) IV. Standby mode).

Aims and objectives

The presented article introduces an innovative hardware and software solution designed to establish seamless secure communication and interoperability between KNX technology (standard EN 50090, ISO/IEC 14543 - Building Automation) and an autonomous mobile robot (AMR) using 5G networks. This AMR was specifically developed to help and support residents of SHC within the domain of social assistance robotics. Its primary purpose is to indirectly monitor occupancy in time forward to detect deviations from the daily routine of elderly residents within the SHC premises and determine the whereabouts of individuals, all without reliance on cameras. This groundbreaking solution leverages the KNX open data protocol to facilitate data connectivity with various other technologies.

The AMR operates on the basis of predefined modes, and it derives follow-up actions from the analysis of operational data. These actions include assessing indoor environmental quality and detecting the opening and closing of windows and doors, all within the context of building automation by KNX technology. A novel method of \(\textrm{CO}_2\,\)waveform prediction with nonlinear autoregressive NN with exogenous inputs (NARX) with different network settings using LMA, BRA and SCGA based on measured operational variables is proposed and applied in the KNX technology system in SHC, such as indoor relative humidity, indoor temperature, interior lighting curve, binary information from magnetic contacts designed to monitor the opening and closing of windows (W1, W2, W3) - (Window 1, Window 2, Window 3) or doors (D1, D2, D3, D4) - (Door 1, Door 2, Door 3, Door 4). Furthermore, within the KNX technology system in SHC, a new method of predicting the \(\textrm{CO}_2\,\)forward flow using NN non-linear model of the time series of inputs and outputs (using the LMA) is proposed and applied using measured operational and technical waveforms of variables such as indoor relative humidity, indoor temperature, indoor lighting intensity waveform, binary information from magnetic contacts designed to monitor the opening and closing of windows (W1, W2, W3) - (Window 1, Window 2, Window 3) or doors (D1, D2, D3, D4) - (Door 1, Door 2, Door 3, Door 4) and then transferring this information in advance to the AMR. The objectives of the research work can be divided into three basic parts.

  • Objective 1: AMR uses an indirect method of determining the location of a person that eliminates the need for cameras. This is achieved by linking AMR with KNX technology and monitoring the opening and closing of windows and doors (W1-W3 and D1-D4) in the SHC within IoT using 5G networks.

  • Objective 2: AMR indirectly monitors the presence of people in a room by forward prediction of occupancy determination, without the use of cameras, using the measured values of the reference waveform \(\textrm{CO}_2\,\)to learn the neural network (NN) model. This is achieved by predicting the \(\textrm{CO}_2\,\)concentration waveform in advance using dynamic nonlinear autoregressive NN models (NARX) and then feeding this information to the AMR in advance. This method provides an indirect way of detecting the occupancy of monitored rooms in time-forward to detect deviations from daily routine of elderly residents in the SHC.

  • Objective 3: AMR indirectly monitors the occupancy of the room with a forward view, without the use of cameras and without the use of a reference waveform \(\textrm{CO}_2\,\)to learn the Nonlinear Input-Output (NIO) Neural Network (NN) models. This is achieved by predicting the concentration \(\textrm{CO}_2\,\)using an NN time series nonlinear input output model in advance and then transmitting this information to the AMR. This method provides an indirect way to detect the occupancy of monitored rooms in time to detect deviations from the daily routine of elderly residents in SHCs.

Related work

The following sections deal with published articles in the field of occupancy and occupancy monitoring in buildings and the use of this information to optimize building operations and energy management. In addition, various occupancy prediction methods have been reviewed. Relevant AI-based frameworks for occupancy detection and monitoring have also been explored.

Comparison with state-of-the-art approaches for occupancy estimation and \(\textrm{CO}_2\,\) sensing.

The Candanedo and Feldheim research used data recorded from light, temperature, relative humidity, and \(\textrm{CO}_2\,\)sensors as a means to detect occupancy and a digital camera to determine field occupancy for training supervised classification models. The trained and tested models used were Random Forest (RF), Gradient Boosting Machines (GBM), Linear Discriminant Analysis (LDA), and Classification and Regression Trees (CART)40. Haidar et al. present a new method for selecting the data collection period and the corresponding sensors (\(\textrm{CO}_2\,\), humidity, occupancy) for a building occupancy prediction model (Random Forest) with a satisfactory accuracy of at least 90%, using eight sensors collecting data at 20-minute intervals or five sensors collecting data at 15-minute intervals41. Brodie et al. used multiple formalistic linear regression modeling and artificial neural network modeling to analyze cost-effective and opportunistic data streams from a smart building to develop occupancy estimates for HVAC control purposes using available data from Wi-Fi access points, \(\textrm{CO}_2\,\)sensors, PIR motion detectors, and electrical load meters and plugs42. Paige Wenbin Tien et al. present a vision-based deep learning approach to detect and recognize occupant activities in building spaces. Based on a convolutional neural network, a model was developed to detect occupancy activity using a camera. An average detection accuracy of 80.62% was achieved43. Table  2 synthesizes the main characteristics required from the presented work with the characteristics of other related works with the prediction of building occupancy.

Table 2 Comparison with state-of-the-art approaches for occupancy estimation and \(\textrm{CO}_2\,\)sensing.
Full size table

Studies investigating the ANNs with the NARX algorithm implementation

Short-term load forecasting is crucial for the operation planning of an electrical grid. Forecasting the next 24 h of electrical load in a grid allows operators to plan and optimize their resources. The purpose of the study by Buitrago et al. was to develop a more accurate short-term load forecasting method that uses non-linear autoregressive artificial neural networks (ANN) with exogenous multivariate input (NARX). Using the proposed framework, an average absolute percentage forecast error of the order of 1% was achieved, representing an improvement 30% over the average error using forward artificial neural networks (ANNs), ARMAX, and state-based methods44. Accurate calculations and predictions of the heating and cooling loads in buildings play an important role in the development and implementation of energy management plans for buildings. A study by Jee-Heon Kim et al. aimed to improve the accuracy of cooling load prediction using an optimized nonlinear autoregressive exogenous neural network model (NARX). In predictive models of cooling loads, removal of missing values and adjustment of structural parameters were shown to help improve the predictive performance of the neural network model45. Forecasting peak load demand is important in building unit sectors, as climate change, technological development, and energy policies are causing an increase in peak demand. The ANN model with external variables (NARX) in the Yunsun Kim et al. study worked best for 1-h to 1-d ahead forecasting46. Predicting building energy consumption is essential for planning and managing energy systems. In the study of Koschwitz et al., the data-driven thermal load forecasting performance of \(\varepsilon\)-SVM Regression (\(\varepsilon\)-SVM-R) based on a Radial Basis Function (RBF) and a polynomial kernel is compared with the result of two non-linear autoregressive exogenous recurring neural networks (NARX RNN) of different depths. The evaluation of the resulting predictions shows that the NARX RNNs produce higher accuracy than (\(\varepsilon\)-SVM-R) models, in combination with comparable computational effort47. The main purpose of energy prediction is to estimate the potential energy savings in a building from various energy management and retrofit programs. The multiple linear regression (MLR) model and non-linear autoregression with exogenous input involving the artificial neural network (NARX-ANN) are widely used for building energy consumption prediction. Although NARX-ANN offers simplicity, there are some limitations in its architecture, which are often associated with optimization problems. MLR prediction is suitable for a system that behaves linearly. Optimizing NARX-ANN with the particle swarm optimization (PSO) technique for energy prediction and basic energy modeling has been found to achieve better results in terms of error measurements48. Table 3 summarizes the main characteristics described in the presented works, together with those of other related studies that implement artificial neural networks using the NARX algorithm.

Table 3 A study describing the application areas of neural networks with the implementation of the NARX algorithm Variables-occupancy monitoring SH.
Full size table

The presented study introduces a novel and unique approach that involves the use of an AMR in SHC within IoT with 5G networks. This AMR is installed for the location of the residents, monitoring the presence of individuals in time to detect deviations from the daily routine of the elderly residents, and determining the occupancy within the monitored spaces. In particular, this is achieved through an indirect methodology, effectively eliminating the need to rely on cameras. Furthermore, the AMR’s operational modes trigger subsequent actions based on measured operational variables. These actions encompass the evaluation of indoor environmental quality and the surveillance of security aspects within the SHC, particularly the the monitoring of the opening and closing of windows and doors. All this can be included in the SHC automation, using secure standard KNX technology. The predictive aspect relies on the utilization of the nonlinear auto-regressive NN with exogenous inputs (NARX) and NN time series nonlinear input output model. The content of the article is summarized in a few sentences.

  • NARX neural network (NN) models with LMA, BRA, and SCGA were used to predict \(\textrm{CO}_2\,\)with reference \(\textrm{CO}_2\,\)for occupancy time and presence monitoring to detect deviations from the daily routine of elderly residents in Smart Home care.

  • Nonlinear NN input-output time series models with LMA were used to predict the progression of \(\textrm{CO}_2\,\)without reference \(\textrm{CO}_2\,\)with time-forward occupancy and presence monitoring to detect deviations from the daily routine of elderly residents in Smart Home care.

  • For the time-forward prediction of the \(\textrm{CO}_2\,\)course, the operational technical values (indoor relative humidity course, indoor temperature course, indoor illuminance course, binary information from magnetic contacts was used to monitor the opening and closing of windows (W1-W3) and doors (D1-D4)) measured within the building automation control using KNX technology.

  • Introduction of secure communication possibilities between KNX building automation technology and AMR within IoT with 5G networks.

Description of the used mathematical methods

To train the neural network (NN) model with NARX, three optimization algorithms were used in the experiments: LMA50 (Table 4), (Table 5), BRA51 (Table 4) and SCGA (Table 4). To train a non-linear input-output NN model, LMA was used in the experiments50 (Table 6), (Table 7). The quality of diagnostic tools presented by artificial intelligence methods can be specified by many different requirements. The reason for using only LMA was the sufficient precision and speed in learning the nonlinear input-output NN model compared to. Quality is usually measured using the mean squared error (MSE) and the Pearson correlation coefficient (Tables 4 - 7), the mean absolute percentage error (MAPE) (Table 5) and the Bland - Altman difference plot (Fig. 9d). The Bland-Altman plot is a data plotting method used to analyze the agreement between two different waveforms.

Non-linear Auto Regressive (NARX) dynamic neural network (NN)

The NARX NN model is used to monitor and time-forward predict the occupancy of the SHC room in advance using non-electrical variables measured within the SHC automation using KNX technology. For the NARX NN model, a reference waveform of concentration \(\textrm{CO}_2\,\)and 10 measured operational variables temperature indoor, relative humidity indoor, illuminance indoor, and binary magnetic contacts (W1-W3, D1-D4) were used. The defining Eq. (1) for the NARX NN model is:

$$\begin{aligned} \begin{aligned} \textit{y(t)} =&f(y(t - 1), y(t - 2), ... , y(t - n_y),&x(t - 1), x(t - 2), ... , x(t - n_x)). \end{aligned} \end{aligned}$$
(1)

where the next value of the dependent output signal y(t) is regressed to the previous values of the output signal and the previous values of the independent (exogenous) input signal (Fig. 1).

Fig. 1
figure 1

NARX Neural Network (NN) model with 10 input variables (T, rH, E, D1, D2, D3, D4, W1, W2, W3), 1 input reference variable \(\textrm{CO}_2\,\)and with 1 output variable time-forward prediction of \(\textrm{CO}_2\,\).

Full size image

Nonlinear input-output (NIO) Neural Network (NN)

The nonlinear input-output neural network model (NIO NN) was used to monitor and predict the occupancy of the SHC room in advance using non-electric variables measured within SHC automation using KNX technology without reference waveform \(\textrm{CO}_2\,\). For the nonlinear input-output (NN) model, 10 measured operational variables were used: temperature indoor, relative humidity indoor, illuminance indoor, and binary magnetic contacts (W1-W3, D1-D4). These are two series, an input series x(t) and an output/target series y(t). The selected model provides a prediction of the value of y(t) from previous values of x(t), but without knowing of previous values of y(t). The selected input/output model can be written as follows (Fig. 2):

$$\begin{aligned} y(t) = t(x(t - 1), \dots , x(t - d)). \end{aligned}$$
(2)
Fig. 2
figure 2

Nonlinear input-output Neural Network (NN) model with 10 input variables (T, rH, E, D1, D2, D3, D4, W1, W2, W3) and 1 output variable time-forward prediction of \(\textrm{CO}_2\,\).

Full size image

Data analysis procedure

A detailed description of the data analysis process is shown in the block diagram in Fig.3. Data were sent within the interoperability between KNX technology and BACnet technology from the SpaceLYnk gateway (with the Wiser vizualisation software) to the Desigo CC software tool, where measured data are visualized and archived. The measured data was transferred to a PC with the MATLAB sw tool. The first step was the preprocessing and normalization of the measured data. In the second step, the corresponding NN model was designed. For the prediction of \(\textrm{CO}_2\,\)from the input measured values x(t) (the quality of the indoor environment in room 220 (temperature T (\(^{\circ }\text {C}\)), \(\textrm{CO}_2\,\)(ppm), relative humidity rH (%), light intensity E(lx), the opening/closing status of windows (W1-W3) in rooms 220 and 217 and doors (D1-D4) in rooms 220, 213, 215, 216, 217) for the transmission of the secured information from KNX technology to AMR in time advance, the NARX NN model was designed. A reference waveform \(\textrm{CO}_2\,\)was used to learn the NARX NN. To predict \(\textrm{CO}_2\,\)from input measured values x(t) (the quality of the indoor environment in room 220 (temperature T (\(^{\circ }\text {C}\)), \(\textrm{CO}_2\,\)(ppm), relative humidity rH (%), light intensity E(lx), the opening/closing status of windows (W1-W3) in rooms 220 and 217 and doors (D1-D4) in rooms 220, 213, 215, 216, 217) for the transfer of the secured information from KNX technology to AMR in time advance, a NIO NN model was designed without the need to use a \(\textrm{CO}_2\,\)reference waveform. The next step was to divide the acquired data between training (learning), validation, and testing of the learned network. The recommended ratios are 70% for learning, 15% for validation of the values, and 15% for testing the network. Learning and validation are used to learn the network and determine when learning is complete to avoid overfitting. The test data then independently check the network. The fourth step is the design of the neural network architecture. The number of hidden neurons and delays are specified. More neurons can solve a more complex problem, but with fewer neurons a more compact solution is available. The fifth step was to learn the NARX NN model and the NIO NN model. LMA, BRA, and SCGA were used as learning algorithms. The sixth and penultimate step is to evaluate and optimize the NARX NN model and the NIO NN model (Fig. 3). To obtain the optimal NN model for the prediction of \(\textrm{CO}_2\,\)in advance, the above steps were repeated with the change of the described parameters. If the NARX NN model performed well against the learning data but was significantly worse against the test data, this indicated overfitting and it was appropriate to reduce the number of neurons. If even learning was not accurate, we tried to provide more input data to the NN models.

Fig. 3
figure 3

Block diagram - processing of the measured quantities with NARX NN and NIO NN models.

Full size image

Experimental part

The experimental part describes in detail the technologies used for data communication between KNX and the AMR technology in SHC. Furthermore, this section describes in detail the implementation procedure and the fulfillment of the objectives of the presented article.

Description of technologies used in practical experiments in HEALTH.Lab - Smart Home Care

First, the object itself is described as HEALTH.Lab - Smart Home Care.

HEALTH.Lab - SHC description

It is an SHC apartment, called HEALTH.Lab - Test bed (Healthcare Energetics Automation Living and Telemedicine House https://healthlab.vsb.cz/en) at VSB–TU Ostrava (Czech Republic). Specifically, the rooms are living room with kitchen 220 (area 56,14 \(\hbox {m}^2\)), bedroom 217 (area 19,68 \(\hbox {m}^2\)), bathroom and toilet 216 (area 10,17 \(\hbox {m}^2\)), hall 215 (area 13,63 \(\hbox {m}^2\)), entrance hall 213 (area 12,83 \(\hbox {m}^2\)), utility room 214 (area 29,61 \(\hbox {m}^2\)). The floor plan can be seen in Fig. 4.

Fig. 4
figure 4

a) Floor plan with positioned sensors and bus buttons for controlling operational and technical functions in Smart Home Care. b) Picture from the Living room, AMR in III.mode - be nearby.

Full size image

The visualization software tool Wiser integrates the visualization, archiving, and control of operational and technical functions in the SHC using the SpaceLYnk KNX module from the KNX technology side. The software tool ETS 6 was used to program the KNX devices within the KNX technology. The KNX sensors MTN6005-0001 (temperature, relative humidity, \(\textrm{CO}_2\,\)) and MTN630719 (illuminance) were used to measure indoor environmental quality. The location of the sensors was chosen according to EN ISO 16000-26. The area of Living Room 220, where the measurements were made, was less than 50 \(\textrm{m}^2\), therefore, only one measuring point at a height of 1.5 m and 1 m from the wall could be used. This is a space ventilated under HVAC control. The \(\textrm{CO}_2\,\)concentration in the room was assumed to be the same at all points. For the measurements, a five-minute interval was chosen to record the values in the SpaceLYnk controller. The measurements were carried out for one month, March 15 - April 14 2022. For the security area of the SHC apartment, magnetic contacts SA203 were connected to the KNX module of binary inputs MTN 644592 to monitor the opening and closing of windows (W1, W2, W3) and doors (D1, D2, D3, D4) in the SHC (Fig. 5). Building automation control and monitoring are implemented using the interoperability of KNX technology with BACnet technology (through the SpaceLYnk KNX / BACnet gateway) (Fig. 3, Fig. 5).

Fig. 5
figure 5

Design of KNX technology communication implementation with AMR within IoT in SHC.

Full size image

Autonomous mobile robot

In this study, we have chosen the MiR100 Autonomous Mobile Robot (AMR), developed by the Danish company MiR.

Robot specification

The selected AMR model stands as the smallest member in the MiR robot lineup, measuring 890 mm in length, 580 mm in width, and 352 mm in height, with a substantial payload capacity of 100 kg. Its compact form factor makes it suitable for deployment in domestic and residential settings, while its impressive payload capability ensures its efficacy in tasks ranging from assisting individuals with limited mobility to transporting larger loads.

AMR and KNX interoperability in 5G networks

The integration of autonomous mobile robots (AMRs) in home care and their cooperation with 5G or KNX technology represents a transformative shift towards more efficient, responsive, and adaptable healthcare and service delivery systems. This type of robot is also widely deployed in places that are harmful or inaccessible to humans52. There is also no problem sharing space with people who behave like cobots53. The use of 5G technology to enable remote control of AMR in challenging environments, such as hospitals, shows the potential of these technologies to support critical healthcare operations54. The independence from local Wi-Fi limitations, thanks to 5G, opens up new possibilities to deploy AMRs in areas where conventional connectivity solutions might falter (Fig. 6).

Fig. 6
figure 6

Block diagram of AMR and KNX interoperability within IoT in SHC.

Full size image

Integrating the MiR100 Autonomous Mobile Robot (AMR) with the KNX building automation system showcases the advanced interoperability facilitated by modern industrial 5G networks. The MiR100, with its built-in REST API over the HTTP GET/POST protocol, enables efficient communication for operational data exchange and task assignments. In a 5G-enhanced setup, the default Wi-Fi connectivity of the AMR can be replaced or augmented by a 5G connection, taking advantage of the high-speed, low-latency communication that 5G networks offer. This ensures more reliable and secure data transmission between the MiR100 and the KNX system. To facilitate this integration, a communication server, now potentially enhanced by 5G connectivity, acts as the intermediary between the KNX system and the MiR100. This server can connect directly to a 5G network, bypassing traditional Wi-Fi limitations (stability, coverage, bandwidth) and offering enhanced security features inherent to 5G such as advanced end-to-end encryption, network slicing for dedicated bandwidth and isolation, and robust identity management and access control mechanisms. The KNX technology connects to this communication server through a LAN connection, managed by a SpaceLYnk logic controller that supports various connectivity options. By integrating 5G, the SpaceLYnk module can now also take advantage of 5G’s capabilities for IoT platforms through MQTT over 5G, offering improved reliability and security for building automation tasks (Fig. 7). The role of the 5G network in this ecosystem allows for real-time, secure communication between the AMR and KNX technology, facilitated by the communication server’s translation between the REST API and the MQTT protocol, supported by 5G’s ultrareliable low latency communication (URLLC) and enhanced Mobile Broadband (eMBB) capabilities. This ensures that mobile robot tasks are executed with minimal delay and maximum security (Fig. 7).

Fig. 7
figure 7

System architecture and communication flows within IoT.

Full size image

Industrial 5G architecture

5G is an industry standard for mobile radio communications that offers a huge bandwidth and transmission capacity, real-time or high subscriber capacity, depending on the requirements, with extremely high reliability and security. Unlike previous generations of mobile networks such as 3G and 4G, industrial 5G also meets industrial requirements for the first time, which can guarantee intelligent wireless communication between machines and applications. With industrial 5G, for example, private networks can also be realized, and therefore all requirements for flexible and forward-looking network connectivity in mobile or highly flexible applications can be met. Industrial 5G finds its application precisely in the field of Industry 4.0. Due to its high degree of flexibility, versatility, usability and efficiency, 5G enables the realization of Industry 4.0, smart manufacturing plants, and the Industrial Internet of Things (IIoT). Within the VSB-TUO campus, some laboratories and spaces are covered by the 5G industrial network, including the Smart Factory and HEALTH.Lab with Smart Flats. This industrial network is used for experimental activities that support science and research within the IoT. In general, it is one of the many technologies that makes it possible to conduct world-class scientific experiments and development activities. The question was asked how to ensure secure data transfer between KNX technology for building automation and AMR for one building or for several buildings within Smart Cities. Another question was how to ensure the transfer of the necessary information between KNX technology and AMR in advance? Based on the question posed, the task was to design a suitable method to create a NN-based model to predict the \(\textrm{CO}_2\,\)flow from the measured data from the SHC to AMR for the secure transfer of information about the presence of people and the occupancy in time to detect deviations from the daily routine of elderly residents of the SHC monitored area - room 220. In the experimental part, a technical solution is presented that involves the use of AMR to locate people in the SHC, to monitor their presence, and to assess the occupancy of the monitored SHC areas. This is achieved by an indirect method that completely eliminates the need for cameras. AMR operates on the basis of predefined modes and derives subsequent actions from the analysis of measured operational data. These actions include evaluating indoor environmental quality and detecting the opening and closing of windows and doors, all in the context of building automation using KNX technology. The experimental part is divided into three main parts:

  • Part 1 (Objective 1): AMR uses an indirect method to determine the location of a person, eliminating the need for cameras. This is achieved by linking AMR with KNX technology and monitoring the opening and closing of windows and doors (W1-W3 and D1-D4).

  • Part 2 (Objective 1): AMR indirectly monitors the presence of people in the room in time to detect deviations from the daily routine of elderly residents, without the use of cameras. This is achieved by predicting the concentration \(\textrm{CO}_2\,\)using the NARX NN in advance and then transmitting this information to the AMR. This approach provides an indirect way of detecting the occupancy of monitored rooms in the SHC.

  • Part 3 (Objective 1): AMR indirectly monitors the presence of people in the room in the future to detect deviations from the daily routine of elderly residents, without the use of cameras. This is achieved by predicting the concentration \(\textrm{CO}_2\,\)using the Nonlinear Input-Output NN model in advance and then transmitting this information to the AMR. This method provides an indirect way of detecting the occupancy of monitored rooms in the SHC.

In the framework of the newly proposed technological system, a new method for \(\textrm{CO}_2\,\)progress prediction in Part 2 with nonlinear autoregressive NN with exogenous input (NARX) with learning algorithms using LMA. In Part 3 a non-linear input-output NN model with LMA was used. Measured operational variables such as indoor relative humidity waveform, indoor temperature waveform, indoor lighting waveform, binary information from magnetic contacts designed to monitor the opening and closing of windows (W1, W2, W3) - (Window 1, Window 2, Window 3) or doors (D1, D2, D3, D4) - (Door 1, Door 2, Door 3, Door 4) are used as input variables of NN models.

Part 1: KNX - AMR - Location of the position of the person in the SHC (objective 1)

AMR locates a person’s position indirectly (without the use of cameras) based on AMR’s secure communication with KNX technology within a 5G network. Operational and technical variables were measured using KNX technology in the SHC in living room 220 (binary magnetic contacts to monitor windows and doors D1-D4, W1-W3, indoor temperature, indoor illuminance, indoor relative humidity) during the period from 15 March 2022 to 14 April 2022 (Fig. 8).

Fig. 8
figure 8

Measured operational and technical variables in normalized format using KNX technology in SHC in Living Room 220 (binary magnetic contacts for monitoring doors D1-D4 and windows W1-W3, temperature indoor T, illuminance indoor E, relative humidity indoor rH) in the period from 15.03.2022 to 14.04.2022.

Full size image

Part 2: KNX - AMR monitors room occupancy indirectly ahead of time by predicting \(\textrm{CO}_2\,\)concentration using NARX’s NN (objective 2)

In the selected living room 220 SHC (Fig. 4), the presence of people in the selected room is monitored indirectly (without the use of a camera) by means of KNX field measurements with the use of KNX technology, based on the prediction of the concentration waveform \(\textrm{CO}_2\,\)using the NARX NN model, and then it is assumed that the information of the predicted waveform \(\textrm{CO}_2\,\)is transmitted to the AMR (indirect determination of the occupancy of the monitored rooms in the SHC). The reference concentration waveform \(\textrm{CO}_2\,\)was measured using the KNX MTN6005-0001 sensor. The \(\textrm{CO}_2\,\)measurement range is 300 ppm to 9999 ppm. The precision is for measured \(\textrm{CO}_2\,\)values from 300 to 1000 ppm; ±120 ppm, for measured \(\textrm{CO}_2\,\)values from 1000 to 2000 ppm; ±250 ppm and for measured \(\textrm{CO}_2\,\)values from 2000 to 5000 ppm; ±300 ppm.

Implementation of the practical part 2 (objective 2)

The prediction model used in Part 2 was the NARX NN model with external (exogenous) input55. It can predict future values of a time series based on its past values and other supporting data that are gathered at the input of the network56. The AMR monitors the presence of occupants in the room indirectly (without using a camera) based on the prediction of \(\textrm{CO}_2\,\)concentration in advance \(\Delta\)t (Fig. 9c) using the NARX NN. The practical procedure for the implementation and learning of NN NARX along with the verification of the learned NN NARX with LMA, BRA, and SCGA on the test data is performed in the steps (Fig. 3). The prediction model used was the NARX NN model from MATLAB R2020b. Subsequently, the measured \(\textrm{CO}_2\,\)waveform values were divided into three sets, where 70% of the data is used to train the network, 15% for validation to better generalize the network and stop training when the generalization stops improving, and the remaining 15% for a test set to evaluate network performance during and after training. The calculated values (MSE and R) of NN NARX with LMA, BRA, and SCGA for the Number of Hidden Neurons 10 with the number of delays d2 - d10 are shown in (Table 4).

Table 4 Calculated values (MSE, R) of NARX NN models with LMA, for Number of Hidden Neurons 10 with number of delays d2 – d10.
Full size table
Table 5 New calculated values (MSE, R, MAPE) of NARX NN models with LMA for Number of Hidden Neurons = 10 with number of delays d = 2 and d = 457.
Full size table

Based on the results obtained by R, MSE (Table 4), the LMA learning algorithm (owing to the speed and precision of the NN calculation) was selected to predict the concentration waveform \(\textrm{CO}_2\,\)in advance of time using NARX NN (Table 4). Subsequently, an NN trained with the LMA is created, and the mean square error (MSE) method is used to measure the error. To achieve adequate results, the NARX NN model needs more input measured variables that provide better information in terms of the location of the monitored persons in the SHC.

From the measured values of the \(\textrm{CO}_2\,\)waveform in the time interval from 22 March to 26 March 2022, the presence of people and the occupancy of the monitored room 220 (living room) can be determined (Fig. 9a):

  1. I.

    \(t_1\) - 7:45 22.3.2022 arrival, \(t_2\) - 9:15 22.3.2022 departure, the occupancy time of living room 220 on 22.3.2022 was t = 1 hour 30 minutes; \(t_3\) - 12:00 22.3.2022 arrival, \(t_4\) - 16:50 22.3.2022 departure, the occupancy time of living room 220 on 22.3.2022 was t = 4 hours 50 minutes (total occupancy time of living room 220 on 22.3.2022 was t = 6 hours 20 minutes);

  2. II.

    \(t_5\) - 8:10 23.3.2022 arrival, \(t_6\) - 12:45 23.3.2022 departure, the total occupancy time of living room 220 on 23.3.2022 was t = 3 hours 35 minutes;

  3. III.

    \(t_7\) - 7:35 24.3.2022 arrival, \(t_8\) - 8:40 24.3.2022 departure, the occupancy time of living room 220 on 24.3.2022 was t = 1 hour 5 minutes; \(t_9\) - 9:50 24.3.2022 arrival, \(t_{10}\) - 12:00 24.3.2022 departure, the occupancy time of living room 220 on 24.3.2022 was t = 2 hours 10 minutes (total occupancy time of living room 220 on 24.3.2022 was t = 3 hours 15 minutes);

  4. IV.

    \(t_{11}\) - 7:00 25.3.2022 arrival, \(t_{12}\) - 12:30 25.3.2022 departure, the total occupancy time of living room 220 on 25.3.2022 was t = 5 hours 30 minutes;

  5. V.

    \(t_{13}\) - 7:20 26.3.2022 arrival.

From the predicted values of the \(\textrm{CO}_2\,\)concentration waveform with ahead of time using NARX NN with LMA (calculated values of \(\textrm{MSE} = 3.322\cdot 10^{-5}, \textrm{R} = 99.913\%, \textrm{MAPE} = 0.0565\) for number of hidden neurons 10 with number of delay d=2 with input values including T, rH, E, D1-D4), (Table 5), the presence of persons and the occupancy of the monitored rooms - living room 220 can be determined in the time interval 22.3.2022 - 26.3.2022 as follows (Fig. 9a):

  1. I.

    \(t_1\) - 7:40 22.3.2022 arrival, \(t_2\) - 9:05 22.3.2022 departure, the occupancy time of living room 220 on 22.3.2022 was t = 1 hour 25 minutes; \(t_3\) - 11:55 22.3.2022 arrival, \(t_4\) - 16:45 22.3.2022 departure, the occupancy time of living room 220 on 22.3.2022 was t = 4 hours 50 minutes (total occupancy time of living room 220 on 22.3.2022 was t = 6 hours 15 minutes);

  2. II.

    \(t_5\) - 8:05 23.3.2022 arrival, \(t_6\) - 12:40 23.3.2022 departure, the total occupancy time of living room 220 on 23.3.2022 was t = 3 hours 35 minutes;

  3. III.

    \(t_7\) - 7:30 24.3.2022 arrival, \(t_8\) - 8:30 24.3.2022 departure, the occupancy time of living room 220 on 24.3.2022 was t = 1 hour; \(t_9\) - 9:40 24.3.2022 arrival, \(t_{10}\) - 11:50 24.3.2022 departure, the occupancy time of living room 220 on 24.3.2022 was t = 2 hours 10 minutes (total occupancy time of living room 220 on 24.3.2022 was t = 3 hours 10 minutes);

  4. IV.

    \(t_{11}\) - 6:50 25.3.2022 arrival, \(t_{12}\) - 12:25 25.3.2022 departure, the total occupancy time of living room 220 on 25.3.2022 was t = 5 hours 35 minutes;

  5. V.

    \(t_{13}\) - 7:15 26.3.2022 arrival.

When comparing the precision of determining occupancy and occupancy time of the monitored living room 220 using the reference waveform of the concentration \(\textrm{CO}_2\,\)and the waveform of the concentration predicted \(\textrm{CO}_2\,\)before time using the NARX NN LMA, the results were as follows: “For the time interval I, the difference in occupancy time between the reference waveform \(\textrm{CO}_2\,\)and the waveform predicted \(\textrm{CO}_2\,\)was 5 min. For time interval II, the time difference of occupancy was 0 minutes. For time interval III, the time difference of occupancy was 5 minutes. For time interval IV, the time difference of occupancy was 5 minutes.

Fig. 9
figure 9

Display of a) measured reference and predicted \(\textrm{CO}_2\,\)concentration waveform using NN NARX with LMA (10 neurons, d = 2), Table 5); b) detail of measured reference and predicted \(\textrm{CO}_2\,\)concentration waveform using NN NARX with LMA (10 neurons, d = 2); c) larger detail of measured reference and predicted \(\textrm{CO}_2\,\)concentration waveform using NN NARX with LMA (10 neurons, d = 2); d) Bland-Altman plot comparing predicted and reference \(\textrm{CO}_2\,\)waveform.

Full size image

Part 3: KNX - AMR monitors room occupancy indirectly ahead of time by predicting \(\textrm{CO}_2\,\) concentration using NIO NN (objective 3)

In the selected living room 220 SHC (Fig. 4), the presence of people in the selected room is monitored indirectly (without the use of a camera) by means of KNX field measurements with the use of KNX technology, based on the prediction of the \(\textrm{CO}_2\,\)concentration waveform using the NIO NN model, and then the information of the predicted \(\textrm{CO}_2\,\)waveform is assumed to be transmitted to the AMR (indirect determination of the occupancy of the monitored rooms in the SHC) without the reference \(\textrm{CO}_2\,\)concentration waveform.

Implementation of the practical part 3 (objective 3)

The prediction model used in Part 3 was the NIO NN models. One of the key aspects of NIO NN is the ability to model nonlinear relationships between inputs and outputs. This nonlinear relationship is ensured through the use of nonlinear activation functions in hidden layers that transform linear combinations of input into complex patterns. The AMR monitors the presence of occupants in the room indirectly (without using a camera) based on the prediction of \(\textrm{CO}_2\,\)concentration in advance \(\Delta\)t (Fig. 10) using operating variables measured from the KNX technology (Fig. 8). The practical procedure for the implementation and learning of the NIO NN models along with the verification of the learned NIO NN models with LMA on the test data is performed in the steps (Fig. 3). The prediction model used was the NIO NN model from MATLAB R2020b. The calculated values (MSE and R) of the NIO NN models with LMA, for the Number of Hidden Neurons 10 and 20 with the delay number d2 - d100 for the input values W1-W3, D1-D4, rH, E, T are in Table 6.

Table 6 Calculated values (MSE, R) of NIO NN models with LMA for Number of Hidden Neurons 10 and 20 with number of delays d2 – d100 for input values W1-W3, D1-D4, rH, E, T.
Full size table

The calculated values (MSE and R) of the NIO NN models with LMA, for the Number of Hidden Neurons 10 and 20 with the delay number d2 - d100 for the input values rH and T are in Table 7.

Table 7 Calculated values (MSE, R) of NIO NN models with LMA for Number of Hidden Neurons 10 and 20 with number of delays d2 – d100 for input values rH, T.
Full size table
Fig. 10
figure 10

Display of a) measured reference and predicted \(\textrm{CO}_2\,\)concentration waveform using NIO NN with LMA for Number of Hidden Neurons 10 with number of delay d100 for input values rH, T (Tab. 7); b) detail of measured reference and predicted \(\textrm{CO}_2\,\)concentration waveform using NIO NN with LMA (10 neurons, d = 100).

Full size image

Discussion

This article follows the study of Vanus et al.57, where the ways of AMR interaction with people in a monitored SHC space are described in detail, using information about the location of monitored people, presence of people or occupancy of the SHC space by robot interaction with building automation using KNX technology: “I. Mode - robot invisible, II. Cleaning mode, III. Mode - be nearby, IV. Standby mode”. The presented article is mainly a study of the technological possibilities of secure communication of KNX technology with AMR using 5G network and its possible use in multiple SHC buildings in cities within the Smart City platform. The advantage of KNX technology is that this standard is widespread worldwide, it is a robust open bus system, it is easily extensible to new technologies within interoperability and it is also modifiable. KNX technology was introduced to global markets in 1990. AMR has the advantage of being the standard in industrial applications. A novelty is the use of AMR in human-machine interaction in the SHC home environment. For the secure operation of KNX technology, backup power supplies and internal and external lightning protection must be provided. In terms of operation, KNX is stable, reliable, robust and safe. The AMR system is battery powered. If the battery is discharged below a defined limit, the AMR is automatically navigated to a docking station where it is automatically recharged. In the event that the AMR cannot be recharged, it can be programmatically set to move to a safe area before it becomes discharged due to complete battery depletion. If the AMR battery is fully discharged, the user must then connect it to a charger to restore power. In the present study, the NARX NN prediction model with Levenberg-Marquardt learning algorithm was used to predict the \(\textrm{CO}_2\,\)waveform. It is a very fast learning model, which can achieve a high percentage of success in predicting the \(\textrm{CO}_2\,\)concentration trend (Table 5), (Fig. 9). The NARX NN prediction model allows the optimization steps to search for the ideal setting to solve a particular problem. Also due to their combinability, they are very time-consuming, because after each single setting, the network has to be re-learned and evaluated whether the change made yielded a better result or just the opposite. The best result in success rate of predicting the future \(\textrm{CO}_2\,\)concentration was for 10 input variables (T, rH, E, D1-D4, W1-W3) to NARX NN with LMA (Number of Hidden Neurons 10, number of delays d2, \(\textrm{MSE} = 3.322\cdot 10^{-5}, \textrm{R} = 99.913\%, \textrm{MAPE} = 0.0565\)) and for 7 input variables (T, rH, E, D1-D4, W1-W3) to NARX NN with LMA (Number of Hidden Neurons 10, number of delays d4), \(\textrm{MSE} = 2.984\cdot 10^{-5}, \textrm{R} = 99.922\%, \textrm{MAPE} = 0.0821\) (Table 5). Of course, such high success rate can only be achieved in certain cases and with proper optimization of the prediction model. NARX can predict the future values of a time series based on its past values and other supporting data that come together at the input of the network. Specifically, the technical indicators should provide the network with an even better prediction success rate because it has more data and the connections between them. This was reflected in its tested model series, where it provided more consistent results. However, the variety of settings and the number of optimization steps increases in this case, as it is possible to additionally optimize and modify the calculation of technical indicators. After the analyses of the two tested prediction models, the NARX NN model with LMA appears to be the best model for predicting the future \(\textrm{CO}_2\,\)concentration path to determine the presence of people in time-forward to detect deviations from daily routine of elderly residents in the SHC monitoring area. This is due to its greater variability, the larger amount of input data on which it can perform technical analyses, and more consistent results in prediction compared to the NAR NN model. When properly tailored and optimized for a specific problem, it can provide valuable information to SHC occupants and HVAC control engineers when making occupancy decisions for the SHC in monitored area.

The calculated values (MSE and R) of NIO NN with LMA, for the Number of Hidden Neurons 10 and 20 with the number of delays d2 - d100 for the input values W1-W3, D1-D4, rH, E, T are in Table 6. The best values of the NIO NN models were MSE = \(1.826 \cdot 10^{-3}\) and R = 95.29% for 20 hidden neurons and d = 90. For 10 hidden neurons and d = 100 were the best values, MSE = \(2.624 \cdot 10^{-3}\) and R = 93.74%. The model learning time was also an important criterion for selecting the best model (Fig. 10). The NIO NN with the LMA model for 10 hidden neurons had a shorter learning time than the models for 20 hidden neurons, where the learning time was on the order of tens of minutes (Table 6). The calculated values (MSE and R) of NIO NN with LMA, for the Number of Hidden Neurons 10 and 20 with the number of delays d2 - d100 for the input values rH and T, are in Table 7. The best values of the NIO NN model were MSE = \(2.366 \cdot 10^{-3}\) and R = 94.18% for 20 hidden neurons and d = 90. For 10 hidden neurons and d = 100 were the best values, MSE = \(3.028 \cdot 10^{-3}\) and R = 92.57%. The model learning time was also an important criterion for selecting the best model (Fig. 10). The NIO NN with the LMA model for 10 hidden neurons had a shorter learning time than the models for 20 hidden neurons, where the learning time was on the order of minutes (Table 7).

Benefits of using NARX NN and NIO NN models:

  • Highly customizable - These models can be adapted to different inputs and parameters, allowing flexible prediction.

  • Dynamic control - the ability to respond to changes in occupant behavior and external conditions ensures dynamic optimization of household functions.

  • Improving comfort and efficiency - Systems can improve the overall energy efficiency of the home while improving occupant comfort by adapting to their needs.

Disadvantages of using NARX NN and NIO NN models:

  • Need for quality data - The effectiveness of these models depends on the availability and quality of historical data and external input.

  • Complexity of implementation - These models can be challenging to set up and train correctly, especially if the system is complex or contains many variables.

In general, the NARX NN and NIO NN models can play a key role in providing a secure transfer of occupancy information of monitored spaces in SHCs in a time-ahead manner between building automation technology (KNX) and AMR. Another potential use in SHC is to improve automation, efficiency and comfort in SH, as they enable prediction and optimization of operational functions based on historical and current data.

Predicting the presence of people in time to detect deviations in the daily routine of elderly residents in the SHC can be very useful. If the prediction is done correctly, it can bring benefits, especially in the areas of comfort, energy efficiency, or safety. The advantages of advance presence prediction could be as follows:

  1. 1.

    Efficient energy management 34,35,36,37,38,39 - Preheating or cooling the room, optimizing lighting consumption, energy savings in combination with smart grid.

  2. 2.

    Increased comfort and personalization of automatic adaptation to the environment - From the point of view of usability testing of the proposed system, the autonomy of AMR and KNX technology systems for building automation and their mutual secure data communication for faster determination of AMR mode (I. Mode-Robot Invisible, II. Mode-Cleaning Mode, III. Mode-Be Nearby (Fig. 4), IV. Standby mode). The intuitive control and user-friendliness of KNX building automation technology is ensured by the use of a decentralized concept and proven by tens of hundreds of thousands of applications worldwide over the last 30 years.

  3. 3.

    Safety aspects and risk prevention - Predictive security measures: if the system anticipates that someone should be in a certain room and detects unusual patterns (e.g., sudden movement when no one should be there), it can trigger security measures. This can prevent burglaries or detect other security threats. Simulating presence while occupants are absent (if occupants are on vacation, the prediction can simulate movement on a normal schedule, which can deter would-be burglars). As part of security, it is necessary to protect the data transmission and IP layer from attack by an attacker through the communication layer or from physical removal and replacement by an unwanted element. KNX has defined two standards, AN 158 KNX Data Security and AN 159 KNX IP Secure. The former is defined in the framework of EN 50090-4-3 and ISO/IEC 14543-3. The latter is defined in the framework of EN 50090-4-4 and ISO/IEC 14543-3-4. KNX Data Secure ensures that, regardless of the type of transmission medium, selected messages sent from KNX devices can be both authenticated and encrypted. The safety functions of the mobile robot are managed using the use of the Sick safety PLC system, known for its excellence compared to other systems, for example, the industrial PC. This robust safety system guarantees a comprehensive level of safety during robot operation. The mobile robot is inherently capable of fully autonomous operation and is programmed to strictly adhere to safety protocols, rendering it incapable of performing any actions that violate established safety rules.57.

  4. 4.

    Providing comfort for elderly or special needs individuals58,59,60 - customization based on daily routines; for example, the system can raise the temperature and turn on lights for an elderly person or person with medical conditions before they enter a room, providing a comfortable and safe environment; Detection of deviations from routine: if a person repeatedly fails to enter a room at the usual time, the system can send an alert of a potential problem, which can be particularly useful when monitoring the health and well-being of vulnerable people.

Advanced technology is needed to make the prediction of the future in time to detect deviations from the daily routine of elderly residents reliable: “Machine learning and historical data analysis - prediction algorithms can analyze historical patterns of resident movement to predict where and when people will be present based on temporal and spatial patterns.; Sensor data - a combination of motion sensors, temperature sensors, rH sensors, \(\textrm{CO}_2\,\)sensors can provide enough information for accurate prediction.; Integration with personal assistants or smart calendars - SH can be integrated with calendars or personal assistants to provide even more accurate predictions based on planned activities.”

The challenges and limitations of the proposed method are as follows:

  1. 1.

    Prediction accuracy - Enhanced predictive accuracy is possible using algorithm optimization or filtration of predicted waveforms with adaptive algorithms or with wavelet transformation57.

  2. 2.

    Ethical and privacy issues - The use of cameras and biometric data for prediction may raise privacy concerns. A solution may be to anonymize the data and ensure that the presence information is not stored unnecessarily. Municipalities and cities are addressing social issues such as inclusion and safety58. Many commercially available home care systems provide remote monitoring functionality, but these systems require someone on the other end of the remote connection to pay attention59. Shak et al. found in a survey of 87 seniors aged 62 to 90 years that a large proportion of the target group used smartphones. Seniors with vision, hearing or mobility limitations appear to be more open to technical support. However, there are privacy concerns60. One possible solution to privacy concerns is to store the personal information of seniors on a local server, preventing the disclosure of their personal information and ensuring security59 . In our article, we describe a newly proposed method to predict occupancy in advance without the use of cameras.

  3. 3.

    Technology and cost requirements - Advanced sensors and machine learning systems can be expensive to install and maintain. The solution may be the implementation of standard, commonly used technologies and open, affordable systems, with an emphasis on ensuring interoperability for mutual data communication and exchange. KNX technologies for building operations management and AMR meet these requirements, including the possibility of providing interoperability with data processing capabilities using neural networks for predicting monitoring data (e.g. occupancy) within an IoT platform.

Conclusion

In this study, a novel and unique approach involving the use of AMR in secure communication with KNX technology using a 5G network within IoT was presented. AMR is used for indirectly locating people, monitoring their presence and detecting occupancy in time-forward to detect deviations from daily routine of elderly residents in SHC. This is achieved through an indirect methodology that effectively eliminates the dependency on cameras. In addition, AMR operational modes trigger follow-up actions (Mode - robot invisible, Cleaning mode, Mode - be nearby, Standby mode)57 based on measured operational and technical variables using KNX technology. The article also presented a newly proposed \(\textrm{CO}_2\,\)concentration prediction method for determining occupancy and presence of users in SHC monitored spaces with time advance using NARX neural networks models. Levenberg-Marquardt algorithm, Bayesian regularization algorithm and Scaled Conjugate Gradient algorithm were used as learning algorithms. After comparing the learning algorithms, the LMA learning algorithm with the best results \(\textrm{MSE} = 2.984 \cdot 10^{-5}, \textrm{R} = 99.922(\%), \textrm{MAPE} = 0.0821\) for NARX NN (Number of Hidden Neurons 10, number of delays d4) for 7 inputs to NARX NN including T, rH, E, D1-D4 was selected in terms of learning rate and prediction accuracy (Table 5). The AMR performs the following key functions, such as determining the location of a person, monitoring the presence of a person in the SHC (based on secure data communication with KNX technology and data processing using NARX NN). Experimental Part 1 (objective 1) described the possibilities of using AMR to determine the location of a person by an indirect method without the use of cameras in the context of connecting AMR with KNX technology and monitoring the opening and closing of windows (W1, W2, W3) or doors (D1, D2, D3, D4). In experimental Part 2 (objective 2), the AMR indirectly monitors the presence of people in a room with forward prediction of the determination of occupancy, without the use of cameras, using the measured values of the \(\textrm{CO}_2\,\)reference waveform to learn the NN model. This is achieved by predicting the \(\textrm{CO}_2\,\)concentration waveform in advance using the NARX NN and then feeding this information to the AMR in advance. This method provides an indirect way of detecting the occupancy in time-forward to detect deviations from daily routine of elderly residents in SHC within the SC using 5G networks. In experimental Part 3 (objective 3), the AMR indirectly monitors the occupancy of the room with a forward-looking view, without the use of cameras and without using a reference \(\textrm{CO}_2\,\)waveform to learn the NN model. This is achieved by predicting the \(\textrm{CO}_2\,\)concentration using a NN time series nonlinear input output model in advance and then transmitting this information to the AMR. This method provides an indirect way to detect the occupancy in time-forward to detect deviations from daily routine of elderly residents in SHC within SCs with 5G networks. The possibilities of using indirect occupancy monitoring with time advance without cameras using secure communication between KNX and AMR technology using 5G network were described. The presented results were achieved using MATLAB software tool with time-forward prediction of \(\textrm{CO}_2\,\)concentration waveform using NARX NN and NIO NN. This newly proposed method provides a possible indirect way of detecting the occupancy of monitored spaces in SHC with time-forward transmission of occupancy information to AMR. Predicting the presence of people in advance can bring significant benefits in a SHC, especially in terms of comfort, savings and safety. However, the usefulness of this approach is subject to the accuracy of the prediction, the available technologies and the protection of privacy. If designed correctly, this prediction brings a proactive and adaptive approach to home management, increasing comfort and saving operational costs.