Dynamic ventilation certificate for smart universities using artificial intelligence techniques

The issue of room ventilation has recently gained momentum due to the COVID-19 pandemic. Ventilation is in fact of particular relevance in educational environments. Smart University platforms, today widespread, are a good starting point to offer control services of different relevant indicators in universities. This study advances a Ventilation Quality Certificate (VQC) for Smart Universities. The certificate informs the university community of the ventilation status of its buildings and premises. It also supports senior management's decision-making, because it allows assessing preventive measures and actions taken. The VQC algorithm models the adequacy of classroom ventilation according to the number of persons present. The input used is the organisation's existing data relating to CO2 concentration and number of room occupants. AI techniques, specifically Artificial Neural Networks (ANN), were employed to determine the relationship between the different data sources included. A prototype of value-added services was developed for the Smart University platform of the University of Alicante, which allowed to implement the resulting models, together with the VQC. The prototype is currently being replicated in other universities. The case study allowed us to validate the VQC, demonstrating both its usefulness and the advantage of using pre-existing university services and resources.


Introduction
A distinctive feature of universities has always been their promotion of creativity, innovation, social and intellectual development and collaboration.They usually transfer these innovative developments to their environment, rather than apply them to themselves.However, the opportunity to develop complete, large-scale solutions for university campuses allows innovation to be systematised more globally.
In this line, IT platforms have been developed to support Smart University projects that are capable of monitoring, controlling and even managing the proper functioning of campus mechanisms, with a direct impact on the university community [1][2][3] .The quality of indoor air and classroom ventilation plays an important role in the wellbeing of the members of this community in particular and in educational environments in general.
Indoor Air Quality (IAQ) is a notable indicator allowing to measure the wellbeing of the occupants of an enclosed space [4][5][6][7] .This metric has gained even more relevance in the context of the COVID pandemic.Indeed, the outbreak has raised awareness of the need to be prepared for this sort of contingency and that temporary measures are wholly insufficient: not only are reliable solutions required but also fast and affordable ones, that can last over time.
The immediate effects of indoor pollution include allergic reactions, headaches, dizziness, and fatigue, while the effects of longterm exposure can lead to serious respiratory illness and heart disease [8] .Furthermore, aerosols accumulated in closed rooms are the main transmitters of viral diseases [5] .In this context, room ventilation has proven to be an efficient solution, both to reduce aerosols that can cause viral contagion, and to dissipate harmful gases [ 9 , 10 ].
The current maturity of Smart University platforms, and the speed at which they are evolving, call for the application of more specific and context-focused standards.Therefore, elements such as indoor air quality, and more specifically, classroom ventilation, is a necessary indicator to measure the wellbeing of both students and teachers [10][11][12] .
Based on the above, in the present work, we sought to enrich Smart University platforms with a dynamic quality certificate that measures classroom ventilation.The idea was to take advantage of the maturity of the technologies and the large volume of information that is already available thanks to the data collected daily by the institutions' monitoring platforms and extensive databases (DB).
The proposal is specified in the definition of a model that allows to generate a certificate that evaluates the quality of classroom and building ventilation (named Ventilation Quality Certificate −VQC).This model is effective, quick to implement and deploy, sustainable over time, inexpensive, low maintenance, based on artificial intelligence (AI) algorithms and dynamically updated.The certificate enriches the information obtained from the data included in current Smart University platforms metrics.The proposed VQC is based on CO 2 measurements and the institution's existing knowledge of its own environment: the number of people connected to the Wi-Fi network; the structure and organisation DB; specifications of hall and classroom volumes, etc. Thanks to the AI techniques applied, and taking advantage of the existing information, it is possible to evaluate both the number and the characteristics of the occupants in a given enclosed space as well as its ventilation efficiency.
The proposed certificate was implemented via a case study at the University of Alicante.The necessary services were incorporated into the university's Smart University platform architecture (called SmartUA ) [13] .Once the validation phase will be completed, the prototype will be transferred throughout the current year to ten public universities, part of an initial consortium to implement a shared, nationwide Smart University platform.
The rest of this paper is structured according to the steps of the study.Section 2 presents an overview of the state of the art.Based on the conclusions obtained, Section 3 explains the research methodology, analysing the input variables ( Section 3.1 ) and describing the proposed VQC model ( Section 3.2 ).In Section 4 , the main elements of the VQC implementation are detailed based on a case study.Section 5 offers a description of the results and their analysis, and Section 6 presents the main findings.

Background and related work
It has been demonstrated that particle and gas concentrations in a room can be harmful to its occupants [ 7 , 14 , 15 ].This accumulation is directly related to the number of persons present and the premises' ventilation conditions.
Carbon dioxide (CO 2 ) is a colourless and odourless gas that is formed in any process that involves the combustion of carboncontaining substances.It is considered to be a simple asphyxiant that basically acts by displacement of oxygen.In high concentrations, CO 2 can cause discomfort such as headaches, breathing problems or dizziness.In non-industrial indoor environments, the main sources of CO 2 are human respiration and tobacco smoke [16] .
Several studies show that in the absence of any other polluting sources (other than people), CO 2 concentration in indoor environments can be used as a marker of the environment's odoriferous load due to human presence [17] .However, it is important to note that human respiration CO 2 emissions are related to that of other human metabolism products (water, biological aerosols, particles, alcohols, aldehydes, etc.) called bioeffluents.
According to other works [18] , studying a ventilation system's operation basically involves measuring ventilation flow.This data can be obtained by directly measuring airflow, but it is often easier to make estimates by measuring CO 2 levels.Therefore, CO 2 concentration can be considered an interesting value both to control CO 2 presence and to study ventilation [ 16 , 17 ].
Various studies [19][20][21] have focused on finding suitable methods to obtain CO 2-based occupancy control, in order to estimate the number of occupants present in a room.Most of them take into account the amount of CO 2 generated by a person.A person's CO 2 emission has in fact been defined in different studies [22][23][24][25] .A healthy person, on a normal diet, generates an amount of CO 2 that can be expressed as ( Eq. ( 1) ): Where: RQ: respiratory rate BMR: basal metabolic rate in Megajoules(MJ)/day M: metabolic activity expressed in MET MET: the approximate amount of oxygen consumed per minute, by a person in a resting state As can be observed ( Eq. ( 1) ), this measure is highly conditioned by various factors: on the one hand, the amount of energy that the body needs to perform its basic vital functions, called Basal Metabolic Rate ( BMR ), which is dependent on age and sex; on the other, the Respiratory Rate ( RQ ) which some authors set at 0.85, while others, following the recommendations of [ 24 , 26 ], use 0.83.Metabolic activity ( M ) also influences this indicator.Some studies have collected different measures of this metric according to the work or action being carried out [27] .In addition, different metrics were defined to determine this value in school-age children, whose weight and height are also factors to consider.Other results, such as those obtained in [25] , show that these values may differ according to the region.In addition, [22] presents experiments under different conditions with the aim of obtaining a measurement of the amount of CO 2 based on experimental measurements.
All these investigations show the interest in the matter and allow us to conclude that CO 2 monitoring is a suitable measure both to control occupancy and ventilation.The different mechanisms proposed to measure ventilation, air quality and CO 2 concentration have clearly shown the existence of multiple factors of influence.Yet no consensus as to how these factors interrelate has been reached, so a black box model could be a good solution in this field.
Artificial Neural Networks (ANNs) have proven to be powerful and robust tools regarding both classifications and the calculation of regressions [28][29][30][31] .ANNs have been widely used over the years in multiple contexts for both linear and nonlinear relationships between input and output variables.Many previous studies argue that the tendency to use ANNs owes to their embedded strengths, their rapid processing prediction abilities [ 28 , 32 ], as well as their capacity to determine the relationships between different variables in the absence of any initial assumption or postulate.In addition to their more general advantages, several works have used these types of techniques in the field of occupancy control [ 29 , 30 , 33 ] and indoor air quality [ 31 , 34 , 35 ].
The control of CO 2 presence and concentrations, as well as ventilation, is of particular interest in educational environments [36][37][38] , where presential attendance is usually required and prolonged.Establishing measures to control these indicators can have a direct impact on the quality of life of teachers and students.And beyond this, the goal is to provide mechanisms that synthesise the information on the effectiveness of the actions taken.Such mechanisms, such as a classroom ventilation certificate, could support decisionmaking, by helping to determine the actions that are the most effective at meeting the set objectives.
Smart City platforms have become a technological benchmark [ 39 , 40 ] in which numerous indicators have been implemented in relation to cities [ 41 , 42 ].Moreover, there has been great progress in the more specific field of Smart Universities [ 2 , 3 , 43 , 44 ].Yet, despite the current maturity of these platforms, no concrete indicators, centring on the specific problems of the university environment, have hitherto been advanced.
The overview of the state of the art allowed us to conclude the following: that classroom ventilation in universities is a basic requirement to guarantee the community's health and wellbeing; that it is necessary to define metrics adapted to academic environments, specifically, indicators for Smart Universities that allow monitoring the measures adopted towards this end; that a good ventilation measure is CO 2 concentration, which is also influenced by various factors such as the number of occupants coexisting in an enclosed space, and more specific characteristics such as the sex and age of those involved; and that there is no widely agreed definition of the relationship and role of the variables involved in the process.The study also shows that ANNs can be a very effective and efficient tool to solve this type of problem.

Research methodology
The present study proposes a Ventilation Quality Certificate ( VQC ).This indicator for Smart Universities is designed to support decision-making and allows monitoring the measures adopted by the governing and management bodies.The VQC is based on combining information on the number of occupants and CO 2 concentration in an enclosed area, in order to obtain a measurement of its ventilation.This information is offered to the community and senior management in the form of a classroom ventilation certificate and a quality seal that is also updated daily according to behaviour over time.
Obtaining the number of occupants of a given space is feasible and viable using technological tools.The task is also facilitated by the existence of a running Smart University platform.By deploying a sensor network that incorporates inexpensive, effective, and easy to implement probes, a room's CO 2 concentration can be measured easily.
On the basis of the above, our hypothesis is that a Ventilation Quality Certificate, based on both CO 2 concentration and room occupancy, constitutes a valid tool to evaluate the effectiveness of the measures adopted to ensure adequate room ventilation, depending on the number of occupants.The analysis and control of this VQC allows, on the one hand, to offer the community dynamic and updated information on the classrooms' ventilation situation, and on the other, to obtain valuable information to measure the results.Therefore, it allows to analyse the effects of a certain strategy and to thus guarantee that the university's senior management makes effective and timely decisions, which is the ultimate goal.
In short, this VQC allows to evaluate and manage ventilation, and therefore provides the Smart University with the ability to provide safe premises.It is thus perfectly aligned with the objectives of these types of platforms.

Selection of input characteristics
To define the VQC, we took classroom occupancy and CO 2 concentration data into account.
As a first measure of occupancy, the proposed solution was based on the hypothesis that people are increasingly connected to the network, especially in university environments, where it is common for there to be at least one connected device per person.In this way, the number of people will be controlled by verifying the number of Wi-Fi connections to the room's access points [45][46][47] .
However, the number of Access Point (AP) connections in a classroom does not always correspond exactly to the number of people in the classroom.First of all, there are basal connections, due to devices that are constantly connected to the AP, regardless of the people who are in the classroom-e.g., desktop computers, printers, sensors or smart devices present in the classroom.Secondly, occupants can individually account for more than one connection to the AP because they may be connecting multiple devices at the same time: laptop, mobile phone, tablet, smartwatch, etc.
Nor is it a trivial matter to determine which parameters to include in the estimation of the number of people based on Wi-Fi connections.To do this, information was collected, categorised and parameterised, in order to establish and weigh the different variables that can influence the relationship between the number of Wi-Fi connections and the number of occupants who coexist in a room.These variables were contrasted using the qualitative Delphi technique [48] , specifically through a normative Delphi study [49] through which the values, preferences or beliefs of a group of experts were checked to select the most significant variables for these purposes [50] .
Based on the previous study, a detailed classification, specification, and weighting of the analysed variables was performed.The selected variables were defined using the results of 150 questionnaires and 50 individual interviews with experts (in charge of different university management departments: quality, health and occupational risks, space management, academic management, and IT services).Behaviour patterns -including the factors that could influence the number of connections to the network -were modelled based on these variables.As a result, the following parameters were found to influence the fluctuation of the estimation of the number of people depending on the connections to the AP: time, type of teaching and type of university subject.
On the other hand, as with Wi-Fi connections, the direct CO concentration data already generated value for the Smart University platform.Different studies have established the limits of CO levels that can be harmful to humans in indoor environments [ 1 , 12 , 22 ].It is therefore highly relevant to have sensors capable of measuring the concentrations of this gas in order to design alert mechanisms [8] .
Regardless of its basic benefits and based on prior studies [ 19 , 20 , 34 ], the number of occupants of given enclosed premises can also be predicted via CO 2 readings, along with other related variables, such as the room's volume or the age of the people present.
The parameters involved in this second occupancy measure, and which therefore become the model's input variables, were selected based on a previous work dedicated to this question [ 4 , 16 , 22 ].The ease of obtaining data from the organisation's DB without violating data protection rights was also taken into account.Thus, added to the CO 2 concentration measured by sensors specifically designed for this purpose, was the classrooms' surface area and height.Indeed, a room's dimensions are a determining factor, as previously observed [ 5 , 22 ].Data such as physical form, diet, sex and age, have also been identified in previous studies [ 16 , 22 , 25 ] and can be indirectly contained in some variables that are easy to consult in the institution's databases such as: type of subject, season of the year, type of degree and academic year.
A total of 10 input characteristics were identified based on: the information collected from experts on the analysis of Wi-Fi connections; prior research; and the influence of several variables on the generation of CO 2, and, therefore, its concentration.Table shows the analysis of the inputs and the discretisation or normalisation of nominal data where necessary.

Developing the ventilation quality certificate (VQC)
To measure classroom ventilation, a classifier with supervised training was proposed.It is capable of using the input variable values and classifying the ventilation situation as adequate or inadequate .This certificate is constantly updated, dynamically, based on the measurements.It is available for both individual classrooms and for the building in general in the form of a quality seal.
We sought to obtain a reliable model, that would also be lightweight, and which could be embedded into an IoT device with limited resources and be supported by several previous studies [ 28 , 30 , 32 , 33 ].Artificial Neural Networks were selected as a classification technique for the proposed certificate model.The VQC could be dynamically calculated after having measured the sensors, but a daily update of the classroom certificate was proposed.Each classroom would be assigned the certificate and the seal obtained from the measurements of the previous day.In this way, both teachers and students could have an idea of how their classroom / building complied with the ventilation measures.Regardless of whether the certificate is displayed for a complete day, the sensors and the classifier process data every 15 min, and this data and the alerts generated by the system can be consulted by those responsible for the classrooms or buildings, allowing them to take the appropriate measures and to improve ventilation if necessary.
Thus, to calculate the ventilation certificate, a neuronal classifier described as the g () function converts the inputs identified in Table 1 into a value of 0 or 1, as appropriate, according to the inadequate ventilation or adequate ventilation, respectively ( Eq. ( 2) ).
This function evaluates ventilation conditions according to the input data every 15 min, and the ventilation quality certificate uses those measurements to classify the classroom conditions as favourable, average , or unfavourable , depending on the case.The measurements are taken during what are considered to be active presential teaching hours at the university, and the VQC uses all these measurements to give a final assessment of the classroom's ventilation ( Eq. ( 3) ): Threshold fav , Threshold med : values that are defined according to the classroom's set objectives, identifying the average of favourable measurements to consider the classroom's conditions as favourable The f v () function uses the classifications generated by the ANN every 15 min implemented in g (), and determines the average that describes the behaviour within the window of hours identified as teaching times ( Eq. ( 4) ) Where: h i: measurement start time hf: measurement end time g i (): value obtained by the function g() at the time of measure-ment i. quantMeasures: number of measurements made between the start (h i ) and end (h f ) of estimated presential teaching times quantMeasures = ( h f − − hi ) *  4   In addition, the Building Ventilation Quality Certificate ( VQCb ) behaves in a similar manner to the VQC , since the possible results of Favourable, Average and Unfavourable are refreshed every day.The VQC b is calculated using the ratio between the number of classrooms with good ventilation (i.e., with favourable VQC ) and the number of occupied classrooms in the building during presential teaching hours ( Eq. ( 5) ) Where: f vB (): calculates the number of classrooms in the building with C V = Favourable CAO: number of classrooms in use Threshold favB , Threshold medB : values that are defined according to the objectives established for the building, identifying the necessary proportion of favourable classrooms to consider the building's conditions as favourable To select the ANN capable of classifying the classroom ventilation situation according to the sensor measurements ( g ()), we used a set of 60,0 0 0 input records that had been previously selected as representative of the widest range of possibilities.The dataset was normalised using each set's mean and standard deviation, which allowed to obtain results with much fewer errors than those obtained with the raw data.The dataset was randomly divided into an 80:20 ratio to generate the training and test sets.
Once the dataset's data was prepared, relu () was used as an activation function of the intermediate layers to generate the ANN model.This decision was based both on the learning speed and on the fact that many libraries are optimised for use, which facilitates using the ANN in an embedded IoT device.For the output layer, the classification being binary, the best results were obtained using sigmoid() as an activation function, and binary_crossentropy as a measure of loss compared to the results obtained with softmax() and sparse_categorical_crossentropy, respectively.
To select the optimal ANN for the problem, we trained the different network architectures and measured the accuracy and loss results related to both training and validation.In addition, we persistently selected the lightest network that met the accuracy objectives the whole time, as this would facilitate using it within an IoT device.    1 ); two completely connected hidden layers of 100 and 70 neurons, respectively; and the output layer with a single output modelled by the sigmoid() function and therefore with values between 0 and 1.
The selected model presented accuracy values very close to 1, with good training and validation results.Fig. 2 shows the ANN training graphs for g () models for both accuracy and loss.
As shown in the graph in Fig. 2 (a), the accuracy moves closer to 1 (100%) as we advance in the training and validation periods (xaxis): the accuracy increases at each training stage until reaching 98.9% accuracy after epoch 50.The accuracy metric computes the frequency of the matches between the target value and the output value.This frequency is ultimately returned as binary accuracy: an operation that divides total by count.
Moreover, the loss is calculated using the binary_crossentropy function.Binary cross entropy compares each predicted probability in order to target the class output which can be either 0 or 1.It then calculates the score based on logarithms that penalises the probabilities based on the distance with the expected value.The ideal value is 0, since 0 would describe that the results obtained are always exactly as expected.As can be observed in graph b, as training times increase (x-axis), the loss (y axis) decreases with a clear tendency towards 0 that culminates with a loss of 0.122 for training and 0.135 for validation.The similar behaviour in both sets (training and validation) indicates that an overfitting problem occurred during training.

Case study
The ANN used to certify the classroom ventilation conditions ( VQC ) were implemented through the open-source library Tensor-   Flow for machine learning, and the resulting model was deployed on the Smart University platform of the University of Alicante (SmartUA)-that has been running since 2014 [13] .Within this platform, over the years, value-added services have been developed with a great impact on the wellbeing of the university community.
A point of maturity has been reached that makes it almost imperative to add indicators that allow evaluating the efficiency of the decisions taken and the measures adopted.The proposed VQC is of great value to SmartUA, especially in the light of the new social situation caused by the pandemic.This task of implementing the VQC was facilitated by the scalable model on which the SmartUA architecture is based.Following the platform's service delivery methodology [51] , we identified the system's main components and their relationships.The most basic components consist of both the network and CO 2 sensors that are being deployed in classrooms and enclosed spaces, forming a network of sensors connected by a LoRa network.These sensors transmit the collected information to the Smart University platform through a Service Oriented Architecture (SOA) based on the M2M (Machine-to-Machine) communication protocol of the MQTT message queue type.Three value-added services were created using this information.The first provides congestion information using the Wi-Fi connections; the second gives CO 2 concentration information; and the third, specific information from the VQC .These services offer general information that can be consulted by university community members, but they also offer information that support the decision-making of the university's senior management and management teams.Thanks to the new indicators, the governance teams will be able to determine the most appropriate strategy in each situation, and the technical and management teams will be able to adopt the measures that maximise performance.In this way, the rules and measures can be tightened or relaxed as the case may be, and therefore, directly influence classroom conditions.This creates a continuous improvement cycle ( Fig. 3 ) Based on the implementation proposal, three main processes were identified to develop the three proposed value-added services: a Wi-Fi connection control service, a CO 2 concentration control service, and the ventilation control VQC service.
To deploy the value-added services, the components were separated into the different layers defined in the architecture and that support the platform's service provision model [51] .This was done ensuring that the information displayed in the presentation layer complied with the current anonymisation, security and data protection regulations ( Fig. 4 ).The monitoring layer contains the CO 2 and network sensors, as well as the generated traffic files.This information is obtained by the business layer through scripting components that are located in the ingest sublayer to be stored in the databases located in the storage layer.The processing layer contains the different software components responsible for using the data retrieved from the Data Access Objects, and the ANN model responsible for evaluating the ventilation situation.These duly processed data are offered as a service and used by the components of the presentation layer to be shown to users.
Taking advantage of the benefits of SmartUA's architecture, which allows different services to reuse components, the VQC service does not add components to the monitoring layers or to the ingest and storage sublayers.Instead, it uses the components deployed for Wi-Fi and CO 2 services.
To implement the Wi-Fi service, the sensors already run by the university were employed to give students and teachers access to the Wi-Fi network.To design CO 2 sensors, a study was performed on the needs and solutions of both CO 2 sensors and controllers currently offered on the market.Based on the study, several proofs of concept were conducted with different devices.The Senseair S8 LP sensor and the LILYGO® TTGO T-Beam controller were used for the final testing ( Fig. 5 ).
The CO 2 detection sensor was located in the centre of the room both in the deployment and test phases, in accordance with [52] , who recommended that it be kept away from any heat source, such as radiators or walls.Moreover, given that research has shown that sensor measurements vary little according to different positions in the room [53] , measurements made in the central position were considered to be representative of the room's CO 2 concentration.
Once the components were identified and the CO 2 and Wi-Fi services implemented, they were deployed on the SmartUA platform for use by the university community.Since the service corresponding to the VQC , was in an experimental phase, it was deployed in a series of classrooms that had been especially selected for the tests.
To implement and train the ANN used to classify the classroom situations, function g (), TensorFlow was used for the facilities it offers and for the need to minimise time to market .The ANN model, once trained, was exported and added to the SmartUA architecture.The code developed for training and testing is available in the public repository [54] .

Results and discussion
In Sections 3.2 , the ANN model conferring the adequate ventilation certificate to both the classrooms and the campus buildings was trained and validated.This section focuses on the validation of the VQC during the Smart University platform case study, where it was implemented as a series of value-added services.
All three services were implemented and deployed within the platform.The Wi-Fi connection service was offered to the community in the form of heat maps that represent the Wi-Fi network's geographical density and state of use-at any time or historically ( Fig. 6 ).
As described in the proposal, the raw values obtained from CO 2 sensors allowed to obtain air quality information based on the concentrations of this gas.This service offers the possibility of consulting classroom CO 2 concentrations.It also presents graphs that show the evolution of this concentration ( Fig. 7 ).
To measure the effectiveness of the proposed VQC , tests were performed based on data designed and collected for these specific purposes.Two university classrooms were used and laboratory measurements were taken with different amounts of occupants under various ventilation conditions.Furthermore, to apply Eqs.T hreshol d medB := 0 , 55 → if 55% of the classrooms are considered f a v ourable, the building is considered a v erage i f it does not r each T hr eshold f a v B (6) To perform the validation, a dataset of 10 0 0 records taken from the two test rooms at different times and during different classes were used.A total of 10% of the data corresponded to problematic ventilation situations.The effectiveness of the certificate achieved with the test data attained 97.8%, with 0.7% of false positives and 1.5% of false negatives, as can be observed in the scatter plot of expected data and data estimated by the application ( Fig. 8 ).
After the experimental results, the service was deployed on the SmartUA platform and integrated into the dashboard used by the university's prevention services and senior management.
As illustrated by the colour patterns in Fig. 9 , the temporary results of the evaluation of the conditions can be analysed by the service every 15 min to enable taking any relevant measures.This figure illustrates how classrooms 2 and 3 always remained within acceptable parameters (green), while classroom 1 presented poor ventilation dynamics (Yellow) and even highly deficient ventilation at times (red).
In addition, the community can consult the information visually through an interactive campus map that displays the VQCs, superimposing the buildings' Quality Seals ( Fig. 10 ).
The services were very well received by the university community, as demonstrated by the considerable rise in the tool's usage statistics since the services were implemented, compared to data of the years prior to its application ( Fig. 11 ).The year 2020 was left out of the graph because the data were not representative, due to the COVID lockdown.
Moreover, satisfaction surveys were conducted with the senior management to uncover their perceptions of the advantages of the implemented services.A summary of the indicators and their evaluation by the respondents is shown in Table 2 .
The university's prevention services controlled agglomerations more closely during the pandemic, especially indoors, thanks to the alarms generated by the platform.
Senior management notably used the prototype to measurebased on air quality -how suitable classrooms were for their occupants over extended periods of time.This analysis led to adjusting classroom assignments in order to find optimal attendance solutions according to the different classes.Although the classrooms had already been assigned using class enrolment data, the tool allowed to refine the solution based on real data and measurements.
Finally, the computer service took advantage of the prototype as a measure to evaluate the distribution of WIFI access points in the university, causing its restructuring and expansion in different areas.
The results obtained by the VQC model in terms of effectiveness, together with the usage statistics and the satisfaction surveys, validated the initial hypothesis.According to this starting hypothesis, the certificate was proposed as a tool to evaluate the effectiveness of the measures applied to guarantee adequate classroom ventilation.

Conclusions
Air quality, and more specifically, the control of classroom ventilation quality is a focus of interest in university environments.The technological maturity of Smart University platforms reveals the potential of such platforms to create university-centred indicators of classroom and building conditions.
Having identified the problem as well as the opportunity presented by Smart University IT platforms, we proposed the elaboration of a Ventilation Quality Certificate ( VQC ) capable of evaluating classroom and campus building ventilation conditions according to number of occupants.To this end, AI algorithms were advanced on which the classroom certificates as well as the building certificates are based, both relying on an ANN-implemented classifier.The classifier input was defined using the data sources relating to both CO 2 concentration and a classroom's number of occupants.We applied: connection data obtained through the Wi-Fi access points; CO 2 concentrations measured by smart sensors; general data, such as the time and season of the year; classroom data, e.g., surface area and height; and specific data of the occupants extracted from the organisation's databases, for example, type of university subject, type of degree, academic year and type of teaching.
Neural networks were found to represent a mature and essential instrument to develop VQC.Indeed, in addition to their demonstrated high effectiveness as classifiers, they facilitate development tools that are optimised for IoT devices, with processing speeds that allow using them in real time.
The proposal was implemented via the application of the VQC on the University of Alicante's Smart University platform (Smar-tUA).To this end, three value-added services were designed, and their components were incorporated into the platform's architecture thanks to the SmartUA's service provision model.The implementation of these three services led to an additional security layer that helped to implement the appropriate control measures of agglomerations, indoor air quality and ventilation.
The tests performed both on the neuronal classifier and the ventilation certificate service generally confirmed the effectiveness of the proposed models.The results obtained in the experimentation phase were supported by the statistics of use of the services offered and by the results of the surveys conducted with the university's top management.
The main lesson learned is that senior management's commitment is essential to tackle projects of this magnitude.Only by being perfectly aligned with the university's strategy is it possible to implement projects that involve different organisational units, diverse interests, and the necessary collaboration of researchers and technicians from various fields.
In the short term, we plan to combine Wi-Fi connection information with CO 2 concentration measurements to obtain more accurate results regarding the number of occupants of a closed room.In the medium and long term, this information will be used to manage incidents, not only related to ventilation, but also to water and electricity consumption.We also envisage adding occupancy information through obfuscated images and vision techniques.The longer-term objective is to eventually enrich the Smart University platform with new KPIs specific to the university environment that can be used as governance and management tools.

Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Fig. 1 (
Fig.1 (a)shows the main architectures assessed and Fig.1(b) details the selected architecture composed of: an input layer with 10 neurons (inputs identified in Table1); two completely connected hidden layers of 100 and 70 neurons, respectively; and the output layer with a single output modelled by the sigmoid() function and therefore with values between 0 and 1.The selected model presented accuracy values very close to 1, with good training and validation results.Fig.2shows the ANN training graphs for g () models for both accuracy and loss.As shown in the graph in Fig.2(a), the accuracy moves closer to 1 (100%) as we advance in the training and validation periods (xaxis): the accuracy increases at each training stage until reaching 98.9% accuracy after epoch 50.The accuracy metric computes the frequency of the matches between the target value and the output value.This frequency is ultimately returned as binary accuracy: an operation that divides total by count.

Fig. 4 .
Fig. 4. General Framework of the IT Architecture.

Fig. 6 .
Fig. 6.Heat map showing the concentration of people in the UA based on Wi-Fi consumption.

Fig. 8 .
Fig. 8. Graphs of expected data and data estimated by the application during testing.

Fig. 9 .
Fig. 9. Measurements allowing the C v service to assess the ventilation status.

Table 1
Input variables.

Table 2
Indicators identified by the university's senior management.