Data mining methods of healthy indoor climate coefficients for comfortable well-being

The German government has set itself ambitious climate protection targets: By 2020, greenhouse gas emissions have to fall 40% below the level at the state of 1990. Therefore, the building sector is aimed by this project as it is one of the large CO₂-producing sectors. The described work is performed with our partners from the renovation management of Bielefeld Sennestadt. The inhabitants live in pre-renovation buildings and they produce more carbon dioxide as normal, whilst losing heating energy caused by poor building state. After renovation and isolation, buildings lose their energy by wrong heating and ventilation habits.

Machine learning algorithms will help to change the heating and ventilation habits of residents by telling them recommendations to get better room climate and to aim lower CO₂ emission. We created machine learning algorithms that are basically learned by healthy air quality information according to common DIN standards [DIN EN ISO 7730; ASR… 2012; Detlef et al. 2013; DIN EN 13779] and further extended learned by personal feelings about air quality inside the apartments. Therefore, an algorithm AI has been developed, which is not only able to decide air quality according to German DIN standards AI_DIN, which tells that the indoor climate is healthy or not, but also further with AI_extended if the occupant feels comfortable or not.

These algorithms could prevent unnecessary CO₂ emissions by telling recommendations to residents and influence their heating and ventilation habits in general. Also, they learn a correct and more energy-efficient ventilation and heating behaviour to reduce the CO₂ emission after a renovation process.

The Smart Monitoring system provides an interactive user interface inside the apartments. Residents can view information about actual measured data of the air quality in their apartments and also some statistical evaluations. The developed intelligent algorithms provide a prediction of how an occupant is normally feeling by interpreting the actual sensor data. It includes a preliminary processing of measured datasets, their physical relations, statistical analysis, data reduction, targeted learning of German DIN standards and the extended learning of individual indoor air climate feeling. The project is supported by ministry of North Rhine Westphalia (2016–2020) under Nr. 322-FH Struktur.

The measuring system was developed for the purpose of promoting energy awareness amongst residents and to examine the ageing residential buildings before and after their renovation. It also shows the collected data to occupant’s and provide collecting, managing and transferring Big Data by intelligent algorithms to support scientific work on collected data. Within an interdisciplinary team, a cost-effective measuring system was developed. In addition to expandability, scalability and maintainability, Smart Monitoring also provides secure data communication of measured sensors with strongly cryptographic algorithms.

As seen in Figure 1, the measuring system in apartments consists of a set of sensor computers (Raspberry Pi’s). Each of them collects measurement data of sensors and actuators and sends them by POST request to a RESTful Webservice interface, which is controlled by main server (Figure 1, left side). The main server handles any type of data and saves it into a local database. The amount of sensor data and standalone computers is scalable. It is possible to have an unlimited amount of measuring systems that send encrypted data from each main server of an apartment to the main ‘FH-Server’ (Figure 1, right side). We have a 1:n relationship between a data receiving server and a scalable amount of data serving sensor computers in an apartment. Also, we could have a scalable amount of data collected in apartments, which provide sending encrypted data to the FH-Server by calling a special web service.

Figure 1

The FH-Server hosts a web application for data collection, analysis and visualisation. The focus of backend application was an efficient communication and storage of all data types. The handling of large datasets and data analysis algorithms allows us to support big data amounts in our system. All developed backend applications use standard open source technologies of Java EE, such as EJB, JNDI, JPA, JAXB and JAX-RS. JAX-RS are used with RESTful web service interfaces to manage the communication. In the JSON-based client server communication via Web service for high security level, a cryptography support was implemented [Behrens et al. 2017].

To measure datasets in different apartments, a standalone measuring system based on Raspberry Pi with Debian Linux is used (figure 2). We use sensors to measure datasets of the relativity humidity RH_in, pressure p, air temperature T_air, inside temperature of the outer wall T_wall, inside temperature of the heating T_heat, carbon dioxide level CO₂ and window ventilation states ST_window, and A weather station measures the outside relativity humidity RH_out, outside temperature T_out and outside wind speed W_out, as shown in Figure 2 [Behrens et al. 2017; Sonntag… 1990].

Figure 2

To measure the personal room climate feelings, a visual display unit (figure 3) is located inside monitored apartment room. Residents reported their air climate feelings AQF by pressing one of three feedback buttons with smiley icons – green, yellow and red. Feeling states good AQF_good, neutral AQF_neutral and bad AQF_bad are generated, which are used for supervised machine learning algorithms.

Figure 3

After pressing a Feedback button as input data, the user gets a view of actual measured air quality data. For training AI_DIN, we interpreted indoor air quality based on a given DIN standard as summarised in Table 1.

The German DIN standard [DIN EN ISO 7730; ASR… 2012; Detlef et al. 2013; DIN EN 13779] gives information about how the actual measurement data in a room should be (Table 1) to have a healthy room climate by interpreted categorisations ranges: AQF_good, AQF_neutral and AQF_bad. DIN standard deals with the sensors RH_in [ASR… 2012], T_air [Detlef et al. 2013], CO₂ [ASR… 2012; DIN EN 13779] and T_wall [DIN EN ISO 7730]. Interpreting air quality state and room climate feeling as good, neutral or bad was made by sensor data fusion, getting multidimensional input feature vector to supervised learning algorithms AI_DIN and AI_extended. Sensor data fusion means that all records made by the measuring system are merged by matching the timestamp. The AI_DIN algorithm is mainly used to train the neural network on almost all given datasets collected by sensors and then it serves as a base to learn the AI_extended algorithm by extending itself. Finally, AI_extended is based on the already learned AI_DIN as a base.

This method allows the algorithm to learn air quality (AI_DIN) from all measured datasets (high frequency by sensors) and to further correct itself as a new branch by learning air feelings (AI_extended) of inhabitants (low frequency by input).

Supervised machine learning comes with a data mining process described in this section. We collected data from apartments and laboratory. After collecting datasets, we started pre-processing by filtering them within matching timestamps, cleaning up faulty data and normalising them (Figure 4).

Figure 4

After pre-processing, we made data reduction and data breakdown to get data according to heating season (winter) and the remaining data (summer). The measuring system collected datasets at a rate of every minute in the winter period and at a rate of every half-minute in the summer period. The ranges are given in Table 2. Overall, we have used 129,047 merged records in the winter period (approximately September to March) and 138,914 merged records in the summer period (approximately April to June) to train the algorithms described next.

Basically, the overall supervised machine learning AI_DIN and AI_extended was made by using classified and randomly picked 70% of training data and 30% of testing data to create the AIs that are based on neural network approaches. The neural network used to train the algorithms was developed within this project and the chosen topology contained five layers. Between the first layer (an input layer) and the last layer (an output layer), we used three hidden layers, of which the first holds 8 neurons, the second holds 16 neurons and the third holds 8 neurons.

Input feature vector (ST_window, RH_in, p, T_air, W_out, T_out, RH_out, CO₂, T_heat and T_wall) and target output vector (AQF_good, AQF_neutral and AQF_bad) were used, and a correlation matrix for each dataset is presented in Figures 5 and 6. The correlation matrix provides information about the relationship of sensor data to each other. The darker connection fields are more correlated than the lighter connection fields.

Figure 5

Figure 6

To differentiate the learned AIs, we named them by alias. AI_DIN-winter learned DIN standard at heating season, AI_{extended-winter} learned resident feelings at heating season, AI_DIN-summer learned DIN standard at not heating season and AI_{extended-summer} learned resident feeling at not heating season.

The first step was to create an intelligent AI_DIN that learns the DIN standard air climate quality classes. The second step was to use the already learned AI_DIN to create a second AI_extended that was additionally based on the interpretation of personal room climate feeling states (AQF_good, AQF_neutral and AQF_bad). In the conclusion, we got two kinds of AIs, AI_DIN and AI_extended.

AI_DIN represents the actual state of the air quality inside an apartment for healthy purpose and AI_extended the actual state of personal feeling of an apartment resident. The classification experiments show the suitability of the learned algorithms for their determination (Figures 7 and 8).

Figure 7

Figure 8

The classification experiment 1 (winter) shows the algorithm AI_DIN-winter, which has learned the air quality conditions, and it made 126614 correct interpretations towards 129047 given datasets.

The classification experiment 2 (winter) shows the extended algorithm AI_{extended-winter}, which has learned the air feeling of residents, and it made 1,265 correct interpretations towards 1,274 given datasets.

Similar classification experiments (Figures 9 and 10) were performed on the not heating season (summer) by learning the AI_DIN-summer and extended learning of the AI_{extended-summer}. The classification experiment 1 (summer) shows the algorithm AI_DIN-summer, which has learned the air quality conditions, and it made 126,614 correct interpretations towards 129,047 given datasets.

Figure 9

Figure 10

The classification experiment 2 (summer) shows the extended algorithm AI_{extended-summer}, which has learned the air feeling of residents, and it made 1,265 correct interpretations towards 1,274 given datasets.

The developed smart monitoring system is based on RaspberryPi platform for measuring the air quality and wall humidity in apartment houses and delivers big sensor data. It recognises actual air quality in an apartment room by further telling us the actual personal air climate feeling of the residents.

In future works, a new academic research project based on the smart monitoring system developed in Bielefeld-Sennestadt for active assistance of the users during renovation process, called Environ and Approved, is needed. The benefit of this work is to use deep learned artificial intelligence for assistance purpose inside of apartments. Therefore, an apartment could be equipped with the algorithm AI_DIN (as the base algorithm) to serve information about actual air quality in the beginning and the resident could teach the second algorithm AI_extended (as a branch) about his/her personal air feelings by sending input data. In addition to information on measured data, residents will also receive suggestions of how to interact with their environment to get better room climate and to reduce CO₂ emissions.