*3.1. Qualitative Analysis*

A survey was conducted to explore the current state of PPM at universities in North America. The survey, which consists of ten questions, was developed based on five aspects (process, cost plan, budget allocation, scheduling, and decision making) of PPM. For example, two survey questions were designed to investigate the current practice and workflow of PPM in universities. The survey was distributed to facility managers who were registered as a member of the APPA at twelve universities. When collecting responses from the universities, the responses with incomplete information were excluded. In addition, out of a total of ten questions, only five questions were analyzed and presented in this study because the remaining five questions were related to personal information, data availability, etc. Table 2 summarizes the five important questions, multiple answers provided for each question, and the corresponding number of responses.


#### **Table 2.** Survey result.

Analyzing the survey results led to the following three main observations: First, the progress of the PPM work assignments was mainly monitored based on the reports generated by CMMS (No. 1 in Table 2). This suggests that CMMS has been mainly adopted by at least half of the facility managers in universities in order to automatically monitor PPM work progress. Second, work, set-up, clean-up, and documentation were identified as the most significant four factors included in the PPM work order estimates; they accounted for 78% of the responses in question No. 3 in Table 2. Third, it was found that most of the university facility managers (76%) responded that the prioritization strategy is the most critical component to improve the effectiveness of the current PPM practice, as illustrated in question No. 5 in Table 2.

Phone interviews were conducted to understand the current status (e.g., types of management systems, maintenance components, and data recorded) of facility management and investigate practical issues in higher education institutions (i.e., universities) in North America. Compared to the survey analysis illustrated in the previous section, the focus of the interview was on exploring the overall FM practice, not being limited to the PPM. A flyer was created and distributed to facility managers who were registered as a member of APPA at thirty-five universities. As a result, twelve participants were recruited for a phone interview which was conducted from November 2019 to January 2020. A total of thirteen questions (three for planning and definition, six for data quality and variables, one for prioritization, and three for methodology) were developed and asked to respondents during the interview. (Additional survey questions can be developed in the future for a more comprehensive understanding of the current status of facility management practice at universities.) The phone interview took approximately 30 min, and each interview was recorded and transcribed digitally.

In this study, seven interview questions were excluded for further analysis since they were associated with definitions of terminologies, willingness to offer raw data, and personal information. As a result, responses to the remaining six important questions were analyzed and presented in Table 3.


**Table 3.** Phone interview result.

It was observed that scheduled maintenance (42%) and PPM (41%) were two major organized maintenance plans adopted in most universities. Within each university, building systems and components were classified based on Uniformat (42%) and MasterFormat (33%). The maintenance task was performed on an individual component tracked. Additionally, it was found that a work order was mostly recorded at the end of the activity (58%) by the technician (83%) using CMMS (75%). The result of the interviews is assumed to reflect the recommended practices of the operation perspective in the facility management at the referenced higher education institutions.

## *3.2. Data Driven Analysis for Qualitative Data*

Two natural language processing (NLP) techniques (topic modeling and sentiment analysis) were applied to the collected interview transcriptions containing a significant amount of textual data (over 50,000 words) to reveal important latent information that was not able to be captured during the interview. NLP techniques have been increasingly used as a quantitative method to derive meaningful insights such as keywords [29], topics [30], and sentiment [31] from a set of textual data (e.g., transcripts) obtained from the interview. Previous studies have demonstrated the efficacy and potential of applying NLP techniques, addressing limitations (e.g., time-consuming, subjective, and error-prone) that reside in qualitative approaches such as interviews and surveys. In other words, conducting NLP analysis provides an opportunity to find unexpected observations or insights based on semantic and syntactic similarities that can be observed within textual data comprising interview transcriptions.

Raw data, interview transcriptions, from 12 universities were preprocessed through the following steps: removing stop words (e.g., "the", "am" and "a") and noises (e.g., blanks and punctuation), word stemming, and tokenization.

Latent Dirichlet Allocation (LDA)—one of the well-established topic modeling approaches—was adopted to identify keywords and prevalent topics in the interview. LDA allows for identifying patterns that can be observed within textual data without a tedious labeling process [32]. In general, LDA produces a couple of topic groups, each of which consists of corresponding keywords. Labeling topic group (naming) relies on human interpretation and judgment [33]. As a result, two topics were identified based on the semantic similarity of keywords in Table 4, which implies that the focus of respondents during the interview was on two aspects of PPM and the maintenance system. Another interesting observation was that Archibus, an integrated platform system for infrastructure and building management [34] frequently appeared during the interviews, which suggests that it was one of the most widely used software in the universities.

**Table 4.** Topics and keywords identified from the interview.


Sentiment analysis was further conducted to identify the facility managers' degree of positiveness or negativeness towards the use of PPM. Note that it was assumed that PPM was the main subject of the phone interview since it was identified as the main topic, as illustrated in Table 4. For the analysis, a large number of tokenized words derived from the previous LDA analysis were used as input to the well-established pre-trained Python module, Valence Aware Dictionary and sentiment Reasoner (VADER) [35]. VADER allows for quantitatively assessing the level of sentiments for the given texts. As a result, it provided a sentiment score between 0 and 1, where 0 indicates complete negative sentiment and 1 denotes complete positive sentiment. The criteria for positive (0.7~1.0), neutral (0.4~0.7), and negative (0.0~0.4) range was set based on the previous studies [36,37].

The results revealed that five universities (B, E, H, I, and L in Figure 3) responded that they were using the PPM (No. 1 in Table 3) showed positive sentiment scores. This finding supports that the universities are willing to adopt PPM with the effectiveness and advantages of the PPM.

**Figure 3.** Sentiment analysis results based on phone interview transcriptions.

#### *3.3. Facility Management Unified Classification Database (FMUCD)*

Over the years, higher education institutions in North America have employed many classification systems (e.g., Uniformat II, UniformatTM, OmniClassTM, and MasterFormat®) to classify building systems, construction, and maintenance activities. As illustrated in Figure 4, Uniformat II [23] provides a more specific facility management structure with three levels (level 1-major group elements, level 2-systems, and level 3-subsystems). For example, in the figure, level 1 includes shell, interiors, services, etc. Regarding "Services" at level 1, it can be divided into HVAC, plumbing, electrical, conveying, and fire protection. For HVAC at level 3, it consists of heating, cooling, distribution systems, controls & instrumentation, terminal & package units, energy supply, etc.

**Figure 4.** Uniformat II classification [23].

This study established a descriptive code entitled Facility Management Unified Classification Code (FMUCO) in the database. The purpose of the FMUCO is (1) to compile the current data from universities to create Mega data and (2) to conduct the data-driven analysis to explore the current status of the facility management in higher education institutions. The FMUCO code is created by combining Uniformat II with generic descriptions of building components from Whitestone cost reference [19] shown in Figure 5.


**Figure 5.** Facility Management Unified Classification Code (FMUCO).

As illustrated in Figure 5, the proposed descriptive code is composed of an 8-digit code; the first three digits describe the system code, the next two digits define the subsystem, and the last three digits are the abbreviation of the component description. The FMUCO has 543 descriptive codes, new elements can be added in the future. This classification method permitted the collected data for each university, which varied significantly in terms of data type, data points, and data attributes (e.g., work order description, cost information, labor hours, etc.), to be managed for the study. Data preprocessing was performed to develop a structured and organized database shown in Figure 6. This preprocessing step included but was not limited to identifying common data attributes, cleaning noisy data, and removing unnecessary data.

#### **Figure 6.** FMUCD structure.

#### *3.4. Quantitative Analysis*

The database developed for this study allowed identification of critical information and risks involved in the facilities management at the component level. Three types of data-driven approaches were adopted for quantitative analysis: (1) statistical comparison analysis, (2) risk-profile analysis, and (3) outlier analysis. Statistical comparison analysis (1) was conducted to explore the current trend of PPM and UPM for the referenced universities. At this stage, the ten systems (e.g., HVAC, electrical, plumbing, conveying, fire protection, etc.) were compared to identify the highest number of work orders and labor hours associated with PPM and UPM at the system level. (2) Risk-profile analysis was conducted on the top three systems to distinguish the risks in the subsystem level of UPM. The risk profiles for top systems aimed to provide basic knowledge to the facility managers about the subsystems with a high probability of getting a UPM work order. The outlier analysis (3) was conducted to identify components with a high risk of generating UPM work orders.
