Understanding structural health monitoring data to support decision-making processes and service life management of mass timber buildings. A preliminary study on use of data scaffolding

Structural health monitoring (SHM) can be used to support decision-making processes leading to improved building management plans. With advanced engineered wooden buildings on the rise, SHM has emerged as a critical tool to document behaviour of these structures. However, monitoring data need to be easily accessible and understandable to support informed decisions. This study explores the potential benefits of an intentionally designed data scaffolding to support interpretation of SHM data for timber structures. Results suggest that appropriate scaffolding (e.g. pinpointing critical parameter ranges) can enable lay users to better interpret monitoring data of timber structures. Results also suggest that, for more highly educated users, access to graphs linked to metadata can be beneficial. Overall, it appears that documentation related to the monitored phenomena and context, as well as more explicit reference to building performance requirements, could improve data-driven decision making for all knowledge levels for the building maintenance sector.


Introduction
Structural Health Monitoring (SHM) refers to methods and techniques used to evaluate the progression of physical phenomena within a structure and its surroundings over time; progressions which can lead to a diversion from a structure's designed behaviour. SHM data are observed, measured, recorded, processed and analysed to evaluate if a structure behaves as intended. Information from SHM is commonly used to support different types of actions: (a) immediate actions to assess and limit damage in case of abrupt changes in a structure due to a major event; (b) immediate or deferred actions to control phenomena negatively affecting the safety and serviceability of a structure, (c) future actions to support service life planning and (d) validate designs and models (Riggio and Dilmaghani 2020). Naturally, different stakeholders may be interested in accessing and evaluating SHM data depending on the type of actions/ decisions that correspond to their given responsibilities. For instance, a contractor may utilize data to optimize construction schedules and demonstrate the compliance to the required performance. A facility manager may utilize the performance data to optimize maintenance plans. An insurance company and a buyer would like to know the physical state of a property.
In recent years, several SHM programs have been implemented in mass timber buildings to help provide benchmark data for these relatively novel construction systems (Morris et al. 2011;Leyder et al. 2015;Kordziel et al. 2019;Lanata 2019;Granello and Palermo 2020;Baas et al. 2021). Mass timber buildings incorporate engineered wood products, such as glue-laminated timber, cross-laminated timber, laminated veneer lumber, etc. as structural elements, thus serving as an alternative to more conventional and common construction systems using steel and concrete. According to the Mass Timber Report (Anderson et al. 2021), concerns about long-term durability and uncertainties during construction are some of the major obstacles that stand in the way of more widespread adoption of this technology. A recent survey on SHM of timber buildings highlighted that most monitoring projects reported in the last decade were aimed at controlling conditions of timber buildings during construction and/or shortly after commissioning (Riggio and Dilmaghani 2020). Hygrothermal monitoring, for example, measures heat, air and moisture transmission and storage, all of which can have a significant influence on the performance and durability of materials and building components.
Enabling such monitoring is critical, as moisture changes of mass timber elements during both construction and while in service can indicate the need for preventative measures to control the response of timber structures and thereby ensure more reliable long-term performance. Moisture content (MC) within the hygroscopic range (typically 0-28%) affects nearly every physical and mechanical property of the material (Glass and Zelinka 2010;Shmulsky and Jones 2011). It can cause dimensional changes (i.e. shrinking and swelling), subsequently affecting building tolerances (Arda Buyuktaskin et al. 2019), while also inducing internal stresses within a wooden member, which can lead to the propagation of cracks (Björngrim et al. 2016;Franke et al. 2018). High or fluctuating levels of MC can have impacts on the long-term deformation of structural timber elements, due to a combination of viscoelastic and mechanosorptive creep (Grossman 1976;Muszyński et al. 2005;Toratti 1992). Exposure to prolonged high levels of moisture (caused for instance by precipitations during construction, leaks or high relative humidity conditions) can lead to mould and fungal decay (Zabel and Morrell 1992). In addition, factors such as chemical interactions and corrosion, also triggered by high MC, can affect the performance of timber connections (Hall and Flock 2008;Niklewski et al. 2018;Sinha et al. 2020).
Available MC data during construction can help contractors to develop the most suitable strategy to protect vulnerable locations from wetting, and to develop dry-out schedules. Contractors can likewise use MC data to plan construction activities, such as the completion of vapour-resistant assemblies, as those for fire encapsulation and some types of building enclosures (Chang et al. 2020). High moisture levels detected during service can help facility managers and building owners to detect possible leaks or condensation build-up in assemblies and adopt immediate remedial measures. For these reasons, MC data are necessary to make critical decisions across all stages of a mass timber building life cycle.
The most common technique to measure wood MC is the resistance method, which measures the electric resistance between two probes embedded into the material. Observed resistance is positively correlated with wood MC, and this method provides a local estimate of MC values. However, the quality of data may be affected by many factors, such as electrical interference, direct water contact, condensation, salt contamination, imperfect sensor calibration or installation, among others (Riggio et al. 2014;Dietsch et al. 2015;Baas et al. 2021).
SHM also analyses structural phenomena, such as deformations, displacements, crack development and accelerations, to evaluate a structure's ability to carry designed loads. Changes in the relative linear position of two or more observed points, which can be indicative of many phenomena including deformations, displacements and moisture-induced movement (i.e. shrinkage and swelling), can be monitored with sensors such as string potentiometers or linear variable differential transformers. String potentiometers are relatively simple sensors, composed of a tensed steel measuring cable wound on a cylindrical spool coupled with a rotational sensor which creates an electrical signal proportional to the cable extension. These measurements can be affected by multiple factors, such as effects of temperature on instrumentation, natural frequencies of sensor components, temporary supports and other load applications during construction, among others (Baas et al. 2021). String potentiometers have been used to monitor vertical movement in multistory timber buildings caused by compression loads and shrinkage (for instance, Serrano et al. 2010;Fast et al. 2016;Wang et al. 2016;Baas et al. 2019). Depending on its magnitude, vertical movement in multistory timber buildings can affect installation of non-structural elements, for example, mechanical equipment attached to the walls or columns, partition walls and facade systems such as curtain walls . The availability of this SHM information during construction can allow adjustments to account for the measured geometrical changes, and also be used to inform and refine future designs.

From data to informed decisions
For a SHM system to be considered successful, it must be able to produce quantifiable value to its stakeholders. Research on the value of information (VOI) of SHM investigates the interest of decision makers in investing in a monitoring system to access information that can reduce losses, due to damage and poor performance, over the lifespan of a structure (Straub and Faber 2005;Zonta et al. 2015). In this context, the value of a SHM system is defined as the difference between the expected cost of operating a structure with the monitoring system and the expected cost of operating the structure without the system (Straub et al. 2017;Giordano et al. 2020). While extensive research on VOI of SHM highlights decision maker's cost-benefit analysis relative to these monitoring projects, limited research has been conducted on the accuracy and use of information gathered from sensor data by different users. This is a critical gap in knowledge that very likely can impact not only the base effectiveness of SHM systems, but also the VOI produced.
Several authors agree that there is a high risk that the great amount of data gathered through sensors are significantly underutilized by building stakeholders, as it is often not accessible or not comprehensible (e.g. Ciribini et al. 2017;Napolitano et al. 2018). Data alone are of no use unless it leads to some form of robust understanding. Monitoring data need to be easily accessible and understandable to support informed decision and guide relevant stakeholders to choose the action that is most effective for a given situation Longman et al. 2023).

The interaction of knowledge and graph sensemaking
As sensor data may be represented in different forms (i.e. numeric, binary, text), in most cases it is preprocessed by a monitoring electronic system and plotted in the form of graphs. These graphs take this raw abstract data and represent observed relationships in Cartesian coordinates, usually amongst two variables: typically, the measured parameter and the time of the measurement. As such, graphs have both topological and typological qualities, and can be used to represent both quantitative and qualitative aspects of an observed natural phenomenon (Bowen and Roth 2005).
Since graphs are the most common way of presenting data by most monitoring platforms, investigating graph data sense-making is key to understanding how SHM can be used to inform decisions on a built asset. While the construction and interpretation of graphs are common practices in science, it is important to understand that this is often not easy for all users (especially low-knowledge or lay users; Roth and McGinn 1998). The extreme variety of interest and expertise of mass timber stakeholders and SHM data users (which includes facility owners, designers and contractors) exemplifies the need to find a reliable and robust means of presenting monitoring data that is accessible and effective for users who differ in what they know and understand about wooden structures.
Work in the field of graph comprehension identifies three levels of cognitive tasks inherent to graph understanding: (1) elementary levelto read a specific value; (2) intermediate levelto read relationships or trends in a graph (comparing set of values in spatial and/or temporal intervals) and (3) an advanced level to read beyond what is presented in a graph (correlating attribute changes by observing spatial and temporal distributions in relation to other attributes) (Curcio 1987;Carswell 1992;Carpenter and Shah 1998). Shah and Hoeffner (2002) reviewed how viewers comprehended graphs, and concluded that there are three fundamental factors that influence viewers' interpretations and should be considered together. Those factors are the visual characteristics of a graph (e.g. format, colour, use of legend, size, etc.), a viewer's knowledge about graphs and a viewer's expectations about the content of the data in a graph. As emphasized by Shah and Hoeffner (2002), a users' existing knowledge and proficiency are critical predictors of how well they use graphic visualizations.
One might expect that higher levels of expertise could result in a lower impact of the graph formats on interpretation (e.g. Pinker 1990). However, it is also possible that higher levels of expertise result in an increase in the effect of graph format on performance, as these well-known format structures strongly activate prime expectations and biases within these high-knowledge individuals (e.g. Shah and Freedman 2011). This is consistent with other work which has also suggested that prior knowledge can affect graph usage. For example, Reingold et al. (2001) suggested that there are significant differences in how experts and lay users interpret and extract information from these visual displays. Experts' extensive knowledge affects not only what they notice, but also how they interpret information from data (Shah and Freedman 2011;Canham and Hegarty 2010).
While there are some studies on how people read and make sense of sensor data from building, these examples are largely focused on domestic energy data visualizations (e.g. McCalley and Midden 2002;Yun et al. 2010;Herrmann et al. 2018). There is a paucity of research related to interpretation and use of SHM data. In the SHM field, data users, such as building owners and managers, are often not experts in all the areas and phenomena related to the monitoring project. In this case, they may need assistance (i.e. scaffolding) to correctly interpret data and consequently take actions based on the acquired information.
Scaffolding, a common instructional practice in which novice learners receive guiding help or support from a more capable user, has been shown to produce reliable and robust effects on learning in multiple contexts (Kim et al. 2018). In fact, one common means of implementing scaffolding is the use of appropriately designed visualizations or visual supports (Wang et al. 2013;Yuan et al. 2017). In these cases, users are given robust visualizations that attempt to facilitate the recognition of patterns or relationships that would otherwise be very difficult for novice learners to identify on their own. In short, these visualizations produce a kind of 'cognitive offloading' (Patterson et al. 2014), in which processing what would normally be very arduous for novice learners is instead offloaded to an external representation like visualization. The reduction in effort thus frees cognitive resources for the user, allowing them to focus on the critical relationships between data points or conceptual units.
The majority of the commonly used commercial building information dashboards used for SHM are typically tailored to show and report data from a specific set of sensors and are often not designed to support interpretation of complex phenomena, such as relationships among parameters. In the last decade, efforts have been made to develop effective ways to visualize and communicate SHM data, and especially so, to contextualize data. Most of these studies propose approaches to visualize the monitored system/building using live camera streams , Augmented Reality (Mustapha et al. 2018, De Amicis et al. 2019 or Virtual Reality (Mustapha et al. 2018;Napolitano et al. 2018), which is then linked to data and metadata (for instance, technical drawings, etc.). None of the studies reported above, however, have used scaffolds as a useful intervention for improving the comprehension of specific physical phenomena depicted by SHM data.

Purpose and contribution of this study
While many studies argue the potential of SHM data to better inform decision-making processes of key stakeholders in the architecture, engineering and construction sector, there is still little evidence that the information embedded in SHM data is fully exploited. The current study is designed to explore how individuals derive information from SHM data of mass timber buildings, and how this information can influence decision making.
In particular, the following research questions were investigated: R1: Is scaffolding useful to enhance sense-making of sensor data and support decisions related to the maintenance of timber buildings? R2: Does the proposed scaffolding differently affect data sense-making and decision making of different users based on their level of expertise or prior knowledge?
It is hoped that the results of this initial study can be used for the development of effective data communication tools and strategies, thereby allowing for dataenabled decision making for the design, construction and maintenance of timber buildings.

Methods
An experimental investigation was conducted to examine how different respondents interpreted sensor data from two monitoring scenarios of timber buildings, i.e. hygrothermal and structural monitoring. This task was administered alternatively with (or without) scaffolding. The types of decisions made by the respondents based on the information inferred by the data were also examined. This investigation was approved by a university Institutional Review Board for research on human subjects prior to its conduct.

Target population and sampling frame
This study considers two target populations: lowknowledge users and more highly educated professionals in mass timber construction. Participants were recruited among faculty, technical staff and students in the Department of Wood Science and Engineering and the School of Civil and Construction Engineering at Oregon State University, as well as among members of the Forest Products Society (an international technical association which involves members from all segments of the forest products industry). Undergraduate students enrolled in wood science or engineering programs represented a sample of low-knowledge data users. Despite the fact that some of these students have some basic conceptual Distribution of expert and non-expert respondents between scaffolded and nonscaffolded version of the questionnaire How many years of work experience do you have? Is scaffolding useful to enhance sensemaking of sensor data?

Data interpretation
In which location did the moisture accumulated the most? a,b Distribution of correct answers among experts/non-experts and with or without scaffolding. Response time Can you select from the following graph, which column experienced the greatest vertical movement and in which period? b,c Is scaffolding useful to support decision making based on sensor data?
Decision making Based on this data, if you were responsible for maintenance of this building, what action would you take?
Relationship between data interpretation and decision (distribution among experts/ non-experts and with or without scaffolding). Response time Why did you decide to do nothing? d Is scaffolding useful to enhance data sense-making and support decision making based on sensor data?

Data interpretation and decision making
Which part of the graph did you use to make that decision? b knowledge of wooden structures, their coursework provides little direct instructions on the monitored phenomena. While certainly not completely ignorant of wood construction and engineering, we feel that this was an appropriate baseline group versus a true novice who has absolutely no experience with wood structures or engineering. Such true novices would likely struggle with the task, not necessarily due to the inadequacy of the support tool, but rather because they have absolutely no knowledge to bootstrap to even begin to understand how to use the tool or manage wood structures. Wood science professionals and engineers working in the forest products sector or at the university were representative of the more highly educated group.

Questionnaire and data collection
A web-based questionnaire was developed in the platform Qualtrics. The questionnaire included a total of 13 questions. Question types were multiple choice either verbal or based on clickable images. Questions focused on demographic information, and questions to assess respondents' comprehension and use of the data (i.e. data analysis and information utilization). Table 1 presents the list of questions, related research questions and type of analysis. The survey presented two SHM scenarios showing hygrothermal monitoring data (i.e. wood MC data) and structural monitoring data (i.e. displacements). The source of the MC data is a hygrothermal monitoring study conducted during the construction of the G. Peavy Forest Science Complex, at Oregon State University reported by . Each MC plot in the graph reported the name of the monitored locations (for instance 'upper corner 1', 'upper corner 3', etc.) given by the researchers who conducted the monitoring study ; the location was a cross-laminated timber shear wall, partially exposed to outdoor conditions during construction. However, no explanation was given in the questionnaire about that specific location of each sensor. Therefore, the respondents did not know 'where' specifically the data were collected and 'what' structural system (i.e. a shear wall) was monitored. In addition, the plots reported the timeframe when data were collected (i.e. from December 2017 to May 2018). However, the questionnaire did not include any additional information about events related to the building during this timeframe that could have had an impact on the hygrothermal conditions of the building itself. Specifically, respondents did not know that the reported data referred to a period when the building was under construction, and, therefore, the monitored locations were partially exposed to the weather.
The source of the displacement data is a structural monitoring study conducted during construction of the Brock Commons at the University of British Columbia, Vancouver, reported by Mustapha et al. (2018). The graph with the displacement data reported the level of the building where the data were collected (i.e. third, fourth or fifth floor), however, no information was provided in the questionnaire on the total number of stories, or other details of the building itself. It is worth mentioning that the monitored building, Brock Commons, is 53-metres-tall and has 18 stories; at the time of construction, one of the concerns was to evaluate the impact of vertical movement due to shrinkage and compression of wood components (Mustapha et al. 2018). Also in this case, the graph reported dates when data were collected (i.e. from November 2016 until June 2017), but not the events that occurred during that period of time. Respondents, therefore, did not know that data were collected during construction, and that vertical displacement sensors were disturbed by shoring used to temporarily support some elements.
Questions were formulated to evaluate: (1) how SHM data were interpreted and (2) how SHM data were used to make decisions (see complete questionnaire in Supplementary Material). For each scenario, participants were first asked to select a portion of a graph (clickable image) corresponding to their chosen answer. Respondents could choose more than one portion of the graph if they felt multiple portions were relevant. A follow-up question asked respondents to choose a type of action to be performed in the example building based on the data portrayed in the graph. In case respondents chose to do nothing based on the shown data, a follow-up question also asked participants to justify that choice. Participants were also timed in their responses to these questions to provide additional insight into how quickly and effectively they might be reasoning with the provided data.
Scenarios were portrayed in two different versions: with and without scaffolds. The two versions were randomly distributed to the respondents. In those questions augmented with scaffolds, visual aids or additional information were presented along with the plotted data to guide respondents and support understanding and interpretation of the presented data (Supplemental Material). Figure 1 shows the non-scaffolded and scaffolded versions of the graph describing MC values in different locations in a timber building.  recognized that most of the literature does not thoroughly discuss irregularity in MC readings, and commonly just the raw data are presented. To scaffold comprehension of the MC graph of the survey, the MC data were post-processed by including techniques such as the use of superimposed moving averages. This technique eliminates the presence of spikes and artefacts that were not representative of the real phenomenon (i.e. water absorption/desorption into/from the wood), but rather noise introduced by conditions such as electromagnetic interferences. In addition, a horizontal line was added to the plot to define a threshold at a common MC value (>20% MC). This threshold was chosen since wood is susceptible to mould development at a MC greater than 20% (Zabel and Morrell 1992). Rainfall data also complemented the MC graphs in the scaffolded version of the questionnaire.
Timber buildings are susceptible to vertical movement due to moisture-induced dimensional changes (shrinkage and swelling), as well as short-term and time-deferred deformation of timber members (e.g. columns, beams, floor panels). To scaffold comprehension of the displacement graph, a table published by Mustapha et al. (2018) was used along with the graph to present column shrinkage estimate calculations. Additionally, the initial position of the column was shown at the zero point on the graph, and the direction of the vertical shortening and elongation of the column were also indicated using arrows. Both a graph and table were used to present the data, since the graph showed the relationships among variables, while the additional table increased the level of detail by showing the discrete data values (Johannessen T and Fuglseth A 2014; Figure 2).

Data analysis
Descriptive and statistical data analysis of the questionnaire were conducted using Microsoft Excel (Microsoft 2016). Tableau software (version 2020.4) was used for visualization of results.  Responses were analysed in terms of their correctness, accuracy and speed. Accuracy was analysed by aggregating 'correct' responses as those selecting the relevant area in the graphs. It is worth mentioning that most of the respondents selected more than one region on each graph. To analyse these data, all the regions selected by a respondent were captured in the data analysis. Regions were coded in a 2-level coding scheme as 'irrelevant' and 'relevant'. Response accuracy was evaluated for each response by calculating the percentage of 'relevant' responses relative to the total number of selected regions. All the questions of the survey, except for the demographic questions, included a hidden timing question to track the response time spent for each question.
Relationships between time and accuracy were measured for each response, and for both versions of the survey (with and without scaffolding). Relationships between the data interpretation and subsequent choice on the type of action were analysed. Results from the scaffolded and non-scaffolded versions were compared to evaluate the effectiveness of scaffolding. Effect sizes (e.g. Cohen's d, η 2 ) are also reported as necessary. Consistent with Cohen (1988), d values of ∼.20 are considered small effects, ∼.50 are considered medium effects and >.80 are very large effects.

Demographics
One hundred-twenty people participated in the questionnaire. The research examined responses from a total of 108 respondents (N = 108), excluding responses from participants who were clearly not engaged with the questionnaire (response times > 3 SDs from mean). Respondents' age ranged from 19 to 71 years old. Figure 3 shows age ranges among more highly educated and low-knowledge respondents. Forty-eight (n = 48) respondents accessed the scaffolded version while 60 (n = 60) viewed the nonscaffolded version. Among the scaffolded group, 15 respondents were more highly educated and 33 lower in knowledge, while in the non-scaffolded group, the highly educated were 27 and the lowknowledge participants were 33. A more complete description of level of expertise (highest completed degree and years of experience) and field of expertise is represented in Figures 4 and 5, respectively. Figure 6 summarizes how respondents answered the question 'in which location moisture accumulated the most?'. The vast majority of respondents selected the correct response (i.e. upper tendon 4), and this did not vary across scaffolding groups (higher education: non-scaffolded 89%, scaffolded 87%; lower knowledge: non-scaffolded 79%, scaffolded 94%; χ 2 = 3.34, p = .07). However, considering the lower knowledge group alone, scaffolding was particularly effective for helping these individuals, increasing the percentage of right answers by 15 percentage points and, even more interesting, by decreasing the mean response time from 51 to 39 s (t(33) = 2.05, p = .04, d = .51), which is a medium-sized effect. It is worth   noting that the presence of scaffolding did not change the mean response time for the more highly educated group from 77.37 s, in the non-scaffolded version, to 64.67 s in the scaffolded version (t(40) = 1.18, p = .25). Figure 7 shows the type of decision that respondents would have made if they were responsible for the maintenance of the example building, based on the presented moisture data. In terms of actions selected, there was no difference between groups in terms of 'checking for leakage' (high education: non-scaffolded 41%, scaffolded 60%; low knowledge: non-scaffolded 58%, scaffolded 49%; χ 2 = .00, p > .05), or 'fungal decay'  (high education: non-scaffolded 7%, scaffolded 13%; low knowledge: non-scaffolded 24%, scaffolded 12%; χ 2 = .36, p = .55). However, there was a difference in checking for 'stains or mould' (high education: nonscaffolded 33%, scaffolded 13%; low knowledge: nonscaffolded 12%, scaffolded 39%; χ 2 = 9.12, p < .01). None of the low-knowledge respondents chose to 'do nothing' (high education: non-scaffolded 15% and scaffolded 13%). Three respondents (one high education and two low knowledge), all within the non-scaffolded group, responded that the 'readings of the sensor did not look correct and probably the sensor needed to be replaced'. It is worth noting that the MC graphs presented in the two versions of the questionnaire ( Figure  1) show relatively higher MCs (i.e. above 20%) in some locations, and for a limited period of time; conditions that are in general insufficient for the development of fungal decay. The last portions of the MC plots, corresponding to the last 3-4 monitored months (depending on the locations), did not show moisture increases, even if raw data in the non-scaffolded version of the questionnaire exhibited some spikes.
When asked what portions of the plot they used to decide the type of action, most respondents selected more than one region. However, it appears that most participants were able to identify at least 1 correct region in the graph, and this did not vary by group (high education: non-scaffolded 70%, scaffolded 73%; low knowledge: non-scaffolded 76%, scaffolded 79%; χ 2 = 1.56, p = .21).
As already reported, six respondents, all of them from the highly educated group, responded that they would not take any action based on the MC data; four of them in the non-scaffolded group and two in the scaffolded group. Figure 8 shows the reasons provided for this choice (based on question 7 -Supplemental Material), and the corresponding mean response time. Response time ranged between 25 and 31 s. The need for more information on how the data were collected was the reason identified by three highly educated respondents in the nonscaffolded group and by one in the non-scaffolded group. Another respondent from the highly educated group without scaffolding needed to compare these data with other published results to understand if there was a situation of concern. Interestingly, the answer 'the graph is not clear enough' was only given by one highly educated respondent in the scaffolding group. Figure 9 shows the number of times each region of the graph (Figure 10) was selected to indicate the data identifying the column where the greatest vertical movement occurred, and the percentage of respondents choosing those regions. As it can be observed, most respondents (64%) selected region #7, and this did not vary by education level or scaffolding, χ 2 = 2.71, p = .10. In many cases, respondents selected more than one region in the graph. Only three respondents answered correctly, selecting areas #3 and/or #5. Two of those respondents were from the highly educated group who answered non-scaffolded questions. One respondent was a low-knowledge respondent in the scaffolded group. Mean response time was shorter (66 s) for responses selecting area #7 and longer (160.33 s) for those selecting the correct regions of the graph (#3 and #5, respectively; t(67) = 3.55, p < .01, d = 2.09), which was also a very large statistical effect. Figure 11 shows the types of actions chosen by the respondents based on the displacement data and mean response times. The action for which highly educated respondents overwhelmingly spent most time to make a decision was 'Check for damage in the partition walls close to the location of the greatest vertical movement' (average time 428.86 sec.). 'Checking for damage in Figure 9. Number of times each region of the displacement graph was selected and the percentage of respondents choosing each region. Figure 10. String potentiometer data: graph regions selectable by respondents.

Displacement graphs
the column experiencing the greatest vertical movement' was the type of maintenance action selected by most respondents (35% of total respondents).
Eleven respondents who opted for this action selected area #10 in Figure 10, while 18 respondents selected area #7 to make that decision, as it can be seen in Figure 12. Figure 13 shows different types of decisions made by respondents selecting area #7. The most commonly selected options, after option #1, which was the first choice among the highly educated respondents in this subgroup, were to 'shore up the column experiencing the greatest vertical movement to avoid problems of instability' (17%) and to either 'check for damage in the partition walls' or 'shore up the floor in the location of the greatest vertical movement' (14% for both).
Just three of all respondents decided to do nothing ( Figure 11). As shown in Figure 14, two of them stated they needed to compare these data with other published results to understand if there was a situation of concern (average time 31 s). One of them needed more information on how the data were collected (average time 9 s). The actions chosen by the three respondents selecting the correct region of the graph identifying the column experiencing the largest vertical movement (i.e. areas #3 and #5) were: to check for damage in the column (one high educationnon-scaffolded) and to check for damage in the partition walls (one low knowledgescaffolded and one high educationnon-scaffolded).

Discussion
The scope of this study is to evaluate usefulness of some scaffolding approaches in supporting SHM data sense-making, and consequent decision making, of individuals with different levels of education and knowledge on mass timber technology.
One of the first observations that can be made is that, in both monitoring scenarios (hygrothermal and structural), there was a prevalent answer to the question asking to identify areas in the graphs that described the phenomenon of interest (i.e. moisture accumulation and column vertical movement, respectively). The answers from a majority of respondents, regardless of their level of education and presence or not of scaffolding, identified areas with the highest positive values. This response was correct in the case of the hygrothermal monitoring scenario (i.e. it Figure 11. Type of maintenance action based on the displacement data as indicated by respondents and mean response times. correctly identified the plot showing the highest accumulation of moisture). However, this type of response was not correct in the structural monitoring scenario, as it did not correctly identify the real displacement phenomenon (i.e. vertical movement of the column) which was characterized by shortening due to shrinkage and axial compression. Instead, high positive displacement values represented a 'disturbance' caused by shoring during construction. It is worth noting that a correct interpretation of the displacement data should have considered the behaviour of the material and of the structural system. In particular, large positive displacements were improbable due to the fact that longitudinal swelling (and shrinkage) of wood is generally quite small (i.e. average values for shrinkage range between 0.1% and 0.2% depending on the MC; Glass and Zelinka 2010). In addition, the largest positive displacement values were recorded on the third floor, where columns were subject to higher axial forces, which would cause shortening due to compression, rather than elongation. So, misinterpretation of the graphs, in case of the structural monitoring was due to lack of understanding of actual physical phenomena as represented by the data. The general poor graph sense-making of the string pot data suggests that scaffolding could have been more effective if more explicit instruction on some fundamental aspects of the material and structural system behaviour was provided. In other words, it might be worthwhile to support and augment SHM data users with relevant information or guidance built into the system (e.g. material behaviour quick tips, etc.).
Scaffolding improved performance in terms of accuracy of low-knowledge respondents in addressing  the question describing the moisture accumulation phenomenon in the hygrothermal monitoring scenario. Interestingly speed of response of more highly educated respondents was lower, and these effects were not trivial as evidenced by medium to large effect sizes. This is consistent with expertise literature, which has shown that experts often are more deliberate in both their representation of the problem (e.g. Chi et al. 1981), and also when it comes time to make a decision (e.g. Moxley et al. 2012). Carefully thinking through the situation and also the subsequent solution naturally should take more time, and this deliberate thinking usually translates into higher and more accurate performance by those with more knowledge.
Another interesting finding involves the relationship between data interpretation and consequent decision-making process. In the hygrothermal monitoring scenario, the most chosen action was to check if there was any leakage in the building. It is worth noting that the last portion of the MC plots, which was the most recent data corresponding to the last 3-4 monitored months (depending on the locations), did not show moisture increases, even if raw data in the non-scaffolded version of the questionnaire exhibited some spikes. A possible conclusion that could have been drawn from the data was that, if there was any leakage in the monitored locations, it happened at the beginning of the monitoring period and was resolved during the second half, as the locations were able to dry out. In this case, looking for leakage could have not been necessary or beneficial. In the scaffolded version of the questionnaire, MC graphs were complemented with rainfall data. From these data, it was possible to observe that, even during weather conditions which were comparable (or wetter) to those characterizing the period of higher MC values, the location which showed higher moisture accumulation kept drying out in the second half of the monitoring period, thus suggesting that, if there was a leakage, it was resolved. This suggests that data were not carefully considered across the temporal dimension, in relation to the evolution of certain phenomena over time (wetting and drying) and, in the case of the presence of data scaffolding, in combination with other influencing factors (rainfall and leaks). This is consistent with other research on graph usage, which has indeed found that participants tend to misuse or ignore temporal scales within graphs (Zikmund-Fisher et al. 2005), or tend to struggle to interpret trends (Maichle 1994;Herrmann et al. 2018). For instance, Herrmann et al. (2018) recommended alternatives to time-series visualizations, such as summary overviews and aggregated data, to better enable building users to make decisions about their energy usage through readings from smart metres.
While it seems that respondents paid more attention to the y-axis dimension of the data (the value of the measured parameter) rather than to the x-axis (the occurrence and duration in time), it also appears that they considered the magnitude of the data (maximum positive value) more than the amplitude (value and direction). This was particularly evident in the responses given for the structural monitoring scenario. In this case, not only did a majority of respondents chose the area with the highest positive value (which was not related to the physical phenomenon of interest), but also selected follow-up actions that were inconsistent with their response (shoring up the column or the floor, therefore causing an additional displacement in the positive vertical direction). This suggests a possible misinterpretation of the data itself.
High-education respondents also questioned data more frequently than low-knowledge users. Twelve high education vs. four low knowledge participants responded that 'the readings of the sensor do not look correct and probably the sensor needs to be replaced'. In addition, only highly educated respondents chose the option of 'doing nothing' when asked what action they would take based on the monitoring data, and motivated their choice with the need of getting more information on how data were collected or on comparable data from other studies. This is somewhat consistent with behaviour of more highly educated users in other information-seeking contexts, whereas they are often more skeptical and likewise more rigorous in their evaluation of sources of information (Brand-Gruwel et al. 2017;Kohnen and Mertens 2019).
Given these potential issues regarding the use of visualizations and data tools, it might be worthwhile to provide additional tool flexibility that allows all users to customize and adjust their visualizations to best serve their specific needs. For example, allowing users to adjust visualization characteristics like timescale, range of values or compare across multiple sensors/data streams simultaneously might enable more appropriate understanding and decision making relative to the presented data. This ability to customize visualizations and interfaces is, to a certain extent, already implemented in some commercial dashboards and is a common suggestion in most human-computer interaction or usability literature, as such customization not only often makes interfaces more usable and accurate, but also increases the likelihood that these tools will be used in the future (Shneiderman and Hochheiser 2001;Febretti and Garzotto 2009). Different timescales could be linked to metadata related to the building service life to support data trend comprehension and better associate data to physical phenomena.
Related to the above point on customization, it might also be worthwhile to consider the cognitive characteristics of users and their interest in the SHM data, and how that might likewise impact interaction with, and use of, any data visualization tool or scaffold. Research has suggested that there are multiple cognitive facets that might impact the effective use of visualizations (Yi 2012;Ziemkiewicz et al. 2012), and catering to these differences (e.g. prior knowledge, personality, interest, etc.) might enable more reliable and robust visualization use. For example, in the current study more educated users appear to request more information, and are somewhat reticent to implement changes without additional information. Allowing this type of users the flexibility to query the system, and thus satisfy their informationsearching needs could ultimately allow them to perform at a much higher level, with much more confidence. Conversely, individuals who are low in knowledge might need to be protected or guided in such a way that they are less susceptible to the unintentional side effects of visualized data (e.g. focusing on peaks, rather than a more holistic view of the data). Creating sets of pre-canned 'view modes', presets or dashboards, which emphasize certain characteristics like time or amplitude (and not others), might provide a more palatable way for novices to view data (Elias and Bezerianos 2011;Vazquez-Ingelmo et al. 2019). In short, as, by definition, novices are not entirely clear on what is the best way to approach data, providing a cognitive bootstrap of sorts that pre-formats the visualization might allow these novices to use the presented information more effectively. Additionally, it will hopefully limit the amount of incorrect interpretations or errors these novices make. Previous research has suggested that such pre-formatted visualization options can significantly increase data reasoning performance (North and Shneiderman 2000). As the current study demonstrated that low knowledge users, in particular, were more likely to go along with the suggestions made by the scaffolding, having well-vetted and reliable presets might further ensure that correct interpretations of data are made.
The type of scaffolding provided in both monitoring scenarios was not particularly effective in supporting decision-making activities. As the utilization of monitoring data have a variety of motivations, depending on the specific data user, the information derived from the data should be tailor-fit for the intended purpose (Glaser and Tolman 2008). The fact that the provided scaffolding did not refer explicitly to requirement levels against which data could be judged (for instance, maximum allowed vertical column shortening, etc.) made the selection of a course of action among the different proposed alternatives challenging.

Conclusions, limitations and future research
This study presented results of an investigation into the interpretation and utilization of SHM data of mass timber structures. Sensor data were presented in the form of line graphs and, in an alternate version of the survey, with scaffolding aimed at pointing out the influence of some physical phenomena related to this type of structures. The study showed a positive effect of scaffolding in supporting data interpretation of novices. More highly educated respondents often expressed the need for more information about the data. The temporal dimension of data and data trends were in general misinterpreted by the majority of the respondents.
Respondents had very limited information of the monitored context (e.g. details of the building and location/characteristics of the monitored point, specific timing of the monitoring with respect to the building life cycle) and this limited information may have influenced their responses. Also, the type and clarity of the scaffolding used in this survey may have impacted respondents' performance. The presented data are representative of two monitoring scenarios, and are not exhaustive of different data types from SHM projects. In a complex, long-term monitoring project, data are often heterogeneous, multiscale and multi-temporal, and often represent different phenomena and events, some of them correlated and some not. Thus, it is possible that the cognitive demands in a real monitoring task might well exceed what was explored in this study. Further, real-world studies might also induce additional demands not otherwise captured here given the simple online nature of this study. This could include additional environmental and other demands beyond just cognitive concerns.
The proposed survey was conceived as an initial step towards a more comprehensive study investigating different types of visualization formats and scaffolding techniques to support comprehension and utilization of SHM data of a monitored timber building, the Forest Science Complex at Oregon State University. However, conclusions from this study can be generalized and applied to other buildings and types of structures.
Characterizing the application domain represents a considerable challenge for the visualization designer due to the variety of data involved, the phenomena of interest and the different uses of the derived information. Since the visualization designers may not have sufficient domain knowledge to extract or even understand the different user's needs, tailored 'user requirement' tasks could be implemented to more effectively use graph visualization in SHM-driven decision-making processes.
Preliminary conclusions drawn from this study suggest that customization of visualization tools should support both independent data-query and preset view modes. Visualization tools should facilitate investigation of the temporal dimension of data to support trend analysis, for instance by dynamically changing time scale. While signal post-processing techniques can be applied to sensor data to eliminate noise and outliers, it is important to provide enough information to more highly educated users to understand the reasons of certain anomalies and, in general, on how the data are collected. This requires the use of dashboards supporting different types of data postprocessing and updatable metadata, including information about the building, its material and construction system, and events related to its life cycle (e.g. construction, repair works, etc.).

Disclosure statement
No potential conflict of interest was reported by the author(s).