Developing a diagnostic assessment tool to evaluate damage in buildings

Abstract

Mining activities at the Groningen gas field are causing earthquakes which result in building damage. This has started a discussion on what type of damage is caused by earthquakes. These discussions are typical in the forensic engineering field, especially in complex cases. The problem is the lack of regulations, in terms of standardizations and uniformization.
To solve that problem, experts can be provided with an independent tool which can contribute to the investigation of the cause of building damage. The tool can help to indicate potential damage causes. This will support the findings of experts. Also, it can draw attention to overlooked damage causes.
The tool is based on relations found in a database of damage cases that have been determined earlier. The database consists of damage cases in the Groningen province. Not all available damage cases were incorporated in this thesis, because processing the damage reports to a database was a labour-intensive job. The analysed dataset consists of 1830 damage cases in 49 buildings. Experts were able to determine the cause of damage in 1180 of these damage cases, which results in a ratio of 64.4% known cases. Only the known cases where applied in the analysis. The buildings were located in seven different areas in the province of Groningen.
Each analysed damage case consists of a damage cause and a description. A description has been structured in 191 characteristics. These characteristics have been categorised into three types: building characteristics, context characteristics and damage characteristics. Building characteristics say something about the function, materials and size of the building. Context characteristics explain the sub soil, vibration sources and external forces in the surrounding of the building. Damage characteristics describe how damage is presented in terms of position, location and shape of damage.
Whether the found relations can be deployed in practice, depends on how useful those relations are. Useful is defined as reliable and meaningful. Reliable is how a found pattern performs according to a test, mostly measured in terms of accuracy or coefficient of determination. Meaningful is whether the found relations are logical to be explained by literature or plausible damage situations. The pattern recognition can introduce some relations and can provide them with a reliability value. However, if the relations are not explainable, they do not mean anything for use in practice.
The relations in these data were found by deploying pattern recognition methods. Two algorithms were utilized as a pattern recognition method: decision tree and linear regression. A decision tree algorithm splits the data into groups by applying thresholds on case characteristics. These thresholds can be made visual in a decision tree figure. Linear regression tries to obtain a target value by means of a linear relation of characteristics. Therefore, the linear regression algorithm determines the slope value of each characteristic.
Classification analyses were done with decision trees on six damage cause categories. The results of that type of analysis were capable of determining if or which damage was caused by a certain cause. Linear regression was performed in order to find regression relations where the technical attributability of a damage cause could be calculated for each case. In the more complex task of regression analysis, only three damage cause categories were suitable for finding a relation.
To determine whether damage was caused by earthquakes, earthquake load in terms of PGV is an important characteristic. Also, the age of a building and trees has a possible significant influence on the occurrence of earthquake damage, according to the found pattern. A relation between those last two characteristics and earthquake damage is not described in literature. Besides that, this decision tree pattern seems to be the most useful pattern for in practice.
Another interesting finding is that hindered deformation mostly occurred at the inside of a building. Combined with other characteristics, a pattern on this damage cause performed with the highest score in this thesis. It has an accuracy of 77%. This means that 77% of the cases in the test set were correctly predicted by the produced classification decision tree. However, the found relation with the characteristics is not always explainable or meaningful so as to be applied in practice. More conclusions of classification analysis are shown in Table 1.
<Table 1, See abstract in report>
The presented findings above are classification relations. Regression analyses were difficult to execute. A desired positive coefficient of determination (R2) could not be reached without subjective interference in the pattern recognition. The best regression result was obtained on damage caused by earthquakes. It had a R2 of 0.48. Which means that 48% of the data was describable in a linear relation. More conclusions of regression analysis are shown in Table 2.
<Table 2, See abstract in report>
It has been interesting to study the relation between characteristics and damage causes. However, the results are not of decisive importance. The building and context characteristics supported by literature were not always selected or applied properly by the pattern recognition. Also, the potential of damage characteristics was not recognized by the algorithms. Nonetheless, the results of earthquake related damage seem promising. They even indicate characteristics which may be worth investigating more closely.