BAM Dissertationen ohne Nummerierung
Filtern
Erscheinungsjahr
- 2023 (2)
Dokumenttyp
- Dissertation (2) (entfernen)
Sprache
- Englisch (2)
Referierte Publikation
- nein (2)
Schlagworte
- Building floor (1)
- DLP (1)
- DMD (1)
- Defect characterization (1)
- Defektcharakterisierung (1)
- Ground penetrating radar (1)
- Moisture measurement (1)
- NDT (1)
- Nondestructive testing (1)
- Structured illumination (1)
Organisationseinheit der BAM
- 8.0 Abteilungsleitung und andere (2) (entfernen)
In the event of moisture deterioration, rapid detection and localization is particularly important to prevent further deterioration and costs. For building floors, the layered structure poses a challenging obstacle for most moisture measurement methods. But especially here, layer-specific information on the depth of the water is crucial for efficient and effective repairs. Ground Penetrating Radar (GPR) shows the potential to generate such depth information. Therefore, the present work investigates the suitability of GPR in combination with machine learning methods for the automated classification of the typical deterioration cases (i) dry, (ii) wet insulation, and (iii) wet screed.
First, a literature review was conducted to identify the most common methods for detecting moisture in building materials using GPR. Here, it especially became clear that all publications only investigated individual time-, amplitude- or frequency features separately, without combining them. This was seen as a potential aspect for innovation, as the multivariate application of several signal features can help to overcome individual weaknesses and limitations.
Preliminary investigations carried out on drying screed samples confirmed the profitable use of multivariate evaluations. In addition to the general suitability and dependencies of various features, first limitations due to possible interference between the direct wave and the reflection wave could be identified. This is particularly evident with thin or dry materials, for which the two-way travel times of the reflected radar signals become shorter.
An extensive laboratory experiment was carried out, for which a modular test specimen was designed to enable the variation of the material type and thickness of screed and insulation, as well as the simulation of moisture deteriorations. The data collected revealed clear differences between dry and deteriored structures within measured B-scans. These deviations were to be detected with the newly introduced B-scan features, which evaluate the statistical deviation of A-scan features within a survey line. In this way, deteriorations to unknown floor structures are recognized, regardless of the material parameters present. In a subsequent training and cross-validation process of different classifiers, accuracies of over 88 \% of the 504 recorded measurements (252 different experimental setups) were achieved. For that, the combination of amplitude and frequency features, which covered all relevant reflections of the radar signals, was particularly beneficial. Furthermore, the data set showed only small differences between dry floors and deteriored screeds for the B-scan features, which could be attributed to a homogeneous distribution of the added water in the screeds. The successfully separation of these similar feature distributions raised the suspicion of overfitting, which was examined in more detail by means of a validation with on-site data.
For this purpose, investigations were carried out at five different locations in Germany, using the identical measurement method like in the laboratory. By extracting drilling cores, it was possible to determine the deterioration case for each measurement point and thus generate a corresponding reference. However, numerous data had to be sorted out before classification, since disturbances due to underfloor heating, screed reinforcements, steel beams or missing insulation prevented comparability with the laboratory experiments. Validation of the remaining data (72 B-scans) achieved only low accuracy with 53 \% correctly classified deterioration cases. Here, the previously suspected overfitting of the small decision boundary between dry setups and deteriored screeds within the laboratory proved to be a problem. The generally larger deviations within (also dry) on-site B-scans were thus frequently misclassified as screed deterioration. In addition, there were sometimes strongly varying layer thicknesses or changing cases of deterioration within a survey line, which caused additional errors due to the local limitation of the drilling core reference. Nevertheless, individual on-site examples also showed the promising potential of the applied signal features and the GPR method in general, which partly allowed a profound interpretation of the measurements. However, this interpretation still requires the experience of trained personnel and could not be automated using machine learning with the available database. Nevertheless, such experience and knowledge can be enriched by the findings of this work, which provide the basis for further research.
Future work should aim at building an open GPR data base of on-site moisture measurements on floors to provide a meaningful basis for applying machine learning. Here, referencing is a crucial point, whose limitations with respect to the moisture present and its distribution can easily reduce the potential of such efforts. The combination of several reference methods might help to overcome such limitations. Similarly, a focus on monitoring approaches can also help to reduce numerous unknown variables in moisture measurements and increase confidence in the detection of different deterioration cases.
Laser-based active thermography is a contactless non-destructive testing method to detect material defects by heating the object and measuring its temperature increase with an infrared camera. Systematic deviations from predicted behavior provide insight into the inner structure of the object. However, its resolution in resolving internal structures is limited due to the diffusive nature of heat diffusion. Thermographic super resolution (SR) methods aim to overcome this limitation by combining multiple thermographic measurements and mathematical optimization algorithms to improve the defect reconstruction.
Thermographic SR reconstruction methods involve measuring the temperature change in an object under test (OuT) heated with multiple different spatially structured illuminations. Subsequently, these measurements are inputted into a severely ill-posed and heavily regularized inverse problem, producing a sparse map of the OuT’s internal defect structure. Solving this inverse problem relies on limited priors, such as defect-sparsity, and iterative numerical minimization techniques. Previously mostly experimentally limited to one-dimensional regions of interest (ROIs), this thesis aims to extend the method to the reconstruction of two-dimensionalROIs with arbitrary defect distributions while maintaining reasonable experimental complexity. Ultimately, the goal of this thesis is to make the method suitable for a technology transfer to industrial applications by advancing its technology readiness level (TRL).
In order to achieve the aforementioned goal, this thesis discusses the numerical expansion of a thermographic SR reconstruction method and introduces two novel algorithms to invert the underlying inverse problem. Furthermore, a forward solution to the inverse problem in terms of the applied SR reconstruction model is set up. In conjunction with an additionally proposed algorithm for the automated determination of a set of (optimal) regularization parameters, both create the possibility to conduct analytical simulations to characterize the influence of the experimental parameters on the achievable reconstruction quality. On the experimental side, the method is upgraded to deal with two-dimensional ROIs, and multiple measurement campaigns are performed to validate the proposed inversion algorithms, forward solution and two exemplary analytical studies. For the experimental implementation of the method, the use of a laser-coupled DLP-projector is introduced, which allows projecting binary pixel
patterns that cover the whole ROI, reducing the number of necessary measurements per ROI significantly (up to 20x).
Finally, the achieved reconstruction of the internal defect structure of a purpose-made OuT is qualitatively and qualitatively benchmarked against well-established thermographic testing methods based on homogeneous illumination of the ROI. Here, the background-noise-free two-dimensional photothermal SR reconstruction results show to outclass all defect reconstructions by the considered reference methods.