Abstract:
Unsupported excavations are very common in Pakistan for foundation construction. Design aids or statutory laws to ensure safe excavation practices do not exist. Unsupported excavations pose settlement, distortion, and pit face failure threats to adjacent buildings. The study intends to find the range of influence zone of unsupported excavations. Data was collected from 44 sites in Mirpur Azad Kashmir, Pakistan. The parameters collected included excavation depth, distance of structures from excavation; foundation width and load of the adjacent building; unit weight, cohesion, and elastic modulus of the soil. Three Machine Learning (ML) models i.e., Multi-layer Perceptron (MLP), Gene Expression Programming (GEP) and Decision Tree (DT) were trained to predict if a combination of the seven parameters was safe against stability damage. MLP model showed maximum accuracy of 89%. The average of the influence ranks from Garson 1991 method and local sensitivity analysis deemed unit weight and elastic modulus as the least influential parameters. The remaining 5 parameters were taken as variables to be modelled in PLAXIS to produce design charts. The most recurrent values from the dataset were taken as inputs. 15 design charts were made incorporating cohesion values 5, 10 and 15 kPa; foundation widths of 7.5, 10, 13.5,16 and 20 meters and loads of 10, 20 and 30 kPa. To get comparable design charts, MLP model was provided with the same inputs as well. Design charts from both approaches were validated against 14 case studies. The accuracy of PLAXIS and MLP design charts proved to be 78.6% and 85.7% respectively. The MLP model and design charts are recommended for safe practice of unsupported excavation for residential units in soft homogenous soils.
