Research On Real-time Adjusting Of Self-compacting Concrete Workability Based On Deep Learning

The workability characteristics of self-compacting concrete(SCC)is strongly affected by the fluctuating in situ conditions,especially the fluctuating moisture content of SCC mixtures.It has been proved accurate using deep learning model as a predictor of SCC workability,including slump flow value(SF)and V-funnel time(VF).The deep learning approach provided an alternative to SCC workability tests,which could be further utilized for estimating the adjusting amount of material automatically.Thus a smart mixer could be developed.In this thesis,both image processing and model computation methods were used for a detailed study on the combination of SCC experiments and deep learning,which was essential and significant within the research.Deep learning approach needs a large amount of data for training to obtain a better generalization ability.However,only limited data was presented via SCC mixing experiment.In this thesis,every mixing cycle was regarded as a separate SCC mixing experiment,based on the characteristics that SCC videos contained several constant mixing cycles.Thus,nearly 1,500 image sequences were generated using only 31 experiments.The approach of data augmentation was further used to expand the data amount to the size of 4,000 to 12,000.The model performance proved the data to be sufficient,and the data preparation to be practical.The thesis presented a procedure of pre-processing,consisting of converting RGB images to grayscale,affine transformation,extracting region of interest and histogram equalization,preventing training from overfitting problem.The redundant information was thus erased.The thesis also presented a down-sampling operation to the SCC image sequences,using various time resolution.This operation was used to reduce computation complexity and training time consumption.A further quantitative analysis indicated that spatial and temporal features belonged to different SCC videos existed differences.Convolutional neural networks and recurrent neural networks were chosen as parts of model structure to carry out corresponding tasks.The proposed model was then built by combining convolutional and recurrent parts together.The hyper-parameters were optimized through mini-numerical experiments.The deep learning model was trained after data prepared and model built.The accuracy and feasibility were validated via running the proposed model on testing set.The model performances were compared,varying the output-space dimensions and time resolutions.Results showed that a given approach was preferred to obtain better model performance.The model with SF and VF as outputs had a better performance than one with SF as output only.The coefficient of determination was up to 0.95.The SCC workability characteristics,especially the slump flow values were chosen as an index to estimate the moisture content of SCC mixture because of the linear relationship.A method for estimating the overall moisture of an SCC mixture to guide batch weight adjusting.The method includes a laboratory-based concrete test and an in situ concrete mixing process.The overall moisture estimates,determined by slump flow values,were confirmed to be reliable through verification experiments.The proposed method may serve as an alternative to manual or automated moisture measurement.