Deep neural network compression for artificial intelligent of Things (AIoT)

Abstract

Stay cable is a very important part of a cable-stayed bridge. Structural Health Monitoring (SHM) is used for cable damage detection as the cable may fail due to external factors. A Long Short-Term Memory Fully Convolution Network (LSTM-FCN) method is used to detect cable damage by measuring cable forces by recognizing biased patterns from intact conditions. This would help solve the pattern recognition problem for cable damage detection through Time Series Classification (TSC). The TSC is trained and validated using data collected from the stay cables, setting the segmented data series as input and cable ID as class a label. Two series were investigated, and raw time series of cable and cable ratios were considered. Further reducing the size of the model, pruning is done through Constant Sparsity and Polynomial Decay and comparing the two to observe which would take up the smaller size while maintaining or improving accuracy. The model is further quantized after pruning to further compress its size for interference while making them faster and keeping as accurate as possible through Post-Training Quantization and Aware Training Quantization. An additional method of reducing size was also used such as Depthwise Separable Convolution 1D is used to replace Fully Convolution Network (FCN). This study proposes a methodology that would reduce the memory of models by sieving through redundant parameters to ensure better and faster interference while still maintaining or improving the accuracy of the model.