Design and Implementation of an FPGA-Based DNN Architecture for Real-time Outlier Detection

Deep neural networks (DNNs) have recently become the standard tool for solving practical problems in various applications, including timely data analysis and near real-time accurate decision-making. DNNs have proven effective in outlier detection, one of sensor networks' primary motivating data analysis applications. Despite the great potential of deep neural networks, their computational resource requirements create a vast gap when it comes to the fast processing time required in real-time applications using low-power, low-cost edge devices. Special care must be taken into account when designing DNNs computational units. This work proposes an FPGA-based Deep Neural Network (DNN) architecture for real-time outlier detection in time series data. The proposed architecture integrates a fine-tuned Autoencoder network and a Long short-term memory (LSTM) network to predict and detect outliers in real-time. The hardware accelerator of the integrated networks combines serial-parallel computation with matrix algebra concepts to reduce computational complexity and enhance the throughput. Experimental results on the resource-constrained Xilinx PYNQ-Z1 board using an open-source sensor network dataset show that the proposed architecture can efficiently analyze and detect outliers in real-time. The implemented design achieves 0.22 ms average latency and 1GOPS throughput. The proposed design's low latency and 94mW power consumption make it suitable for resource-constrained edge platforms.