Embedding principal component analysis inference in expert sensors for big data applications

Abstract

The increasing relevance of big data applications in fields as the Internet of Things (IoT) and Industry 4.0 implies that sensors are requested to be secure and accurate. In the last years, sensors are evolving toward complex monitoring functionalities, increasing the complexity of data, meaning that the analysis stage is usually performed away from the sensor layer, i.e., the fog or the cloud. This separation entails issues for response time and security. As a possible way to address this data analysis closer to the edge, embedded machine-learning (ML) techniques have shown to be a good solution, leading to expert sensors. Feature extraction tools, as principal component (PC) analysis (PCA), might offer a solution to reduce the amount of data transmitted through the network, adding additional security because information is not transmitted as raw data. However, PCA is time-consuming and therefore, it should be carefully optimized according to the hardware used in the sensor device. This chapter proposes to embed the PCA inference stage in a low-cost field-programmable system on chip (SoC) (FPSoC) while performing a design space exploration for a general PCA inference problem. To this end, the authors analyze metrics, such as latency, scalability, and usage of hardware resources. The resulting architectures are compared to a multicore OpenMP approach to be executed in an ARM processor, analyzing the advantages of using the FPSoC implementation in speedup.