Storage Swapping Dual-Processor Edge Device Architecture for Acoustic Non-Real-Time Multi-Model Inference Applications

Edge devices are critical components of any Internet of Things (IoT) application that collects data and performs initial processing. Processing often involves classifying a time series for known patterns. Currently, most applications perform classification as quickly and continuously as possible, which is well-suited to real-time use cases. However, continuous processing consumes power continuously. For applications that do not require real-time processing, using multiple processors on a single device can distribute the data collection and processing workload. The key issue with such an architecture is sharing storage for recent data and properly timing the classification process to run periodically. In this paper, a switched storage-swapping mechanism is used to reduce energy consumption by recording data in a single storage and sharing it between a continuous, low-power recording processor and a periodic classifier processor. Additionally, to further reduce processing, data is stored in segments, which are processed individually. An electronic feature detection (EFD) mechanism is discussed that identifies traits of target patterns using purely electronic components to tag segments of interest for classification. These further reduce energy and time-related costs, even compared to a single-processor architecture with low- and high-power modes corresponding to data collection and classification, respectively. We use an audio classification problem, augmented with an RC filter as the EFD. This architecture enables the creation of complex prototype devices for initial deployments that can support multiple machine learning and deep learning models of any size, before moving to dedicated commercial devices that use the best-performing models. With this approach, specific models may be chosen in specific contexts to improve accuracy.