Edge-Compatible Convolutional Autoencoder Implemented on FPGA for Anomaly Detection in Vibration Condition-Based Monitoring

Estimating the state of the health of machines using vibration signals analyzed with complex deep-learning models poses memory and processing speed constraints for edge implementation, whereas cloud-based computing poses high latency, high cost of data transmission and storage, and privacy threats. Moreover, in most practical settings, a large amount of data may be available or easily obtainable for machines in healthy/normal operating condition, and not in anomalous conditions. Thus, there exists a need to develop lightweight unsupervised learning based algorithms that can be implemented on an embedded platform for data processing and estimating state of health. In this letter, we present a lightweight convolutional autoencoder (CAE) implemented on a low-cost FPGA platform (PYNQ-Z2, Xilinx Zynq-7020 SoC) to discern anomalies in vibration data sets. We present an extensive performance analysis of various encoding methods to represent 1-D time-series as images and the feasibility analysis of FPGA-implementation of CAE. The optimized model has less than 10 000 trainable parameters and shows the highest accuracy (
$>$
88% on CPU,
$>$
85% with fixed point arithmetic on FPGA) using scalogram encoding when trained and tested on the Airbus SAS helicopter accelerometer dataset.