Real-Time Redis Deep-Learned Detection at the Edge
Speaker(s): Ben Jacobs
We have developed an approach for using deep-learned image-processing models over low power, performance constrained edge devices for sensing and control in manufacturing contexts. We have applied this technology to the problem of classifying failures in injection-molded, liquid bottling line. Failures include both bottle-production failures as well as filling failures. Such a vision-based edge sensor has great potential over a wide range of industrial applications, since it (1) may be applied post-hoc to existing manufacturing lines, (2) can be reprogrammed by retraining so it applies over an enormous space w/o new software or hardware development, (3) can cover very wide ranges of failures since these learned models are sensitive to many details of the product.
A solution is needed that won’t impact the speed of production and is a cost-effective remedy that can perform in near real-time, identify and optimize while dealing with intermittent connectivity issues.
Analytics Fire engineers opted for a RedisEdge implementation running on an Nvidia Jetson Nano for real-time defect detection and rejection at key stages of the manufacturing process. This solution provided the flexibility to scale and created cheap autonomous monitoring sensors that could handle intermittent connectivity issues found within the plant.
Our deep-learned models are inherently noisy and generate significant false positives and false negatives. But classification errors on successive frames of the video feed are decorrelated enough that combining a sequence of many such predictions can greatly reduce both classes of errors simultaneously. We use RedisTimeSeries to derive a running prediction signal then show results optimizing window size, sample weight decay by time, and prediction thresholds. We will show the analysis we performed in optimizing these tradeoffs in order to generate a stable and accurate prediction signal.
By leveraging the speed and efficiency of Redis at the edge, Analytics Fire was able to simplify the management of tensor models, reliably moving large amounts of data, and most importantly create a self-reliant edge device capable of making near real-time decisions.