Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

VARADE: a Variational-based AutoRegressive model for Anomaly Detection on the Edge

Alessio Mascolini, Sebastiano Gaiardelli, Francesco Ponzio, Nicola Dall'Ora, Enrico Macii, Sara Vinco, Santa Di Cataldo, Franco Fummi

TL;DR

VARADE introduces a variationally grounded autoregressive model tailored for edge-based anomaly detection in Industry 4.0. By predicting the next time step as a Gaussian distribution through a light convolutional architecture, and optimizing via a joint reconstruction and KL-divergence loss, the model provides a robust anomaly score based on predictive uncertainty. In a real industrial case with a KUKA manipulator, VARADE outperformed lightweight baselines in AUC-ROC while maintaining low latency and reasonable power and memory use on both Jetson Xavier NX and Jetson AGX Orin platforms. The approach demonstrates the practicality of on-device, real-time anomaly detection for complex, multivariate streams, with direct implications for edge AI deployment in smart factories.

Abstract

Detecting complex anomalies on massive amounts of data is a crucial task in Industry 4.0, best addressed by deep learning. However, available solutions are computationally demanding, requiring cloud architectures prone to latency and bandwidth issues. This work presents VARADE, a novel solution implementing a light autoregressive framework based on variational inference, which is best suited for real-time execution on the edge. The proposed approach was validated on a robotic arm, part of a pilot production line, and compared with several state-of-the-art algorithms, obtaining the best trade-off between anomaly detection accuracy, power consumption and inference frequency on two different edge platforms.

VARADE: a Variational-based AutoRegressive model for Anomaly Detection on the Edge

TL;DR

VARADE introduces a variationally grounded autoregressive model tailored for edge-based anomaly detection in Industry 4.0. By predicting the next time step as a Gaussian distribution through a light convolutional architecture, and optimizing via a joint reconstruction and KL-divergence loss, the model provides a robust anomaly score based on predictive uncertainty. In a real industrial case with a KUKA manipulator, VARADE outperformed lightweight baselines in AUC-ROC while maintaining low latency and reasonable power and memory use on both Jetson Xavier NX and Jetson AGX Orin platforms. The approach demonstrates the practicality of on-device, real-time anomaly detection for complex, multivariate streams, with direct implications for edge AI deployment in smart factories.

Abstract

Detecting complex anomalies on massive amounts of data is a crucial task in Industry 4.0, best addressed by deep learning. However, available solutions are computationally demanding, requiring cloud architectures prone to latency and bandwidth issues. This work presents VARADE, a novel solution implementing a light autoregressive framework based on variational inference, which is best suited for real-time execution on the edge. The proposed approach was validated on a robotic arm, part of a pilot production line, and compared with several state-of-the-art algorithms, obtaining the best trade-off between anomaly detection accuracy, power consumption and inference frequency on two different edge platforms.
Paper Structure (15 sections, 7 equations, 3 figures, 2 tables)

This paper contains 15 sections, 7 equations, 3 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Background and related works
  3. Methods
  4. Proposed solution
  5. Derivation of the loss function
  6. Baseline solutions
  7. Implementation details
  8. Industrial case study
  9. Kuka anthropomorphic manipulator
  10. Data stream characterization
  11. Experimental setup
  12. Experimental results
  13. Jetson Xavier NX
  14. Jetson AGX Orin
  15. Conclusions and future works

Figures (3)

  • Figure 1: Architecture of VARADE. Current ($t_0$) and past time steps ($t_{-1} \dots t_{-T}$) are processed by a cascade of convolutional layers and a final linear projection. The output is the estimated probability distribution of the next time step, $P(t_{1})$.
  • Figure 2: Case study setup. The KUKA manipulator is instrumented with different sensors: 7 accelerometers with six axes (one for each joint) and one single-phase power meter. The sensors are connected directly with an embedded board for detecting different classes of anomalies.
  • Figure 3: Inference Frequency vs Accuracy of the anomaly detection models (identified by the marker color). Marker shape represents the adopted edge device (square for Jetson Xavier NX, and triangle for Jetson AGX Orin), while marker size is proportional to power consumption.