ADEPOS: A Novel Approximate Computing Framework for Anomaly Detection Systems and its Implementation in 65-nm CMOS

Sumon Kumar Bose, Bapi Kar, Mohendra Roy, Pradeep Kumar Gopalakrishnan, Lei Zhang, Aakash Patil, Arindam Basu

Introduction

As the Internet of Things (IoT) shifts toward edge computing, sensor nodes are increasingly required to process data locally to minimize the energy costs of wireless transmission. However, edge devices operate under extremely tight energy budgets. Anomaly detection—critical for applications like industrial predictive maintenance and medical monitoring—presents a unique computational profile: data is "normal" for the vast majority of the time. Traditional systems waste significant energy by running high-precision, complex models continuously on this routine data.

This paper introduces ADEPOS (Anomaly Detection Based Power Savings), a framework designed to be "computationally lazy" during normal operations and "highly alert" only when potential anomalies are detected. By combining algorithmic innovations with custom 65-nm CMOS hardware, ADEPOS achieves up to an 8.95x reduction in energy consumption while maintaining 100% detection accuracy.

The ADEPOS Framework

Dynamic Network Scaling (DNS)

At the core of ADEPOS is Dynamic Network Scaling. Rather than utilizing a single, monolithic neural network, the system employs an ensemble of "Base Learners" based on mini-Extreme Learning Machines (ELMs). The system initiates detection using a minimal subset of learners (e.g., one or three). If these learners signal a potential fault or disagree, the system dynamically activates additional learners to confirm the finding. This hierarchical approach ensures that high computational costs are only incurred when the data warrants closer scrutiny.

Figure 3: Flowchart of the Dynamic Network Scaling (DNS) logic, showing how the number of base learners (N_BL) is adjusted based on fault detection.

Approximate Computing and Neuron Generation

To further reduce the workload, ADEPOS utilizes lower precision and complexity during normal operation. A key mathematical innovation is the Hidden Neuron Generation (NG) technique. By subtracting the outputs of physical neurons, the system creates "virtual" neurons with distinct random weight distributions. This allows a small physical network to mimic the performance of a much larger one with minimal additional Multiply-Accumulate (MAC) operations, significantly reducing the memory and computational footprint.

Hardware-Software Co-design

The energy efficiency of ADEPOS is not purely algorithmic; it relies on tight integration with custom hardware. The authors fabricated a test chip in 65-nm CMOS featuring an OpenMSP430 microprocessor and a custom on-chip DC-DC buck converter.

Figure 4: System block diagram showing the integration of the custom BUCK converter with the microprocessor (uP) to enable Dynamic Voltage and Frequency Scaling (DVFS).

This architecture supports Dynamic Voltage and Frequency Scaling (DVFS), allowing the hardware to drop its operating voltage (in 50mV steps) and frequency when the software selects lower-complexity models. The chip reached its minimum energy operating point at 720mV and 8.5MHz.

Experimental Results

The framework was validated using two real-world datasets:

  • NASA Bearing Dataset: Vibration data for industrial machinery maintenance.
  • UPenn and Mayo Clinic Seizure Detection: EEG data for identifying epileptic seizures.

On the NASA dataset, ADEPOS achieved an 8.95x energy reduction compared to fixed high-complexity models. For seizure detection, energy savings ranged from 6.56x to 13.1x depending on patient-specific data. Crucially, because the system always scales back to full precision to confirm a final alarm, it maintained 100% detection accuracy across all tests.

Conclusion and Future Work

ADEPOS demonstrates that significant energy savings in IoT can be achieved by optimizing for the most common case: normal data. While the system currently faces some memory overhead from storing inactive base learners, the NG technique helps mitigate this. Future work may explore deeper ASIC integration and power-gating specific chip regions to push energy efficiency even further, making sophisticated anomaly detection viable for the smallest battery-powered sensors.

Summary updated

Abstract

To overcome the energy and bandwidth limitations of traditional IoT systems, “edge computing” or information extraction at the sensor node has become popular. However, now it is important to create very low energy information extraction or pattern recognition systems. In this paper, we present an approximate computing method to reduce the computation energy of a specific type of IoT system used for anomaly detection (e.g. in predictive maintenance, epileptic seizure detection, etc). Termed as Anomaly Detection Based Power Savings (ADEPOS), our proposed method uses low precision computing and low complexity neural networks at the beginning when it is easy to distinguish healthy data. However, on the detection of anomalies, the complexity of the network and computing precision are adaptively increased for accurate predictions. We show that ensemble approaches are well suited for adaptively changing network size. To validate our proposed scheme, a chip has been fabricated in UMC 65nm process that includes an MSP430 microprocessor along with an on-chip switching mode DC-DC converter for dynamic voltage and frequency scaling. Using NASA bearing dataset for machine health monitoring, we show that using ADEPOS we can achieve 8.95X saving of energy along the lifetime without losing any detection accuracy. The energy savings are obtained by reducing the execution time of the neural network on the microprocessor.
BibTeX
@article{bose2020adepos,
  author = {Sumon Kumar Bose and Bapi Kar and Mohendra Roy and Pradeep Kumar Gopalakrishnan and Lei Zhang and Aakash Patil and Arindam Basu},
  title = {{ADEPOS: A Novel Approximate Computing Framework for Anomaly Detection Systems and its Implementation in 65-nm CMOS}},
  journal = {IEEE Trans. Circuits Syst. I Regul. Pap.},
  year = {2020},
  doi = {10.1109/tcsi.2019.2958086},
}
Figure