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 UMC65nm 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.
Introduction
Internet of Things (IoT) devices for industrial monitoring and medical sensing are severely constrained by battery life. While edge computing reduces the need for energy-intensive data transmission to the cloud, running complex neural networks locally remains a challenge. Most anomaly detection systems spend over 99% of their operational lifetime monitoring "healthy" data. ADEPOS (Anomaly Detection Based Power Savings) leverages this observation by using a "lazy" computing approach: performing low-precision, low-complexity work when systems are healthy and only scaling up resources when an anomaly is suspected.
The ADEPOS Framework
The core of ADEPOS is a Dynamic Network Expansion (DNS) strategy using an ensemble of "Base Learners" (BLs). These learners are implemented as Extreme Learning Machines (ELMs), which are computationally efficient because they use random weights for the hidden layer and only require training for the output layer.
The system starts by processing data with a minimal number of BLs. If the initial learners agree the data is healthy, the system remains in a low-power state. If a potential anomaly is detected, the framework dynamically activates additional learners to confirm the finding. This ensures that high computational precision is only applied at the decision boundaries where it is most needed.
Furthermore, the authors employ a "virtual neuron" technique. By subtracting the outputs of physical neurons, the system can generate additional hidden neurons without the need for expensive multiplication operations, significantly reducing the Multiply-Accumulate (MAC) count.
Hardware Implementation and Co-design
To demonstrate the framework, a custom chip was fabricated in a 65nm CMOS process. The design features an MSP430-like microprocessor integrated with an on-chip switching-mode DC-DC buck converter. This integration is crucial for Dynamic Voltage and Frequency Scaling (DVFS). During periods of healthy data monitoring, the hardware can physically lower its supply voltage and frequency to save power, ramping up only when the ADEPOS algorithm triggers a higher-complexity state.
Experimental Results
The ADEPOS framework was validated using the NASA Bearing Dataset for predictive maintenance and the UPenn/Mayo Clinic Seizure Detection Challenge for medical monitoring.
- Energy Efficiency: The system achieved an 8.95X reduction in energy consumption on the NASA dataset compared to fixed-size networks. For seizure detection, it achieved a 6.56X average energy saving while maintaining 100% sensitivity.
- Memory Footprint: The virtual neuron generation technique reduced the program memory footprint by 64% to 69%, which is vital for memory-constrained edge devices.
- Hardware Performance: The on-chip buck converter maintained a stable output ripple of 3.8mV even during voltage scaling, ensuring microprocessor reliability at near-threshold power levels.
Conclusion
ADEPOS provides a robust framework for energy-efficient edge-based anomaly detection. By combining algorithmic approximate computing with hardware-level voltage scaling, the system achieves nearly an order of magnitude improvement in energy efficiency without sacrificing accuracy. Future research will focus on extending this dynamic scaling approach to 2D data like edge vision and optimizing for leakage power in advanced CMOS nodes.
Summary updated