[Paper Review] Bandwidth and Energy Ecient Decentralized Sequential Change Detection

Georgios Fellouris, George V. Moustakides|arXiv (Cornell University)|Jan 1, 2013

Advanced Statistical Process Monitoring35 references3 citations

TL;DR

This paper proposes a decentralized sequential change detection scheme that uses random-time, low-bit transmissions from sensors to a fusion center, where parallel CUSUM tests detect changes efficiently. It proves second-order asymptotic optimality, showing performance loss relative to full-observation detection is bounded as false alarm rates approach zero, even under fixed communication rates.

ABSTRACT

The problem of decentralized sequential change detection is considered, where an abrupt change occurs in an area monitored by a number of sensors; the sensors transmit their data to a fusion center, subject to bandwidth and energy constraints, and the fusion center is responsible for detecting the change as soon as possible. A novel sequential detection rule is proposed that requires communication from the sensors at random times and transmission of only low-bit messages, on which the fusion center runs in parallel a CUSUM test. The second-order asymptotic optimality of the proposed scheme is established both in discrete and in continuous time. Specically, it is shown that the inicted performance loss (with respect to the optimal detection rule that uses the complete sensor observations) is asymptotically bounded as the rate of false alarms goes to 0, for any xed rate of communication. When the rate of communication from the sensors

Motivation & Objective

• To address the challenge of detecting abrupt changes in sensor networks under strict bandwidth and energy constraints.
• To design a decentralized detection scheme that minimizes communication while maintaining detection efficiency.
• To establish theoretical performance bounds showing that the proposed method approaches optimal detection as false alarm rates decrease.

Proposed method

• Sensors transmit data to the fusion center at random times, using only low-bit messages to reduce bandwidth and energy use.
• The fusion center runs parallel CUSUM tests on incoming messages to detect changes in real time.
• The detection rule is designed to be second-order asymptotically optimal in both discrete and continuous time settings.
• Communication is modeled as a Poisson process, enabling random transmission scheduling that reduces average communication rate.
• The scheme leverages the memoryless property of the CUSUM test to maintain detection performance with minimal data exchange.
• Theoretical analysis uses asymptotic optimality criteria to compare performance against the ideal case with full sensor data.

Experimental results

Research questions

• RQ1 How can decentralized change detection be made both bandwidth- and energy-efficient without sacrificing detection speed?
• RQ2 What is the fundamental performance limit of such schemes when communication is constrained to a fixed rate?
• RQ3 Can a random, low-bit transmission strategy achieve near-optimal detection performance in the asymptotic regime of low false alarm rates?
• RQ4 How does the performance of the proposed scheme compare to the optimal centralized detection rule using full sensor observations?
• RQ5 What is the asymptotic behavior of the detection delay and false alarm rate trade-off under communication constraints?

Key findings

• The proposed scheme achieves second-order asymptotic optimality, meaning its detection delay performance approaches that of the optimal centralized rule as the false alarm rate tends to zero.
• The performance loss due to communication constraints is asymptotically bounded, regardless of the fixed communication rate.
• The use of random-time, low-bit transmissions maintains detection efficiency while significantly reducing energy and bandwidth usage.
• The CUSUM test applied in parallel at the fusion center ensures strong detection performance despite limited data.
• The theoretical analysis confirms that the scheme is robust to communication rate limitations in both discrete and continuous time models.
• The results hold under general conditions on sensor observations and communication processes, demonstrating broad applicability.