Abstract
In modern industrial monitoring, the volume of sensor data often exceeds the capacity for full real-time observability due to bandwidth and computational constraints. This paper presents a novel approach to Partially Observable Multi-sensor Sequential Change Detection (POMSCD) by integrating Bayesian inference with a Combinatorial Multi-Armed Bandit (CMAB) framework. Our method leverages inter-sensor correlations to infer the state of unobserved variables and employs an adaptive sampling strategy to prioritize sensors that maximize detection power. Experimental results on synthetic and real-world semiconductor manufacturing data demonstrate that our CMAB-based approach significantly reduces the Average Detection Delay (ADD) compared to standard sampling heuristics.
Introduction
The rapid growth of the Internet of Things (IoT) and complex industrial systems, such as power grids and semiconductor fabrication lines, has led to environments monitored by hundreds or thousands of sensors. The primary goal is sequential change detection: identifying an anomaly or "change-point" as quickly as possible after it occurs. However, full observability is frequently unattainable due to limited battery life in wireless nodes, restricted communication bandwidth, or processing bottlenecks.
The challenge lies in determining which subset of $m$ sensors to monitor out of $p$ total sensors at each time step to minimize detection delay. This paper addresses this "Partially Observable" challenge by treating sensor selection as an online optimization problem.
Methodology
Online Weighted Bayesian Change Detection
A key innovation of this work is the move away from treating sensors as independent entities. By exploiting the underlying correlation structure of the system, the proposed model can update the posterior probability of a change even for sensors that are not currently being sampled.
The system utilizes a Bayesian posterior estimator with time-decayed weights, ensuring that recent observations have a greater impact on the test statistic, $\Lambda(n)$. This statistic is derived from a confidence region of the system's state; when it exceeds a predefined threshold, an alarm is triggered.
Adaptive Sampling via CMAB
To solve the selection problem, the authors formulate it as a Combinatorial Multi-Armed Bandit (CMAB). In this setup:
- Arms: Each sensor represents an "arm" that can be pulled.
- Reward: The reward is defined as the "detection power"—specifically, the non-centrality parameter of the test statistic's distribution.
- Strategy: A Upper Confidence Bound (UCB) algorithm is employed to balance exploration and exploitation. Exploitation focuses on sensors showing signs of deviation, while exploration ensures that no part of the system remains unmonitored for too long, preventing "blind spots."
Experimental Results
Synthetic Data Performance
The researchers conducted extensive simulations with $p=10$ and $p=100$ sensors to evaluate the Average Detection Delay (ADD). The CMAB approach was compared against several baselines, including Random sampling (RAND), Top-r Adaptive Sampling (TRAS), and Numerical Adaptive Sampling (NAS).
As shown in the results, the CMAB strategy significantly outperforms traditional methods. For instance, in a 100-sensor environment with a change magnitude of $\delta=1$, the CMAB approach achieved an ADD of approximately 8.91, while random sampling lagged behind at 190. This highlights the efficiency of the UCB-based exploration strategy in high-dimensional spaces.
Real-World Application: SECOM Dataset
The framework was also validated using the SECOM dataset from the UCI Machine Learning Repository, which contains 1,567 samples with 31 variables from a semiconductor process. The CMAB strategy maintained its superior performance even under tight sensing constraints (e.g., monitoring only 3 or 5 variables), demonstrating its practical utility in real-world industrial settings.
Conclusion
This research demonstrates that modeling sensor selection as a CMAB problem, combined with Bayesian correlation modeling, provides a robust solution for anomaly detection in partially observable systems. By mathematically balancing the need to investigate suspicious signals with the need to maintain global awareness, the proposed method achieves near-optimal detection speeds. Future work will focus on reducing the computational complexity for extremely large-scale systems and extending the framework to non-Gaussian data distributions.
Summary updated