Abstract
In this paper, we consider the problem of acoustic source localization by acoustic sensor networks (ASNs) using a promising, learning-based technique that adapts to the acoustic environment. In particular, we look at the scenario when a node in the ASN is displaced from its position during training. As the mismatch between the ASN used for learning the localization model and the one after a node displacement leads to erroneous position estimates, a displacement has to be detected and the displaced nodes need to be identified. We propose a method that considers the disparity in position estimates made by leave-one-node-out (LONO) sub-networks and uses a Markov random field (MRF) framework to infer the probability of each LONO position estimate being aligned, misaligned or unreliable while accounting for the noise inherent to the estimator. This probabilistic approach is advantageous over naïve detection methods, as it outputs a normalized value that encapsulates conditional information provided by each LONO sub-network on whether the reading is in misalignment with the overall network. Experimental results confirm that the performance of the proposed method is consistent in identifying compromised nodes in various acoustic conditions.
Introduction
Modern smart environments utilize Acoustic Sensor Networks (ASNs) to perform tasks such as source localization for voice-controlled interfaces or industrial monitoring. While traditional localization methods rely on rigid physical models, state-of-the-art learning-based techniques adapt to the specific acoustic characteristics of a room, such as echoes and geometry. However, these models are highly sensitive to the spatial consistency of the network. If a sensor node is accidentally moved or rotated after the training phase, the learned spatial "map" becomes invalid, potentially doubling localization errors. Maintaining system integrity requires an automated way to detect such misalignments and identify the specific displaced nodes.
Methodology
The proposed framework combines semi-supervised localization with a graphical model for robust failure detection.
SSGP Localization
The system employs Semi-Supervised Gaussian Process (SSGP) localization based on Relative Transfer Functions (RTFs). This approach allows the model to learn the acoustic environment using a minimal set of labeled sound sources (known positions) supplemented by a larger volume of unlabeled data, which is more practical for real-world deployment.
Leave-One-Node-Out (LONO) and MRF Framework
To identify a displaced node, the system generates $M$ sub-networks, where $M$ is the total number of nodes. Each sub-network produces a position estimate by excluding one node at a time. The core innovation lies in applying a Markov Random Field (MRF) with Fully Connected Latent Variables (FCLVs) to these estimates. Instead of relying on simple error thresholds, the MRF treats the status of each sub-network as a latent state—either Aligned, Misaligned, or Unreliable. Through a message-passing scheme, the nodes reach a consensus: if the sub-network excluding Node A is stable while others are erratic, Node A is identified as misaligned with high probability.
Experimental Setup
The method was evaluated in a simulated 6m x 6m x 3m room equipped with four sensor nodes, each containing two microphones. The training dataset consisted of 5 labeled and 300 unlabeled speech sources. To test robustness, the system was subjected to varying reverberation times ($T_{60}$ = 0.2s, 0.4s, and 0.6s) and random node displacements involving both translation and rotation.
Results
The MRF-based detector demonstrated significant advantages over simpler thresholding methods. By incorporating prior knowledge of error distributions (e.g., half-normal for aligned states), the system remains robust to the inherent noise of acoustic environments.
As shown in the performance comparison, the MRF method consistently outperforms the naïve detector. For instance, at a reverberation time of $T_{60} = 0.4$s, the MRF achieved an Area Under the Curve (AUC) of 0.82, compared to just 0.62 for the naïve approach. Furthermore, the system's sensitivity scales with the magnitude of the displacement.
The probabilistic output naturally reflects the degree of misalignment, allowing even small shifts to be captured more reliably than with fixed thresholds. This sensitivity ensures that the probability of movement detection increases proportionally with the size of the shift.
Conclusion
This research presents a robust probabilistic framework for maintaining the calibration of ASNs. By leveraging the consensus of sub-networks through an MRF, the system can identify displaced sensors across various acoustic conditions. Future work will focus on extending this model to dynamic scenarios with moving sound sources and developing methods to detect the simultaneous displacement of multiple sensor nodes.
Summary updated