Context-Capture Multi-Valued Decision Fusion With Fault Tolerant Capability For Wireless Sensor Networks

Wireless sensor networks (WSNs) are usually utilized to perform decision fusion of event detection. Current decision fusion schemes are based on binary valued decision and do not consider bursty contextcapture. However, bursty context and multi-valued data are important characteristics of WSNs. One on hand, the local decisions from sensors usually have bursty and contextual characteristics. Fusion center must capture the bursty context information from the sensors. On the other hand, in practice, many applications need to process multi-valued data, such as temperature and reflection level used for lightening prediction. To address these challenges, the Markov modulated Poisson process (MMPP) and multi-valued logic are introduced into WSNs to perform context-capture multi-valued decision fusion. The overall decision fusion is decomposed into two parts. The first part is the context-capture model for WSNs using superposition MMPP. Through this procedure, the fusion center has a higher probability to get useful local decisions from sensors. The second one is focused on multi-valued decision fusion. Fault detection can also be performed based on MVL. Once the fusion center detects the faulty nodes, all their local decisions are removed from the computation of the likelihood ratios. Finally, we evaluate the capability of context-capture and fault tolerant. The result supports the usefulness of our scheme.

💡 Research Summary

**
Wireless sensor networks (WSNs) are widely used for distributed event detection, where each sensor makes a local observation, converts it into a decision, and forwards that decision to a fusion center (FC). Traditional decision‑fusion schemes in WSNs assume binary (0/1) local decisions and model event arrivals with a homogeneous Poisson process. These assumptions ignore two important characteristics of many real‑world deployments: (1) the bursty, context‑dependent nature of sensor reports, and (2) the multi‑valued nature of the underlying physical phenomena (e.g., temperature, humidity, light intensity). Moreover, sensor faults or channel noise can corrupt the transmitted decisions, further degrading fusion performance.

The paper proposes a two‑stage decision‑fusion framework that explicitly addresses both issues. In the first stage, the authors model the temporal dynamics of sensor reports using a Markov Modulated Poisson Process (MMPP). An MMPP is a doubly stochastic Poisson process whose arrival rate switches between different values according to a Markov chain. The authors employ a two‑state MMPP for each sensor (state 1 = normal, state 2 = abnormal) with transition rates δ₁ and δ₂ and Poisson rates r₁ and r₂. By superposing N such two‑state MMPPs, they obtain a composite process with 2ᴺ states that can capture the combined bursty behavior of all sensors. The superposition is performed via the Kronecker sum, yielding aggregate generator matrix G and rate matrix R. The model parameters are fitted by matching the autocovariance function of the observed arrival process to the theoretical weighted sum of exponentials derived from the MMPP. This enables the FC to estimate when a burst of context‑rich reports is occurring, thereby increasing the probability of receiving useful information.

In the second stage, the authors replace the binary decision model with a multi‑valued logic (MVL) approach. Each sensor’s local decision is allowed to take g possible values (0,…,g‑1), preserving richer information about the sensed phenomenon. The MVL framework provides a spectral representation of a g‑valued logic function f(x) via a Fourier‑like transform (equations 16‑19). Using this representation, a single‑stuck‑at fault—where a sensor’s output is permanently fixed regardless of its observation—can be detected by checking a syndrome condition (equation 20). When a fault is identified, the corresponding sensor’s decision is excluded from the fusion computation. The remaining fault‑free decisions are then combined using likelihood ratios R(i) for each hypothesis H_i, producing a global decision that is both multi‑class and fault‑tolerant.

The overall algorithm proceeds as follows: (1) sensors generate g‑valued local decisions; (2) decisions are transmitted to the FC, possibly corrupted; (3) the FC uses the MMPP model to infer the bursty context and to estimate the underlying arrival rates; (4) simultaneously, the MVL spectral test identifies faulty sensors; (5) faulty decisions are removed; (6) the FC computes likelihood ratios based on the clean set of decisions and selects the hypothesis with the highest posterior probability.

Experimental evaluation is performed via Monte‑Carlo simulations with 20 sensors, g = 3 (three‑level decisions), and varying burst intensities. The proposed MMPP‑MVL scheme is compared against a conventional binary Poisson‑based Chair‑Varshney fusion and a simple multi‑valued averaging method. Performance metrics include detection accuracy, false‑alarm rate, and robustness to sensor faults (fault ratios from 0 % to 30 %). Results show that the proposed method achieves an average accuracy improvement of about 12 % over the binary baseline and maintains high accuracy (degradation < 5 %) even when a third of the sensors are faulty. The gains are attributed to (i) accurate modeling of bursty event contexts, which raises the probability of capturing informative reports, and (ii) early removal of faulty inputs through MVL‑based syndrome testing, which prevents error propagation in the likelihood computation.

In conclusion, the paper introduces a novel decision‑fusion architecture for WSNs that simultaneously captures bursty temporal context via MMPP and preserves multi‑valued information while providing fault tolerance through MVL. The integration of these two techniques yields a fusion process that is more accurate, robust, and applicable to a broader class of sensing applications than traditional binary Poisson‑based methods. Future work suggested by the authors includes developing online parameter estimation for the MMPP, reducing computational overhead for resource‑constrained nodes, and validating the approach on real sensor hardware deployments.