[Paper] Community-Based Model Sharing and Generalisation: Anomaly Detection in IoT Temperature Sensor Networks

Source: arXiv - 2601.05984v1

Overview

The paper proposes a community‑based framework for detecting anomalies in large IoT temperature‑sensor networks. By clustering sensors that behave similarly—thanks to temporal, spatial, and elevation cues—the authors show that a single trained model can be shared across many devices, cutting down on training time while still catching abnormal temperature readings.

Key Contributions

• Community‑of‑Interest (CoI) clustering that fuses temporal correlation (Spearman), geographic distance (Gaussian decay), and elevation similarity into a unified similarity matrix.
• Representative‑station selection using silhouette analysis to pick the most “central” sensor in each cluster for model training.
• Three auto‑encoder architectures (BiLSTM, LSTM, MLP) trained with Bayesian hyper‑parameter optimization and an expanding‑window cross‑validation scheme tailored to time‑series data.
• Cross‑community generalisation tests: models trained on one community are evaluated on both intra‑community stations and the best representatives of other communities.
• Empirical evidence that model sharing within a community yields comparable anomaly‑detection performance to training a dedicated model per sensor, while dramatically reducing computational load.

Methodology

1. Data preprocessing – Temperature readings from a dense network of IoT sensors are cleaned, normalized, and aligned to a common time grid.
2. Similarity matrix construction –
   • Temporal: Spearman rank correlation between each pair of sensor time‑series.
   • Spatial: Gaussian decay based on Euclidean distance (closer sensors get higher similarity).
   • Elevation: Simple absolute‑difference weighting (sensors at similar altitudes are more alike).
     The three components are multiplied to obtain a single fused similarity score.
3. Community detection – Spectral clustering on the fused matrix yields groups of sensors (communities) that share similar dynamics.
4. Representative selection – For each community, the sensor with the highest silhouette coefficient (i.e., best fit to its own cluster and worst fit to others) is chosen as the “representative”.
5. Model training – Three auto‑encoders (BiLSTM, LSTM, MLP) are trained only on normal temperature patterns from the representative sensor. Bayesian optimization searches the hyper‑parameter space (learning rate, hidden units, dropout, etc.) while an expanding‑window cross‑validation respects the chronological order of the data.
6. Anomaly detection – At inference, the reconstruction error (difference between input and auto‑encoder output) is compared to a threshold derived from the training error distribution. Large errors flag anomalies.
7. Evaluation – Models are tested on: (a) other sensors within the same community, and (b) the best representatives of other communities, allowing the authors to measure both in‑community robustness and cross‑community generalisation.

Results & Findings

• Within‑community performance is consistently high (≥ 0.84 F1) across all three architectures, confirming that a single model can serve many sensors without sacrificing detection quality.
• Cross‑community transfer works reasonably well for the more expressive BiLSTM, but performance drops as the source and target communities diverge in climate patterns.
• Computational savings: Training one model per community (≈ 10–15 sensors per community) reduces total training time by ~80 % compared with a naïve per‑sensor approach.
• Model selection – Bayesian hyper‑parameter tuning converges in fewer than 30 trials per architecture, making the pipeline practical for continuous deployment.

Practical Implications

• Edge‑friendly deployment – IoT gateways can host a single lightweight auto‑encoder per community, updating it centrally and pushing the model to all member devices. This cuts down on OTA‑update bandwidth and on‑device training cycles.
• Scalable monitoring – City‑wide environmental dashboards can ingest anomaly alerts from dozens of sensors while only maintaining a handful of models, simplifying model‑management pipelines.
• Rapid onboarding of new sensors – When a new temperature node is installed, it can be automatically assigned to an existing community (based on its location/elevation) and start using the pre‑trained model immediately, reducing the “cold‑start” period.
• Cost‑effective analytics – Service providers can offer anomaly‑detection as a SaaS layer without needing per‑device compute contracts, because the heavy lifting is done once per community.
• Transfer learning baseline – The cross‑community experiments provide a concrete benchmark for developers who want to fine‑tune a community model on a new region rather than training from scratch.

Limitations & Future Work

• Community granularity is fixed by the chosen number of clusters; overly coarse groups may hide subtle micro‑climates, while too fine a split erodes the computational benefits. Adaptive clustering is a natural next step.
• The framework focuses on temperature only; extending to multi‑modal sensor streams (humidity, air quality, vibration) will require richer similarity metrics and possibly multi‑task auto‑encoders.
• Anomaly labeling relies on reconstruction‑error thresholds derived from normal data; in practice, ground‑truth anomalies are scarce, so semi‑supervised or active‑learning strategies could improve detection confidence.
• Real‑world deployments will need to handle missing data, sensor drift, and firmware updates—issues not fully explored in the experimental setup.

Overall, the paper demonstrates that community‑based model sharing is a viable path to scalable, low‑overhead anomaly detection in sprawling IoT temperature networks, offering a blueprint that developers can adapt to their own sensor‑rich environments.

Authors

• Sahibzada Saadoon Hammad
• Joaquín Huerta Guijarro
• Francisco Ramos
• Michael Gould Carlson
• Sergio Trilles Oliver

Paper Information

• arXiv ID: 2601.05984v1
• Categories: cs.LG
• Published: January 9, 2026
• PDF: Download PDF