Multimodal Learning for Automatic Anomalies Detections in Water Consumption Profiles

Detecting anomalies in household water consumption is critical for sustainable resource management, yet predominant unimodal approaches (e.g., consumption validation checks) ignore synergies between usage patterns and behavioral habits. This paper introduces Multimodal Learning for Automatic Anomaly Detection in Water Consumption, a novel framework integrating time series and structured surveys to enhance contextual accuracy. Our method begins with rigorous preprocessing, including a missing value imputation scheme prioritizing intra-household consistency over citywide averages, followed by a multimodal fusion pipeline combining Recurrent layers, dense networks, and dynamic attention mechanisms to align temporal and survey data hierarchically. Notably, the architecture incorporates a hybrid detector: Isolation Forest guides a discrepancy attention module, dynamically weighting mismatches between reported and observed consumption to mitigate false positives. Our final architecture was validated via unsupervised metrics and expert analysis on households in Pereira, Colombia, and outperforms unimodal benchmarks in detection accuracy. Integrating socio-behavioral attributes (e.g., washing frequency, household size) improves causal interpretability-a breakthrough for actionable environmental policies. The method's scalability in sparse, high-dimensional datasets positions it as a pioneering tool for intelligent urban water monitoring, directly supporting Sustainable Development Goals for efficient resource management.