SensePoints: a deep learning approach based on multi-modal data fusion via multiple parallel convolutions to classify points in smart buildings

• A deep learning approach which fuses text metadata and time series via parallel convolutions to classify the points (and their associated characteristics) in buildings. • State-of-the-art performance on over 11,000 labelled points from 150+ diverse smart buildings, achieving up to 60% improvement in classification accuracy compared to existing approaches. • Enables automated and scalable metadata mapping, unlocking cross-building interoperability for smart energy applications. The rapid proliferation of Internet of Things (IoT) devices in smart buildings has enabled data-driven energy and operational analytics. However, the lack of standardized metadata across buildings remains a major barrier to scalability and interoperability of such applications. Manually mapping the names of thousands of heterogeneous sensing and control points to a unified semantic schema like Brick is labor-intensive, error-prone, and fundamentally non-scalable. This paper presents SensePoints , a novel deep learning framework that automates the classification of heterogeneous points by fusing textual metadata and time series data. SensePoints employs a multi-headed convolutional neural network in which each modality is processed through an independent convolutional branch, preserving modality-specific features prior to fusion. Textual point names are encoded using BERT-based semantic embeddings, while time series behaviour is represented using complementary statistical and frequency-domain features. The framework simultaneously predicts both Brick class labels and their associated characteristics properties, enabling richer semantic building models. SensePoints is evaluated using more than 11,000 labelled points from 150+ real-world buildings. It achieves an F score of 0.75 for Brick class labels and up to 0.88 for key characteristics properties. It also outperforms representative benchmarks including autoencoder-based neural networks, tree-based time series classifiers, as well as attention-enhanced recurrent neural networks and a recent vision transformer-based model. Overall, SensePoints delivers up to 60% improvement in classification accuracy over existing methods. Results demonstrate that modality-aware convolutional fusion significantly improves accuracy and generalization across heterogeneous buildings, enabling scalable and interoperable IoT-driven analytics for intelligent building management.