A fuzzy deep neural network accurately classifies household appliances in real time using symbolic data and multi-label AI.
Recognizing Household Appliances with AI: Managing Uncertainty and Imbalanced Data with Fuzzy Intelligent Networks
In a world increasingly focused on sustainability and energy efficiency, knowing exactly which electrical devices are operating in a home is no longer just a technical detail, it’s the key to building smart grids, reducing costs, and optimizing consumption. However, identifying multiple appliances at the same time based on their current and voltage profiles is anything but simple, especially when the data is imbalanced and the signals overlap.
Fuzzy Deep Learning for Multi-Label Classification
This is where a new approach comes into play, based on deep neural networks and fuzzy logic, designed to directly address the challenge of multi-label classification in the presence of class imbalance. At the core of the solution is a Fuzzy Deep Neural Network (FDNN) that combines the powerful feature extraction of neural networks with the flexibility of fuzzy logic in handling ambiguity and class overlap.
High-Frequency Data Meets Symbolic Compression
The model operates on high-frequency time series, recorded from real household appliances in American homes. Each data sequence is transformed into a symbolic representation using the Symbolic Aggregate approXimation (SAX) technique, which reduces data dimensionality and significantly speeds up training while preserving high performance.
Flexible Labels for Complex Realities
Unlike traditional approaches, here each appliance can belong to multiple categories at once, with partial degrees of membership. A fridge, for example, may be classified both as an “electronic device” and as “motor-assisted,” depending on its use. This more realistic view allows the system to adapt better to complex scenarios and to learn effectively even from uncertain labels.
Outperforming Classic AI: Accuracy and Speed
The results are clear: the FDNN model outperforms well-known approaches like LSTM and CNN in both accuracy and robustness, achieving excellent performance even on minority classes. The training time reduction enabled by symbolic representation makes this technology suitable for real-time applications, such as home energy monitoring, predictive maintenance, or optimized load management.
A Scalable Vision for Smart Energy Systems
This kind of intelligent classification is just the beginning. In the future, the same paradigm could be extended to optimize consumption at the individual level, aggregate data across energy communities, or even reduce operational costs in cloud environments. The secret? Combining the adaptability of fuzzy logic with the power of deep learning and the efficiency of symbolic representations.
Authors
F. Succetti, A. Rosato, M. Panella
