Nonexclusive Classification of Household Appliances by Fuzzy Deep Neural Networks

Most machine learning classification systems are designed to make definitive decisions. An observation belongs to one category or another, leaving little room for uncertainty. While this assumption simplifies the learning process, it rarely reflects how real-world systems actually behave. Household electrical appliances are a perfect example. Many devices perform multiple functions simultaneously, combining electronic components, motors, heating elements, and other mechanisms that produce power consumption patterns with significant similarities. As a result, different appliances can generate electrical signatures that overlap, making strict classification both difficult and, in some cases, conceptually incorrect. To address this challenge, a new methodology introduces a learning framework capable of recognising that an appliance may legitimately belong to multiple categories at the same time, assigning different degrees of membership instead of forcing a single definitive label.

The proposed approach integrates fuzzy logic with deep neural networks to model uncertainty directly within the classification process. Instead of producing a single output class, the model estimates how strongly an appliance belongs to each possible category, reflecting the intrinsic ambiguity of its operating behaviour. One of the key innovations lies in eliminating the need for manually designed fuzzy inference rules. Traditionally, fuzzy systems rely on expert knowledge to define how different inputs should be interpreted. Here, that role is delegated to a deep learning architecture capable of automatically learning the relationships between power consumption time series and their corresponding fuzzy membership values. This combination allows the flexibility of fuzzy reasoning to benefit from the representation learning capabilities of modern neural networks, producing a model that adapts directly from data while preserving the interpretability associated with fuzzy classification.

Deep neural networks often require considerable computational resources and long training times, particularly when processing sequential data. To overcome this limitation, the proposed architecture introduces a randomisation strategy that significantly reduces the number of trainable parameters. The first convolutional layer of the network is randomly initialised and kept fixed throughout the training process, allowing subsequent layers to focus on learning higher-level temporal representations. Although this simplification reduces the computational complexity of the model, it preserves a high level of predictive performance while dramatically accelerating the learning procedure. The resulting architecture combines a randomised one-dimensional convolutional layer, a Long Short-Term Memory network capable of modelling temporal dependencies, a fully connected layer, and a final sigmoid stage that generates fuzzy membership values for each class. This design enables efficient learning without sacrificing the ability to capture the dynamic characteristics of electrical consumption signals.

The methodology is evaluated using a large real-world dataset containing electricity consumption measurements collected from domestic appliances operating under normal household conditions. Rather than relying on artificially generated signals, the experiments considered appliances performing their everyday functions across multiple homes over an extended period. Each appliance is described through power consumption time series and assigns fuzzy membership values according to different semantic groupings. Some categories reflect the physical characteristics of the devices, while others represent their functional role inside the home. Because many appliances naturally satisfy multiple descriptions simultaneously, fuzzy labels provide a far more realistic representation than conventional mutually exclusive classes. The proposed model is compared with fully trainable deep neural networks and standard LSTM architectures commonly adopted for time-series classification. The experiments demonstrate that the randomised fuzzy deep neural network achieves competitive, and in several scenarios superior, prediction accuracy while requiring substantially shorter training times.

Accurately identifying household appliances from their electricity consumption profiles has implications that extend well beyond classification itself. Understanding how devices operate enables more sophisticated demand-side energy management, supports intelligent user profiling, and contributes to more efficient smart grids and energy communities. Perhaps even more importantly, this work demonstrates that embracing uncertainty can improve the quality of artificial intelligence models. Instead of forcing complex physical systems into rigid categories, the proposed approach allows ambiguity to become part of the learning process itself. As intelligent energy infrastructures continue to evolve, combining fuzzy reasoning with efficient deep learning architectures offers a promising direction for building systems that are not only faster and more scalable, but also better aligned with the complexity of the real world.