Abstract
Modern industrial monitoring requires accurate fault diagnosis (FD) under conditions of data scarcity and high computational constraints. This paper introduces the Reservoir-Embedded-Directional Network (REDNet), a framework that shifts the diagnostic task from raw data space to a "model space." By representing multi-dimensional time-series instances as fitted models, REDNet captures complex intra- and inter-dimensional correlations. Our evaluation shows that REDNet maintains over 90% F1-score with as few as 10 samples per class and achieves training speeds up to 100 times faster than state-of-the-art deep learning architectures.
Introduction
Fault Diagnosis (FD) is essential for maintaining the safety and efficiency of industrial systems. However, practitioners face significant challenges: the rarity of labeled fault data, the complexity of cross-sensor correlations, and the prohibitive computational cost of training deep learning models like Transformers or GRUs for edge deployment. Existing model-oriented methods often rely on Echo State Networks (ESNs), which typically process data in a single temporal direction, potentially missing critical cross-sensor dynamics.
Methodology: The REDNet Framework
The proposed Reservoir-Embedded-Directional Network (REDNet) addresses these challenges by leveraging the efficiency of reservoir computing and the representational power of model-based features.
Multi-Directional Reservoirs
Unlike standard ESNs, REDNet employs multiple reservoirs aligned with different correlation directions (e.g., time and cross-sensor). This allows the network to capture dynamics both along the timeline and across different variables simultaneously. The internal weights of these reservoirs are fixed and random, which eliminates the need for expensive backpropagation.
The Model Space Shift
The core innovation lies in how data is represented. Instead of classifying raw sensor readings, a unique REDNet is fitted to each data instance. The resulting output weights—obtained via a fast, one-shot ridge regression—serve as the feature vector in a new "model space." Because different system states (healthy vs. specific faults) produce distinct dynamical patterns, their corresponding models naturally form separable clusters.
Experimental Results
The researchers evaluated REDNet across four major benchmarks: Gearbox Fault Diagnosis (GFD), Triaxial Bearing Vibration (TBV), Server Machine Dataset (SMD), and Case Western Reserve University (CWRU) bearing data.
Performance in Data-Scarce Scenarios
REDNet demonstrated remarkable robustness to small training sets. When training data was limited to just 10 samples per category, REDNet maintained an F1-score above 90%, significantly outperforming GRUs and Transformers, which require much larger datasets to generalize effectively.
Computational Efficiency
The "one-shot" training nature of REDNet provides a massive speed advantage. On the GFD dataset, REDNet completed training in 1.2 seconds, compared to 104.9 seconds for a GRU and 193.7 seconds for a Swin-Transformer—a nearly 100x improvement.
Conclusion
REDNet offers a parsimonious and high-speed alternative for industrial fault diagnosis. By focusing on system dynamics rather than raw values and utilizing a multi-directional model space, it overcomes the hurdles of data scarcity and computational overhead. Future work will focus on automated hyperparameter optimization and extending the framework to handle even more complex, high-dimensional correlation structures.
Summary updated