Abstract
Deep learning has revolutionized intelligent fault diagnosis, yet its reliance on large-scale annotated data remains a bottleneck. Automated annotation strategies often introduce noisy labels, which trigger a self-reinforcing "confirmation bias" during training. This paper introduces Multi-granularity Evidence Labels (MgEL), a framework inspired by quantum entanglement and collapse. By performing feature-space fusion between entangled sub-distributions, MgEL constructs a superposition state to derive robust auxiliary labels. This approach suppresses confirmation bias and maintains high accuracy even under extreme noise (up to 90%), offering a reliable solution for low-cost, data-driven industrial diagnostics.
The Challenge: The Echo Chamber of Noisy Labels
In industrial settings, manual labeling of vibration signals for machinery fault diagnosis is prohibitively expensive. Engineers frequently turn to automated systems to generate training datasets. However, these automated "teachers" are imperfect, often assigning incorrect labels to complex or borderline samples.
Standard "Learning with Noisy Labels" (LNL) techniques attempt to mitigate this by using the model's own predictions to filter or correct the data. This creates a dangerous feedback loop known as confirmation bias. If a model is initially trained on corrupted data, its predictions will likely be biased. Using these biased predictions to "clean" the dataset only reinforces the errors, leading to a total collapse in performance as noise levels increase.
Methodology: Quantum-Inspired Robustness
To break the cycle of confirmation bias, the authors propose MgEL. While implemented on classical hardware, MgEL adopts the mathematical metaphors of quantum mechanics—specifically entanglement and superposition—to manage label uncertainty.
1. Entangled Partitioning
MgEL splits each training batch into two "entangled" subsystems, A and B. When evaluating a sample in subsystem A, the model is restricted to finding its nearest neighbors within subsystem B. This cross-subspace querying acts as a structural safeguard, preventing the model from simply memorizing local noise patterns in its immediate vicinity.
2. Superposition and Evidence Fusion
The labels of the identified neighbors form a "superposition state," representing a probability distribution of potential true classes. From this state, MgEL derives two key components:
- Pseudo-label ($y'_i$): The class with the highest amplitude (most frequent among neighbors).
- Collapsed Label ($\bar{y}_i$): A label sampled probabilistically from the distribution, representing an uncertainty-aware observation.
These are then fused with the original noisy annotation using an "evidence mass" calculation. This mass is dynamically adjusted based on the estimated global noise level of the dataset, ensuring that the model trusts its neighbors more when the original labels are likely corrupted.
3. The Coherence Safeguard
The final layer of protection is the Coherence Mechanism. Before a label is committed for training, MgEL checks for consensus between the original label, the pseudo-label, and the collapsed label. If at least two agree, the fused label is used. If all three disagree, the model acknowledges its uncertainty and reverts to the superposition state, effectively deferring the decision until the feature representations improve in later epochs.
Experimental Results
The framework was tested on three real-world bearing fault datasets (Single, Multiple, and Damage) under both symmetric and asymmetric noise conditions.
Performance in Extreme Noise
MgEL demonstrated remarkable resilience in high-noise scenarios. In the Single dataset under 90% symmetric noise—where nearly all labels are wrong—MgEL achieved 23.02% accuracy. While this sounds low, it is more than triple the performance of the next best baseline, JoCoR (7.42%), which effectively failed to learn.
Comparative Superiority
Across all tests, MgEL consistently outperformed seven state-of-the-art baselines. On average, it surpassed the strongest competitors by 7% to 16% in accuracy. Visual analysis using t-SNE confirmed that MgEL maintains clean, separable feature clusters even when the input labels are highly chaotic, whereas baseline methods produce overlapping, uninformative latent spaces.
Conclusion and Future Outlook
MgEL represents a significant step forward in making deep learning practical for industrial environments where perfect data is a luxury. By replacing blind self-trust with a quantum-inspired evidence fusion strategy, it successfully navigates the pitfalls of confirmation bias.
Future work will focus on refining the entanglement schemes for smaller or highly imbalanced datasets and exploring the integration of temporal logic to further enhance label reliability in time-series diagnostic data.
Summary updated