Information Bottleneck Driven Visual Fault Diagnosis Method Utilizing Manifold Learning With Inner-Autoencoder

Traditional fault diagnosis methods often rely on limited visualization techniques, typically reducing monitoring data to simplistic curves and static thresholds, which can lead to significant information loss and limit comprehensive fault assessment. To address these issues, this article presents a visualization-driven fault diagnosis scheme under the framework of the Information Bottleneck theory, which integrates Inner-Autoencoder with Uniform Manifold Approximation and Projection for unsupervised dimensionality reduction. The Inner-Autoencoder first extracts essential residual variables from high-dimensional data, forming minimum sufficient statistics that facilitate effective manifold learning. These latent representations are clustered within a joint training structure, enhancing separability and generating virtual labels that reinforce high-dimensional data similarity, thereby improving visualization fidelity. Furthermore, all the aforementioned steps are implemented within the framework of Information Bottleneck theory, specifically optimized from both compression and prediction aspects to ensure that the dimensionality reduction results have regular boundaries and to maintain consistency with the high-dimensional data as much as possible in fault diagnosis. Validation on induction motor and steel plate fault datasets demonstrates the proposed method’s superior visualization clarity and diagnostic accuracy, underscoring its potential as a robust solution for advanced fault analysis and diagnosis.