TLDR: The Dual-Granularity Guidance Network (DGGN) is a novel framework for Few-Shot Class-Incremental Fault Diagnosis (FSC-FD) in industrial systems. It addresses catastrophic forgetting of old knowledge and overfitting on scarce new data by employing two parallel feature learning streams: a fine-grained stream for capturing class-specific features from limited new samples, and a coarse-grained stream for preserving general, class-agnostic knowledge. These streams are dynamically fused using a multi-semantic cross-attention mechanism. DGGN also incorporates a Boundary-Aware Exemplar Prioritization strategy to mitigate forgetting and a decoupled Balanced Random Forest classifier to counter decision boundary bias caused by data imbalance. Extensive experiments on TEP and MFF datasets demonstrate DGGN’s superior diagnostic performance and stability compared to state-of-the-art methods.
In the complex world of industrial systems, ensuring operational reliability and safety is paramount. This often relies on quickly and accurately identifying equipment failures, a field known as fault diagnosis. While advanced data-driven methods have made significant strides, they face a critical challenge: real-world industrial environments are dynamic. Fault patterns evolve, and new types of failures emerge constantly, often with very limited data available for these new faults. This scenario is known as Few-Shot Class-Incremental Fault Diagnosis (FSC-FD).
FSC-FD presents a dual problem for diagnostic models: they must continuously learn about new fault classes from just a handful of samples without “catastrophically forgetting” the knowledge of older, previously learned fault types. Additionally, with so little new data, models are prone to “overfitting,” becoming too specialized to the scarce new examples and failing to generalize effectively.
To tackle these significant hurdles, researchers have introduced a novel framework called the Dual-Granularity Guidance Network (DGGN). This innovative approach explicitly separates the learning of features into two parallel streams, each focusing on a different “granularity” of information.
Dual-Granularity Representation: The Core Innovation
The DGGN framework is built upon the idea that not all knowledge is the same. It distinguishes between two types of features. First, a fine-grained representation stream is dedicated to capturing highly specific and discriminative features unique to new, limited fault samples. It uses a specialized component called the Multi-Order Interaction Aggregation (MOIA) module to extract detailed, class-specific information from these scarce data points. Second, in parallel, a coarse-grained representation stream focuses on modeling and preserving general, “class-agnostic” knowledge—information that is shared across all fault types, old and new. This stream provides a stable foundation, representing what broadly constitutes a “fault” without being tied to a specific type.
These two distinct representations are then dynamically combined using a multi-semantic cross-attention mechanism. This mechanism allows the stable, general knowledge from the coarse-grained stream to guide the learning of the fine-grained features. This guidance is crucial for preventing overfitting on the limited new data and for reducing conflicts between features of old and new classes.
Mitigating Forgetting and Imbalance
Beyond the dual-granularity feature learning, DGGN incorporates additional strategies to enhance its performance and stability. To combat catastrophic forgetting, DGGN uses a smart sample replay strategy called Boundary-Aware Exemplar Prioritization (BAEP). Instead of randomly selecting old samples to remember, BAEP prioritizes “exemplars” that lie near the decision boundaries in the feature space. These boundary samples are more challenging for the model and, by replaying them, the model is forced to learn more robust and generalizable representations, improving its ability to distinguish between classes. Furthermore, real-world fault data often suffers from class imbalance, where some fault types are much rarer than others. To counter the bias this can introduce in classification decisions, DGGN employs a decoupled Balanced Random Forest (BRF) classifier. This classifier is specifically designed to handle imbalanced datasets, ensuring fair and accurate predictions across all fault types.
Also Read:
- Adapting Anomaly Detection for Dynamic Medical Imaging
- Smart Security for Robotic Systems: Detecting Anomalies in ROS2 Cyber-Physical Systems
Validation and Impact
The effectiveness of the DGGN framework was rigorously tested through extensive experiments on two prominent datasets: the Tennessee Eastman Process (TEP) benchmark, which simulates chemical processes, and the real-world Multiphase Flow Facility (MFF) dataset. These experiments covered challenging scenarios, including class imbalance and long-tailed distributions (where some classes have very few samples).
The results consistently demonstrated that DGGN achieves superior diagnostic performance and stability compared to existing state-of-the-art FSC-FD approaches. For instance, in class-imbalanced settings, DGGN showed significant improvements in average class-wise accuracy. Its ability to maintain stable performance throughout the incremental learning process highlights its success in mitigating catastrophic forgetting and feature conflicts.
Ablation studies further confirmed the independent contributions of each component within DGGN. Removing the MOIA module, the class-agnostic model, or the multi-semantic cross-attention mechanism, or omitting knowledge transfer, all led to noticeable declines in accuracy, underscoring the collaborative strength of the framework.
This research marks a significant step forward in developing truly adaptive intelligent diagnosis systems for industrial applications. The code for DGGN is publicly available, fostering further research and development in this critical area. You can find the research paper here.
Future work will focus on optimizing the model architecture for greater efficiency, developing dynamic memory management methods, and extending the framework to more complex cross-device or cross-domain scenarios, further enhancing its real-world applicability.
