TLDR: A new framework, Hierarchical Knowledge Guided (HKG), improves fault intensity diagnosis in industrial systems by explicitly incorporating hierarchical relationships between fault classes. Unlike traditional methods, HKG uses Graph Convolutional Networks and a novel re-weighted correlation matrix to guide deep learning models, leading to superior performance on real-world cavitation and bearing datasets, especially in detecting early-stage faults.
Maintaining the health of complex industrial machinery is crucial for efficiency and safety. A key aspect of this is Fault Intensity Diagnosis (FID), which involves identifying the severity of faults within mechanical devices. Traditional FID methods often rely on a ‘chain of thought’ approach, treating problems as direct input-output mappings. However, this overlooks the inherent relationships and dependencies that exist between different fault types or ‘target classes’. For instance, various stages of cavitation (like incipient, constant, or choked flow) are not isolated events but are interconnected.
To address this limitation, researchers have introduced a novel framework called Hierarchical Knowledge Guided (HKG) fault intensity diagnosis. This framework takes inspiration from the ‘tree of thought’ paradigm, which acknowledges and leverages the hierarchical dependencies among target classes. Unlike previous methods that might focus solely on complex network designs, HKG explicitly embeds this hierarchical knowledge into deep learning models, making it adaptable to various representation learning techniques.
How HKG Works
The HKG framework operates with two main components: a feature representation learning stream and a hierarchical knowledge learning stream. The first stream uses advanced deep learning methods, such as Convolutional Neural Networks (CNNs) and Transformers, to extract detailed features from raw signals like acoustic or vibration data. These signals are first pre-processed using techniques like sliding windows and time-frequency transforms (e.g., STFT) to make them suitable for analysis.
The innovation lies in the hierarchical knowledge learning stream. Here, the framework constructs a hierarchical ‘tree’ of fault classes, representing their relationships (e.g., ‘cavitation’ as a parent to ‘incipient cavitation’). This hierarchical structure is then used with Graph Convolutional Networks (GCNs). GCNs are particularly good at learning from graph-structured data, making them ideal for understanding the connections between different fault classes.
A critical element developed within HKG is the ‘re-weighted hierarchical knowledge correlation matrix’ (Re-HKCM). This matrix is derived from a data-driven statistical correlation matrix but is refined by embedding the hierarchical knowledge. This process helps to explicitly model inter-class dependencies and, importantly, mitigates common issues in GCNs like ‘over-smoothing,’ where distinct features between nodes can become blurred. The framework also uses word embeddings, like GloVe, to represent the semantic relationships between classes, further enhancing the GCN’s ability to learn meaningful connections.
The GCN then uses this enriched information to generate a set of ‘interdependent global hierarchical classifiers.’ These classifiers are not trained in isolation but are designed to work together, reflecting the hierarchical nature of the faults. Finally, these hierarchical classifiers are applied to the deep features extracted from the signals, allowing the entire model to learn end-to-end and make more accurate fault intensity predictions.
Also Read:
- Anticipating System Failures: A New Approach to Risk-Based Health Management
- Unlocking Graph Neural Network Decisions with Class-Level Explanations
Experimental Success
The effectiveness of HKG was rigorously tested on four real-world datasets. These included three cavitation datasets provided by SAMSON AG, covering different types of cavitation (incipient, constant, choked flow) and non-cavitation states, as well as a public dataset of bearing vibration signals from Paderborn University. The results consistently showed that HKG outperformed recent state-of-the-art FID methods across various evaluation metrics like accuracy, precision, recall, and F1-score.
Notably, HKG demonstrated significant improvements in diagnosing ‘incipient cavitation,’ which represents early-stage faults and is often challenging to detect. The framework proved its versatility by enhancing the performance of different deep learning backbones, including various ResNet, DenseNet, MobileNet, ViT, and Swin architectures. Ablation studies further confirmed the importance of HKG’s core components, such as the Re-HKCM and the optimal number of GCN layers, in achieving these superior results. The method also showed robustness even when dealing with downsampled signals, which is important for real-world applications with varying sensor capabilities.
In conclusion, the Hierarchical Knowledge Guided framework offers a robust and adaptable solution for fault intensity diagnosis in complex industrial systems. By explicitly incorporating hierarchical knowledge, it provides a more nuanced understanding of fault classes, leading to more accurate and reliable predictions. For more details, you can refer to the full research paper.
