A General Fault Prediction Framework based on Relationship Mining and Graph Neural Network

In the field of equipment health management, fault prediction is a valuable and challenging task. Previous fault prediction methods are generally designed for specific mechanical equipment with the specific requirements of data type (e.g., vibration acceleration signals), and they cannot be applied to different mechanical equipment. In this paper, a fault prediction modeling framework based on relationship mining and graph neural network is proposed, which can be widely used in different types of mechanical equipment for fault prediction. First, based on alarm records and maintenance records, the Apriori algorithm is used to analyze the relationship between alarm signals and fault events. Then, the alarm track graph (ATG) for a certain period (e.g., seven days) is constructed based on the key alarm signals. Finally, the graph convolutional network (GCN) is used to predict the fault occurrence for a future certain period (e.g., three days). As can be expected, the modeling approach can effectively mine the correlation between the alarm records and maintenance records, and thus make reliable predictions of future faults. Since the above process utilizes only two common data types, i.e. alarm records and maintenance records, the proposed prediction framework can be easily implemented in a practical background.