Zero-Shot Compound Fault Diagnosis Method Using Semantic Construction and Embedding

Compound fault diagnosis has been a challenge in rotating machinery health management. Current compound fault diagnosis methods rely on a substantial number of labeled samples, which are impractical to obtain for every type of compound fault and pose safety risks when collected in industrial settings. To address these issues, we have built a bearing fault dataset that includes compound faults under multiple conditions. We propose a zero-shot compound fault identification model, which can be established solely using vibration data from single faults under no-load conditions and is suitable for multiple unknown load conditions. First, we design a compound fault logical semantic (CFLS) construction method based on statistical indicators from vibration data of single faults under no-load conditions. Second, we develop a multihead self-attention (MHSA)-based feature extraction approach that extracts features from vibration data under various load conditions. Third, we train a semantic embedding model using the semantics and features derived from single faults. By inputting the extracted features of compound faults into the model, we can get the compound fault dynamic semantic (CFDS) under multiple load conditions. Finally, the compound faults are identified by using semantics regression to compare the Euclidean distance between CFLS and CFDS. This method has been validated in two cases using both the self-built HDU dataset and the Paderborn dataset. The results show that it can recognize compound faults across multiple unknown domains and the identification accuracy has reached 78.17%.