Improved Fault Diagnosis Model Based on Bootstrap Your Own Latent Algorithm for a Multistage Centrifugal Pump

Insufficient data in machinery poses a significant problem in Prognostic and Health Management research due to extended durations of error-free machinery operation. The resulting data scarcity negatively impacts the performance of supervised training models. However, during the product testing stage in real-industrial applications, it is relatively easier to obtain nominal operating condition data that includes healthy and various faulty data, while other operating conditions only provide healthy data. This nominal operating condition data can provide some faulty information that can generalize to other conditions, thereby addressing the data insufficiency challenge. To this end, we propose a fault diagnosis model that employs nominal operating condition data and extra healthy data from other working conditions to be improved. Our method utilizes a contrastive learning approach called BYOL to train the fault diagnosis model. At first, we pre-train the BYOL using nominal operating condition data and extra healthy data from other operating conditions. Then fine-tune the entire network using only nominal operating condition data. Our experiments and comparison results demonstrate the effectiveness of our approach, yielding improved classification accuracies in different operating conditions.