Structural Rotor Fault Diagnosis Using Attention-Based Sensor Fusion and Transformers

Despite attention models' (AM) success in diverse domains, their application in failure detection and predictive maintenance (FDPM) field is limited. The existing literature of complex rotating machinery (RM) systems with multiple sensors pose the following challenges in applying AM and transformer networks: i) lack of proper fault-specific embedding representation for a long sequence of vibration data, ii) inability to provide adaptive weightage to sensor segments based on its fault sensitivity in the sensor fusion, and iii) failure in incorporating symptomatic fault features in fault decision-making. Hence, we propose an FDPM framework to address such inadaptability issues in diagnosing structural rotor faults (SRF), which is the root cause of most RM issues. The proposed framework facilitates the use of symptomatic fault features by extracting the distinctive frequency components (DFC) from the vibration spectrum. A combined feature representation is generated by bagging the DFC and time-domain features to endorse the most discriminative capability within fewer dimensions. Subsequently, a multi-sensor fusion is proposed to create the embedding representation using attention, assisted with fault pattern-based ranking to ensure the relative importance of fused sensor vectors and their fault sensitivity. With this reduced dimension-embedding, the transformers with multi-head self-attention capture the different aspects of dependencies even from short-length sequences, thereby lessening the execution time. Two recurrent transformers are also utilized to capture the local dependency, and their performance is compared with general transformers on the Meggitt and MaFaulDa datasets. The results demonstrate state-of-the-art performance in SRF diagnosis with more than 99.0% accuracy on both datasets.