A rotating machinery fault diagnosis method based on large language models and frequency-domain feature-guided data augmentation

To address fault-sample scarcity, class imbalance, and mode collapse in conventional generative models, this study proposes large-model-based frequency-domain data augmentation for rotating machinery fault diagnosis. The absolute-mean-threshold method first extracts transient impact features from vibration signals. Fast Fourier transform and envelope-spectrum analyses then extract representative frequency-domain descriptors for different fault mechanisms. Time-domain statistics, impact features, and dominant spectral descriptors are encoded in a structured prompt that conditions GPT-4o to generate candidate numerical sequences. Because GPT-4o does not perform an explicitly controllable Fourier analysis during conversational generation, each candidate is subsequently evaluated using an external fast Fourier transform or envelope-spectrum program. Candidates that violate the predefined spectral peak-matching criterion are discarded and regenerated through a generation–validation–feedback procedure. In the two evaluated case studies, the generated data showed visually comparable time-domain waveform patterns, amplitude ranges, and dominant spectral peaks relative to the reference signals. Under the adopted class-imbalanced settings, large-model-based frequency-domain data augmentation improved the downstream classification performance of the convolutional neural network–based classifiers. These results support the use of large-model-based frequency-domain data augmentation for fault-data augmentation in the evaluated bearing and rotor case studies.