VLM4FD: A Multimodal Vision-Language Model for Cross-Condition Bearing Fault Diagnosis

To address the limited diagnostic capability of mechanical equipment under varying operating conditions, we proposes a multimodal large-model-based diagnostic method named VLM4FD. First, the one-dimensional time-series signal is mapped to a two-dimensional time–frequency spectrogram using the Short-Time Fourier Transform (STFT), thereby supporting subsequent processing by the multimodal large model. Second, statistical methods are employed to extract characteristic indicators of the signal in both the time domain and the frequency domain, which are then converted into structured textual descriptions. Finally, a Vision-Language Model (VLM) is used to perform cross-modal fusion of image features and textual statistical measures, and a three-stage training strategy is adopted to endow the VLM with fault diagnosis capability. Experimental results demonstrate that VLM4FD exhibits excellent generalization performance under cross-condition scenarios.