FD-DDLLM: A large language model enhanced dual-domain transformer for intelligent fault diagnosis

Abstract

To overcome the issues of inadequate multi-scale feature representation and limited model interpretability in equipment fault diagnosis under complex operating conditions, this paper proposes a large language model enhanced dual-domain transformer for intelligent fault diagnosis, which integrates time–frequency features with text semantic enhancement. Initially, the raw vibration signals are transformed using the short-time Fourier transform, converting one-dimensional time-domain signals into two-dimensional time–frequency representations. Time-domain features are extracted using a Time Series Transformer, while frequency-domain features are obtained through a Vision Transformer. A dynamic adaptive weight learning-based feature fusion and alignment mechanism is then introduced to facilitate deep cross-domain modeling and enhance information complementarity between heterogeneous time–frequency features. In addition, domain knowledge pertinent to fault diagnosis is incorporated into a large language model through low-rank adaptation fine-tuning, thereby establishing a multimodal diagnostic framework with semantic reasoning capabilities. This approach effectively mitigates the pronounced “black-box” nature and limited interpretability associated with traditional deep learning-based methods. Experimental results indicate that the proposed method achieves diagnostic accuracies of 98.7 and 98.4% on the Case Western Reserve University and Northeast Forestry University datasets, respectively. Ablation studies further confirm the efficacy of each key module in enhancing diagnostic performance.