New fault modes keep emerging during the longterm operation of industrial equipment, demanding continual learning for incremental diagnosis and maintenance decision-making. However, few-shot samples and class imbalance in practical engineering pose challenges to effective continual learning. To address these issues, a continual learning-based large language model (CL-LLM) is proposed for incremental fault diagnosis and maintenance decision-making under complex conditions. A wide-kernel deep convolutional network embedded with multi-scale efficient channel and spatial attention is designed. It adopts parallel convolution branches to synchronously capture the transient impact and periodic modulation features of vibration signals. Discriminative representations are extracted by adaptively screening key fault frequency bands. An incremental learning strategy combining random memory replay and a scalable linear classifier is proposed to retain a small number of historical category samples for subsequent training. During the incremental learning phase, the weights of historical categories are fixed. Only classification layer parameters are expanded to mitigate the catastrophic forgetting of CL-LLM in long-term continual learning. A projection adaptation layer is designed to map numerical fault features to the embedding space of the large language model, constructing a cross-modal alignment framework for vibration signals and text semantics. A low-rank adaptation strategy is adopted to fine-tune the low-rank matrices of attention layers for domain adaptation, enabling CL-LLM to output class-incremental diagnosis and maintenance decision results. Experiments are implemented on a public equipment operation and maintenance dataset and the self-built gearbox test rig, followed by comparative analysis and ablation studies. The results show that the proposed CL-LLM performs excellently in few-shot and class-imbalanced scenarios. Its learning forgetting rate for incremental fault diagnosis and maintenance decision-making is significantly lower than that of comparative methods.