关键词: 高压换流阀, 声纹诊断, 自监督学习, 因果知识, 注意力机制, 少标签样本学习

Abstract: To address the urgent need for operational status monitoring of high-voltage converter valves and the challenges of scarce labeled samples and insufficient model interpretability in acoustic fault diagnosis,a novel acoustic fault diagnosis method for high-voltage converter valves that integrates time-frequency self-supervised representation learning and a causal knowledge-guided attention mechanism is proposed. Firstly,a time-frequency predictive network framework is adopted,and a time-frequency dual-stream encoder structure is designed. Through self-supervised learning tasks,robust initial time-frequency feature representations are extracted from unlabeled acoustic data. Secondly,a causal knowledge graph of the fault-acoustic feature relationship for high-voltage converter valves is constructed and quantified as causal guidance parameters. These parameters are used to adjust the generation of cross-correlation matrices in time-frequency self-supervised learning,inducing the model to focus on features associated with specific faults. Finally,a small number of labeled samples are utilized to fine-tune the diagnostic network.Experimental results demonstrate that,under the condition of few labeled samples,the proposed method outperforms baseline models in effectively improving both the accuracy of acoustic fault diagnosis for high-voltage converter valves and the interpretability of model decisions.

Key words: high-voltage converter valve, acoustic diagnosis, self-supervised learning, causal knowledge, attention mechanism, few-shot learning