摘要：

随着能源转型和碳中和的全球发展趋势, 储能变流器关键组件的稳定性变得至关重要。特别是其功率器件和散热器在实际运行中的稳定性直接关系到整个系统的可靠性。关注储能变流器功率模组振动信号的故障诊断问题, 传统诊断方法处理复杂信号时往往面临挑战, 需要频繁地调整参数。此外, 由于储能变流器的工作环境复杂, 现有深度学习诊断方法的性能也不尽如人意。为此, 提出一种基于大模型知识和通道注意力网络的储能变流器功率模组故障诊断方法LLMCAN。首先通过预训练的大规模语言模型, 在特征提取过程中利用丰富的领域知识, 增强模型对复杂功率模组振动信号的分析能力。其次引入通道注意力网络使模型能够自适应学习信号中不同通道之间的关系, 提高故障诊断的准确性。在包含1 000条真实工况数据的储能变流器信号数据集上进行验证, 其中包括正常工况和9种故障模式。实验结果表明, 该方法在多种度量指标下均显示出优越性能, 其中诊断准确率高达99.8%, 远超传统方法, 为储能变流器功率模组的故障诊断提供一个高效、准确的解决方案。

关键词: 储能变流器, 故障诊断, 深度学习, 注意力机制, 信号处理

Abstract:

Owing to the global trends of energy transition and carbon neutrality, ensuring the stability of energy storage inverters as key components is crucial. Specifically, the stability of power devices and heat sinks during actual operation is directly related to the reliability of the entire system. This study primarily focuses on the fault diagnosis of the vibration signals in the power module of energy storage inverters. Traditional diagnostic methods often face challenges when dealing with complex signals, and require frequent parameter adjustments. Moreover, owing to the complicated working environment of energy storage inverters, existing deep learning diagnostic methods exhibit unsatisfactory performance. This study proposes a fault diagnosis method for the power modules of energy storage inverters based on a Large Model Knowledge and Channel Attention Network (LLMCAN). First, a pre-trained Large Language Model (LLM) is utilized in the feature extraction process to leverage rich domain knowledge, thereby enhancing the analytical capability of the model for complex vibration fault signals from the power module. Second, the introduction of the channel attention network enables the model to adaptively learn the relationship between different channels in the signal. This further improves the accuracy of the fault diagnosis. The LLMCAN is validated on an energy storage inverter signal dataset containing 1 000 real-world operating data points, including normal operating conditions and nine fault modes. Based on extensive experiments, this method demonstrates excellent performance under various metric indicators, with a diagnostic accuracy rate of 99.8%. This accuracy significantly surpasses that of traditional methods. Thus, an efficient and accurate solution for the fault diagnosis of the power modules of energy storage inverters is provided.

Key words: energy storage inverter, fault diagnosis, deep learning, attention mechanism, signal processing