基于多模态信息异构融合的液体脉冲神经网络模型,用于可解释的旋转机械故障诊断
A liquid-impulse neural network model based on heterogeneous fusion of multimodal information for interpretable rotating machinery fault diagnosis[
ES评分 0
| DOI |
10.1016/j.ymssp.2026.113923 |
| 刊名 |
|
| 年,卷(期) |
2026, 246() |
| 作者 |
|
| 作者单位 |
南华大学
|
| 摘要 |
针对旋转机械故障诊断中动态特征衰减、多模态融合效率低以及诊断可解释性不足的问题,本文提出了一种可解释的多模态异构融合液态脉冲神经网络(LINN)模型。首先,构建了一个基于微分方程的液态编码层,通过分块反馈机制对非平稳信号中的时序动态演化特征进行建模。此外,引入了多通道泄漏积分-激发(MC-LIF)脉冲神经元,通过结合交替梯度和膜电位衰减策略来增强瞬态冲击特征的提取。最后,设计了一种注意力引导的多模态融合机制,以实现时频特征的自适应整合和贡献可解释性量化。在高噪声和变负载条件测试中,LINN仅使用410万个参数就实现了超过98.7%的准确率,以及88.64%的跨条件泛化准确率。消融实验验证了液态层和脉冲机制在增强动态建模和噪声免疫性方面的关键作用,而基于时频域注意力(TFDA)的可解释性分析进一步揭示了模型对关键时频模态贡献的敏感响应。该方法为复杂工况下的智能诊断提供了一种高精度、强泛化和可解释性的有效解决方案。 消融实验验证了液层和脉冲机制在增强动态建模和噪声免疫性方面的关键作用,而基于时频域注意力(TFDA)的可解释性分析则进一步揭示了模型对关键时频模态贡献的敏感响应。该方法为复杂工况下的智能诊断提供了一种高精度、强泛化能力和可解释性的有效解决方案。 试一试: 专家审校
|
| Abstract |
For the problems of dynamic feature attenuation, low efficiency of multimodal fusion and insufficient diagnostic interpretability in rotating machinery fault diagnosis, this paper proposes an interpretable multimodal heterogeneous fusion liquid impulse neural network (LINN) model. First, a liquid state coding layer based on differential equations is constructed to model the timeseries dynamic evolution features in non-stationary signals via a chunked feedback mechanism. Moreover, a multi-channel leaky integrate-and-fire (MC-LIF) impulse neurons are introduced to enhance the extraction of transient shock features by combining alternative gradient and membrane potential attenuation strategies. Finally, an attention-guided multimodal fusion mechanism is designed to realize adaptive integration and contribution interpretability quantification of time-frequency features. In the high-noise and variable-load condition tests, LINN achieves more than 98.7% accuracy with only 4.1 M parameters and 88.64% cross-condition generalization accuracy. The ablation experiments verify the key role of liquid layer and impulse mechanism in enhancing dynamic modelling and noise immunity, and the interpretability analysis based on time-frequency domain attention (TFDA) further reveals the sensitive response of the model to key time-frequency modal contributions. The method provides an effective solution with high accuracy, strong generalization and interpretability for intelligent diagnosis under complex working conditions.
|
| 关键词 |
旋转机械故障诊断 ; 液脉冲神经网络 ; 多通道漏电集成与火灾(MC-LIF)脉冲神经元 ; 注意力引导融合 ; 时频域关注
|
| KeyWord |
Rotating machinery fault diagnosisLiquid-impulse neural networkMulti-channel leaky integrate-and-fire (MC-LIF) impulse neuronsAttention-guided fusionTime-frequency domain attention
|
| 基金项目 |
|
| 页码 |
113923-113923 |
eshun Y, Yingkui G, Haidon S, et al..
基于多模态信息异构融合的液体脉冲神经网络模型,用于可解释的旋转机械故障诊断 [J].
Mechanical Systems and Signal Processing.
2026; 246; ().
113923 - 113923.