基于可视化高维张量空间分解的动力装置振动一致性溯源研究

The bigger discrete vibration of the power system, caused by the defect of processing and assembly of set of parts, will directly affect the noise index of equipment. Thereby, how to ensure the vibration consistency of power system is an urgent needed. Due to the multi-source excitation and coupling vibration, the vibration signals are mixed and unable to assess and component discriminant problem. The key of vibration consistency tracing of power system is the demixing and identification problem of variable condition and multi-source excitation response. To realize this target, the source tracing of vibration consistency for a power system based on visual high-dimensional tensor factorization is proposed in the project. With the excitation source in group components based on the analysis of the vibration transmission and coupling rule, the vibration tensor space of power system can be constructed by reasonable distribution of sensor and the working condition of performance parameters. On the basis of the visual analytics, Non-negative Tensor Factorization is adopted to extract the feature sub-spaces of the different vibration source. According to the visual embedded low-dimensional subspace, it is available to evaluate the vibration states with the feature fusion of vibration source, meanwhile, the visual correlation analysis of influencing factors is adopted to find the problem set of parts, the tracing of power system vibration anomalies can be realized effectively, which can supply the power system of equipment tests and operation safety.

由于组部件的加工、装配质量等方面缺陷，会导致动力装置的振动出现较大离散，直接影响到噪声指标。如何保证动力装置振动的一致性成为迫切需要解决的问题。而动力装置的多工况、多源激励振动耦合使得振动信号混合，无法准确评估和判别问题组件。因此，动力装置振动一致性溯源的关键是多工况、多振源激励响应的解混和辨识问题。本项目提出了基于可视化高维张量空间分解的动力装置振动一致性溯源评估研究。在关键组部件激励源振动传播与耦合规律的基础上，通过高维特征张量空间来表征振动与工况性能信息的变化，进而利用可视化张量的非负分解来感知振源的特征子空间，在可视化子空间中通过振源特征的融合评估振动离散状态，并根据影响因素的可视化关联分析来寻找问题组部件，实现动力装置振动异常溯源，从而保持和提高动力装置的运行质量。

由于动力装置结构的高耦合性，振动信号传递路径的复杂性，多工况环境及固液耦合等因素，使得在其振动过程中无法准确评估和判别问题组件。为此，本项目在部件激励源振动传播与耦合规律的基础上，通过复杂网络理论建立了测点与特征信息的一致表征；采用优化参数的短时傅里叶变换时频谱构建出高维特征张量空间，有效表征了装置的振动状态；研究了时频张量的非负分解模型，通过稀疏约束项有效提取张量数据的局部特征；提出了基于稀疏约束多通道时频张量的振动特征分离方法和基于相空间重构时频张量分解的冲击故障特征感知方法，分解出有效的振动特征，实现特征的解混；利用分解矩阵的聚类特性，提出了基于非负张量分解聚类的振动状态评估和溯源方法，根据样本聚类和核张量的可视化关联实现状态评估与异常溯源。