结合软约束的演化数据流模糊聚类算法

多源局部放电检测中，不同类型的局放信号同时存在且不断变化使得信号的分离更具挑战，而这种情况同样存在于许多数据流的聚类分析场景中。为了能够适应类簇内的不均匀密度和类簇间的重叠边界问题，同时对数据流的漂移和演化进行及时跟踪，提出了一种结合软约束的实时数据流模糊聚类算法。算法引入2种模糊性软约束来描述微簇距离和密度上的不确定度，通过阈值划分出核心微簇、边界微簇和离群微簇;在类簇边缘使用模糊隶属度，给予微簇分属不同类簇的可能性，保证类簇的完整性并提高聚类效果;使用两阶段的流程结构和2种时间窗口模型，赋予算法具有对可变化数据流的适应能力和更低的时间空间占用率。在多种数据集上的实验表明，该算法相比同类型算法在聚类效果上提升了1%~3%，且平均运行时间缩短5%~20%，在实际硬件平台的测试中也验证了算法的聚类分离性能。

Fuzzy clustering algorithm for evolving data streams combined with soft constraints

In multi-source partial discharge (PD) detection, distinct PD signals exist simultaneously and change constantly, which makes signal separation more challenging. This situation also exists in many scenarios of clustering analysis of data streams. To adapt to the heterogeneous density within the cluster and overlapping borders between clusters, and to track the drift and evolution of data streams in time, this paper proposes a real-time data stream fuzzy clustering algorithm combined with soft constraints. Firstly, two fuzzy soft constraints are introduced to describe the uncertainty in the distance and density of micro-clusters (mc). These micro-clusters are divided into core-mc, border-mc and outlier-mc based on thresholds. Secondly, fuzzy membership degrees are used at the edge of the clusters to estimate the possibility of mc belonging to different types of clusters, ensuring the integrity and improving the clustering effect. Finally, the method uses a two-stage procedure and time window models to endow the algorithm with adaptability to changing data streams and lower memory occupancy. Experiments on various dataset show that the clustering effect of this algorithm is improved by 1%~3% and the average runtime is shortened by 5%~20% compared with counterparts. Separation performance is also verified in the hardware platform test.