Scholarlink

<p>Self-organizing map is a popular unsupervised technique used to organize data into a structured two or three dimensional map and thus helpful to users for identification of clusters, outliers that are present in a high-dimensional data space. Interpretable self-organizing map (iSOM) is an advanced algorithm designed to overcome the challenge of folds and self-intersections commonly encountered in conventional SOM implementation. For environmental applications such as air and water quality monitoring, or climate pattern recognition, data are often collected from diverse sources at remote geographic locations. Distributed clustering allows local processing of these environmental datasets, reducing communication overhead and preserving data privacy. In this manuscript, the iSOM is suitably modified to perform distributed clustering in wireless sensor network (WSN) using winner neurons only. The resultant algorithm, termed as distributed interpretable self-organizing map (DiSOM), shares the winning neurons with the neighborhood sensor nodes in diffusion mode of mutual cooperation. This helps in reducing high-dimensional unlabeled data into meaningful clusters, making it easier to interpret and visualize trends. The proposed algorithm is tested on seven real-life WSN datasets to perform distributed clustering: (a) air quality management dataset of Delhi, (b) weather station monitoring dataset of Canada, (c) water quality monitoring dataset of Thames river, (d) Washington Cook Agronomy farm dataset, (e) Algerian forest fire dataset, (f) Intel Berkeley research lab dataset, and (g) land mines detection dataset. To evaluate the quality of learning of the&#xa0;proposed DiSOM, topographic error (TE) and quantization error (QE) are used as fitness measures. The convergence curves of TE and QE are lower, visualized weight positions have less folds, and winning neurons are more dispersed on the 2D grid map, as observed in the simulation results of all the seven datasets. The quality of clustering is measured with Silhouette index and Dunn Index, which reveal better performance of the&#xa0;proposed DiSOM over DSOM, Distributed PSO (DPSO), and K-means (DK-means) algorithms. Further, Kruskal–Wallis and Wilcoxon rank-sum-based non-parametric statistical tests reported that the proposed DiSOM algorithm values pass the null-hypothesis compared to the existing three benchmark algorithms.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Distributed clustering with interpretable self-organizing maps in wireless sensor network for environmental data analysis

  • 期刊论文

  • 发表日期：

    2026年3月27日

查看全文
  • Anita Panwar,
  • Satyasai Jagannath Nanda

摘要

Self-organizing map is a popular unsupervised technique used to organize data into a structured two or three dimensional map and thus helpful to users for identification of clusters, outliers that are present in a high-dimensional data space. Interpretable self-organizing map (iSOM) is an advanced algorithm designed to overcome the challenge of folds and self-intersections commonly encountered in conventional SOM implementation. For environmental applications such as air and water quality monitoring, or climate pattern recognition, data are often collected from diverse sources at remote geographic locations. Distributed clustering allows local processing of these environmental datasets, reducing communication overhead and preserving data privacy. In this manuscript, the iSOM is suitably modified to perform distributed clustering in wireless sensor network (WSN) using winner neurons only. The resultant algorithm, termed as distributed interpretable self-organizing map (DiSOM), shares the winning neurons with the neighborhood sensor nodes in diffusion mode of mutual cooperation. This helps in reducing high-dimensional unlabeled data into meaningful clusters, making it easier to interpret and visualize trends. The proposed algorithm is tested on seven real-life WSN datasets to perform distributed clustering: (a) air quality management dataset of Delhi, (b) weather station monitoring dataset of Canada, (c) water quality monitoring dataset of Thames river, (d) Washington Cook Agronomy farm dataset, (e) Algerian forest fire dataset, (f) Intel Berkeley research lab dataset, and (g) land mines detection dataset. To evaluate the quality of learning of the proposed DiSOM, topographic error (TE) and quantization error (QE) are used as fitness measures. The convergence curves of TE and QE are lower, visualized weight positions have less folds, and winning neurons are more dispersed on the 2D grid map, as observed in the simulation results of all the seven datasets. The quality of clustering is measured with Silhouette index and Dunn Index, which reveal better performance of the proposed DiSOM over DSOM, Distributed PSO (DPSO), and K-means (DK-means) algorithms. Further, Kruskal–Wallis and Wilcoxon rank-sum-based non-parametric statistical tests reported that the proposed DiSOM algorithm values pass the null-hypothesis compared to the existing three benchmark algorithms.