Abstract
The global stress distribution and state parameter analysis of the building's main structure is an urgent problem to be solved in the online state assessment technology of building structure health. In this paper, a stress curve clustering algorithm of fiber Bragg grating stress sensor based on density clustering algorithm is proposed. To solve the problem of large dimension and sparse sample space of sensor stress curve, the distance between samples is measured based on improved cosine similarity. Aiming at the problem of low efficiency and poor effect of traditional clustering algorithm, density clustering algorithm based on mutual nearest neighbor is used to cluster. Finally, the classification of the daily stress load characteristics of the sensor is realized, which provides a basis for constructing the mathematical analysis model of building health. The experimental results show that the stress curve clustering method proposed in this paper is better than the latest clustering algorithms such as HDBSCAN, CBKM, K-mean++, FINCH and NPIR, and is suitable for the feature classification of stress curves of fiber Bragg grating sensors.
Introduction
Modern structural health monitoring for large-scale infrastructure increasingly relies on Fiber Bragg Grating (FBG) sensors. These optical devices reflect specific wavelengths of light corresponding to strain or temperature changes, providing high-accuracy, real-time data. However, the sheer volume of data generated—a "data deluge"—presents a significant analytical challenge for engineers attempting to assess structural integrity.
Engineers need to analyze daily stress curves, which represent the structural "heartbeat" over a 24-hour period. These curves are high-dimensional and often reside in sparse data spaces, making traditional clustering methods like K-means or DBSCAN inadequate. Such methods frequently struggle with noise, non-spherical data distributions, and the need for manual parameter tuning that does not scale well to complex building environments.
Methodology
To address these challenges, this research proposes a specialized multi-stage pipeline designed to classify daily stress curves effectively. The process focuses on identifying patterns of structural health by prioritizing the "shape" of the stress data over raw magnitude.
The process begins with Preprocessing and Dimensionality Reduction, utilizing Principal Component Analysis (PCA) to condense 24-hour curves (often containing hundreds of data points) into essential attributes while preserving the underlying shape. A key innovation is the use of Improved Cosine Similarity instead of standard Euclidean distance. This metric is more robust for high-dimensional data, focusing on the relative trends and patterns of the stress curves.
The core of the approach is the Mutual Nearest Neighbor (MNN) Clustering algorithm. Unlike standard density-based clustering, MNN only connects points if they are mutually among each other’s $k$-nearest neighbors. This "backbone" strategy identifies the skeleton of a cluster first, making the algorithm highly resistant to noise and preventing "bridge" points from incorrectly merging distinct structural states.
Experimental Setup and Results
The algorithm was validated using 15 standard benchmark datasets and real-world FBG data collected over three months. Each real-world sample consisted of a 24-hour stress curve sampled every two minutes. The proposed method was compared against several state-of-the-art algorithms, including HDBSCAN, CBKM, K-means++, FINCH, and NPIR.
The results demonstrate that the MNN-based approach consistently outperforms its competitors. On real-world sensor data, it achieved the highest Calinski-Harabasz (CH) index, indicating superior cluster separation and density. Furthermore, as shown in the performance metrics across various benchmark datasets, the proposed algorithm maintains high Normalized Mutual Information (NMI) and Adjusted Rand Index (ARI) scores, even on complex datasets like "Spiral" where traditional centroid-based methods fail.
Conclusion
This research provides a robust framework for categorizing the daily stress load characteristics of buildings. By prioritizing shape patterns and utilizing a mutual nearest neighbor density approach, the algorithm offers a reliable basis for mathematical models of building health. Future work will focus on making the $k$-parameter adaptive and optimizing the algorithm's computational complexity for large-scale applications in civil engineering.
Summary updated