Issue 78

D. S. Lobanov et alii, Fracture and Structural Integrity, 78 (2026) 207-219; DOI: 10.3221/IGF-ESIS.78.12

Analysis of the correlation coefficient matrices revealed similar patterns for samples cut along the weft and along the warp (Fig. 11). The strongest correlation, of around 0.8, is observed between the amplitude and duration of the signals. This indicates that the accumulation of significant damage within the body, followed by the formation of main cracks or the fracture of reinforcing fibres, is accompanied by the release of a large amount of elastic energy. This leads both to high amplitude values and to a prolonged decay period of relaxation oscillations within the material’s structure. A significant correlation, within the range of 0.7, is also observed between duration and rise time, indicating a proportional relationship between the increase in the total signal duration and the time taken to reach peak values. Physically, this reflects the dynamics of the development of large defect sources and shows that it takes longer for stresses to rearrange around them than for small local events. Meanwhile, the median frequency is negatively correlated with all other data, which may indirectly indicate that during the material’s destruction process, low-frequency oscillations are generated, whilst all others occur at higher frequencies.

Figure 11: Correlation coefficient matrices: 1 – Warp; 2 – Weft.

Figure 12: Signals clustered by different sample types according to the « amplitude–duration–rise time parameters ».

Thus, the amplitude, duration and rise time of the AE signals were selected for data clustering. Analysis of the dataset using the silhouette index and the Davis-Boldin index showed that, for most samples, division into four clusters would be optimal Clustering was performed for each sample type sequentially using the k-means and GMM methods. Fig. 12 shows the percentage ratio between the numbers of signals in each cluster for individual sample types. It is worth noting that the distribution proved to be uneven; in each group of samples, a single dominant cluster can be identified. However, for samples without defects and those with defects created by a cylinder and a sphere, the dominant clusters do not depend on

215