交联聚乙烯电缆绝缘层机械性能的光谱表征与预测方法研究

Objective The elongation at break (EAB) and tensile strength (TS) are key indicators for characterizing the insulation condition and mechanical strength of cross-linked polyethylene (XLPE) cable. The current condition monitoring techniques involve complex sample preparation processes and lack timeliness, which severely restricts their application in on-site operational condition assessment. Near-infrared (NIR) spectroscopy is a non-destructive detection method based on the analysis of molecular vibration absorption characteristics. It offers prominent advantages such as fast speed, low sample requirements, and simple operation, making it the most promising optical detection approach. Preliminary research has identified characteristic absorption peaks in the NIR range of XLPE that are highly correlated with aging, which can serve as a basis for evaluation. However, studies on NIR detection of XLPE are still in the exploratory stage. On one hand, existing research remains at the descriptive level of spectral feature phenomena, lacking explanation for the underlying mechanisms of spectral changes. This results in models lacking mechanistic guidance, making it difficult to extract truly aging-relevant feature parameters from spectral data. On the other hand, mainstream models mostly rely on traditional chemometric algorithms, which have limited capability in mining high-dimensional and nonlinear spectral data, thus failing to establish an accurate mapping relationship between spectral features and aging conditions. Consequently, the detection accuracy and reliability fail to meet engineering application requirements. Based on this, this paper proposes a new NIR-based method for detecting the mechanical properties of XLPE, focusing on the study of material degradation mechanisms at the microscopic level, feature extraction, and the development of high-performance prediction models. Methods By conducting accelerated thermal aging tests, we systematically collect XLPE cable insulation samples at different aging stages. We use a portable spectrometer and an efficient acquisition method to obtain the NIR spectra of the XLPE samples. Tensile tests are employed to calibrate the data. Based on Savitzky-Golay (S-G) smoothing and standard normal variate (SNV) transformation, the spectra are pre-processed to remove the interference of noise and surface scattering of the samples. Through two-dimensional correlation spectroscopy analysis and quantum chemical calculations, the patterns and microscopic mechanisms of the evolution of NIR spectra during the aging process are revealed. Eight key features within the 900‒2100 nm band are identified and their group assignments are obtained, which are regarded as the core spectral features related to aging. On this basis, a diversified subspace ensemble model with strong expression ability for high-dimensional and nonlinear spectral data is proposed. After that, we compare the modeling effects of different base learners, select the Gaussian process regression (GPR) model as the optimal learner, and use the Bayesian optimization algorithm to optimize the random subspace dimension d, the number of base learners l and the penalty coefficient λ in the model, so as to achieve high-precision prediction of EAB and TS. Results and Discussions The results show that when l=25, d=48, and λ=0.12, the EAB prediction model performs best (root-mean-square error (RMSE) is 11.6,average relative error (MAPE) is 4.5%, coefficient of determination (R2) is 0.997). When l=27, d=45, and λ=0.15, the TS prediction model performs best (RMSE is 0.57, MAPE is 2.2%, R2 is 0.996). The average relative errors of the proposed model for predicting EAB and TS are both less than 5%. Compared with single learner modeling, the errors are reduced by nearly 60%, demonstrating superior mechanical property prediction ability. To comprehensively evaluate the model performance, the prediction effects of multiple ensemble learning methods are compared. The experimental results indicate that the method in this paper has the lowest RMSE and MAPE among all the compared methods, showing excellent prediction performance when dealing with high-dimensional vector data. To further verify the effectiveness of the method, 12 real aged XLPE cables with different operating times are selected. Their insulation layers are sliced, and the spectral data are collected using a NIR spectrometer. The constructed model is used to evaluate the mechanical properties, and the results are compared with the measured values from the tensile test. The results show that the prediction errors of the model for the EAB and TS of the real aged samples are both less than 10%. Conclusions To meet the need for rapid detection of the mechanical properties of XLPE cable insulation materials, this paper combines experimental analysis and theoretical calculation to reveal the correlation mechanism between NIR spectral features and the aging state of the materials, and establishes a high-precision mechanical property prediction model. The main conclusions are as follows. 1) By combining two-dimensional correlation spectroscopy analysis and quantum chemical calculation, the key spectral response mechanisms in the XLPE aging process are obtained. The study finds that there are 8 autocorrelation peaks in the 900‒2100 nm band. Among them, characteristic peaks such as 1217 nm and 1706 nm are sensitive to the main-chain breakage, and characteristic peaks such as 1394 nm, 1420 nm, and 1751 nm mainly reflect the accumulation of oxidation products such as aldehydes, carboxylic acids, and esters. 2) A diversified subspace ensemble modeling strategy based on the random perturbation mechanism is proposed, which effectively solves the problems of information loss in traditional methods and non-linear modeling. The RMSE values of the model for predicting EAB and TS are 11.6 and 0.57 respectively, and the relative errors of the predictions for real aged samples are all less than 10%. 3) The method proposed in this paper can complete the spectral acquisition of samples and the high-precision prediction of mechanical properties within 30 s, significantly improving the efficiency and enabling on-site detection. It provides an efficient and feasible technical means for the rapid and in-situ assessment of the insulation state of XLPE cables.