Objective The stress concentration is an indicator for defect formation and final failure. The research on in-service inspection of stress status is an important criterion of healthy monitoring in metal components and structures. Laser ultrasonic is a promising method for stress measurement. In this paper, a laser ultrasonic system for stress measurement is built. In previous work, a single feature is used to evaluate the stress status of the component structure. There commonly are inherent limitations for stress analysis by using a single feature. Taking full advantage of features from different domains is promising to improve the accuracy of stress measurement. As different features may have high correlation with each other, it is significantly important to select features to eliminate the redundant information and reduce dimensionality of the dataset. Moreover, it is important to select the optimal machine learning method to build the stress prediction model. In this work, a multi-feature fusion network combining principal component analysis (PCA) and support vector machine (SVM) algorithm is proposed to analyze laser ultrasonic for assessing and predicting the stress status in materials. Besides, the performance of different regression models is compared. Methods Laser is used to generate the ultrasonic wave. The principle of laser ultrasonic stress measurement is based on acoustic effect. The velocity of ultrasonic wave is highly correlated with the stress status of materials. The experimental setup includes laser generation device, data acquisition card and stress applying system, as sketched in Fig. 1. YAG laser is used to induce ultrasonic wave. The acoustic emission sensor with frequency band between 60 and 400 kHz is used to collect the acoustic signal. By rotating the screw, a tension stress is applied on the sample. The amplitude of stress is shown by a stress indicator. The relationship between the laser ultrasonic and stress is obtained by applying different stress. To accurately evaluate the stress, a multi-feature fusion benchmark model is proposed for stress prediction, which is presented in Fig. 2. The original signal is filtered by Chebyshev filtering. Five features extracted from time or frequency domain are used to construct j×5 feature maps. Then these 10 features are combined to establish a j×10 map. By using PCA, the dimensionality of feature matrix is reduced from 10 to 5. The five principal components are used as input of SVM model to build the stress prediction model. Compared with the traditional regression model of single feature, multiple linear regression, Bayes and radom forest, the R2 value of the proposed model is the highest. Results and Discussions The filtered laser ultrasonic signal is presented in Fig. 3(b). The envelope of ultrasonic signal is calculated and shown in Fig. 4(a). The energy of frequency spectrum is concentrated between 150 and 500 kHz [Fig. 4(b)]. The time delay of the wave packet is increased with the increase of tension [Fig. 5(a)]. Although the relationship between delay time and stress is linear, the measurement result is unstable [Fig. 5(b)]. Different features are extracted for stress characterization. As the features have multicollinearity, the PCA method is utilized to reduce the dimensionality of feature maps. The cumulative contribution rate of principal components is shown in Fig. 6. Accordingly, we select the first five principal components to train the stress prediction model. It is a critical step to select an appropriate kernel function. By comparing the stress prediction results [Fig. 7], the radial basis function (RBF) kernel function is found to be optimal. To verify the superior performance of the proposed method, the stress prediction results by using different regression models are shown in Fig. 9. The errors of the stress prediction by using single feature model, multiple linear regression model, and Bayes model are relatively high. The random forest and SVM methods are more robust than other regressive approaches for stress measurement. From Fig. 11(a), it is seen that the R2 values by using SVM model in the training set and test set are 0.996 and 0.96, respectively. Moreover, the root mean square error (RMSE) by using SVM model is the lowest among all the prediction models. Conclusions In this work, multi-order statistical characteristics of laser ultrasonic from time and frequency domain are investigated for stress characterization. Chebyshev filter is designed to reduce the noise of the laser ultrasonic signal. As a result, the signal-to-noise ratio of the signal and the reliability of the stress prediction model are significantly improved. The feature map is constructed by extracting different order statistical characteristics from time and frequency domain. The multicollinearity of different features is analyzed by correlation analysis. The dimensionality of the feature maps is decreased from 10 to 5 based on PCA method. The redundancy and complexity of the stress prediction model are reduced. A lightweight feature fusion network based on the combination of PCA and SVM is proposed to build the stress prediction model. It is verified that RBF is the optimal kernel function. High precision stress evaluation of metal components can be realized based on laser ultrasonic time-frequency statistical feature fusion combining PCA and SVM. © 2024 Science Press. All rights reserved.
