With the economy’s development and the enhancement of people’s living standards, the need for daily-use ceramic is increasing day by day. To meet various life scenes and requirements, there are not only low-cost products in the market but also high-end products. In the trend of product diversification, there exists the phenomenon of low-priced goods being counterfeited as high-priced goods, which damages the consumers’ rights and interests. Due to the difference in the producing area and the technology, the quality and price are set differently. It is difficult to identify with the naked eye due to its slight difference in appearance, so we must take reasonable measurements to select and identify products strictly.

The laser-induced breakdown spectroscopy (LIBS), a fast and emerging element detection approach, has the benefits of simple sample preparation and fast detection speed.The LIBS can be done in situ and can be employed for the identification of various kinds of daily-use ceramics. For this purpose, LIBS integrated with a generalized regression neural network (GRNN) is used for the classification and identification of various types of daily-use ceramics. All the 40 daily-use ceramic samples employed in this experiment are purchased from the market. The places of origin are Chaozhou City in Guangdong Province, Jingdezhen City in Jiangxi Province, Liling City in Hunan Province, and Suzhou City in Jiangsu Province. Before we get start, the samples are cleaned, dried, and then fragmented. The four various categories of daily-use ceramic samples are excited using LIBS technology. It is worth noting that a non-painting small fragment demonstrating a flat surface and without colored glaze is selected to produce the LIBS emission spectrum. LIBS spectra of ceramic samples are separated into training sets and test sets at random. The GRNN classification model is trained using training sets, and the test sets are employed to confirm the generated classification model.

The experiment primarily consists of two aspects: first, discussing the impact of spectral data extraction on model effectiveness; second, optimizing the model’s performance. To discuss the impact of spectral data extraction on model effectiveness, two spectral data inputs are proposed: the whole spectral range and various primary characteristic spectral lines of the main elements. The ceramic body’s main elements include Fe, Mg, Si, Ca, Ti, K, Al, and Na. The main characteristic spectral lines are screened from the National Institute of Standards and Technology Atomic Spectra Database. Based on the LIBS spectra produced in the experiment and the screening principle of characteristic spectral lines, 28 characteristic lines of various elements are reserved (Table 1). To optimize the classification model’s performance and enhance the accurate rate of recognition, the model with the set of 28 characteristic spectral lines of the primary elements of the ceramic body as the input is improved. Before modeling, the spectral data are adjusted and then the abnormal spectral data are excluded using the Mahalanobis distance to reduce the adverse effect of poor spectra on the GRNN classification model. Then, the principal component analysis is employed to further simplify the LIBS spectral data. As shown in Fig. 6, the four major factors’ accumulative contribution rate to total variation accounts for more than 95%, maintaining most of the information, and achieving the purpose of simplifying the data. After finishing the GRNN model optimization, the two kinds of daily-use ceramics’ average spectra from Suzhou and Chaozhou are computed, and the difference in the various elements’ contents is discussed.

As demonstrated in Table 2, the analyses show that the screening of valid data for the LIBS spectrum can significantly increase the modeling efficiency and the speed by about 45 times. On this basis, the spectral data are normalized, the abnormal spectra are screened out, and then the principal component analysis is employed to optimize the network’s performance. As demonstrated in Figs. 4 and 5, the prediction accuracy of GRNN increases from 94.5% to 97.44% after being screened by Mahalanobis distance. As demonstrated in Fig. 7, the prediction accuracy of GRNN after principal component analysis (PCA) can reach 100%. The process is repeated and the results are averaged. The findings are summarized in Table 3, the single prediction accuracy for the second set can reach 100%, and the average prediction accuracy can reach 99.74% after normalization, the data filtering with Mahalanobis distance and principal component analysis. Figure 8 describes the difference between the two ceramics’ average spectra from Suzhou and Chaozhou, the spectral line intensities of Ca, Al, K, Mg, and Si are visibly different.

In this research, the LIBS integrated with GRNN for the classification and identification of various types of daily-use ceramics is proposed. It has gained excellent findings. Taking the whole spectral range as the input of the GRNN classification model takes longer time than taking various main characteristic spectral lines of the primary elements as the input of the GRNN classification model. The model’s poor performance is attributed to data redundancy. The modeling speed is increased by 45 times after the screening of spectral lines. On this basis, spectral data are normalized, abnormal spectra are screened out using Mahalanobis distance, and then the principal component analysis is employed to optimize the network’s performance. The accuracy of classification is further enhanced. On all these counts, the correct classification of daily-use ceramic from their LIBS spectra integrated with GRNN can be achieved.