Foodborne pathogenic bacteria pose a serious threat to food safety and global public health, making the development of sensitive and accurate detection methods imperative. Traditional detection methods, such as plate count techniques, molecular biology assays, and immunological assays, require intricate procedures, present potential false outcomes, and demand skilled operators. The rapid development of information technology and the widespread application of intelligent technologies have created new opportunities for efficient and accurate bacteria detection and identification. This study aims to design a comprehensive experiment for bacteria identification based on machine learning, integrating professional knowledge of food microbial detection with machine learning–based data processing techniques. The ultimate goal is to guide students to transform intelligent technologies into innovative tools for solving practical problems in the food industry and to cultivate their ability to address complex engineering challenges creatively.

Multicolor carbon quantum dots (CQDs) were synthesized via a one-pot solvothermal method at 180 ℃ for 6 h, utilizing o-phenylenediamine as the precursor in the presence of different acid reagents (e.g., citric, boric, terephthalic, tartaric, and nitric acids) to modulate fluorescence emission. Following synthesis, the products were purified through dialysis and freeze-drying. The optical and surface properties of the CQDs were characterized using fluorescence spectroscopy and zeta potential analysis. For the sensing assay, bacterial suspensions were standardized to an OD₆₀₀ of 1 and incubated with the CQDs for 4 h. The differential adsorption of CQDs on bacterial surfaces was elucidated via confocal laser scanning microscopy (CLSM). Fluorescence responses were quantitatively recorded using a microplate reader across five detection channels (Ex/Em: 339/421, 387/541, 399/572, 543/605, and 616/639 nm). Subsequently, linear discriminant analysis (LDA) was performed in Python to process the multivariate fluorescence data for dimensionality reduction and bacterial classification.

The synthesized CQDs, denoted as B-, G-, Y-, O-, and R-CQDs, exhibited excitation maxima at 339, 387, 399, 543, and 616 nm, respectively. The B-, G-, and Y-CQDs displayed single emission peaks (421, 541, and 572 nm), respectively, whereas the O- and R-CQDs exhibited characteristic dual-emission peaks at 605/650 nm and 639/680 nm. The CQDs showed quantum yields ranging from 12.96% to 28.51% and distinct zeta potentials (5.02–7.53 mV). CLSM imaging and spectral analysis confirmed that different bacterial species selectively accumulated CQDs to varying extents, particularly in the green and yellow fluorescence channels, thereby generating unique fluorescence fingerprints. LDA demonstrated robust discrimination capability, with the first two canonical discriminants accounting for 93.3% of the total variance (F1: 83.6%; F2: 9.7%) and producing well-separated clusters within 95% confidence ellipses. Confusion matrix analysis verified 100% classification accuracy for both the training and testing datasets. Furthermore, the sensing array retained its efficacy in tap water samples, where the first, second, and third discriminant functions explained 74.4%, 22.0%, and 3.1% of the variance, respectively, achieving 100% identification accuracy without complex pretreatment.

This experiment successfully combines food microbial detection with intelligent data processing technology. It not only verifies the feasibility of using CQD-based fluorescence fingerprints combined with the LDA algorithm for bacterial identification in food safety applications but also provides an effective educational platform for students. The project helps students master the application of machine learning in food safety detection, enhances their innovative thinking and practical skills, and lays a solid foundation for their future involvement in solving practical engineering problems in the food industry.