As an important component of the atmospheric environment, bioaerosols have a profound effect on environmental quality, climate change, and human health. As environmental and public health problems intensify, the monitoring and identification of bioaerosols have attracted widespread attention. However, traditional bioaerosol identification methods, such as microbial culture and molecular biology techniques, are slow and complex. We combine attenuated total reflection Fourier transform infrared (ATR-FTIR) spectroscopy with one-dimensional convolutional neural network (1D-CNN) to leverage the high sensitivity, non-invasive and real-time advantages of spectroscopic technology, as well as deep learning powerful capabilities in feature extraction and classification of complex spectral data, and build an efficient and accurate bioaerosol identification model.

Bioaerosol samples, including three types of bacteria and three types of fungi, are used as the research object, and high-quality infrared absorption spectrum data are collected using a Fourier transform infrared spectrometer with an attenuated total reflection (ATR) accessory. To improve data quality, preprocessing techniques such as wavelet packet transform and Savitzky-Golay filtering are used for baseline correction and noise filtering. On this basis, a 1D-CNN model, including a convolution layer, a pooling layer, a dropout layer, and a fully connected layer, is constructed to utilize its powerful feature extraction and classification capabilities for the fast and accurate identification of bioaerosols. The effectiveness and superiority of the model are fully verified through reasonable data set division, multi-angle performance evaluation, and comparison with traditional machine learning methods. A mixed sample test plan of different concentrations is designed to further evaluate the model's generalization ability in complex environments.

Through comparative analysis of test set recognition accuracy, the 1D-CNN model proposed in this paper performs exceptionally well in the bioaerosol recognition task, significantly better than the traditional support vector machine (SVM) method. In identifying six bioaerosol samples, the accuracy of the 1D-CNN model reaches 100%, while the SVM achieves only 95%, fully demonstrating the advantages of convolutional neural networks in feature extraction and classification of complex spectral data. The generalization ability and robustness of the 1D-CNN model are further evaluated through methods such as confusion matrix analysis (Fig. 4) and cross-validation (Table 2). We also design tests with mixed samples of Aspergillus at different concentrations to simulate the real-world complexities. Experimental results show that the proposed method performs well in recognition tasks with subtle features, maintaining high accuracy and demonstrating the practicability and scalability of the method.

To achieve rapid and accurate identification of bioaerosols, we propose a new method based on 1D-CNN and ATR-FTIR. By applying the 1D-CNN deep learning model to feature extraction and classification of ATR-FTIR spectral data, the method achieves 100% accuracy in identifying six common bioaerosol samples, demonstrating significantly better performance than the traditional SVM method. In addition, the constructed model shows high recognition accuracy in cross-validation and low-concentration sample testing. This study illustrates the great potential of combining deep learning technology with ATR-FTIR spectroscopy for rapid and accurate bioaerosol identification, providing a new technical approach for environmental monitoring and public health protection.