The student approached the topic independently and responsibly throughout the year. He initiated his own data collection, carried out in several phases so that the data could also be evaluated with respect to potential changes over time. Partial results of the assignment were already successfully presented by the student this year in the form of the conference paper Classification of Normal and Anomalous Air Using BME688 Sensor Data (EEICT). He analyzed and preprocessed the collected data using available tools and selected three architectures for training an odor anomaly detector.

The text of the thesis itself can, in places, be criticized mainly for its noticeable inspiration by language models, especially in the formatting style, chapter titles, etc. In terms of content, the theoretical part is sufficient. My main criticism concerns the practical part, where the student sometimes does not provide the reader with sufficient information about the steps and experiments being performed, e.g., why classification into several classes is carried out, what the justification for the selection of the models is, for what purpose the weekly “normal air” data are analyzed, etc., or this information is lost in an unclear amount of text.

The thesis assignment was fulfilled, and based on the shortcomings mentioned above, I grade the thesis with 85 points. Points proposed by supervisor: 85

Student describes the theory and state of the art in olfactory sensing and machine learning methods suitable for anomaly detection in this field. The work focuses mainly on Random Forest and autoencoder-based approaches, specifically a dense autoencoder, a dense autoencoder using convolutional features, and an LSTM autoencoder. A more detailed description is devoted to the Bosch BME688 sensor used in the thesis, including its measurement configuration and heater profile. The thesis also addresses calibration, feature extraction, sensor validity, and data preprocessing. In the final part, the thesis presents the performance results of the individual models and discusses their suitability for practical anomaly-detection scenarios.
The quality of the thesis is reduced by its formal presentation. In many places, the text is written in an inconsistent style, and at times it is overly popularizing or colloquial, which is not appropriate for a diploma thesis. The theoretical part also contains some technical inaccuracies. For example, in the section on Random Forest, the explanation is mixed with a description of Isolation Forest. In addition, the dense autoencoder using convolutional features is evaluated in the experimental part, but it is not sufficiently introduced in the theoretical part. The citation style is not fully consistent; in some places, sources are cited as [6], [27], while elsewhere they are cited as [1, 21]. Some abbreviations and technical terms are introduced inconsistently. I would also appreciate a greater use of graphical elements or concise schematic explanations instead of long textual descriptions.
The assignment was fulfilled, and despite the above-mentioned reservations, I recommend the thesis for defence with the grade B, corresponding to 81 points. Topics for thesis defence:

1. The thesis uses the fixed heater profile heater_354 throughout the experiments. Did you test other BME688 heater profiles or duty-cycle configurations? If not, how could the choice of heater profile influence the separability of individual substances and the performance of the anomaly detector?
2. Did you test dynamic transitions between substances, for example clean air - substance A - substance B? If not, how would the proposed method handle residual contamination or a change from one anomalous substance to another without full chamber ventilation and recalibration?

Points proposed by reviewer: 81