Prediction of gas mixture reactivity based on detonation pipe vibrations

The aim of this paper is to make an approach of creation a machine learning system predicting gas mixture composition being burned in a process pipe, based on pipe vibrations measurements. Task is divided into two parts: performing an experiment to get a necessary experimental data, and developing prediction algorithm. First, the basic principles of machine learning and signal processing are presented. Machine learning is the subfield of computer science that focuses on creating algorithms that can learn from provided data and perform predictions, either classification or regression. Signal processing is a general statement for all activities performed on information in form of a signal. In this particular work the emphasis is put on Fourier transform. After introduction, a brief description of the pipe response to internal detonation and pressure load is provided. It is of most significance, since the sensors used in the experiment base on pipe vibrations. Finally, the experimental part is described. The experiment consisted of performing a series of hydrogen-air explosion in pipes, with various hydrogen concentration. Measurement is performed with three sensors: piezoelectric sensor, knock combustion sensor - both measuring vibrations of pipe - and a pressure sensor, measuring pressure. This data is fed to a machine learning algorithm, that works as follows: first, measurement from a sensor is interpolated using b-splines. Then it transposes data from time domain to frequency domain using Fourier transform. Afterwards it is merged into one array. The set is divided into training and scoring sets, using cross-validation techniques. Training sets are used to feed classificator: SVM, SGD, naive Bayes, logistic regression, linear SVC, Ada Boost, perceptron. From this algorithm the prediction score of each classificator is derived and arranged with each other. It appears, that the algorithms used in conjunction with piezoelectric sensor give the score averaging to 50 %. The analysis of frequency spectrum is needless, since there is not enough features. The best classifiers are Perceptron, Naive Bayes and Support Vector Machine. Data from pressure sensor give much better results, with accuracy even up to 90 %. Fourier transform boosts the accuracy of classifiers. The best one is logistic regression. Therefore prediction of gas mixture reactivity based on detonation pipe vibration is possible.

---

Opracowanie rekordu w ramach umowy 509/P-DUN/2018 ze środków MNiSW przeznaczonych na działalność upowszechniającą naukę (2018).