Abstract:

In order to address false alarms or alarm failure caused by a single technology in fire detection, a combustion experimental platform was designed and built. PM10 and CO, concentration of which increased rapidly in the flue gas after fire, were selected as characteristic parameters.Then, data processing was conducted for them, and a fire detection model was established by adopting six machine learning algorithms, including logistic regression(LR), linear discriminant analysis (LDA), kNN algorithm, classification and regression tree(CART), naive Bayes, and support vector machine (SVM). Finally, the model's classification performance was assessed. The results show that among six algorithms, kNN features higher evaluation accuracy, recall rate, F₁ value and kappa value, with its accuracy of assessment reaching as high as 95.2%, making it possible to accurately identify combustion state. This method can accurately detect fire by classifying rapidly changing concentrations of PM10 and CO in combustion products.

Key words: machine learning, fire detection, PM10, classification algorithm, k-nearest neighbor algorithm(kNN)