• Use of machine learning to reduce false alarms

      Ali, Muhammed Usman (University of Bedfordshire, 2020)
      Machine learning is adopted widely in many sectors including healthcare, automotive and finance where machine learning use cases include disease detection, predictive maintenance, and fraud detection. During 2017/2018, around 40%(226,000) of the incidents attended by fire and rescue service were false alarms. Therefore, this thesis is focused towards the application of machine learning on fire alarm systems data to address the rising problem of false alarms. The fire alarm system on site gathers the data about different events which can be utilised to conduct the experiments with machine learning. Therefore , to address this problem five different classification machine learning models including Logistic Regression, Support Vector Machines, Naïve Bayes Classifier, Decision Trees and Random Forests have been used to experiment with data gathered from fire alarm system. The performance of the different machine learning models is evaluated using different methods such as precision, recall, f1- score, confusion matrix, k-fold cross validation and mean accuracy to find the best suited models for reducing false alarm rates. Experiments were conducted on data gathered from the fire alarm system, 10-fold cross validation results indicated Naïve Bayes Classifier detecting 51 out of 53 Fires correctly but with a high misclassification rate and low mean accuracy of 61%. The remaining four models failed in classifying any fires correctly with 0.00 recall, still achieving overall accuracy in the range of 97-98% due to high imbalance in the dataset. The Cohen Kappa value of 0.0 was achieved by models indicating poor agreement in the decisions made. Machine learning models exhibited better performance on the new test data with incorporated temperature data, models achieved higher recall in the range of 0.70 to 1.00 during 10-fold cross validation as well as higher Cohen Kappa scores in the range 0.73 to 0.88 indicating substantial agreement in the decisions made by the machine learning models. The results on fire system data indicated machine may not be that effective due to poor correlation between the features in the data and high imbalance in the data. However, much better results are achieved by incorporating some additional sensors such as temperature into the fire alarm system data.