ANOMALY BEHAVIOR ANGKOT DETECTION OF ANGKOT USING AUTOENCODER LONG SHORT TERM MEMORY AND ID3 BASED ON THE TRAVEL HISTORICAL DATA

• Main Author: Rin Nurmalasari, Rin
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/55196
• Institution: Institut Teknologi Bandung

There are several types of public transportation in Indonesia, such as Angkot/Angkutan Kota (City Transportation) becoming the focus of this study. The angkot data used in this study was those collected from the applications, namely SEMUT and GPS Tracker, installed in every angkot. This GPS sensor data can be used for facilitating the monitoring of all activities and behaviors done by angkot. This study would create a model that could detect anomaly behavior angkot done by angkot. An anomaly is defined as abnormal behavior. The model was constructed to provide information if the anomaly behavior angkot was identified. Before conducting data analysis, the data taken from SEMUT should pass the data cleaning process, such as parsing, correction and standardization, enhancement, matching, and consolidation. The pre-processing was thereupon conducted including attribute selection, Exploratory Data Analysis (EDA), determining data training, and data testing. After conducting the EDA stage, the activities and behaviors done by angkot could be observed and defined as anomaly behavior angkot. Subsequently, the data separation for the normal behavior in the angkot to construct the anomaly behavior angkot detection using the Long-Short Term Memory (LSTM) autoencoder. This model was used for training the normal data without any label. The result from the model evaluation had a gradual drop in the loss MAE value. The final value in the training reached 0.0099. The output from the result of the anomaly behavior angkot detection using the long-short term memory autoencoder was the data label. The data would be automatically labeled with true or false. The true label was provided if the loss MAE value was more than the threshold value (0.1), indicating that there was anomaly behavior angkot. Meanwhile, the false label was given when the loss MAE in the tested data was less than the threshold value of 0.1, indicating that the angkot had normal behavior. After testing and checking by comparing the output of the model and the checking result using a manual method using the data exploration, the Long-Short Term Memory autoencoder had an accuracy of 89.6%. Further, the anomaly behavior angkot using the Iterative Dichotomiser Three (ID3) method was constructed. The anomaly behavior angkot done by the angkot consisted of off-Track, waiting time, and Speeding up. This training process was done by data normalization and nondata normalization. The evaluation technique for the training and testing was 10- fold cross-validation. To know the difference between the ID3 performance using normalization and the ID3 performance without normalization, the t-test was used.iv Based on the testing, since the sig. value of ? ?, namely sig. (p-value) of 0.001 or less than 0.05, the H0 was rejected and H1 was accepted (a significant difference identified). It can conclude that the performance of the ID3 using normalization is higher than the performance of the ID3 without normalization. The ID3 using normalization had an accuracy of 98%, while the ID3 without normalization had an accuracy of 95%. Furthermore, manual testing was applied to the models by comparing time of the model with time of the data exploration. The result of the checking and testing had an accuracy of 91.2%. Lastly, the result of EDA and the model is presented in the form of a web dashboard.