ELECTRICAL contain different types of anomalies and showthat LSTM

ELECTRICAL ENGINEERING DEPARTMENTINDIAN INSTITUTE OF TECHNOLOGY ROORKEEProposal OnAnomaly Detection of Temporal Data using LSTMCourse: EEN-300 Industry Oriented ProblemTitle :Modelling Time Series Data with Recurrent Neural Networks using LSTM unitsand Detecting Anomalies.Objectives 😕 To do a comparative study of different sequential data models? To study LSTMs, their complex architecture and ability to learn long-range dependencies? To build prediction models for temporal data and develop an anomaly detection method? To experiment with real-world datasets that contain different types of anomalies and showthat LSTM RNNs are suitable for general purpose time series modelling and anomalydetection.Abstract :The primary reason for the success of deep learning models is their ability to learn high-levelrepresentations but one major assumption of these models is the independence among datasamples. However, ANNs treat each data sample individually and thereby lose the benefit thatcan be derived by exploiting this sequential information.RNNs process the input sequence one element at a time and maintain a hidden state vector whichacts as a memory for past information. They selectively retain relevant information allowingthem to capture dependencies across several time steps thus allowing them to utilize both currentinput and past information while making future predictions.We start by doing a comparative study of different models of sequential data (Hidden MarkovModels HMM, feed-forward neural networks, RNNs with LSTM) and explore the use of Longshort-term memory (LSTM) for modelling sequential data and developing a method for anomalydetection in temporal data.Due to the challenges in obtaining labelled anomaly datasets, an unsupervised approach will beemployed. The model will be trained to learn the normal time series patterns and predict futurevalues. The resulting prediction errors are modelled to give anomaly scores and hence detectanomalies.We will conduct experiments with real-world datasets encompassing not only different kind ofpatterns but also the nature of anomalies and the results will be analyzed.Relevance to Industry :Anomaly detection is important because it often indicates useful, critical, and actionableinformation that can benefit businesses and organizations.It is utilized in a wide array of fields such as:? Fault detection in Industrial Systems? Fault detection in Power System? Health Monitoring (Anomaly detection in Electrocardiograms(ECG) signals)? Fraud Detection for Financial Transactions? Intrusion Detection (identifying strange patterns in network traffic that could signal ahack)? Risk and Uncertainty in the Evaluation Investments (Housing and Stock marketpredictions).Future Scope :This project will aid in the creation of labelled data that can be further exploited to predictpossible patterns and types of anomalies which may occur. If enough labelled data havinganomalies is available then the prediction accuracy can be improved and the real-timeapplication will be possible.Also, if developed further in the right domain and given suitable data set, anomaly detectionsystem can be used to predict anomalies in industries such as power industries (surge in powerdemand), the stock market (share price), weather forecasting (extreme change in weathercondition), etc.Timeframe :Description of Work Supposed date of completionPhase 1 Study LSTMs, their complex architecture Mid-February 2018Phase 2 Build prediction models for temporal data End-March 2018Phase 3 Develop an anomaly prediction method andexperiment on real-world datasetsMid-April 2018Mentor’s Details :Name Designation Contact No. Email-idDr Felix OrlandoMaria JosephAssistant Professor,EED IITR8979897211 [email protected] Members’ Details :S.No Name Enrollment No. Contact No. Email-id1 Deepak Verma 15115045 9717806207 [email protected] Vikas Kumar 15115133 7060864145 [email protected] Yash MukundKant15115140 9075234514 [email protected] of Approval :Guided by:Dr Felix Orlando Maria JosephAssistant ProfessorElectrical Engineering DepartmentSubmitted to:Prof. S.P. GuptaEmeritus ProfessorElectrical Engineering Department