1. Department of Electrical Engineering
University of Arkansas
Hidden Markov Model for
Bad Data Detection
Md Abul Hayat
mahayat@uark.edu
Feb 8, 2019
2. Contents
• Markov Assumption
• State Space Models
– Properties
– Dynamic Linear Model
• Hidden Markov Model
– Introduction
• Problem Formulations
– Mathematical Proofs
• Experimental Results
10. Overview of Data
• 6 Patients, each having 2 signals or time series data
– We denote them as PAT# 33, 34, 35, 36, 37, 39
• Each patient has two signals
– PVP: Peripheral Venous Pressure (Blood Pressure Collected from Vein)
• Weak Periodic Component
– PZO: Peizo-electric Signal (Corresponds to Heart Rate)
• Strong Periodic Component
• Both these signals were sampled at 1000 samples/ second
• I am using data with sampling rate of 100 samples/ second
– 100 samples or data points correspond to 1 second.
11. Hidden Markov Model
• Hidden States (Discrete)
For each sample i, there is a hidden state
– θi = 0, which corresponds to good data and Yi ~ N(μ0, σ0
2)
– Similarly θi = 1, corresponds to bad data with Yi ~ N(μ1, σ1
2)
• Estimating Parameters: Baum-Welch (EM for HMM)
• Estimating States: Viterbi Algorithm
26. Experimental Results
• Hidden States (Discrete)
For each sample i, there is a hidden state
– θi = 0, which corresponds to good data and Yi ~ N(μ0, σ0
2)
– Similarly θi = 1, corresponds to bad data with Yi ~ N(μ1, σ1
2)
• Estimating Parameters: Baum-Welch (EM for HMM)
• Estimating States: Viterbi Algorithm
33. Comments on PVP Data Model Assumptions
• Two distributions should have different mean which are far apart.
• Bad part of data should have significantly large variance.
• The distribution changes after a big noise spike in some cases.
• HMM with Gaussian Distribution is good at removing sudden
spikes from time series data.
• Non-stationary assumption should also be taken care of.
35. HMM Results (PAT# 33 Continued)
– Using HMM for PZO can be used for more granular cleaning of both PVP
and PZO signals
– The distribution is not normal, other distributions can also be used.
– Can be modelled using 3 states.
PZO Signal