Created
October 19, 2016 12:12
-
-
Save smrmkt/2dfe38015198006296ab2722fcb4b73d to your computer and use it in GitHub Desktop.
Anomaly detection for Fitbit
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
install.packages("fitbitScraper") | |
devtools::install_github("trinker/plotflow") | |
library(fitbitScraper) | |
library(ggplot2) | |
library(plotflow) | |
library(dplyr) | |
# Fitbitからのデータ取得 | |
cookie = login(email="www.fitbit.com ログイン用メールアドレス", password="www.fitbit.com ログイン用パスワード") | |
hr_data <- get_intraday_data(cookie, what="heart-rate", date="2016-10-06") | |
names(hr_data) <- c("time", "heart_rate") # 変数名にハイフンが入っているのでリネーム | |
hr_data <- hr_data %>% filter(heart_rate > 0) # Fitbitを取り外しているときは,値が0になるため取り除く | |
hr_plot <- ggplot(data=hr_data, aes(x=time, y=heart_rate)) + | |
geom_line(size=0.8) + | |
ylim(min=0, max=120) + | |
labs(x="Time", y="Heart rate") + | |
theme(axis.text=element_text(size=14), | |
axis.title=element_text(size=16, face="bold")) | |
plot(hr_plot) | |
# 特異スペクトル変換法による異常値判定 | |
w <- 16 # 部分時系列の要素数 | |
m <- 2 # 類似度算出に使用するSVDの次元数 | |
k <- w / 2 # SVD算出に用いる部分時系列数 | |
L <- k / 2 # 類似度を比較する2つの部分時系列群間のラグ | |
get_anomaly_score <- function(d, w, m, k, L) { | |
anomaly_score <- rep(0, length(d)) | |
for(t in (w+k):(length(d)-L+1)) { | |
start <- t - w - k + 1 | |
end <- t - 1 | |
X1 <- t(embed(d[start:end], w))[w:1,] # t以前の部分時系列群 | |
X2 <- t(embed(d[(start+L):(end+L)], w))[w:1,] # 異常度算出対象の部分時系列群 | |
U1 <- svd(X1)$u[,1:m] # X1にSVDを行い上位m次元を抜き出す | |
U2 <- svd(X2)$u[,1:m] # X2にSVDを行い上位m次元を抜き出す | |
sig1 <- svd(t(U1) %*% U2)$d[1] # U1とU2の最大特異値を取得 | |
anomaly_score[t] <- 1 - sig1^2 # 最大特異値の2ノルムを1から引くことで異常値を得る | |
} | |
anomaly_score | |
} | |
hr <- hr_data$heart_rate | |
anomaly_score <- get_anomaly_score(hr, w, m, k, L) | |
as_data <- data.frame(time=hr_data$time, anomaly_score=anomaly_score) | |
as_plot <- ggplot(as_data, aes(x=time, y=anomaly_score)) + | |
geom_line(color="red", size=0.8) + | |
labs(title="", x="Time index", y="anomaly score") + | |
theme(axis.text=element_text(size=14), | |
axis.title=element_text(size=16, face="bold")) | |
combined_plot <- ggdual_axis(lhs=hr_plot, rhs=as_plot) | |
plot(combined_plot) | |
# 特異スペクトル変換のパラメタによる変動 | |
as_data_w08 <- data.frame(time=hr_data$time, anomaly_score=get_anomaly_score(hr, 8, m, k, L)) | |
as_data_w16 <- data.frame(time=hr_data$time, anomaly_score=get_anomaly_score(hr, 16, m, k, L)) | |
as_data_w32 <- data.frame(time=hr_data$time, anomaly_score=get_anomaly_score(hr, 32, m, k, L)) | |
as_plot_w <- ggplot(as_data_w08, aes(x=time, y=anomaly_score)) + | |
geom_line(color="blue", size=0.8) + | |
geom_line(data=as_data_w16, aes(x=time, y=anomaly_score), color="red", size=0.8) + | |
geom_line(data=as_data_w32, aes(x=time, y=anomaly_score), color="green", size=0.8) + | |
labs(title="", x="Time index", y="anomaly score") + | |
theme(axis.text=element_text(size=14), | |
axis.title=element_text(size=16, face="bold")) | |
combined_plot_w <- ggdual_axis(lhs=hr_plot, rhs=as_plot_w) | |
plot(combined_plot_w) | |
as_data_m1 <- data.frame(time=hr_data$time, anomaly_score=get_anomaly_score(hr, w, 1, k, L)) | |
as_data_m2 <- data.frame(time=hr_data$time, anomaly_score=get_anomaly_score(hr, w, 2, k, L)) | |
as_data_m3 <- data.frame(time=hr_data$time, anomaly_score=get_anomaly_score(hr, w, 3, k, L)) | |
as_plot_m <- ggplot(as_data_m1, aes(x=time, y=anomaly_score)) + | |
geom_line(color="blue", size=0.8) + | |
geom_line(data=as_data_m2, aes(x=time, y=anomaly_score), color="red", size=0.8) + | |
geom_line(data=as_data_m3, aes(x=time, y=anomaly_score), color="green", size=0.8) + | |
labs(title="", x="Time index", y="anomaly score") + | |
theme(axis.text=element_text(size=14), | |
axis.title=element_text(size=16, face="bold")) | |
combined_plot_m <- ggdual_axis(lhs=hr_plot, rhs=as_plot_m) | |
plot(combined_plot_m) |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment