第24章: 実践プロジェクト集

# Smart Manufacturing Analytics - 総合パッケージ読み込み library(tidyverse) # データ操作の基盤 library(tidymodels) # 機械学習パイプライン library(torch) # 深層学習 library(modeltime) # 時系列予測 library(plotly) # インタラクティブ可視化 library(shiny) # ダッシュボード構築 library(DT) # データテーブル library(shinydashboard) # ダッシュボードUI library(config) # 設定管理 library(pins) # データバージョン管理 # プロジェクト設定の初期化 config <- config::get() board <- board_local() # データパイプライン設定 data_sources <- list( iot_sensors = "data/iot_sensor_data.csv", production_logs = "data/production_logs.csv", quality_data = "data/quality_measurements.csv", maintenance_records = "data/maintenance_history.csv" ) # モデル設定 model_config <- list( prediction_horizon = 24, # 24時間先まで予測 update_frequency = "hourly", model_types = c("arima", "prophet", "xgboost", "neural_network") )

# 統合データパイプライン関数 create_integrated_dataset <- function() { # IoTセンサーデータの読み込みと前処理 iot_data <- read_csv(data_sources$iot_sensors) %>% mutate( timestamp = ymd_hms(timestamp), # 異常値検出とクリーニング temperature = ifelse(temperature > 1000 | temperature < -50, NA, temperature), pressure = ifelse(pressure < 0, NA, pressure), vibration = ifelse(vibration > 100, NA, vibration) ) %>% # 時間ベースの特徴量エンジニアリング mutate( hour = hour(timestamp), day_of_week = wday(timestamp, label = TRUE), shift = case_when( hour %in% 6:14 ~ "morning", hour %in% 14:22 ~ "afternoon", TRUE ~ "night" ) ) # 生産ログデータの処理 production_data <- read_csv(data_sources$production_logs) %>% mutate( timestamp = ymd_hms(timestamp), efficiency = production_count / target_count, # ローリング統計の計算 rolling_avg_efficiency = slide_dbl(efficiency, mean, .before = 23, .complete = FALSE), efficiency_trend = efficiency - lag(rolling_avg_efficiency, 24) ) # 品質データの統合 quality_data <- read_csv(data_sources$quality_data) %>% mutate( timestamp = ymd_hms(timestamp), quality_score = (100 - defect_rate) * quality_index, quality_category = case_when( quality_score >= 95 ~ "excellent", quality_score >= 85 ~ "good", quality_score >= 70 ~ "acceptable", TRUE ~ "poor" ) ) # 全データソースの時系列結合 integrated_data <- iot_data %>% left_join(production_data, by = "timestamp") %>% left_join(quality_data, by = "timestamp") %>% # 欠損値の補間 arrange(timestamp) %>% fill(everything(), .direction = "down") %>% filter(complete.cases(.)) return(integrated_data) }

# アンサンブル予測モデルの構築 build_predictive_models <- function(data) { # データ分割（時系列を考慮） splits <- data %>% time_series_split( date_var = timestamp, assess = "2 weeks", cumulative = TRUE ) # 特徴量エンジニアリング・レシピ recipe_spec <- recipe(equipment_failure ~ ., data = training(splits)) %>% step_rm(timestamp) %>% # 時系列特徴量の生成 step_lag(temperature, pressure, vibration, lag = 1:6) %>% step_slidify(temperature, pressure, vibration, period = 24, .f = mean, prefix = "mean_24h_") %>% step_slidify(temperature, pressure, vibration, period = 24, .f = sd, prefix = "sd_24h_") %>% # カテゴリ変数の処理 step_dummy(all_nominal_predictors()) %>% # 数値変数の正規化 step_normalize(all_numeric_predictors()) %>% # 欠損値の処理 step_impute_knn(all_predictors(), neighbors = 5) # モデル仕様の定義 rf_spec <- rand_forest( trees = tune(), mtry = tune(), min_n = tune() ) %>% set_engine("ranger", importance = "impurity") %>% set_mode("classification") xgb_spec <- boost_tree( trees = tune(), tree_depth = tune(), learn_rate = tune() ) %>% set_engine("xgboost") %>% set_mode("classification") # ワークフローセットの作成 workflow_set <- workflow_set( preproc = list(recipe = recipe_spec), models = list( random_forest = rf_spec, xgboost = xgb_spec ) ) # ハイパーパラメータチューニング cv_folds <- time_series_cv(training(splits), assess = "1 week", cumulative = FALSE) tuned_models <- workflow_set %>% workflow_map( "tune_grid", resamples = cv_folds, grid = 25, metrics = metric_set(accuracy, roc_auc, precision, recall), verbose = TRUE ) return(list( splits = splits, tuned_models = tuned_models, cv_folds = cv_folds )) }