\u307e\u305a\u306f\u4eca\u56de\u4f7f\u7528\u3059\u308b\u30c7\u30fc\u30bf\u3092pandas\u3067\u8aad\u307f\u8fbc\u3093\u3067\u78ba\u8a8d\u3057\u3066\u898b\u3066\u307f\u307e\u3059\u3002<\/p>\n\n\n\n
import pandas as pd\nfrom matplotlib import pyplot as plt\ndf = pd.read_csv('Seismograph vertical acceleration nmsqsec of the Kobe earthquake recor.csv' ,parse_dates=True)\ndf.columns = ['times','x']\ndf.set_index('times', inplace = True)\nfig, ax = plt.subplots()\ndf.plot(legend=False, ax=ax)\nplt.show()<\/code><\/pre><\/div>\n\n\n\n![\"\"](\"https:\/\/obenkyolab.com\/wp-content\/uploads\/2022\/03\/image-4.png\")
<\/figure>\n\n\n\n\u898b\u305f\u611f\u3058\u3001times\u304c1000\u304f\u3089\u3044\u307e\u3067\u306f\u632f\u5e45\u304c\u843d\u3061\u7740\u3044\u3066\u3044\u3066\u30011500-2500\u306b\u5927\u304d\u304f\u306a\u3063\u3066\u3044\u307e\u3059\u3002\u3002<\/p>\n\n\n\n
1500-2500\u306e\u9593\u306b\u672c\u5f53\u306b\u73fe\u5b9f\u3068\u3057\u3066\u7570\u5e38\u304c\u767a\u751f\u3057\u305f\u306e\u304b\u3069\u3046\u304b\u306f\u308f\u304b\u308a\u307e\u305b\u3093\u304c\u3001\u4eca\u56de\u306f\u4eee\u306btimes\u304c0-1000\u306e\u671f\u9593\u306f\u6b63\u5e38\u30671500-2500\u306f\u7570\u5e38\u3068\u3044\u3046\u98a8\u306b\u8003\u3048\u308b\u3053\u3068\u306b\u3057\u307e\u3059\u3002<\/p>\n\n\n\n
\u305d\u306e\u305f\u3081\u3001\u5b66\u7fd2\u30c7\u30fc\u30bf\u3068\u3057\u30660-1500\u306e\u671f\u9593\u3092\u53d6\u308a\u51fa\u3057\u5b66\u7fd2\u30c7\u30fc\u30bf\u3068\u3057\u30011501-3000\u306e \u533a\u9593\u3092\u691c\u8a3c\u7528\u306e\u7570\u5e38\u30c7\u30fc\u30bf\u3002<\/p>\n\n\n\n
x = df[0:1500]\nx_ano = df[1500:3000]<\/code><\/pre><\/div>\n\n\n\n\u305d\u308c\u305e\u308c\u30d7\u30ed\u30c3\u30c8\u3057\u3066\u78ba\u8a8d\u3057\u3066\u307f\u307e\u3059\u3002Y\u8ef8\u306e\u30b9\u30b1\u30fc\u30eb\u306f-40000~45000\u306e\u7bc4\u56f2\u3067\u540c\u3058\u306b\u3057\u3066\u304a\u304d\u307e\u3059\u3002<\/p>\n\n\n\n
\u6b63\u5e38<\/strong><\/p>\n\n\n\nx.plot(legend=False)\nplt.ylim([-40000, 45000])<\/code><\/pre><\/div>\n\n\n\n![\"\"](\"https:\/\/obenkyolab.com\/wp-content\/uploads\/2022\/03\/image-5.png\")
<\/figure>\n\n\n\n\u7570\u5e38<\/strong><\/p>\n\n\n\nx_ano.plot(legend=False)\nplt.ylim([-40000, 45000])<\/code><\/pre><\/div>\n\n\n\n![\"\"](\"https:\/\/obenkyolab.com\/wp-content\/uploads\/2022\/03\/image-6.png\")
\u6b63\u5e38\u3068\u7570\u5e38\u3067\u632f\u5e45\u306e\u9055\u3044\u304c\u898b\u305f\u76ee\u3067\u3082\u3088\u304f\u308f\u304b\u308a\u307e\u3057\u305f\u3002<\/figcaption><\/figure>\n\n\n\n\u3053\u3053\u3067\u3001\u5b66\u7fd2\u306b\u4f7f\u7528\u3059\u308b\u6b63\u5e38\u30c7\u30fc\u30bf\u3092\u6a19\u6e96\u5316\u3057\u3066\u304a\u304d\u307e\u3059\u3002\u307e\u305f\u3001\u5f8c\u307b\u3069\u691c\u8a3c\u3067\u4f7f\u7528\u3059\u308b\u7570\u5e38\u30c7\u30fc\u30bf\u306b\u3064\u3044\u3066\u3082\u4f75\u305b\u3066\u6a19\u6e96\u5316\u3057\u3066\u304a\u304d\u307e\u3059\u3002\u3053\u3053\u3067\u30dd\u30a4\u30f3\u30c8\u3068\u3057\u3066\u306f\u3001\u7570\u5e38\u30c7\u30fc\u30bf\u306e\u6a19\u6e96\u5316\u306e\u969b\u306b\u6b63\u5e38\u30c7\u30fc\u30bf\u306e\u5e73\u5747\u3068\u6a19\u6e96\u504f\u5dee\u3092\u4f7f\u7528\u3059\u308b\u3053\u3068\u3067\u3059\u3002\u7570\u5e38\u30c7\u30fc\u30bf\u306b\u3064\u3044\u3066\u3001\u7570\u5e38\u30c7\u30fc\u30bf\u306e\u5e73\u5747\u3068\u6a19\u6e96\u504f\u5dee\u3092\u4f7f\u7528\u3057\u3066\u3057\u307e\u3046\u3068\u3001\u6b63\u5e38\u6642\u3068\u306e\u6bd4\u8f03\u3067\u8003\u3048\u305f\u5834\u5408\u306b\u5e73\u7b49\u306b\u8a55\u4fa1\u3067\u304d\u306a\u3044\u3068\u3044\u3046\u554f\u984c\u304c\u3042\u308b\u305f\u3081\u3067\u3059\u3002<\/p>\n\n\n\n
#\u6a19\u6e96\u5316\nx_norm = (x - x.mean())\/x.std()\nx_ano_norm = (x_ano - x.mean())\/x.std()<\/code><\/pre><\/div>\n\n\n\n<\/span>\u5b66\u7fd2\u7528\u30c7\u30fc\u30bf\u306e\u6e96\u5099<\/span><\/h3>\n\n\n\n\u6b21\u306b\u30c7\u30fc\u30bf\u309250\u70b9\u30921\u305a\u3064\u30b9\u30e9\u30a4\u30c9\u3055\u305b\u308b\u95a2\u6570\u3092\u5b9a\u7fa9\u3057\u307e\u3059\u3002t_steps\u304c\u30e2\u30c7\u30eb\u306b\u5b66\u7fd2\u3055\u305b\u308b\u969b\u306e\u5165\u529b\u306e\u6b21\u5143\u306b\u306a\u308a\u307e\u3059\u3002\u3053\u308c\u306f\u524d\u8ff0\u306ekeras\u306e\u30c1\u30e5\u30fc\u30c8\u30ea\u30a2\u30eb\u3092\u53c2\u8003\u306b\u3057\u3066\u3044\u307e\u3059\u3002\u30b9\u30c6\u30c3\u30d7\u6570\u306a\u3069\u306f\u9069\u5f53\u306b\u6c7a\u3081\u3066\u3044\u307e\u3059\u3002<\/p>\n\n\n\n
import numpy as np\nt_steps = 50\ndef create_sequences(df):\n x = []\n for i in range(0, len(df) - t_steps + 1):\n x.append(df[i:i + t_steps].to_numpy())\n x_out = np.array(x)\n return x_out<\/code><\/pre><\/div>\n\n\n\n\u5b9a\u7fa9\u3057\u305f\u95a2\u6570\u3092\u4f7f\u7528\u3057\u3066\u3001\u5b66\u7fd2\u30c7\u30fc\u30bf\u306b\u5bfe\u3057\u3066\u30ea\u30b9\u30c8\u3092\u751f\u6210\u3057\u307e\u3059\u3002<\/p>\n\n\n\n
x_train = create_sequences(df_normalized)\nx_train.shape<\/code><\/pre><\/div>\n\n\n\n(1451, 50, 1)<\/pre>\n\n\n\n<\/span>\u30e2\u30c7\u30eb\u306e\u5b9a\u7fa9<\/span><\/h3>\n\n\n\n\u5b66\u7fd2\u30c7\u30fc\u30bf\u306eshape\u306b\u5408\u308f\u305b\u305f\u6df1\u5c64\u5b66\u7fd2\u30e2\u30c7\u30eb\u306elayer\u3092keras\u306eSequential\u3092\u4f7f\u7528\u3057\u3066\u5b9a\u7fa9\u3057\u307e\u3059\u3002\u30dd\u30a4\u30f3\u30c8\u3068\u3057\u3066\u306f\u30aa\u30fc\u30c8\u30a8\u30f3\u30b3\u30fc\u30c0\u3068\u3059\u308b\u305f\u3081\u3001\u7573\u307f\u8fbc\u307f\u3067\u30b5\u30a4\u30ba\u3092\u5909\u66f4\u3057\u305f\u306e\u3061\u3001\u6700\u7d42\u7684\u306bConv1DTranspose\u3067\u540c\u3058\u30b5\u30a4\u30ba\u306b\u5fa9\u5143\u3057\u3066\u3044\u307e\u3059\u3002\u307e\u305f\u3001\u6700\u9069\u5316\u306b\u306fAdam\u3092\u4f7f\u7528\u3057\u3066\u3044\u307e\u3059\u3002<\/p>\n\n\n\n
model =keras.initializers.Initializer()\nmodel = keras.Sequential(\n [\n layers.Input(shape=(x_train.shape[1], x_train.shape[2])),\n layers.Conv1D(\n filters=32, kernel_size=7, padding="same", strides=1, activation="relu"\n ),\n layers.Dropout(rate=0.2),\n layers.Conv1D(\n filters=16, kernel_size=7, padding="same", strides=1, activation="relu"\n ),\n layers.Conv1DTranspose(\n filters=16, kernel_size=7, padding="same", strides=1, activation="relu"\n ),\n layers.Dropout(rate=0.2),\n layers.Conv1DTranspose(\n filters=32, kernel_size=7, padding="same", strides=1, activation="relu"\n ),\n layers.Conv1DTranspose(filters=1, kernel_size=7, padding="same"),\n ]\n)\nmodel.compile(optimizer=keras.optimizers.Adam(learning_rate=0.001), loss="mse")\nmodel.summary()<\/code><\/pre><\/div>\n\n\n\nModel: \"sequential\"\n_________________________________________________________________\n Layer (type) Output Shape Param # \n=================================================================\n conv1d (Conv1D) (None, 50, 32) 256 \n \n dropout (Dropout) (None, 50, 32) 0 \n \n conv1d_1 (Conv1D) (None, 50, 16) 3600 \n \n conv1d_transpose (Conv1DTra (None, 50, 16) 1808 \n nspose) \n \n dropout_1 (Dropout) (None, 50, 16) 0 \n \n conv1d_transpose_1 (Conv1DT (None, 50, 32) 3616 \n ranspose) \n \n conv1d_transpose_2 (Conv1DT (None, 50, 1) 225 \n ranspose) \n \n=================================================================\nTotal params: 9,505\nTrainable params: 9,505\nNon-trainable params: 0<\/pre>\n\n\n\n\u3044\u3088\u3044\u3088\u3001\u4f5c\u6210\u3057\u305f\u30e2\u30c7\u30eb\u306b\u5bfe\u3057\u3066\u3001\u5b66\u7fd2\u7528\u30c7\u30fc\u30bf\u3092fit\u3057\u3066\u5b66\u7fd2\u3057\u3066\u3044\u304d\u307e\u3059\u3002<\/p>\n\n\n\n
history = model.fit(\n x_train,\n x_train,\n epochs=200,\n batch_size=100,\n validation_split=0.1,\n callbacks=[\n keras.callbacks.EarlyStopping(monitor="val_loss", patience=5, mode="min")\n ],\n)<\/code><\/pre><\/div>\n\n\n\n...\nEpoch 30\/200\n14\/14 [==============================] - 0s 21ms\/step - loss: 0.0273 - val_loss: 0.0182\nEpoch 31\/200\n14\/14 [==============================] - 0s 24ms\/step - loss: 0.0266 - val_loss: 0.0183\nEpoch 32\/200\n14\/14 [==============================] - 0s 20ms\/step - loss: 0.0259 - val_loss: 0.0203<\/pre>\n\n\n\n\u5b66\u7fd2\u66f2\u7dda\u3092\u78ba\u8a8d\u3057\u3066\u307f\u307e\u3059\u3002<\/p>\n\n\n\n
plt.plot(history.history["loss"], label="Training Loss")\nplt.plot(history.history["val_loss"], label="Validation Loss")\nplt.legend()\nplt.show()<\/code><\/pre><\/div>\n\n\n\n![\"\"](\"https:\/\/obenkyolab.com\/wp-content\/uploads\/2022\/03\/image-7.png\")
<\/figure>\n\n\n\nValidation\u30c7\u30fc\u30bf\u306b\u5bfe\u3059\u308b\u640d\u5931\u3082\u30c8\u30ec\u30fc\u30cb\u30f3\u30b0\u30c7\u30fc\u30bf\u306b\u5bfe\u3057\u3066\u4e0b\u304c\u3063\u3066\u3044\u3063\u3066\u3044\u308b\u306e\u3067\u3001\u826f\u3055\u305d\u3046\u3067\u3059\u3002<\/p>\n\n\n\n
<\/span>\u7570\u5e38\u30c7\u30fc\u30bf\u306e\u691c\u51fa<\/span><\/h3>\n\n\n\n\u3053\u3053\u304b\u3089\u4f5c\u6210\u3057\u305f\u30e2\u30c7\u30eb\u3092\u4f7f\u7528\u3057\u3066\u3001\u7570\u5e38\u30c7\u30fc\u30bf\u306e\u691c\u51fa\u3092\u884c\u3063\u3066\u3044\u304d\u307e\u3059\u3002<\/p>\n\n\n\n
\u307e\u305a\u3001\u30e2\u30c7\u30eb\u306b\u5b66\u7fd2\u30c7\u30fc\u30bf\u3092\u5165\u529b\u3057\u3001\u4e88\u6e2c\u5024\u3092\u53d6\u5f97\u3057\u307e\u3059\u3002\u4eca\u56de\u306f\u30aa\u30fc\u30c8\u30a8\u30f3\u30b3\u30fc\u30c0\u306e\u305f\u3081\u3001\u5165\u529b\u306b\u5bfe\u3057\u3066\u4e88\u6e2c\u5024\u3082\u540c\u3058\u3082\u306e\u304c\u51fa\u3066\u304f\u308b\u3088\u3046\u306a\u60f3\u5b9a\u3067\u3059\u3002\u3053\u306e\u4e88\u6e2c\u5024\u306b\u5bfe\u3057\u3066\u3001\u5b9f\u969b\u306e\u5165\u529b\u5024\u3068\u306e\u7d76\u5bfe\u8aa4\u5dee\u5e73\u5747\u3092\u8a08\u7b97\u3057\u3001\u7d76\u5bfe\u8aa4\u5dee\u5e73\u5747\u306e\u6700\u5927\u5024\u3092\u7570\u5e38\u5224\u5b9a\u306e\u95be\u5024\u3068\u3057\u3066\u4f7f\u7528\u3059\u308b\u3053\u3068\u306b\u3057\u307e\u3059\u3002<\/p>\n\n\n\n
x_train_pred = model.predict(x_train)\ntrain_mae_loss = np.mean(np.abs(x_train_pred - x_train), axis=1)\nthreshold = np.max(train_mae_loss)<\/code><\/pre><\/div>\n\n\n\n\u6b21\u306b\u3001\u7570\u5e38\u6ce2\u5f62\u306e\u5165\u3063\u305f\u691c\u8a3c\u7528\u306e\u30c7\u30fc\u30bf\u3092\u30e2\u30c7\u30eb\u306e\u5165\u529b\u306b\u5408\u3046\u3088\u3046\u5b66\u7fd2\u30c7\u30fc\u30bf\u3068\u540c\u69d8\u306b\u30a6\u30a3\u30f3\u30c9\u30a6\u3092\u7528\u3044\u3066\u30c7\u30fc\u30bf\u30bb\u30c3\u30c8\u3092\u4f5c\u6210\u3057\u3001\u4e88\u6e2c\u30e2\u30c7\u30eb\u306b\u5165\u529b\u3057\u307e\u3059\u3002<\/p>\n\n\n\n
\u5f97\u3089\u308c\u305f\u7d50\u679c\u3068\u5165\u529b\u306e\u5024\u306b\u304a\u3051\u308b\u7d76\u5bfe\u8aa4\u5dee\u5e73\u5747\u3092\u8a08\u7b97\u3057\u3001\u7d76\u5bfe\u8aa4\u5dee\u5e73\u5747\u304c\u95be\u5024\u7570\u5e38\u306b\u306a\u3063\u3066\u3044\u308b\u30a4\u30f3\u30c7\u30c3\u30af\u30b9\u3092\u62bd\u51fa\u3057\u3066\u304a\u304d\u307e\u3059\u3002<\/p>\n\n\n\n
x_test = create_sequences(x_ano_norm)\nx_test_pred = model.predict(x_test)\ntest_mae_loss = np.mean(np.abs(x_test_pred - x_test), axis=1)\ntest_mae_loss = test_mae_loss.reshape((-1))\nanomalies = test_mae_loss > threshold<\/code><\/pre><\/div>\n\n\n\n<\/span>\u7570\u5e38\u691c\u51fa\u7d50\u679c\u306e\u78ba\u8a8d<\/span><\/h3>\n\n\n\n\u6700\u5f8c\u306b\u3001\u7570\u5e38\u3068\u5224\u5b9a\u3055\u308c\u305f\u30a4\u30f3\u30c7\u30c3\u30af\u30b9\u3092\u542b\u3093\u3067\u3044\u308b\u30c7\u30fc\u30bf\u30bb\u30c3\u30c8\u306b\u5bfe\u3057\u3066\u8d64\u8272\u3067\u30d7\u30ed\u30c3\u30c8\u3057\u307e\u3059\u3002<\/p>\n\n\n\n
anomalous_data_indices = []\nfor data_idx in range(t_steps - 1, len(x_ano_norm) - t_steps + 1):\n if np.all(anomalies[data_idx - t_steps + 1 : data_idx]):\n anomalous_data_indices.append(data_idx)\n \ndf_subset = x_ano.iloc[anomalous_data_indices]\nfig, ax = plt.subplots()\ndf.plot(legend=False, ax=ax)\ndf_subset.plot(legend=False, ax=ax, color="r")\nplt.show()<\/code><\/pre><\/div>\n\n\n\n![\"\"](\"https:\/\/obenkyolab.com\/wp-content\/uploads\/2022\/03\/image-8.png\")
<\/figure>\n\n\n\n\u3053\u306e\u3088\u3046\u306b\u632f\u5e45\u304c\u5927\u304d\u304f\u306a\u3063\u3066\u3044\u308b\u7b87\u6240\u304c\u8d64\u8272\u3067\u5224\u5b9a\u3055\u308c\u307e\u3057\u305f\u3002<\/p>\n\n\n\n
<\/p>\n\n\n\n
<\/span>\u307e\u3068\u3081<\/span><\/h2>\n\n\n\n\u4eca\u56de\u306fkeras\u306e\u30b5\u30a4\u30c8\u306b\u3042\u308b\u30aa\u30fc\u30c8\u30a8\u30f3\u30b3\u30fc\u30c0\u306e\u30c1\u30e5\u30fc\u30c8\u30ea\u30a2\u30eb\u3092\u53c2\u8003\u306b\u3057\u3066\u3001Kaggle\u306e\u30aa\u30fc\u30d7\u30f3\u30c7\u30fc\u30bf\u3092\u5bfe\u8c61\u306b\u7570\u5e38\u691c\u77e5\u3092\u8a66\u3057\u3066\u307f\u307e\u3057\u305f\u3002\u30d1\u30e9\u30e1\u30fc\u30bf\u306b\u3064\u3044\u3066\u306f\u30c1\u30e5\u30fc\u30c8\u30ea\u30a2\u30eb\u3082\u306e\u3092\u305d\u306e\u307e\u307e\u5229\u7528\u3057\u305f\u90e8\u5206\u3082\u3042\u308a\u307e\u3057\u305f\u304c\u3001\u305d\u308c\u306a\u308a\u306b\u691c\u51fa\u3067\u304d\u3066\u3044\u308b\u3088\u3046\u306a\u611f\u3058\u3067\u3057\u305f\u3002\u4eca\u5f8c\u3082keras\u306e\u30e1\u30bd\u30c3\u30c9\u3092\u3044\u308d\u3044\u308d\u3068\u8a66\u3057\u3066\u307f\u305f\u3044\u3068\u601d\u3044\u307e\u3059\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"
\u6df1\u5c64\u5b66\u7fd2\u3092\u7c21\u5358\u306b\u4f7f\u7528\u3059\u308b\u3053\u3068\u304c\u3067\u304d\u308bkeras\u3092\u4f7f\u7528\u3057\u3066\u6642\u7cfb\u5217\u30c7\u30fc\u30bf...<\/p>\n","protected":false},"author":1,"featured_media":1218,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"_locale":"ja","_original_post":"https:\/\/obenkyolab.com\/?p=4800","footnotes":""},"categories":[7,53],"tags":[],"class_list":["post-4800","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-python","category-53","ja"],"_links":{"self":[{"href":"https:\/\/obenkyolab.com\/index.php?rest_route=\/wp\/v2\/posts\/4800","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/obenkyolab.com\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/obenkyolab.com\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/obenkyolab.com\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/obenkyolab.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=4800"}],"version-history":[{"count":8,"href":"https:\/\/obenkyolab.com\/index.php?rest_route=\/wp\/v2\/posts\/4800\/revisions"}],"predecessor-version":[{"id":4813,"href":"https:\/\/obenkyolab.com\/index.php?rest_route=\/wp\/v2\/posts\/4800\/revisions\/4813"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/obenkyolab.com\/index.php?rest_route=\/wp\/v2\/media\/1218"}],"wp:attachment":[{"href":"https:\/\/obenkyolab.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=4800"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/obenkyolab.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=4800"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/obenkyolab.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=4800"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}