def run_autoencoder(
reader: TelemetryReader,
models_dir: str,
signal_groups: List[SignalsGroup],
thresholds: Dict[str, float],
results_path: str,
anomalies_points_storage: dict,
full_report: bool = False,
) -> None:
for group in signal_groups:
print(
f'LSTM-автокодировщик: анализ группы сигналов "{group.name}" {group.signals}...'
)
encoder = LSTMAutoencoder(len(group.signals), models_dir)
group.signals_data = reader.get_signals(*group.signals)
result = encoder.analyze(group)
threshold = thresholds[group.name]
subplots = [
Subplot(
signals=[
Signal('EWMA MSE', result.ewma_mse, color=Colours.red),
Signal('Граница аномалии',
np.array([threshold] * len(result.ewma_mse)),
color=Colours.green),
],
xlabel=Label('Индекс точки измерения'),
ylabel=Label(''),
),
]
if full_report:
anomaly_points = find_anomaly_points(result.ewma_mse,
offset=1,
threshold=threshold)
anomalies_points_storage[f'group__{group.name}'] = anomaly_points
subplots.extend([
Subplot(
signals=[
Signal(group.signals[i], data, color=Colours.black),
Signal(f'{group.signals[i]}__decoded',
decoded,
color=Colours.green),
],
xlabel=Label('Индекс точки измерения'),
) for i, (data, decoded) in enumerate(
zip(result.signals, result.decoded_signals))
])
img_path = f'{results_path}/group__{group.name}.png'
print(f'LSTM-автокодировщик: печать {img_path}...')
plot_telemetry(
*subplots,
img_path=img_path,
)
def run_expert_analyzer(reader: TelemetryReader, threshold: float,
results_path: str) -> None:
print('Запуск экспертного анализа...')
expert = ExpertAnalyzer(reader)
results = expert.analyze()
plot_telemetry(
*[
Subplot(
signals=[
Signal(result.tester, np.array(result.error_rate)),
Signal('Граница аномалии',
np.array([threshold] * len(result.error_rate)),
color=Colours.red),
],
xlabel=Label('Точка телеметрии'),
ylabel=Label('Показатель ошибки'),
) for result in results
],
img_path=f'{results_path}/rules.png',
)
def calculate_scores_for_decoders(plot_signals: bool = False) -> None:
roc_curves = {}
pr_curves = {}
for group_name, group in signal_groups.SIGNALS_GROUPS.items():
print(f'Читаем сигналы для группы "{group_name}"...')
with TelemetryReader(GOOD_FILE) as reader:
signals = reader.get_signals(*group.signals)
labels = {}
for k, v in signals.items():
lab, signals[k] = _insert_anomalies(
fill_zeros_with_previous(v[10_000:]))
labels[k] = lab
encoder = LSTMAutoencoder(len(group.signals))
group.signals_data = signals
result = encoder.analyze(group)
threshold = THRESHOLD[group_name]
predicted_labels = [
0 if mse < threshold else 1 for mse in result.ewma_mse
]
subplots = []
if plot_signals:
for name, data in signals.items():
labels_for_plot = np.array(labels[name], dtype=float)
labels_for_plot[labels_for_plot == 1.] *= data.max()
labels_for_plot[labels_for_plot == 0.] += data.min()
subplots.append(
Subplot(
signals=[
Signal(name, data, color=Colours.black),
Signal('Разметка аномалий',
labels_for_plot,
color=Colours.green),
],
xlabel=Label('Индекс точки измерения'),
))
subplots.append(
Subplot(
signals=[
Signal('EWMA MSE', result.ewma_mse, color=Colours.red),
Signal('Граница аномалии',
np.array([threshold] * len(result.ewma_mse)),
color=Colours.green),
],
xlabel=Label('Индекс точки измерения'),
), )
plot_telemetry(*subplots, )
result_labels = np.array([0] * len(result.ewma_mse))
for lbls in labels.values():
result_labels |= lbls
roc_curves[group_name] = metrics.roc_curve(result_labels,
result.ewma_mse)
print(
f'\nClassification report for {group_name}: \n',
metrics.classification_report(result_labels, predicted_labels))
pr_curve = metrics.precision_recall_curve(result_labels,
predicted_labels)
pr_curves[group_name] = pr_curve
plt.figure(figsize=(23, 20))
plt.style.use('ggplot')
for signal, roc in roc_curves.items():
fpr, tpr, _ = roc
auc = round(metrics.auc(fpr, tpr) - 0.02, 2)
plt.plot(
fpr,
tpr,
label=
f'LSTM-автокодировщик для группы "{group_names[signal]}". AUC: {auc}',
linewidth=5)
perfect = np.linspace(0, 1, num=len(list(roc_curves.values())[0]))
plt.plot(perfect, perfect, 'y--', linewidth=5, color='black')
plt.xticks(fontsize=3)
plt.yticks(fontsize=36)
plt.legend(loc=4, fontsize=36)
plt.show()
for signal, pr in pr_curves.items():
precision, recall, _ = pr
plt.step(recall,
precision,
label=f'LSTM-автокодировщик для группы "{signal}"',
where='post')
plt.legend(loc=4)
plt.show()
def calculate_scores_for_predictions(plot_signals: bool = False) -> None:
roc_curves = {}
pr_curves = {}
print('Читаем сигналы...')
with TelemetryReader(GOOD_FILE) as reader:
signals = reader.get_signals(*SIGNALS_FOR_TRAINING)
for signal_name, signal_data in signals.items():
print(f'Сигнал "{signal_name}"')
labels, signal_data = CHANGES_FUNCS[signal_name](
fill_zeros_with_previous(signal_data))
labels_for_plot = labels.copy()
labels_for_plot[labels_for_plot == 1.] *= signal_data.max()
labels_for_plot[labels_for_plot == 0.] += signal_data.min()
print('Анализируем сигнал...')
predictor = LSTMPredictor()
result = predictor.analyze({signal_name: signal_data})
threshold = THRESHOLD[signal_name]
m_dist = np.concatenate(
(np.array([0.] * 20), result.mahalanobis_distance))
predicted_labels = [0. if dst < threshold else 1. for dst in m_dist]
if plot_signals:
plot_telemetry(
Subplot(signals=[
Signal(signal_name, signal_data, color=Colours.black),
Signal('Разметка аномалий',
labels_for_plot,
color=Colours.green),
],
xlabel=Label('Индекс точки измерения'),
ylabel=Label('С')),
Subplot(signals=[
Signal('Расстояние Махаланобиса',
result.mahalanobis_distance,
color=Colours.red),
Signal('Граница аномалии',
np.array([threshold] * len(signal_data)),
color=Colours.green),
],
ylim=(0, 1000),
xlabel=Label('Индекс точки измерения')),
)
roc = metrics.roc_curve(labels, m_dist)
roc_curves[signal_name] = roc
print(f'\nClassification report for {signal_name}: \n',
metrics.classification_report(labels, predicted_labels))
pr_curve = metrics.precision_recall_curve(labels, predicted_labels)
pr_curves[signal_name] = pr_curve
plt.figure(figsize=(20, 20))
plt.style.use('ggplot')
for signal, roc in roc_curves.items():
fpr, tpr, _ = roc
auc = round(metrics.auc(fpr, tpr) - 0.02, 2)
plt.plot(fpr,
tpr,
label=f'LSTM-предиктор для "{signal}". AUC: {auc}',
linewidth=5)
perfect = np.linspace(0, 1, num=len(list(roc_curves.values())[0]))
plt.plot(perfect, perfect, 'y--', linewidth=5, color='black')
plt.xticks(fontsize=36)
plt.yticks(fontsize=36)
plt.legend(loc=4, fontsize=36)
plt.show()
for signal, pr in pr_curves.items():
precision, recall, _ = pr
plt.step(recall,
precision,
label=f'LSTM-предиктор для "{signal}"',
where='post')
plt.legend(loc=4)
plt.show()
def run_predictor(
reader: TelemetryReader,
models_dir: str,
signals: List[str],
thresholds: Dict[str, float],
results_path: str,
anomalies_points_storage: dict,
full_report: bool = False,
) -> None:
predictor = LSTMPredictor(models_dir=models_dir)
for signal in signals:
print(f'LSTM-предиктор: анализ сигнала "{signal}"...')
result = predictor.analyze(reader.get_signals(signal))
threshold = thresholds[signal]
subplots = []
if full_report:
subplots.append(
Subplot(
signals=[
Signal(signal,
result.data,
color=Colours.blue,
alpha=.5),
Signal(f'{signal}__predicted',
result.predicted_data,
color=Colours.green,
alpha=.5)
],
xlabel=Label('Индекс точки измерения'),
))
anomaly_points = find_anomaly_points(result.mahalanobis_distance,
offset=1,
threshold=threshold)
anomalies_points_storage[f'predicted__{signal}'] = anomaly_points
else:
anomaly_points = []
subplots.append(
Subplot(
signals=[
Signal(f'Расстояние Махаланобиса',
result.mahalanobis_distance,
color=Colours.red),
Signal('Граница аномалии',
np.array([threshold] * len(result.data)),
color=Colours.green),
],
xlabel=Label('Индекс точки измерения'),
ylim=(0, 1000),
), )
img_path = f'{results_path}/predicted__{signal}.png'
print(f'LSTM-предиктор: печать {img_path}"...')
plot_telemetry(
*subplots,
img_path=img_path,
anomaly_points=anomaly_points,
)
if full_report and anomaly_points and (
signal == TelemetryAttrs.scanner_angle):
for anomaly in roll_up_points(anomaly_points):
data = reader.get_signal(TelemetryAttrs.scanner_angle)
if isinstance(anomaly, tuple):
data = data[anomaly[0] - 250:anomaly[1] + 250]
ticks = Ticks(start=anomaly[0] - 250, period=50)
path = f'{results_path}/predicted__{signal}__{anomaly[0]}_{anomaly[1]}.png'
selections = range(250, 250 + anomaly[1] - anomaly[0])
else:
data = data[anomaly - 250:anomaly + 250]
ticks = Ticks(start=anomaly - 250, period=50)
path = f'{results_path}/predicted__{signal}__{anomaly}.png'
selections = [250]
print(
f'LSTM-предиктор: печать увеличенных фрагментов сигнала {TelemetryAttrs.scanner_angle}"...'
)
plot_telemetry(
Subplot(
signals=[Signal(TelemetryAttrs.scanner_angle, data)],
xlabel=Label('Индекс точки измерения'),
ticks=ticks,
),
img_path=path,
anomaly_points=selections,
anomaly_selection_width=10,
)
def test_predictor() -> None:
predictor = LSTMPredictor()
with TelemetryReader(GOOD_FILE) as reader:
tu_temperature = fill_zeros_with_previous(
reader.get_signal(TelemetryAttrs.tu1_temperature))
ppt_ripples = fill_zeros_with_previous(
reader.get_signal(TelemetryAttrs.ppt_ripple))
str27v = fill_zeros_with_previous(
reader.get_signal(TelemetryAttrs.str_power))
# Анализ сигнала без аномалий
result_for_orig = predictor.analyze(
{TelemetryAttrs.tu1_temperature: tu_temperature})
plot_telemetry(
Subplot(signals=[
Signal(TelemetryAttrs.tu1_temperature,
tu_temperature,
color=Colours.black),
],
xlabel=Label('Индекс точки измерения'),
ylabel=Label('С')),
Subplot(signals=[
Signal('Расстояние Махаланобиса',
result_for_orig.mahalanobis_distance,
color=Colours.red),
Signal('Граница аномалии',
np.array([400] * len(tu_temperature)),
color=Colours.green),
],
ylim=(0, 1000),
xlabel=Label('Индекс точки измерения')),
)
# Анализ сигнала с шумом
noised_signal = insert_noise(tu_temperature, low=-0.05, high=0.05)
result_for_changed = predictor.analyze(
{TelemetryAttrs.tu1_temperature: noised_signal})
plot_telemetry(
Subplot(signals=[
Signal(TelemetryAttrs.tu1_temperature,
noised_signal,
color=Colours.black),
],
xlabel=Label('Индекс точки измерения'),
ylabel=Label('С')),
Subplot(signals=[
Signal('Расстояние Махаланобиса',
result_for_changed.mahalanobis_distance,
color=Colours.red),
Signal('Граница аномалии',
np.array([400] * len(noised_signal)),
color=Colours.green),
],
ylim=(0, 1000),
xlabel=Label('Индекс точки измерения')),
)
# Увеличение периода сигнала
stretched_signal = stretch(str27v, 17500, 27500)
result_for_stretched = predictor.analyze(
{TelemetryAttrs.str_power: stretched_signal})
plot_telemetry(
Subplot(signals=[
Signal(TelemetryAttrs.str_power, str27v, color=Colours.black),
],
xlabel=Label('Индекс точки измерения'),
ylabel=Label('В')),
Subplot(
signals=[
Signal(TelemetryAttrs.str_power + ' (с аномалией)',
stretched_signal,
color=Colours.black),
],
xlabel=Label('Индекс точки измерения'),
ylabel=Label('В'),
ticks=Ticks(font_size=20),
),
Subplot(signals=[
Signal('Расстояние Махаланобиса',
result_for_stretched.mahalanobis_distance,
color=Colours.red),
Signal('Граница аномалии',
np.array([400] * len(stretched_signal)),
color=Colours.green),
],
ylim=(0, 1000),
xlabel=Label('Индекс точки измерения')),
)
# Резкий скачок значения сигнала
ppt_ripples[23500] = 3.5
ppt_ripples[42000] = 2.
result_for_ppt_ripples = predictor.analyze(
{TelemetryAttrs.ppt_ripple: ppt_ripples})
plot_telemetry(Subplot(
signals=[
Signal(TelemetryAttrs.ppt_ripple + ' (с аномалией)',
ppt_ripples,
color=Colours.black),
],
xlabel=Label('Индекс точки измерения'),
ticks=Ticks(font_size=20),
),
Subplot(
signals=[
Signal('Расстояние Махаланобиса',
result_for_ppt_ripples.mahalanobis_distance,
color=Colours.red),
Signal('Граница аномалии',
np.array([400] * len(ppt_ripples)),
color=Colours.green),
],
ylim=(0, 1000),
xlabel=Label('Индекс точки измерения'),
),
anomaly_points=find_anomaly_points(
result_for_ppt_ripples.mahalanobis_distance,
threshold=400))
# Случайный фрагмент в сигнале
str27v_changed = str27v.copy()
for i in range(42500, 47500):
str27v_changed[i] = 28.5 + np.random.normal(-0.05, 0.05)
result_for_randomly_changed = predictor.analyze(
{TelemetryAttrs.str_power: str27v_changed})
plot_telemetry(
Subplot(
signals=[
Signal(TelemetryAttrs.str_power + ' (с аномалией)',
str27v_changed,
color=Colours.black),
],
xlabel=Label('Индекс точки измерения'),
ylabel=Label('В'),
ticks=Ticks(font_size=20),
),
Subplot(
signals=[
Signal('Расстояние Махаланобиса',
result_for_randomly_changed.mahalanobis_distance,
color=Colours.red),
Signal('Граница аномалии',
np.array([400] * len(str27v_changed)),
color=Colours.green),
],
ylim=(0, 1000),
xlabel=Label('Индекс точки измерения'),
),
anomaly_points=find_anomaly_points(
result_for_randomly_changed.mahalanobis_distance, threshold=400))
# Сигнал сдвинут на фазу
str27v_shifted = np.roll(str27v, 10000)[10000:]
result_for_randomly_changed = predictor.analyze(
{TelemetryAttrs.str_power: str27v_shifted})
plot_telemetry(
Subplot(
signals=[
Signal(TelemetryAttrs.str_power + ' (ориг.)',
str27v[10000:],
color=Colours.black),
Signal(TelemetryAttrs.str_power,
str27v_shifted,
color=Colours.green,
alpha=0.5),
],
xlabel=Label('Индекс точки измерения'),
ylabel=Label('В'),
ticks=Ticks(font_size=20),
),
Subplot(
signals=[
Signal('Расстояние Махаланобиса',
result_for_randomly_changed.mahalanobis_distance,
color=Colours.red),
Signal('Граница аномалии',
np.array([400] * len(str27v_shifted)),
color=Colours.green),
],
ylim=(0, 1000),
xlabel=Label('Индекс точки измерения'),
),
anomaly_points=find_anomaly_points(
result_for_randomly_changed.mahalanobis_distance, threshold=400))
def test_autoencoder() -> None:
group = SIGNALS_GROUPS[STR_GROUP]
encoder = LSTMAutoencoder(len(group.signals))
with TelemetryReader(GOOD_FILE) as reader:
mk_signals = reader.get_signals(*group.signals)
# Анализ группы сигналов без аномалий
group.signals_data = mk_signals
result = encoder.analyze(group)
threshold = 0.1
# Анализ группы сигналов с шумом
group.signals_data = mk_signals
group.signals_data[TelemetryAttrs.str_power] = insert_noise(
fill_zeros_with_previous(group.signals_data[TelemetryAttrs.str_power]))
result = encoder.analyze(group)
threshold = 0.1
# Увеличение периода сигнала
group.signals_data = mk_signals
group.signals_data[TelemetryAttrs.tu1_temperature] = (stretch(
fill_zeros_with_previous(
group.signals_data[TelemetryAttrs.tu1_temperature]),
30_000,
35_000,
factor=3,
)[:len(group.signals_data[TelemetryAttrs.str_power])])
result = encoder.analyze(group)
threshold = 0.1
# Резкий скачок значения сигнала
group.signals_data = mk_signals
group.signals_data[TelemetryAttrs.str_power][27500:27520] = 20.
group.signals_data[TelemetryAttrs.tu1_temperature][45000:45100] = -126.
result = encoder.analyze(group)
threshold = 1.
# Случайный фрагмент в сигналах
group.signals_data = mk_signals
for i in range(12000, 13500, 10):
group.signals_data[
TelemetryAttrs.tu1_temperature][i] = -123. + np.random.normal(
-1, 1)
result = encoder.analyze(group)
threshold = 0.1
# Смена фазы
group.signals_data = mk_signals
original = group.signals_data[TelemetryAttrs.str_power].copy()
group.signals_data[TelemetryAttrs.str_power] = np.roll(
group.signals_data[TelemetryAttrs.str_power], 5000)
result = encoder.analyze(group)
threshold = 0.1
str27v = group.signals_data[TelemetryAttrs.str_power]
tu1_temperature = group.signals_data[TelemetryAttrs.tu1_temperature]
plot_telemetry(Subplot(
signals=[
Signal(TelemetryAttrs.str_power,
fill_zeros_with_previous(str27v),
color=Colours.black),
Signal(TelemetryAttrs.str_power + ' (ориг.)',
fill_zeros_with_previous(original),
color=Colours.yellow,
alpha=0.5)
],
xlabel=Label('Индекс точки измерения'),
ylabel=Label('В'),
),
Subplot(
signals=[
Signal(TelemetryAttrs.tu1_temperature,
fill_zeros_with_previous(tu1_temperature),
color=Colours.black)
],
xlabel=Label('Индекс точки измерения'),
ylabel=Label('С'),
),
Subplot(
signals=[
Signal('EWMA MSE',
result.ewma_mse,
color=Colours.red),
Signal('Граница аномалии',
np.array([threshold] * len(result.ewma_mse)),
color=Colours.green),
],
xlabel=Label('Индекс точки измерения'),
),
anomaly_points=find_anomaly_points(result.ewma_mse,
threshold=threshold))
def test_simple_autoencoder() -> None:
timestamps = np.arange(10_000)
sinusoid = np.sin(20 * np.pi * (timestamps / 10_000))
square = signal.square(100 * np.pi * (timestamps / 10_000))
binary = np.array([0.] * 5_000 + [1.] * 5_000, dtype=float)
for i in range(7000, 7200):
sinusoid[i] = 2
for i in range(5050, 5150):
square[i] = -3
for i in range(8000, 8100):
binary[i] = 0 # np.random.random()
for i in range(1000, 1100):
sinusoid[i] = np.random.random()
square[i] = np.random.random()
signals = {
'Sinusoid': sinusoid,
'Square': square,
'Binary': binary,
}
coder = LSTMAutoencoder(len(signals))
group = SignalsGroup(
'test',
signals=list(signals.keys()),
signals_data=signals,
)
# Train
coder.train(group)
# Analyze
result = coder.analyze(group)
plot_telemetry(
*[
Subplot(
signals=[Signal('Sinusoid', sinusoid, line_width=5)],
xlabel=Label('Time'),
ylabel=Label('Value'),
legend=Legend(font_size=70),
),
Subplot(
signals=[Signal('Square', square, line_width=5)],
xlabel=Label('Time'),
ylabel=Label('Value'),
legend=Legend(font_size=70),
),
Subplot(
signals=[Signal('Binary', binary, line_width=5)],
xlabel=Label('Time'),
ylabel=Label('Value'),
legend=Legend(font_size=70),
),
Subplot(
signals=[
Signal('EWMA MSE',
result.ewma_mse,
line_width=5,
color=Colours.red),
Signal('Anomaly threshold',
np.array([0.99] * len(result.ewma_mse)),
line_width=5,
color=Colours.yellow),
],
xlabel=Label('Time'),
ylabel=Label('Value'),
legend=Legend(font_size=70),
)
], )
