industrial-sensor-anomaly-data / generate_data.py
Petsteb's picture
Upload 2 files
dd99a72 verified
"""
Synthetic Data Generator: Industrial Equipment Sensor Anomaly Detection
Generates realistic multivariate sensor data from a simulated manufacturing
plant with 5 equipment units. Embeds 4 distinct anomaly types at ~4% rate
with realistic noise, missing values, temporal correlations, and concept drift.
Reproducible with fixed random seeds.
"""
import numpy as np
import pandas as pd
from datetime import datetime, timedelta
SEED = 42
np.random.seed(SEED)
# ============================================================
# Configuration
# ============================================================
N_EQUIPMENT = 5
EQUIPMENT_IDS = [f"EQ-{str(i).zfill(3)}" for i in range(1, N_EQUIPMENT + 1)]
START_DATE = datetime(2024, 1, 1)
READINGS_PER_EQUIPMENT = 10000 # 1-min intervals => ~7 days each
TOTAL_ROWS = N_EQUIPMENT * READINGS_PER_EQUIPMENT # 50,000
ANOMALY_RATE_TARGET = 0.04 # ~4%
# Sensor baseline profiles per equipment (slightly different characteristics)
EQUIPMENT_PROFILES = {
"EQ-001": {"temp_base": 72, "vib_base": 0.15, "press_base": 150, "rpm_base": 3000, "flow_base": 45},
"EQ-002": {"temp_base": 68, "vib_base": 0.12, "press_base": 155, "rpm_base": 3200, "flow_base": 48},
"EQ-003": {"temp_base": 75, "vib_base": 0.18, "press_base": 145, "rpm_base": 2800, "flow_base": 42},
"EQ-004": {"temp_base": 70, "vib_base": 0.14, "press_base": 152, "rpm_base": 3100, "flow_base": 46},
"EQ-005": {"temp_base": 73, "vib_base": 0.16, "press_base": 148, "rpm_base": 2900, "flow_base": 44},
}
# Operating modes with transition probabilities
OPERATING_MODES = ["startup", "normal", "high_load", "cooldown", "idle"]
MODE_DURATION_RANGE = {
"startup": (15, 45),
"normal": (120, 480),
"high_load": (30, 180),
"cooldown": (20, 60),
"idle": (30, 120),
}
MODE_MULTIPLIERS = {
"startup": {"temp": 0.85, "vib": 1.3, "press": 0.9, "rpm": 0.7, "flow": 0.6},
"normal": {"temp": 1.0, "vib": 1.0, "press": 1.0, "rpm": 1.0, "flow": 1.0},
"high_load": {"temp": 1.25, "vib": 1.5, "press": 1.15, "rpm": 1.2, "flow": 1.3},
"cooldown": {"temp": 0.9, "vib": 0.8, "press": 0.95, "rpm": 0.5, "flow": 0.4},
"idle": {"temp": 0.7, "vib": 0.3, "press": 0.85, "rpm": 0.1, "flow": 0.05},
}
# Anomaly type definitions
# Type 1: Thermal runaway (gradual temperature escalation over 15-40 min)
# Type 2: Bearing degradation (vibration increases + high-freq noise)
# Type 3: Pressure leak (slow pressure drop with flow compensation)
# Type 4: Sensor malfunction (one sensor gives erratic readings - NOT real failure)
ANOMALY_TYPES = ["thermal_runaway", "bearing_degradation", "pressure_leak", "sensor_malfunction"]
def generate_operating_mode_sequence(n_steps, rng):
"""Generate a realistic sequence of operating modes."""
modes = []
current_mode = "startup"
steps_remaining = rng.integers(*MODE_DURATION_RANGE[current_mode])
for _ in range(n_steps):
modes.append(current_mode)
steps_remaining -= 1
if steps_remaining <= 0:
# Transition logic
if current_mode == "startup":
current_mode = "normal"
elif current_mode == "normal":
current_mode = rng.choice(["normal", "high_load", "cooldown"], p=[0.6, 0.3, 0.1])
elif current_mode == "high_load":
current_mode = rng.choice(["normal", "high_load", "cooldown"], p=[0.4, 0.3, 0.3])
elif current_mode == "cooldown":
current_mode = rng.choice(["idle", "normal", "startup"], p=[0.4, 0.4, 0.2])
elif current_mode == "idle":
current_mode = rng.choice(["startup", "idle"], p=[0.7, 0.3])
steps_remaining = rng.integers(*MODE_DURATION_RANGE[current_mode])
return modes
def generate_base_sensor_data(n_steps, profile, modes, rng):
"""Generate correlated base sensor readings with mode-dependent behavior."""
t = np.arange(n_steps)
# Slow ambient drift (diurnal cycle ~1440 min)
ambient_cycle = 3.0 * np.sin(2 * np.pi * t / 1440) + 0.5 * np.sin(2 * np.pi * t / 720)
# Equipment age effect (very subtle degradation over time)
age_factor = 1.0 + 0.0001 * t
temp_mult = np.array([MODE_MULTIPLIERS[m]["temp"] for m in modes])
vib_mult = np.array([MODE_MULTIPLIERS[m]["vib"] for m in modes])
press_mult = np.array([MODE_MULTIPLIERS[m]["press"] for m in modes])
rpm_mult = np.array([MODE_MULTIPLIERS[m]["rpm"] for m in modes])
flow_mult = np.array([MODE_MULTIPLIERS[m]["flow"] for m in modes])
# Core sensors (correlated via physics)
temp_base = profile["temp_base"] * temp_mult * age_factor + ambient_cycle
vib_base = profile["vib_base"] * vib_mult * age_factor
press_base = profile["press_base"] * press_mult
rpm_base = profile["rpm_base"] * rpm_mult
flow_base = profile["flow_base"] * flow_mult
# Add realistic sensor noise
temp = temp_base + rng.normal(0, 0.8, n_steps)
vibration = vib_base + rng.normal(0, 0.01, n_steps)
pressure = press_base + rng.normal(0, 1.5, n_steps)
rpm = rpm_base + rng.normal(0, 15, n_steps)
flow_rate = flow_base + rng.normal(0, 0.8, n_steps)
# Derived/correlated sensors
# Power consumption correlates with RPM and load
power = 0.012 * rpm + 0.3 * flow_rate + rng.normal(0, 0.5, n_steps)
# Coolant temperature lags behind main temp
coolant_temp = np.convolve(temp, np.ones(10) / 10, mode='same') - 5 + rng.normal(0, 0.3, n_steps)
# Acoustic level correlates with vibration and RPM
acoustic_db = 40 + 80 * vibration + 0.005 * rpm + rng.normal(0, 1.0, n_steps)
# Oil viscosity inversely related to temperature
oil_viscosity = 100 - 0.4 * temp + rng.normal(0, 0.5, n_steps)
# Humidity (mostly ambient, weakly correlated)
humidity = 45 + 10 * np.sin(2 * np.pi * t / 1440 + 1.5) + rng.normal(0, 2, n_steps)
return {
"temperature_c": temp,
"vibration_mm_s": np.clip(vibration, 0, None),
"pressure_kpa": pressure,
"motor_rpm": np.clip(rpm, 0, None),
"flow_rate_lpm": np.clip(flow_rate, 0, None),
"power_consumption_kw": np.clip(power, 0, None),
"coolant_temp_c": coolant_temp,
"acoustic_level_db": np.clip(acoustic_db, 20, 120),
"oil_viscosity_cst": np.clip(oil_viscosity, 10, 150),
"humidity_pct": np.clip(humidity, 10, 95),
"ambient_temp_c": 22 + ambient_cycle + rng.normal(0, 0.3, n_steps),
}
def inject_anomalies(sensors, modes, n_steps, rng):
"""Inject 4 types of anomalies into sensor data."""
is_anomaly = np.zeros(n_steps, dtype=int)
anomaly_type = np.full(n_steps, "normal", dtype=object)
# Calculate how many anomaly windows we need
target_anomaly_count = int(n_steps * ANOMALY_RATE_TARGET)
injected = 0
# Track used indices to avoid overlap
used = np.zeros(n_steps, dtype=bool)
# Type 1: Thermal runaway (gradual, 15-40 min windows)
n_thermal = int(target_anomaly_count * 0.30)
thermal_injected = 0
attempts = 0
while thermal_injected < n_thermal and attempts < 500:
attempts += 1
window_len = rng.integers(15, 41)
start = rng.integers(100, n_steps - window_len - 10)
if used[start:start + window_len].any():
continue
# Only inject during normal or high_load (realistic)
if modes[start] not in ("normal", "high_load"):
continue
ramp = np.linspace(0, 1, window_len) ** 1.5 # Accelerating ramp
magnitude = rng.uniform(15, 35)
sensors["temperature_c"][start:start + window_len] += magnitude * ramp
# Correlated: coolant also rises but lagged
lag = min(5, window_len // 3)
sensors["coolant_temp_c"][start + lag:start + window_len] += magnitude * 0.6 * ramp[:window_len - lag]
# Oil viscosity drops as temperature rises
sensors["oil_viscosity_cst"][start:start + window_len] -= magnitude * 0.3 * ramp
# Power increases slightly
sensors["power_consumption_kw"][start:start + window_len] += 2 * ramp
is_anomaly[start:start + window_len] = 1
anomaly_type[start:start + window_len] = "thermal_runaway"
used[start:start + window_len] = True
thermal_injected += window_len
# Type 2: Bearing degradation (vibration spikes, 5-20 min)
n_bearing = int(target_anomaly_count * 0.25)
bearing_injected = 0
attempts = 0
while bearing_injected < n_bearing and attempts < 500:
attempts += 1
window_len = rng.integers(5, 21)
start = rng.integers(100, n_steps - window_len - 10)
if used[start:start + window_len].any():
continue
# Vibration increases with high-frequency noise
vib_increase = rng.uniform(0.3, 0.8)
noise_amp = rng.uniform(0.05, 0.15)
sensors["vibration_mm_s"][start:start + window_len] += vib_increase + noise_amp * rng.standard_normal(window_len)
# Acoustic level increases proportionally
sensors["acoustic_level_db"][start:start + window_len] += 15 + 5 * rng.standard_normal(window_len)
# Slight RPM fluctuation
sensors["motor_rpm"][start:start + window_len] += rng.normal(0, 30, window_len)
is_anomaly[start:start + window_len] = 1
anomaly_type[start:start + window_len] = "bearing_degradation"
used[start:start + window_len] = True
bearing_injected += window_len
# Type 3: Pressure leak (subtle, slow, 20-60 min)
n_pressure = int(target_anomaly_count * 0.25)
pressure_injected = 0
attempts = 0
while pressure_injected < n_pressure and attempts < 500:
attempts += 1
window_len = rng.integers(20, 61)
start = rng.integers(100, n_steps - window_len - 10)
if used[start:start + window_len].any():
continue
if modes[start] not in ("normal", "high_load"):
continue
ramp = np.linspace(0, 1, window_len)
pressure_drop = rng.uniform(8, 25)
sensors["pressure_kpa"][start:start + window_len] -= pressure_drop * ramp
# Flow rate compensates (system tries to maintain)
sensors["flow_rate_lpm"][start:start + window_len] += 3 * ramp
# Power increases as pump works harder
sensors["power_consumption_kw"][start:start + window_len] += 1.5 * ramp
is_anomaly[start:start + window_len] = 1
anomaly_type[start:start + window_len] = "pressure_leak"
used[start:start + window_len] = True
pressure_injected += window_len
# Type 4: Sensor malfunction (erratic readings on ONE sensor - THIS IS A TRAP)
# These look like anomalies but are just sensor failures, NOT equipment failures
# A smart model should learn to distinguish these
n_sensor_mal = int(target_anomaly_count * 0.20)
sensor_mal_injected = 0
attempts = 0
while sensor_mal_injected < n_sensor_mal and attempts < 500:
attempts += 1
window_len = rng.integers(10, 30)
start = rng.integers(100, n_steps - window_len - 10)
if used[start:start + window_len].any():
continue
# Pick a random sensor to malfunction
target_sensor = rng.choice([
"temperature_c", "vibration_mm_s", "pressure_kpa", "motor_rpm"
])
# Sensor gives erratic readings (not physically consistent)
original = sensors[target_sensor][start:start + window_len].copy()
noise_scale = np.std(sensors[target_sensor]) * rng.uniform(2, 5)
# Random spikes and drops (uncorrelated with other sensors)
erratic = original + noise_scale * rng.standard_normal(window_len)
# Sometimes stuck at a value
if rng.random() > 0.5:
stuck_start = rng.integers(0, max(1, window_len // 2))
stuck_len = rng.integers(3, min(10, window_len - stuck_start))
erratic[stuck_start:stuck_start + stuck_len] = original[stuck_start]
sensors[target_sensor][start:start + window_len] = erratic
# Key distinction: OTHER correlated sensors remain normal!
# This is how an agent should detect it's a sensor issue, not equipment
is_anomaly[start:start + window_len] = 1
anomaly_type[start:start + window_len] = "sensor_malfunction"
used[start:start + window_len] = True
sensor_mal_injected += window_len
return is_anomaly, anomaly_type
def add_missing_values(df, rng):
"""Add realistic missing value patterns."""
n = len(df)
sensor_cols = [
"temperature_c", "vibration_mm_s", "pressure_kpa", "motor_rpm",
"flow_rate_lpm", "power_consumption_kw", "coolant_temp_c",
"acoustic_level_db", "oil_viscosity_cst", "humidity_pct"
]
# Random sporadic missing (~1-2% per sensor)
for col in sensor_cols:
mask = rng.random(n) < rng.uniform(0.008, 0.025)
df.loc[mask, col] = np.nan
# Block missing (simulating sensor outages, 5-15 min blocks)
n_blocks = rng.integers(15, 30)
for _ in range(n_blocks):
block_sensor = rng.choice(sensor_cols)
block_start = rng.integers(0, n - 20)
block_len = rng.integers(5, 16)
df.loc[block_start:block_start + block_len, block_sensor] = np.nan
# Correlated missing: when one sensor drops, nearby sensors sometimes drop too
for col in sensor_cols[:5]:
missing_idx = df[df[col].isna()].index
for idx in missing_idx:
if rng.random() < 0.15: # 15% chance of correlated dropout
neighbor = rng.choice([c for c in sensor_cols if c != col])
if idx in df.index:
df.loc[idx, neighbor] = np.nan
return df
def add_concept_drift(sensors, n_steps, rng):
"""Add subtle concept drift — baseline shifts partway through."""
drift_point = rng.integers(n_steps // 3, 2 * n_steps // 3)
drift_ramp = np.zeros(n_steps)
ramp_len = rng.integers(500, 1500)
end_point = min(drift_point + ramp_len, n_steps)
drift_ramp[drift_point:end_point] = np.linspace(0, 1, end_point - drift_point)
drift_ramp[end_point:] = 1.0
# Subtle baseline shift in temperature and pressure
sensors["temperature_c"] += drift_ramp * rng.uniform(1.5, 3.5)
sensors["pressure_kpa"] -= drift_ramp * rng.uniform(1.0, 3.0)
return sensors
def generate_equipment_data(eq_id, profile, rng):
"""Generate complete data for one equipment unit."""
n = READINGS_PER_EQUIPMENT
# Timestamps
timestamps = [START_DATE + timedelta(minutes=i) for i in range(n)]
# Operating modes
modes = generate_operating_mode_sequence(n, rng)
# Base sensor data
sensors = generate_base_sensor_data(n, profile, modes, rng)
# Add concept drift
sensors = add_concept_drift(sensors, n, rng)
# Inject anomalies
is_anomaly, anomaly_type = inject_anomalies(sensors, modes, n, rng)
# Equipment metadata
equipment_age_hours = np.arange(n) / 60 + rng.uniform(500, 5000)
last_maintenance_hours = np.zeros(n)
maintenance_interval = rng.integers(2000, 4000)
last_maint = 0
for i in range(n):
if (i - last_maint) > maintenance_interval:
last_maint = i
maintenance_interval = rng.integers(2000, 4000)
last_maintenance_hours[i] = (i - last_maint) / 60
# Build dataframe
data = {
"timestamp": timestamps,
"equipment_id": eq_id,
"operating_mode": modes,
**sensors,
"equipment_age_hours": np.round(equipment_age_hours, 1),
"hours_since_maintenance": np.round(last_maintenance_hours, 1),
"is_anomaly": is_anomaly,
"anomaly_type": anomaly_type,
}
return pd.DataFrame(data)
def add_leakage_features(df, rng):
"""
Add features that look useful but contain target leakage.
These should be identified and removed by a careful agent.
"""
n = len(df)
# Leakage 1: rolling_anomaly_rate — computed from future labels
# This is a rolling mean of is_anomaly with centered window (uses future data)
df["rolling_anomaly_rate"] = (
df.groupby("equipment_id")["is_anomaly"]
.transform(lambda x: x.rolling(60, center=True, min_periods=1).mean())
)
# Add small noise to disguise it
df["rolling_anomaly_rate"] += rng.normal(0, 0.005, n)
df["rolling_anomaly_rate"] = df["rolling_anomaly_rate"].clip(0, 1)
# Leakage 2: maintenance_priority_score — assigned AFTER anomaly detected
# Higher scores for rows that are actually anomalies (post-hoc label)
df["maintenance_priority_score"] = rng.uniform(0, 30, n)
anomaly_mask = df["is_anomaly"] == 1
df.loc[anomaly_mask, "maintenance_priority_score"] += rng.uniform(20, 50, anomaly_mask.sum())
# Leakage 3: alert_code — categorical that's derived from anomaly_type
def map_alert(atype):
if atype == "normal":
return rng.choice(["NONE", "TEMP_WARN", "VIB_WARN"], p=[0.92, 0.04, 0.04])
elif atype == "thermal_runaway":
return rng.choice(["TEMP_WARN", "TEMP_CRIT"], p=[0.3, 0.7])
elif atype == "bearing_degradation":
return rng.choice(["VIB_WARN", "VIB_CRIT"], p=[0.4, 0.6])
elif atype == "pressure_leak":
return rng.choice(["PRESS_WARN", "PRESS_CRIT"], p=[0.5, 0.5])
elif atype == "sensor_malfunction":
return rng.choice(["SENSOR_ERR", "NONE"], p=[0.6, 0.4])
return "NONE"
df["alert_code"] = df["anomaly_type"].apply(map_alert)
return df
def main():
rng = np.random.default_rng(SEED)
all_data = []
for eq_id, profile in EQUIPMENT_PROFILES.items():
print(f"Generating data for {eq_id}...")
eq_rng = np.random.default_rng(rng.integers(0, 2**31))
eq_data = generate_equipment_data(eq_id, profile, eq_rng)
all_data.append(eq_data)
df = pd.concat(all_data, ignore_index=True)
# Sort by timestamp, then equipment_id
df = df.sort_values(["timestamp", "equipment_id"]).reset_index(drop=True)
# Add unique row IDs
df.insert(0, "reading_id", [f"R{str(i).zfill(6)}" for i in range(len(df))])
# Add leakage features (trap for naive agents)
df = add_leakage_features(df, rng)
# Add missing values
df = add_missing_values(df, rng)
# Print stats
print(f"\nDataset shape: {df.shape}")
print(f"Anomaly rate: {df['is_anomaly'].mean():.4f}")
print(f"Anomaly type distribution:")
print(df["anomaly_type"].value_counts())
print(f"\nMissing values per column:")
print(df.isnull().sum()[df.isnull().sum() > 0])
# Save raw data
df.to_csv("/home/claude/data.csv", index=False)
print(f"\nSaved to /home/claude/data.csv")
return df
if __name__ == "__main__":
df = main()