By the end of this chapter, you will be able to:
Statistical anomaly detection uses math to spot unusual sensor readings. Think of measuring the heights of everyone in a room – someone three times the average height would clearly stand out. Similarly, if a temperature sensor suddenly reads far outside its normal range, statistical methods can automatically flag it as suspicious without needing complex machine learning.
Core Concept: Statistical methods detect anomalies by measuring how far a value deviates from “normal” - either in standard deviations (Z-score) or quartile ranges (IQR).
Why It Matters: Statistical methods are computationally lightweight, require no training data, and can run on resource-constrained edge devices. They catch 80% of anomalies with 10% of the complexity of ML approaches.
Key Takeaway: Use Z-score for Gaussian data, IQR for skewed or bounded data, and adaptive thresholds when “normal” changes over time.
Before diving into this chapter, you should be familiar with:
Statistical approaches form the foundation of anomaly detection in IoT systems. They are computationally lightweight, suitable for edge deployment, and work well when data follows known distributions. Unlike machine learning methods that require training data and significant compute resources, statistical detectors can run on microcontrollers with as little as 16 bytes of RAM. This chapter covers three complementary methods – Z-score, IQR, and adaptive thresholds – each suited to different data characteristics and deployment constraints.
Statistical anomaly detection works by quantifying how unusual a value is relative to historical patterns, using mathematical measures of deviation.
Z-Score Mechanism:
Why this works: In Gaussian distributions, 99.7% of values fall within ±3σ. Readings beyond this are statistically rare (<0.3% probability if data is normal).
IQR Mechanism:
Why this works: IQR uses percentiles (order statistics) rather than mean/std, making it robust to outliers and distribution shape. Works even when data is skewed or has heavy tails.
Adaptive Thresholds Mechanism:
Why this works: Sensor characteristics drift due to aging, temperature, etc. Fixed thresholds become obsolete. Adaptive methods track changing baselines automatically.
Core Concept: Measure how many standard deviations a data point is from the mean.
Formula:
\[Z = \frac{x - \mu}{\sigma}\]
The Z-score threshold directly controls your false alarm rate. For a Gaussian distribution, the probability of exceeding a threshold is:
\[P(|Z| > z) = 2 \times (1 - \Phi(z))\]
where \(\Phi(z)\) is the cumulative distribution function. Concrete thresholds:
Example: A temperature sensor samples every minute (1,440 readings/day). With \(Z > 3.0\), expect \(1,440 \times 0.0027 \approx 4\) false alarms per day even when nothing is wrong. For 100 sensors, that’s 400 false alarms daily. Adjust threshold to \(Z > 3.5\) (0.047% rate) for 68 daily false alarms across the fleet.
Adjust the sensor parameters below to see how Z-score thresholds translate to real detection boundaries and false alarm rates.
When It Works:
When It Fails:
Implementation Example:
import numpy as np
class ZScoreDetector:
def __init__(self, threshold=3.0, window_size=100):
self.threshold = threshold
self.window_size = window_size
self.values = []
def update(self, value):
"""Streaming Z-score anomaly detection"""
self.values.append(value)
# Keep only recent window
if len(self.values) > self.window_size:
self.values.pop(0)
# Need minimum samples
if len(self.values) < 30:
return False, 0.0
mean = np.mean(self.values)
std = np.std(self.values)
if std == 0: # Avoid division by zero
return False, 0.0
z_score = abs((value - mean) / std)
is_anomaly = z_score > self.threshold
return is_anomaly, z_score
# Usage example: Temperature sensor monitoring
detector = ZScoreDetector(threshold=3.0)
# Seed with 30+ normal readings to meet minimum sample requirement
import random
random.seed(42)
for _ in range(35):
normal_temp = 22.0 + random.uniform(-0.5, 0.5)
detector.update(normal_temp)
# Now test with normal and anomalous readings
for temp in [22.1, 22.5, 21.8, 22.3, 22.0]:
anomaly, score = detector.update(temp)
print(f"Temp: {temp}C, Z-score: {score:.2f}, Anomaly: {anomaly}")
# Anomalous reading
anomaly, score = detector.update(-5.0) # Sensor malfunction
print(f"Temp: -5.0C, Z-score: {score:.2f}, Anomaly: {anomaly}")
# Output: Temp: -5.0C, Z-score: ~75.0, Anomaly: TrueImplementation sketch:
abs((value - mean) / std) for each new reading.Temperature example: after learning a normal baseline around 22 C, a -5 C reading produces an extremely large Z-score and is flagged immediately.
Not all unusual values indicate real-world problems. A temperature spike from 22C to -40C is almost certainly a sensor malfunction, not actual weather. IQR is the better choice when your data is skewed or has natural outliers because it uses percentiles rather than mean/standard deviation, making it immune to the distortions that extreme values cause in Z-score calculations. For IoT sensors with physical constraints (humidity 0-100%, voltage 0-5V), combine IQR with hard bounds – any reading outside physical limits is immediately flagged as a sensor fault, not an anomaly.
Core Concept: Identify outliers based on the spread of the middle 50% of data.
Formula:
\[\mathrm{IQR} = Q_3 - Q_1\] \[\mathrm{Lower\ Bound} = Q_1 - 1.5 \times \mathrm{IQR}\] \[\mathrm{Upper\ Bound} = Q_3 + 1.5 \times \mathrm{IQR}\]
Advantages over Z-Score:
IoT Use Cases:
Implementation:
import numpy as np
def iqr_anomaly_detection(data, value):
"""
Detect if a new value is anomalous using IQR method
"""
q1 = np.percentile(data, 25)
q3 = np.percentile(data, 75)
iqr = q3 - q1
lower_bound = q1 - 1.5 * iqr
upper_bound = q3 + 1.5 * iqr
is_anomaly = (value < lower_bound) or (value > upper_bound)
return is_anomaly, lower_bound, upper_bound
# Example: Battery voltage monitoring (normal: 3.3-4.2V)
battery_readings = [3.85, 3.92, 3.78, 3.88, 3.95, 3.82, 3.90,
3.87, 3.93, 3.81, 3.89, 3.86]
# Sorted: [3.78, 3.81, 3.82, 3.85, 3.86, 3.87, 3.88, 3.89, 3.90, 3.92, 3.93, 3.95]
# Q1 ≈ 3.84, Q3 ≈ 3.91, IQR ≈ 0.07
# Check new reading
new_reading = 2.1 # Low battery or sensor fault
anomaly, lower, upper = iqr_anomaly_detection(battery_readings, new_reading)
print(f"Battery: {new_reading}V")
print(f"Expected range: {lower:.2f}V - {upper:.2f}V")
print(f"Anomaly: {anomaly}")
# Output: Battery: 2.1V, Expected range: 3.74V - 4.00V, Anomaly: TrueImplementation sketch:
Q1) and 75th percentile (Q3).IQR = Q3 - Q1.lower = Q1 - 1.5 * IQR and upper = Q3 + 1.5 * IQR.lower or above upper.Battery example: a 2.1 V reading falls well below the expected band, so the detector marks it as anomalous.
Core Concept: Calculate mean and standard deviation over a sliding window to adapt to changing conditions.
Why It Matters for IoT:
Implementation:
from collections import deque
import numpy as np
import random
class AdaptiveThresholdDetector:
def __init__(self, window_size=50, n_std=3.0):
self.window = deque(maxlen=window_size)
self.n_std = n_std
def update(self, value):
"""Add a value and compare it to the rolling baseline."""
if len(self.window) <= 10:
self.window.append(value)
return False, None, None
mean = np.mean(self.window)
std = np.std(self.window)
lower = mean - self.n_std * std
upper = mean + self.n_std * std
flagged = (value < lower) or (value > upper)
if not flagged: # Avoid contaminating the baseline
self.window.append(value)
return flagged, lower, upper
# Example: Indoor temperature monitoring with day/night cycles
detector = AdaptiveThresholdDetector(window_size=50, n_std=2.5)
for temp in [22.7, 23.0, 23.2]:
reading = temp + random.uniform(-0.2, 0.2)
flagged, lower, upper = detector.update(reading)
print(round(reading, 1), flagged, lower, upper)Implementation sketch:
mean ± n_std * std.Window Size Trade-Offs:
| Window Size | Responsiveness | Stability | Best For |
|---|---|---|---|
| Small (10-50) | High - detects sudden changes | Low - sensitive to noise | Fast-changing environments |
| Medium (50-200) | Balanced | Balanced | General IoT monitoring |
| Large (200+) | Low - slow to adapt | High - ignores noise | Stable industrial processes |
Small (10-50)
Medium (50-200)
Large (200+)
Option A: Fixed statistical thresholds (Z-score > 3, IQR bounds) Option B: Adaptive thresholds with moving statistics Decision Factors: Fixed thresholds are simpler to implement and explain (auditable for compliance), work well when normal behavior is stable, and have lower computational overhead. Adaptive methods handle concept drift, seasonal patterns, and changing sensor characteristics, but require tuning window sizes and may temporarily miss anomalies during adaptation periods. Choose fixed for stable industrial processes; choose adaptive for environments with natural variation like building HVAC or outdoor sensors.
Mobile comparison guide: use these cards instead of the desktop matrix.
Z-Score
IQR Method
Adaptive Thresholds
Scenario: Arla Foods, a Scandinavian dairy cooperative, monitors refrigerated transport trucks delivering milk from farms to processing plants across Denmark. Each truck has 4 temperature sensors (front, rear, left wall, right wall) reporting every 30 seconds. Arla needs to detect cooling failures before milk spoils (milk must stay between 2 C and 6 C).
Given:
Step 1: Select detection method
| Method | Suitability | Reasoning |
|---|---|---|
| Z-Score | Good | Temperature is normally distributed around 3.8 C |
| IQR | Acceptable | Would also work, but Z-Score is simpler for Gaussian data |
| Adaptive | Not needed | “Normal” is stable (refrigeration setpoint does not drift) |
Z-Score
IQR
Adaptive
Step 2: Set Z-Score threshold
| Threshold | Temperature Trigger | False Positive Rate | Detection Speed |
|---|---|---|---|
| Z = 2.0 | 4.6 C (3.8 + 2 x 0.4) | 4.6% (too many alerts) | Very early |
| Z = 3.0 | 5.0 C (3.8 + 3 x 0.4) | 0.27% | Early (1.0 C below limit) |
| Z = 4.0 | 5.4 C (3.8 + 4 x 0.4) | 0.006% | Moderate (0.6 C margin) |
| Z = 5.0 | 5.8 C (3.8 + 5 x 0.4) | 0.00006% | Late (only 0.2 C margin) |
Z = 2.0
Z = 3.0
Z = 4.0
Z = 5.0
Selected: Z = 3.0 (triggers at 5.0 C, giving 1.0 C and approximately 15 minutes of warning before the 6 C regulatory limit is reached at typical failure rate of 0.07 C/minute).
Step 3: Calculate business impact
| Metric | Before (Manual) | After (Z-Score) | Improvement |
|---|---|---|---|
| Detection time | At delivery (2-6 hours late) | Within 90 seconds of anomaly | 99% faster |
| Monthly spoilage events | 23 truck loads | 3 truck loads (caught early, rerouted) | 87% reduction |
| Monthly spoilage cost | EUR 276,000 | EUR 36,000 | EUR 240,000 saved |
| False alarms per truck per day | 0 | ~31 (0.27% of 11,520 readings per truck) | High, needs filtering |
| Edge compute cost | 0 | EUR 0 (runs on existing truck gateway MCU) | No hardware needed |
Detection time
Monthly spoilage events
Monthly spoilage cost
False alarms per truck per day
Edge compute cost
Result: A Z-Score threshold of 3.0, running on the existing truck gateway microcontroller with zero additional hardware cost, detects cooling failures within 90 seconds and saves Arla EUR 240,000 per month in prevented milk spoilage. The raw false alarm rate of ~31 per truck per day is addressed by requiring 3 consecutive Z > 3.0 readings (90 seconds) before escalating, reducing false alerts to fewer than 1 per truck per day across the fleet.
Key Insight: Statistical anomaly detection does not require cloud connectivity, ML models, or specialized hardware. A Z-Score calculation consumes fewer than 20 floating-point operations per reading and fits in 16 bytes of RAM, making it deployable on any microcontroller. The key design decision is the threshold: Z = 3.0 provides the sweet spot between early warning (1.0 C margin) and manageable false positives (0.27%) for this cold chain use case.
Experiment with different anomaly detection methods on simulated IoT sensor data. Adjust the threshold and detection method to see how they affect true positive and false positive rates.
Mobile reading guide: Z-score maps to Gaussian edge data, IQR maps to skewed or bounded sensors, and adaptive thresholds map to changing baselines that need window tuning.
How These Concepts Connect:
Foundation Concepts:
When Statistical Methods Fail:
Deployment and Operations:
Edge Computing:
Z-score assumes a roughly bell-shaped distribution. Sensor data with heavy tails or strong skew will produce excessive false alarms. Visualise the distribution first; if non-Gaussian, use IQR or a log-transformation.
Computing Z-score against the overall historical mean ignores drift. A gradual temperature rise will eventually flag every reading as an anomaly. Always use a sliding or exponentially weighted window.
At 1 Hz, a 0.27% false alarm rate means 864 false alarms per day per sensor. Scale your threshold with sampling rate; higher rates often require tighter bounds (4σ or 5σ) plus temporal persistence checks.
Mixing raw ADC counts (0–4095) with calibrated engineering units (°C) will dominate the score with the higher-magnitude variable. Always normalise each sensor channel independently.
Statistical methods provide the foundation for anomaly detection in IoT:
Key Takeaway: Start with statistical methods. They catch 80% of anomalies with minimal resources. Only escalate to ML when statistical approaches fail.
Sammy the Sensor was monitoring the temperature of the school aquarium. Every day, the water was a nice, steady 24 degrees Celsius – give or take half a degree.
One morning, Sammy read 24.3 degrees. “Perfectly normal!” he reported to Max the Microcontroller.
Then he read 24.1. “Still fine!” Then 23.8. “No worries!”
Then suddenly: 15.2 degrees!
“WHOA!” shouted Sammy. “That is WAY different from normal!”
Max the Microcontroller quickly did some math. “The average temperature has been 24.0 degrees, and readings usually only vary by about 0.5 degrees. Your reading of 15.2 is almost EIGHTEEN standard deviations away from normal! That is called a Z-score of 18 – anything above 3 is suspicious!”
Lila the LED turned bright red to alert the teacher. They discovered the aquarium heater had broken!
“But wait,” said Bella the Battery. “What if we had a different sensor watching battery voltage? My voltage readings are not spread out evenly – they cluster near the top when I am full and drop fast at the end. Would Z-score work for me?”
“Great question!” said Max. “For you, we use the IQR method instead. It looks at the middle chunk of your readings and does not care about the shape of the data. It is more robust – like wearing a raincoat that works in any weather!”
Key lesson: Z-score works great when data follows a bell curve (like temperature). IQR works for any shape of data (like battery voltage). Pick the right tool for your data!
| If you want to… | Read this |
|---|---|
| Detect seasonally dependent anomalies | Time-Series Methods |
| Handle complex multivariate patterns | Machine Learning Approaches |
| Integrate methods into a production pipeline | Detection Pipelines |
| Evaluate your detector’s accuracy | Performance Metrics |
| Return to the module overview | Anomaly Detection Overview |