Fire Distinguisher - 9/10

Problem Description: Fire Distinguisher

Context

SafeDrive Technologies, a pioneering company in the field of automotive safety, is developing an intelligent disaster prevention system for modern vehicles. With the increase in incidents caused by overheating and gas leaks, the company needs a Fire Distinguisher - an AI system capable of detecting dangerous anomalies in real-time and classifying them correctly.

You are the Lead Machine Learning Engineer. Your task: to build the detection system that can save lives.

Data Description

The data comes from sensors mounted in experimental vehicles, collected at 1-second intervals:

Input Sensors:

• id: ID associated with the record
• gas_raw: Raw gas concentration detected
• temp_c: Temperature in degrees Celsius
• flame: Binary flame sensor (0 = absent, 1 = present)
• gas_avg: Moving average of gas concentration
• temp_avg: Moving average of temperature
• dGas: Rate of gas change
• dTemp: Rate of temperature change
• sensor_noise: Noise introduced by the sensor
• timestamp_ms: Unix timestamp in milliseconds
• sensor_id: Sensor ID (S001-S005)
• battery_level: Sensor battery level
• vehicle_speed: Vehicle speed
• weather_condition: Weather condition (sunny, rainy, cloudy, foggy)
• gas_sensor_drift: Gas sensor drift
• ambient_temp: Ambient temperature

Target:

• label: Hazard type
  • NORMAL: Safe conditions
  • GAS_LEAK: Gas leak detected
  • OVERHEAT: Dangerous overheating
  • FIRE: Active fire

Provided Files

• train.csv: Training data with all columns (including label)
• test.csv: Test data without the label column

Mission Requirements

Subtask 1: Hazard Profile (20 points)

The safety team wants to understand the distribution and characteristics of hazards in the training data.

Your task: For train.csv, calculate the following 12 statistics:

1. Percentage of readings for each hazard type (4 values: NORMAL, GAS_LEAK, OVERHEAT, FIRE) - rounded to 2 decimal places
2. Average temperature for each hazard type (4 values) - rounded to 1 decimal place
3. Average gas concentration for each hazard type (4 values) - rounded to 1 decimal place

Output: 12 numerical values (4 percentages + 4 average temperatures + 4 average gas concentrations).

Example: For NORMAL: 45.32% of readings, average temperature 32.5°C, average gas 245.3

Subtask 2: Anomaly Score (15 points)

The safety team wants to identify abnormal readings that could indicate dangerous situations. Normal readings form a pattern in the feature space — your task is to measure how far each test reading is from this pattern.

1. Select the most relevant features

• Identify the top 6 numerical attributes that have the strongest Spearman correlation with the hazard type (encoded: NORMAL=0, GAS_LEAK=1, OVERHEAT=2, FIRE=3)
• In case of a tie, use alphabetical order

2. Prepare the data

• Replace missing values with the median (calculated only on train)
• Scale the data using standardization (trained on train, applied to test)

3. Define the NORMAL profile

• Use only records with label == "NORMAL" from train to establish what "normal behavior" means

4. Calculate the anomaly score

• For each reading in test, calculate the Mahalanobis distance from the center of the NORMAL distribution
• Formula: d = sqrt((x - mu)^T * Sigma^(-1) * (x - mu))
  • mu = mean vector of NORMAL records
  • Sigma = covariance matrix of NORMAL records
  • For numerical stability, add a small regularization before inversion: Sigma_reg = Sigma + I * 1e-6, where I is the identity matrix
• Round to 2 decimal places

Output: The anomaly score for each ID in test.csv.

Example:

12,5234,8.9222,6567,1.2332,7891,14.45

Hint: In Python: scipy.spatial.distance.mahalanobis() or manual calculation with numpy.

Subtask 3: Risk Level (15 points)

The raw score obtained is not easily interpreted by the field team. Transform it into risk categories (LOW/MEDIUM/HIGH) using thresholds based on historical data.

1. Calculate thresholds based on train

• Apply the same procedure from Subtask 2 to all records in train
• Calculate the 33rd percentile and 66th percentile of these scores
• These become your thresholds: p33 and p66

2. Classify each test reading

Output: The risk level for each ID in test.csv.

Example:

13,5234,HIGH23,6567,LOW33,7891,HIGH

Subtask 4: Hazard Classifier (50 points)

CORE MISSION! This is the brain of the Fire Distinguisher system. Every second, hundreds of vehicles transmit data, and your system must decide instantly: is everything normal or is there a danger? A mistake in fire detection can cost lives. SafeDrive counts on you to build the most accurate classifier possible.

Your task: For each record in test.csv:

• Predict the label_name (NORMAL / GAS_LEAK / OVERHEAT / FIRE)
• Build a multi-class classification model using the available features

Output: The predicted label (NORMAL, GAS_LEAK, OVERHEAT, or FIRE) for each test reading.

Evaluation

Final score: 100 points maximum.

Submission File Format (submission.csv)

Structure: subtaskID,datapointID,answer

Details:

Subtask 1: 12 rows with fixed statistics

• datapointID 1-4: Percentages for NORMAL, GAS_LEAK, OVERHEAT, FIRE
• datapointID 5-8: Average temperatures for the 4 classes
• datapointID 9-12: Average gas concentrations for the 4 classes

Subtask 2: N rows

• datapointID = id from test.csv
• answer = Mahalanobis anomaly score (rounded to 2 decimal places)

Subtask 3: N rows

• datapointID = id from test.csv
• answer = risk level (LOW, MEDIUM, or HIGH)

Subtask 4: N rows

• datapointID = id from test.csv
• answer = predicted label (NORMAL, GAS_LEAK, OVERHEAT, FIRE)