ML in AOC

Using Machine Learning in Airline Operations Control (AOC)

Applying machine learning (ML) in an Airline Operations Centre (AOC) is a highly promising and complex domain. The AOC functions as the command hub for airline operations, where operators handle an overwhelming amount of real-time data to ensure smooth execution of flights. Supporting these professionals with intelligent systems means freeing them up to focus on strategic decisions rather than routine monitoring.

The Problem: Flight Diversions and Information Delays

Let’s consider a common scenario: when a flight undergoes a diversion or a go-around, the first point of communication typically comes from Air Traffic Control (ATC). However, not all airlines have established direct communication channels with every airport across their network. This delay in acquiring diversion information often results in inefficient logistical decisions.

Imagine an aircraft scheduled to deliver cargo to a city. The cargo is meant to be picked up by trucks from the destination airport. But due to unforeseen circumstances, the flight is diverted to a different airport. The trucks, unaware of this change, are still dispatched to the original airport, leading to wasted resources, time, and fuel. If there were a way to detect such diversions in near real-time—using only ADS-B data (latitude, longitude, altitude, ground speed)—airlines could respond faster, adapt logistics, reduce environmental impact, and save operational costs.

---

The ML Opportunity

There is a clear need to apply ML here. The challenge lies in detecting flight deviations based on limited but continuous flight data. This problem is complex because:

• Data is streamed in real time, not static.

• A single data point (e.g., position at a given second) is not meaningful on its own.

• The sequence or pattern of data points (e.g., changing altitude or heading over time) is what indicates a deviation.

For example, if points 1, 2, 3, and 4 in a data stream individually show normal behavior, a deeper temporal analysis might reveal that a significant course change is taking place—something a human eye or rule-based logic might miss.

---

Why Not Just Use Logic?

It’s a valid question: if we can detect diversions using fixed logical rules (e.g., "if aircraft turns more than 45 degrees from the planned track and descends before destination, then it’s likely a diversion"), why introduce the complexity of ML?

Limitations of Rule-Based Systems:

• Rigid and not adaptive to new patterns.

• Require manual tuning and constant updating.

• Cannot handle ambiguous or borderline cases well.

• Fail to learn from past incidents without reprogramming.

Advantages of ML Over Pure Logic:

• Learns patterns from large datasets (even patterns humans can’t easily spot).

• Generalizes better across different flight paths, aircraft types, or weather conditions.

• Continuously improves as more data becomes available.

---

ML Model Considerations

The standard process of building an ML model typically involves:

1. Data collection: Gather a diverse set of flight records.

2. Data labeling: Identify which flights had deviations.

3. Splitting: Divide data into training (usually ~80%) and testing (~20%) sets.

4. Feature engineering: Add useful indicators (e.g., rate of descent, heading change).

5. Model training: Apply the algorithm to learn from training data.

6. Validation: Measure performance (accuracy, precision, recall) on test data.

In our use case, there are special complexities:

• Data structure: Each "data point" is actually a sequence—e.g., 100 time-steps for a single flight segment.

• Prediction scope: The model must not only detect a diversion but ideally do so early—before it’s confirmed by external sources.

• Sequential logic: The model must consider historical context, not just the current state.

Thus, a model such as an LSTM (Long Short-Term Memory) or GRU (Gated Recurrent Unit) neural network could be well suited, as they are capable of learning from sequences and capturing temporal dependencies.

---

Types of ML: Supervised vs Unsupervised vs Reinforcement

1. Supervised Learning

Used when labeled data is available (e.g., "this flight diverted", "this one didn't").

Pros:

• High accuracy with enough data.

• Easy to interpret performance.

• Useful for prediction and classification.

Cons:

• Requires extensive labeled datasets.

• Can be biased by poor labeling.

Examples: Random Forest, SVM, Decision Trees, Neural Networks.

---

2. Unsupervised Learning

Used when labels are not available, aiming to find structure in the data.

Pros:

• Can reveal hidden patterns.

• Useful for anomaly detection and clustering.

• Helps with dimensionality reduction.

Cons:

• Less precise.

• Harder to evaluate results.

• Difficult with mixed data types (categorical + numerical).

Examples: K-Means, DBSCAN, Hierarchical Clustering.

---

3. Reinforcement Learning

Used in decision-making environments where the algorithm learns by reward and penalty.

Pros:

• Powerful for sequential decision problems.

• Learns from interaction.

Cons:

• Complex to implement.

• Requires simulated environments or long training times.

Not ideal for current application, unless we’re simulating flight operations or learning policy optimization.