💻 computer science

WMGIN: Windowed Multi-Granularity Interpretable Network for Hard Landing Warning

This paper proposes WMGIN, a self-attention-based multi-granularity interpretable deep learning model that processes multi-scale QAR data through windowed and layer-wise fusion strategies to accurately predict hard landing events while maintaining interpretability.

Original authors: Zemin Lin, Yonglian Li, Wei Li, Guolong Li, Zonghua Ye, Tian Zeng, Yan Liu, Zhifeng Wang

Published 2026-10-06

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Zemin Lin, Yonglian Li, Wei Li, Guolong Li, Zonghua Ye, Tian Zeng, Yan Liu, Zhifeng Wang

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The final seconds of a flight are often the most tense, a moment where the smooth rhythm of travel gives way to the precise, high-stakes choreography of touching down. For decades, aviation safety experts have known that the landing phase, though it makes up only a tiny fraction of a flight's total time, accounts for a disproportionate number of accidents. Among these, the "hard landing" stands out as a particularly dangerous event. It occurs when an aircraft hits the runway with too much force, a jolt that can damage the plane's structure and endanger everyone on board. To prevent this, airlines rely on flight data recorders, devices that constantly log hundreds of measurements like speed, altitude, and the pilot's control inputs. The challenge for scientists has long been how to turn this massive, complex stream of numbers into a clear warning before the wheels even touch the ground. Traditional methods often struggle to find the subtle patterns hidden in the data, while newer, more powerful computer systems can predict risks accurately but operate like black boxes, offering no explanation for why they made a specific prediction.

A team of researchers from Zhuhai Xiangyi Aviation Technology Co., Ltd. has developed a new approach to solve this dual problem of accuracy and clarity. They created a system called WMGIN, a specialized computer model designed to watch flight data and predict hard landings while simultaneously explaining its reasoning. The researchers tested their system on a vast collection of real-world data from 44,804 flights of a specific aircraft type, the A321. Their goal was not just to build a model that could spot a dangerous landing, but to build one that could tell a pilot or safety officer exactly which factors—such as the rate of descent or the angle of the nose—were driving the danger at any given moment. The results showed that their new system could identify hard landings with greater reliability than previous methods, and crucially, it could point directly to the specific moments and variables that led to the prediction.

The core difficulty the researchers faced was the nature of the data itself. Flight recorders capture information at different speeds; some measurements, like engine speed, are recorded once per second, while others, like vertical acceleration, are recorded eight times per second. Older computer models often had to smooth out these differences, averaging the fast data to match the slow, which inevitably threw away important details. The new system, WMGIN, avoids this loss by learning to read the raw data at its natural pace. It treats the flight data as a series of time windows, looking at small chunks of information to understand the immediate behavior of the plane. This allows it to see fine details, like a sudden shift in the pilot's control stick, without getting lost in the noise of the entire flight.

To make sense of the complex relationships between different parts of the aircraft, the model uses a strategy that mimics how a human might analyze a situation. It first looks at individual variables, such as the vertical speed or the position of the rudder, to understand their local behavior. Then, it brings these observations together to see how they influence one another. A key innovation in this system is the use of a "class token," a special digital marker that acts as a summary for the entire analysis. As the model processes the data, this marker gathers the most important clues from every part of the flight. By watching how the model assigns weight to this marker, the researchers can see exactly which parts of the flight history mattered most for the final decision. This turns the model from a mysterious predictor into a transparent tool that reveals its logic.

When the researchers put their system to the test, the results were compelling. They compared WMGIN against several other established methods, including older statistical techniques and more modern deep learning models. In nearly every test case, the new system outperformed the others, correctly identifying a higher percentage of actual hard landings. This is a critical distinction in aviation safety; missing a dangerous landing is far worse than falsely flagging a safe one. The system's ability to catch these events was particularly strong when analyzing data from the final thirty-two seconds before touchdown. Beyond the numbers, the system provided a window into the "why" behind the prediction. In one specific case study of a flight that resulted in a hard landing, the model correctly identified that the pilot had descended too quickly from a low altitude and failed to adjust the nose angle in time to soften the impact. The system highlighted the vertical speed and the control inputs as the primary drivers of the risk, matching the findings of human experts who reviewed the same flight.

The study also revealed that the system's design choices were essential to its success. When the researchers removed specific components, such as the layer that merges different time windows or the layer that tracks the relationship between different variables, the model's performance dropped significantly. This confirmed that the ability to see both the small, immediate details and the broader connections between them was necessary for accurate prediction. Furthermore, the system's transparency allowed the team to verify that it was focusing on the right things. It paid close attention to parameters like vertical acceleration and pitch control, which are known to be critical in landing safety, rather than getting distracted by irrelevant data points.

While the study focused specifically on hard landings, the implications for flight safety are broad. The ability to predict an event with high accuracy while also providing a clear, human-readable explanation is a significant step forward. It moves the field away from relying on models that simply output a "yes" or "no" and toward systems that can act as a partner in safety analysis. The researchers noted that their work is limited to a single type of event and a specific aircraft model, but they believe the underlying architecture could be adapted to detect other safety issues, such as runway overruns or tail strikes. By proving that a deep learning model can be both highly accurate and deeply understandable, this work offers a new path for using artificial intelligence to keep the skies safer, turning raw data into actionable insight without sacrificing the clarity that safety professionals need.

Technical Summary: WMGIN for Hard Landing Warning

Problem Statement

Landing is the most critical phase of flight, accounting for a disproportionate share of aviation accidents despite representing only 1% of total flight time. A "hard landing," defined as an event where the aircraft's vertical acceleration exceeds a specific threshold (e.g., 1.5g) after touchdown, poses significant risks to structural integrity and passenger safety.

Current research utilizing Quick Access Recorder (QAR) data to predict hard landings faces two primary limitations:

Predictive Accuracy: Traditional machine learning methods often fail to capture complex feature interdependencies in multi-scale, multivariate time-series data, leading to suboptimal accuracy. Conversely, many deep learning approaches that improve accuracy often neglect the preservation of raw data granularity or fail to model temporal dependencies effectively.
Interpretability: Most existing deep learning models function as "black boxes." While they may achieve high predictive performance, they lack the ability to explain why a prediction was made. This is a critical gap in aviation safety, where understanding the causal mechanisms (e.g., specific pilot actions or aircraft states) is as important as the prediction itself.

Methodology: WMGIN

The authors propose WMGIN (Windowed Multi-Granularity Interpretable Network), a self-attention-based deep learning model designed to balance high predictive accuracy with multi-granularity interpretability. The architecture processes QAR data through three distinct stages:

1. Embedding Layer

To handle the multi-scale nature of QAR data (where parameters have varying sampling frequencies, e.g., 1 Hz to 8 Hz), the model standardizes the temporal dimension. Instead of simple averaging which causes information loss, the embedding layer concatenates high-frequency data points within each second to form a unified feature vector. This preserves the high-frequency details while unifying the time dimension to a fixed length ( $T$ ).

2. Time Window Encoder (TW Encoder)

This module focuses on capturing local temporal dependencies for individual variables.

Windowed Self-Attention: Unlike standard Transformers that use a global receptive field, the TW Encoder employs a window-based self-attention mechanism with a fixed window size (e.g., 2 tokens). This restricts information exchange to adjacent time steps, preserving the "information barrier" necessary for interpretability.
Multi-Layer Fusion: The encoder utilizes a hierarchical design where adjacent time tokens are merged after each window-based Transformer block. This progressively expands the temporal information contained within a single token, allowing the model to learn representations from fine-grained to coarse-grained levels.
Relative Position Encoding (RPE): Learnable RPE matrices are used within windows to capture the relative distance between tokens.

3. Interpretable Hybrid Encoder (IH Encoder)

This module addresses global dynamic correlations among different variables and provides interpretability.

Global Inter-Dimensional Mask: To model the fact that correlations between variables weaken as the time interval increases, the authors introduce a mask constructed from a fixed reference curve based on a deformed Gaussian function ( $y = e^{-2x^2} + 0.001$ ). 31 function values are taken at equal intervals from this curve to serve as initial weight values. These values form a reference vector which is then used to construct time-aware masks for each layer. To match the reduced token counts in deeper layers, each layer's mask is downsampled via AvgPool. This mask constrains the attention mechanism, assigning higher weights to closer time steps and lower weights to distant ones.
Learnable Class Tokens: A unique feature of WMGIN is the introduction of a learnable class token ($cls$) at each layer. Unlike standard attention weights that measure similarity between tokens, the attention weights calculated toward the class token represent the contribution of each token to the final prediction.
Prediction Mechanism: The final prediction relies solely on the concatenated class tokens from all layers. This design ensures that the model's decision-making process is explicitly traceable through the attention weights assigned to these tokens.

Key Contributions

Novel Architecture (WMGIN): The integration of TW blocks (for local temporal patterns) and IH blocks (for global variable correlations) via a multi-layer fusion design. This allows the model to capture both individual variable dynamics and cross-variable interactions at multiple granularities.
Multi-Granularity Interpretability: By aggregating the weights of class tokens across layers, the model reveals the contributions of specific time periods and variables at different levels of abstraction (fine-grained to coarse-grained).
Experimental Validation: Comprehensive evaluation on a real-world dataset of 44,804 A321 flights demonstrates that WMGIN outperforms state-of-the-art baselines (including RF, LSTM, and various Transformer variants) in hard landing prediction while providing actionable insights.

Experimental Results

The study utilized a dataset of 44,804 A321 flights, categorized into 44,115 normal landings and 689 hard landings. The data was split into training, validation, and test sets, with upsampling applied to the training set to address class imbalance.

Performance Metrics: The model was evaluated using Precision, Recall, F1-score, and PR-AUC.
Comparative Performance: WMGIN achieved the highest performance across nearly all metrics and data types (defined by different time windows before touchdown). Specifically, it achieved a Recall of 96.25% and a PR-AUC of 71.53% on Type I data (32s to touchdown), outperforming the next best Transformer-based model (PatchTST).
Ablation Studies: Removing key components (Embedding Layer, Fusion Layer, RPE, or the Global Mask) resulted in significant performance degradation, confirming the necessity of each module.
Interpretability Case Study: Visualization of attention weights on a specific hard landing flight revealed that the model correctly identified critical parameters:
- IVV (Vertical Speed) and Radio Altimeters indicated an abnormal descent rate.
- PITCH/PITCH CMD and ROLL/ROLL CMD highlighted pilot control inputs that deviated from standard procedures.
- The analysis correctly pinpointed that the pilot failed to increase the pitch angle in time to mitigate vertical velocity, a conclusion consistent with domain expertise.

Significance and Claims

The paper claims that WMGIN successfully addresses the trade-off between predictive accuracy and interpretability in flight safety applications. By using learnable class tokens and windowed attention, the model not only predicts hard landings with superior accuracy but also provides a transparent explanation of the underlying factors (specific variables and time windows) driving the prediction.

The authors conclude that this approach offers valuable insights for flight safety analysis, enabling the identification of root causes for hard landings. They acknowledge a limitation: the current model focuses on a single event type (hard landing). Future work is proposed to extend WMGIN as a backbone for multi-classification tasks involving other safety events, such as tail strikes and runway excursions, to explore intrinsic relationships among different flight safety incidents.