工程科学与技术 (Nov 2024)
Aircraft Multi-stage Altitude Prediction Under Satellite Signal Loss
Abstract
Objective Combining an inertial navigation system (INS) and a global positioning system (GPS) in stable conditions of GPS satellite signals offers the most accurate altitude indication, termed inertial satellite composite altitude. When GPS signals are lost or unstable, aircraft must rely only on INS altitude, introducing a discrepancy compared to the composite altitude. This reduction in altitude indication accuracy significantly affects navigation performance. Hence, the predictive recovery of aircraft altitude without GPS signals is crucial. Current research faces challenges in mining high-dimensional flight data and enhancing prediction accuracy. This study proposes a multi-stage altitude prediction model using attention mechanisms and temporal convolutional neural networks (TCNs).Methods The aircraft’s flight stages are determined to facilitate targeted altitude prediction for different flight phases. Traditional clustering algorithms often struggle to capture transitional states in time-series data. Therefore, a fuzzy logic approach is adopted to map ambiguous inputs to explicit output states, enabling the extraction of climb, cruise, and descent phases from the aircraft’s entire flight process. This segmentation aids in better capturing phase-specific features for the prediction model and provides data reserves for the three stages. Addressing the longitudinal nature of aircraft flight parameter time series, a long temporal correlation attention (LTCA) mechanism is designed for feature extraction, enhancing spatiotemporal correlation representation. LTCA efficiently exploited attention mechanisms to extract key features from multi-dimensional flight parameter data samples through adaptive global average pooling (GAP) and one-dimensional convolution, considering global and local information. This approach provided a more effective feature representation for aircraft altitude prediction in the absence of satellite signals. Then, an LTCA–TCN altitude prediction model is constructed using the temporal data processing capability of TCNs. Finally, due to the inability of classical regression model performance metrics to account for error tolerance across different flight phases, a novel evaluation metric called “Score” is proposed to assess the multi-stage altitude prediction capability of the model. This metric considered overestimation and underestimation scenarios compared to ground truth, setting error upper and lower bounds to penalize data points exceeding error tolerance limits. The Score aggregates the prediction scores of each data sample for each stage, yielding an overall score and providing a comprehensive evaluation of the model’s performance across flight phases. Results and Discussions This study conducts a series of experiments on a dataset of atmospheric inertial navigation data for fixed-wing aircraft to predict the inertial satellite composite altitude in the event of satellite signal loss, using inertial pressure altitude as the reference baseline. Several comparative experiments are designed to comprehensively evaluate the LTCA–TCN altitude prediction algorithm proposed herein. Three subsets of data corresponding to different flight phases are utilized for altitude prediction, and the performance is compared to commonly used convolutional neural networks (CNN), recurrent neural networks (RNN), long short-term memory networks (LSTM), gated recurrent units (GRU), TCN, and the LTCA–TCN algorithm proposed in this study. Experimental results showed that the LTCA–TCN algorithm outperforms other comparative algorithms in root mean square error (RMSE) and average Score metrics. It achieves the best RMSE and Score across all three phases. Compared to RNNs and their variants, the LTCA–TCN algorithm yields superior prediction results while maintaining a simpler structure and requiring fewer computational resources. In addition, compared to the baseline TCN algorithm, the proposed LTCA–TCN algorithm reduces RMSE by 1.43 m and lowers Score by 0.08. Particularly in the cruise phase, the RMSE reaches 7.97 m, within a 10 m range, demonstrating high prediction accuracy. Therefore, the LTCA–TCN algorithm indicates significant advantages in multi-stage altitude prediction tasks. Specific GPS satellite signal loss scenarios are simulated to evaluate the LTCA–TCN algorithm’s performance. Corresponding periods of satellite signal loss are set for the ascent, cruise, and descent phases to predict the Inertial Satellite Composite Altitude during these periods. Experimental results indicated that the LTCA–TCN model exhibits good fitting capability, effectively capturing different change modalities with high flexibility and adaptability in each phase. A relative error ratio is calculated to measure the extent to which the predicted altitude is closer to the actual value of the inertial satellite composite altitude compared to the inertial pressure altitude provided by INS. The relative error ratios of the LTCA–TCN model for all three phases are within 1, indicating that the model’s predictions are generally more accurate than the inertial pressure altitude.Conclusions The results indicated that the LTCA–TCN model achieves high prediction accuracy across multiple stages, outperforming commonly used neural network algorithms in the field and providing optimal predictive performance. This model can offer more reliable multi-stage altitude indications for aircraft in the event of satellite signal loss.
