Abstract
An accurate fault detection method is critical in preventing the integrity of integrated navigation system from the abnormal measurements which may occur any time. Here, a genetic algorithm optimization for deep belief network fault detection method is proposed, where the system measuring residuals sequence is used as the input, and the output is the system operating state, such as normal or fault types, in pointwise. The proposed technique extracts the features with various scales, which contain both the local and the general information of the signal sequence, for making a comprehensive and precise classification. To show the validity of the proposed method, simulations based on INS/GNSS integrated navigation system are carried out. The simulation results show that the proposed fault detection algorithm and method is superior to the existing algorithms on the faults detection rate and false alarm rate, and thus, system reliability and navigation precision have been greatly improved.
Introduction
Modern navigation systems, particularly those used in autonomous vehicles and aerospace, rely heavily on integrated navigation architectures. The most common configuration combines Inertial Navigation Systems (INS) with Global Navigation Satellite Systems (GNSS). While this integration offers superior accuracy by leveraging the strengths of both sensors, it also introduces significant vulnerabilities. A single sensor fault—whether a sudden outlier or a subtle drift—can propagate through the data fusion process, leading to critical errors in positioning and velocity.
Traditional fault detection methods, such as the chi-square test, often struggle with the non-linear and time-varying dynamics of real-world movement. Conversely, modern data-driven approaches like standard neural networks often require manual hyperparameter tuning, which is prone to human error and sub-optimal performance. This paper presents GA-DBN, a self-optimizing framework that utilizes Genetic Algorithms to refine Deep Belief Networks for high-precision fault detection.
The GA-DBN Framework
The proposed methodology integrates three core components: a Federal Kalman Filter (FKF), a Deep Belief Network (DBN), and a Genetic Algorithm (GA) for optimization.
Federal Kalman Filter and Residual Generation
The system employs a non-reset FKF for multi-sensor data fusion. The primary output used for fault detection is the "residual" sequence—the mathematical difference between observed sensor data and the model's predictions. These residuals serve as the diagnostic signal; any deviation from expected statistical behavior indicates a potential sensor fault.
Deep Belief Network (DBN) Architecture
The DBN is constructed by stacking multiple Restricted Boltzmann Machines (RBMs). This architecture is particularly adept at unsupervised feature extraction. It can automatically learn the complex "signatures" of various sensor faults directly from the raw residual data, eliminating the need for manual feature engineering.
Genetic Algorithm (GA) Optimization
The performance of a DBN is highly sensitive to its architecture (e.g., number of hidden nodes) and learning parameters (e.g., learning rate, momentum). To automate the selection of these "hyperparameters," the researchers implemented a Genetic Algorithm. The GA treats these parameters as "chromosomes" and evolves a population of networks over successive generations. By selecting the "fittest" individuals—those with the highest detection accuracy—the system converges on an optimal configuration without human intervention.
Experimental Setup and Datasets
To validate the GA-DBN approach, a comprehensive simulation environment was established using high-fidelity MEMS IMU and GNSS specifications. The training set comprised 4,700 data sequences, with 300 additional sequences reserved for validation and testing across four distinct fault types: sudden outliers, positioning error surges, gradual positioning errors (often caused by multipath or jamming), and sensor drifts related to hardware degradation.
Results and Discussion
The GA-DBN was benchmarked against a standard DBN and a Convolutional Neural Network (CNN). The results demonstrate a clear advantage for the optimized approach:
- Detection Accuracy: GA-DBN achieved a success detection rate of 92.6%, significantly higher than the standard DBN (74.5%) and the CNN (81.6%).
- False Alarm Rate: The system maintained an exceptionally low false alarm rate, reaching 0% for gradual faults, whereas the CNN exhibited a 2.7% false alarm rate for outliers.
- Computational Efficiency: Despite the added complexity of the GA optimization phase, the training time (0.76s) remained competitive and was nearly twice as fast as the CNN (1.32s).
A key insight from the results is the system's ability to handle gradual faults. Traditional methods often miss slow sensor drifts because the residual error grows very slowly. The GA-DBN’s multi-scale feature extraction proved robust enough to identify these subtle patterns early.
Conclusion
The GA-DBN framework represents a significant step forward in ensuring the integrity of integrated navigation systems. By automating the design of the neural network through Genetic Algorithms, the researchers have created a system that is both more accurate and more robust than traditional or manually-tuned models. Future work will focus on testing the system with real-world sensor data and expanding the framework to include additional sensors like LiDAR and visual odometry.
Summary updated