Abstract: Prognostics and Health Management (PHM) systems are key technologies for improving system performance and availability while reducing maintenance manpower and support costs. A typical PHM system mainly includes data acquisition, data analysis, fault diagnosis, health assessment, life prediction, and maintenance decision-making. Proper data processing directly impacts fault diagnosis, health assessment, and life prediction; therefore, selecting the correct algorithm is crucial for the entire system. This paper, aiming to improve the accuracy and reliability of fault diagnosis, systematically studies various methods for detecting and diagnosing aircraft faults and discusses the advantages and disadvantages of information fusion technology, wavelet transform, and Bayesian networks.
Keywords: PHM; information fusion; wavelet transform; Bayesian network
1 Introduction
With the increasing complexity and integration of aviation systems, and considering the reliability, safety, and economy of aircraft systems, the role of Predictive Health Management (PHM) technology, with detection technology at its core, is indispensable and represents one of the trends in aviation technology development. The development of PHM technology is driven by system-level integrated applications, integrating functions such as condition monitoring, fault diagnosis, prediction, health management, and maintenance. It enables the correct maintenance activities to be carried out on the correct parts at the correct time, allowing people to determine when to maintain the system based on its current health status before it completely fails, thereby achieving self-service support and reducing usage and maintenance costs. In other words, the research on data processing algorithms is the core of fault detection technology and determines the accuracy of the detection system.
2. Domestic and international development status
Aviation fault diagnosis technology originated in the late 1960s. Since the 1990s, with the deepening research and application of artificial intelligence and data fusion technologies, foreign countries have been developing towards integrated information systems based on data fusion and online diagnostic functions. Fault diagnosis algorithms can be divided into two main categories: mathematical model-based diagnostic algorithms and classification diagnostic methods. The key to applying mathematical model diagnostic methods is establishing relatively accurate mathematical models and fault equations and their solutions. In the past two decades, the US has conducted extensive research on the application of data fusion technology, employing data fusion algorithms that combine artificial neural networks with DS evidence theory. It has also applied signal feature spectrum extraction technology and multiple artificial intelligence technologies combining fuzzy logic and neural networks. Domestic research on multi-sensor data fusion started relatively late, only gradually gaining attention at the end of the 20th century. Most research is based on various types of artificial neural networks for fault diagnosis algorithms. Currently, domestic research on multi-sensor data fusion technology is still in its initial stage compared to foreign fault diagnosis technologies.
3. Basic Principles and Classification of Fault Diagnosis
For example, an aircraft engine generally consists of major components such as a compressor, combustion chamber, and turbine. Each component is composed of several parts and components. From a systems theory perspective, if we consider the engine as a system, then any event or state in which the system or a part of the system cannot or will not perform its intended function is collectively referred to as a fault. If we consider the actual causal process of engine performance degradation (result) caused by engine fault (cause) as the forward process, then the fault diagnosis process is the reverse process of finding the cause of the fault (determining the location of the fault) by using changes in measurable parameters (performance degradation).
Fault equation modeling can generally be categorized into two types: classification-based diagnostic methods and fault equation modeling. Classification-based diagnostic methods, which are not based on mathematical models, are grounded in pattern recognition theory, and the possible fault states of the system being diagnosed are limited. Common classification methods include pattern matching, discriminant function classification, probability and statistics-based classification, neural network classification, and rule-based reasoning classification. Fault equation modeling, on the other hand, requires the establishment of complex mathematical models. The accuracy of the fault model model is highly dependent on the accuracy of the fault diagnosis, and external interference has a relatively large impact on the solution results.
4. Research on Algorithms
3.1 Data Fusion Technology Based on Neural Networks
Due to the complexity of aircraft structures, the severe nonlinearity of models, and the interference of measurement integration errors on fault detection, aircraft fault diagnosis is complex and difficult. Based on these factors, information fusion technology is employed to diagnose faults. For example, voltage, temperature, and energy data from measurement points are appropriately fused through processes such as weighting/voting, Bayesian inference, and Kalman filtering to improve the confidence level of the data sources. A schematic diagram of information fusion technology is shown in Figure 1.
Figure 1. Schematic diagram of information fusion technology
Research on information fusion fault diagnosis methods based on DS evidence theory can effectively improve the accuracy of fusion diagnosis. Applying the fundamental theories of various improved BP networks, radial basis function networks, probabilistic neural networks, SOFM self-organizing feature map networks, and Elman regression neural networks, and conducting research under different normalization methods and noise interference conditions, it is concluded that the equal variance standard normalization method is optimal, and the probabilistic neural network has the strongest anti-interference ability and the highest fault diagnosis accuracy.
3.2 Wavelet Transform
For fault signals containing a large number of time-varying and singular components, traditional analysis methods such as the Fourier transform, which are suitable for stationary signal analysis, will produce significant errors, hindering the extraction of fault signal features. Wavelet transform, as an evolution and extension of the Fourier transform method, possesses excellent time-frequency characteristics. The asymptotic signal instantaneous frequency extraction algorithm based on wavelet ridges can effectively extract fault features from signals whose frequency changes continuously with time. Sometimes, fault signals and normal signals exhibit very similar behavior in the time-frequency domain, making accurate extraction impossible with traditional methods. However, the load fluctuation judgment method based on wavelet transform can clearly distinguish fault signals from normal signals, thereby improving the reliability of fault diagnosis. For high-frequency noise in testing, a floating threshold algorithm based on wavelet packet transform is proposed, which can effectively remove noise while retaining the useful high-frequency components of the signal. For example, the processing of the starting armature current signal using wavelet packets is shown in the following figure:
Figure 2. Amplified waveform of the original signal.
Figure 3. Signal amplification after Fourier transform.
Figure 4. Signal amplification after floating threshold filtering.
As can be seen from Figures 3 and 4, the floating threshold based on wavelet packet transform can more effectively remove noise similar to the original signal, which is extremely important for signal processing.
3.3 Bayesian Networks
Bayesian networks (BNs) are relatively easy to model many recognizable behaviors in the real world. However, their most important strength lies in inference computation regarding behavior, interpretation, and parameter selection. BNs have a solid foundation in probabilistic reasoning, and their conditional independence effectively expresses the correlations between equipment faults. BNs can quickly calculate the probability of fault causes using some fault symptoms, requiring less existing information and capable of inference even with incomplete or uncertain information. BNs have strong learning capabilities; in practice, they can relearn based on new samples to improve their fault diagnosis capabilities and accuracy. In practice, unpredictable external factors may cause data loss or incomplete data acquisition at some measurement points. Literature shows that BNs trained with incomplete data can still maintain a good fault diagnosis rate (80%), and can identify faulty modules even with some missing measurement point data. This verifies that BNs can effectively handle uncertain and incomplete information and draw correct conclusions, a superiority unmatched by other methods. If the network is not ideal, a reasonable network structure can be obtained through structure learning, which has the characteristic of timely modification.
4. Conclusion
This paper addresses the complexity of aircraft data by employing various data processing algorithms. The DS evidence theory-based information fusion fault diagnosis method utilizes multi-sensor information fusion and processing techniques to select and extract various fault features. It studies different theoretical methods for extracting, effectively identifying, and isolating fault features from different fault data. The wavelet packet transform-based floating threshold algorithm effectively handles high-frequency noise; Bayesian networks possess strong learning capabilities for new data samples, effectively improving fault diagnosis capabilities and accuracy.
