Bio-Inspired Fault Diagnosis for Aircraft Fuel Pumps Using a Cloud-Edge System Authors: Yang Miao, Yantang Li, Jun Pan, Zhen Liu, Lei Liu, Zeng Wang, Zijing Wang Published: 2023 ABSTRACT The fuel pump serves as the central component of the aircraft fuel system, necessitating real-time data acquisition for monitoring purposes. As the number of sensors increases, there is a substantial rise in data volume, leading to a simultaneous increase in computational processing for traditional Prognostics and Health Management methods while computational efficiency decreases. In response to this challenge, a novel health monitoring approach for aircraft fuel pumps is proposed based on the collaborative utilization of cloud-edge resources. This approach enables efficient cooperation among the sensor side, edge side, and cloud side to achieve timely fault warnings and accurate fault classification for fuel pumps. Within this method, anomaly judgment tasks are allocated to the edge side, and an anomaly judgment method that integrates the 3σ threshold and “3/5 strategy” is devised. Additionally, a fault diagnosis algorithm, founded on a convolutional auto-encoder, is formulated in the cloud to discern various fault types and severities. Comparative results demonstrate that, in contrast to long short-term memory networks, convolutional neural networks, extreme learning machines, and support vector machines, the proposed method yields improvements in accuracy of 4.35%, 6.40%, 17.65%, and 19.35%, respectively. Consequently, it is evident that the proposed method exhibits notable efficacy in the condition monitoring of aircraft fuel pumps. 1. Introduction The fuel pump assumes a crucial role within the fuel system of aircraft, serving as a fundamental component responsible for delivering the requisite pressure and fuel flow to the engine. Monitoring the operational integrity of the fuel pump is of paramount importance, as it significantly contributes to upholding the stability of the fuel system. The existing methods like simulation analysis [1,2], fault tree analysis [3], and signal pro-cessing [4,5] can be used to analyze the fault mechanism of the fuel pump and its rela-tionship. The emergence of Prognostics and Health Management (PHM) has been inextricably linked with the advent of information technology [6,7,8,9]. It represents a novel approach to health state management that draws on the latest research findings of contemporary information and artificial intelligence technologies, with particular emphasis on equipment health state monitoring, prediction, and management. PHM has served as a catalyst for a pivotal transition in equipment diagnosis, from “post-maintenance” and “scheduled maintenance” to “situational maintenance”. Currently, PHM has found diverse applications in the diagnosis of different devices [10,11,12,13,14]. Especially in the field of aircraft, the application of PHM is much more than common. Che et al. [15] present a PHM model that combines multiple deep learning algorithms for condition assessment, fault classification, sensor prediction, and remaining useful life (RUL) estimation of aircraft systems. Hsu et al. [16] proposed a case study on the implementation of a health assessment and prediction workflow for RUL based on the PHM framework of currently in-service aircraft, which could significantly benefit fleet operators and aircraft maintenance. Yan et al. [17] combined the existing maintenance management experience and the output result of PHM to make more reliable verification for the maintenance of aircraft air-condition systems. The fundamental approach of PHM involves the utilization of sensors dispersed throughout the equipment to gather operational data. This data are then subjected to analysis through relevant algorithms, aiming to extract valuable information that reflects the equipment’s status, identifies fault modes, and evaluates the remaining lifespan of its components. While most of the health monitoring systems mentioned in the aforementioned literature rely on centralized processing of sensor data, the complexity of equipment structures and the increase in key measurement points have resulted in exponential growth in operational data volume. Consequently, traditional PHM systems experience significant delays in centralized data processing. The introduction of edge computing [18] effectively addresses this issue. Cloud-edge collaboration has emerged as a crucial research area within the domain of device health management [19,20]. It entails the use of edge computing to share a portion of the computational load with the cloud. This approach not only mitigates the lengthy network transmission delays associated with uploading computational tasks to the cloud but also enables real-time feedback of device abnormalities to technicians. Erbao [21] et al. have successfully addressed the issues of low efficiency and high energy consumption inherent in traditional online monitoring methods for high-voltage equipment in substations by constructing an integrated cloud-edge system. Qiu et al. [22] have analyzed the relationship between various forms of edge computing and industrial edge computing systems while exploring key technologies within the industrial edge computing framework. Li et al. [23] have designed an intelligent fault diagnosis system based on cloud-edge collaboration, employing a fault diagnosis algorithm that combines short-duration memory networks and multi-scale convolutional networks. This system has been applied to machine tool detection, meeting accuracy requirements while reducing diagnosis time. However, in the aforementioned methods, the monitoring targets primarily consist of individual types of information, such as electrical signals and vibration signals, whereas the monitoring of fuel pumps often involves multiple signal types, like vibration signals and pressure signals. Researchers have also proposed advanced fault diagnosis methods for fuel pumps. For instance, Verhulst et al. [24] developed a fault detection and diagnosis (FDD) scheme to isolate damaged injectors in internal combustion engines. Additionally, Pecho et al. [25] explored the application of a novel technique, Recurrence Quantification Analysis (RQA), for monitoring non-linear transient accelerometer data, specifically for use in auxiliary aircraft hardware health monitoring. A method of intermittent fault diagnosis based on support vector machines (SVM) is presented by Jiang et al. [26] to quickly and effectively diagnose intermittent faults in aircraft fuel pumps. Miao et al. [27,28] addressed the challenges of limited fault data and low diagnostic accuracy of certain new aircraft fuel pumps by deploying the transfer learning method. This approach enabled feature extraction from similar fault data and established an auxiliary dataset to bolster fault diagnosis accuracy, despite the dearth of available data. This study proposes a cloud-edge collaborative monitoring method specifically for aircraft fuel pumps, incorporating multiple types of information to ensure real-time monitoring effectiveness and accuracy. The main work is as follows:Divide the different tasks on the sensor side, the edge side, and the cloud side to achieve collaborative monitoring;The anomaly judgment method and fault diagnosis algorithm are designed for edge-side state detection and cloud fault classification;The feasibility and effectiveness of the proposed method are verified by fuel pump simulation experiments. Divide the different tasks on the sensor side, the edge side, and the cloud side to achieve collaborative monitoring; Divide the different tasks on the sensor side, the edge side, and the cloud side to achieve collaborative monitoring; The anomaly judgment method and fault diagnosis algorithm are designed for edge-side state detection and cloud fault classification; The anomaly judgment method and fault diagnosis algorithm are designed for edge-side state detection and cloud fault classification; The feasibility and effectiveness of the proposed method are verified by fuel pump simulation experiments. The feasibility and effectiveness of the proposed method are verified by fuel pump simulation experiments. 2. Fault Diagnosis Framework for Fuel Pumps Based on Cloud Edge Collaboration 2.1. General Framework A fault diagnosis method for aircraft fuel pumps based on cloud-edge collaboration was established. The methodology comprises three principal components: the cloud side, the edge side, and the sensor side. Each component assumes distinct responsibilities and engages in collaborative relationships, as elucidated in Figure 1. Sensor side: At critical measurement locations of the fuel pump, multiple sensors (e.g., vibration, pressure, etc.) are strategically positioned to capture real-time operating data. These sensors facilitate the collection of comprehensive information regarding the fuel pump’s performance. The acquired real-time data are subsequently transmitted via wired connections to the edge side for further processing and analysis. Sensor side: At critical measurement locations of the fuel pump, multiple sensors (e.g., vibration, pressure, etc.) are strategically positioned to capture real-time operating data. These sensors facilitate the collection of comprehensive information regarding the fuel pump’s performance. The acquired real-time data are subsequently transmitted via wired connections to the edge side for further processing and analysis. Edge side: Upon receiving the uploaded data from each sensor, the location information of the corresponding measurement points is recorded, and the characteristic values are extracted simultaneously. Subsequently, the threshold method is employed to identify anomalies within the dataset. Segments of data that meet the abnormal condition are wirelessly transmitted to the cloud for further analysis. In the ground and airborne visual interfaces, both the abnormal alarm information and the respective positions of the measurement points are presented, providing a comprehensive overview of the detected anomalies. Edge side: Upon receiving the uploaded data from each sensor, the location information of the corresponding measurement points is recorded, and the characteristic values are extracted simultaneously. Subsequently, the threshold method is employed to identify anomalies within the dataset. Segments of data that meet the abnormal condition are wirelessly transmitted to the cloud for further analysis. In the ground and airborne visual interfaces, both the abnormal alarm information and the respective positions of the measurement points are presented, providing a comprehensive overview of the detected anomalies. Cloud side: The fault segment information transmitted by each edge is stored in the historical database, which facilitates subsequent fault history tracing. An intelligent classification model is employed to identify the types of fault data detected by the system. Additionally, uploading more fault data are used to update and train the model parameters, thus enhancing its migration ability. Ultimately, the classification and diagnosis results are presented through the ground and airborne visual interfaces. Cloud side: The fault segment information transmitted by each edge is stored in the historical database, which facilitates subsequent fault history tracing. An intelligent classification model is employed to identify the types of fault data detected by the system. Additionally, uploading more fault data are used to update and train the model parameters, thus enhancing its migration ability. Ultimately, the classification and diagnosis results are presented through the ground and airborne visual interfaces. 2.2. Data Acquisition on Sensor Side In order to develop a specialized cloud-edge collaborative fault diagnosis system for fuel pumps, this study establishes a ground-based experimental setup that closely emulates the operational characteristics of an authentic aircraft fuel system. The schematic representation of this apparatus is presented in Figure 2. Given the imperative need to capture vibration and pressure data, a pressure sensor is installed at the oil outlet. Owing to the constrained space within the fuel pump, the direct installation of a vibration sensor poses a challenge. As a result, the vibration sensor is positioned on the motor connected to the fuel pump. The arrangement of the vibration sensor is depicted in Figure 3. 2.3. Anomaly Detection on Edge Side 2.3.1. Detection Index The sensor side is responsible for collecting a range of signals, which are subsequently transmitted to the edge side for preliminary data analysis. On the edge side, an initial analysis is conducted, utilizing the vibration and pressure signals obtained from the sensor side as the primary basis for establishing distinct detection indices. The direct analysis of the original vibration signals presents certain drawbacks that primarily stem from the large volume of signals collected within a given time frame. These concerns include an increased computational burden at the edge side due to the sheer quantity of original signals, as well as delayed upload speed and subsequent high latency in real-time cloud monitoring when transmitting fault data for pattern recognition at the cloud level, owing to the increased data volume. In order to address these concerns, time-domain indicators are adopted as the fundamental criteria for anomaly detection, allowing for the upload of sample data to the cloud. A comprehensive analysis evaluating the sensitivity and stability of time-domain features is presented in Table 1. Based on the information presented in the aforementioned table, it can be observed that the kurtosis factor and margin factor exhibit high sensitivity and are particularly responsive to the occurrence of faults. On the other hand, the root-mean-square demonstrates good stability and effectively assesses changes in vibration signals. The comprehensive performance of the peak and peak-to-peak indices is superior. Taking into account the overall computational complexity, five key indices, namely peak, peak-to-peak, root mean square, kurtosis factor, and margin factor, are selected as the primary characteristics for determining abnormal conditions in fuel pump assessment. The peak and peak-to-peak indices capture the maximum amplitude of the vibration wave pattern, making them well-suited for detecting surface stripping faults. These types of faults generate signals with significant values within a short time period, which can be mathematically described by Equations (1)–(3):(1)Xmax=max(xi) (2)Xmin=min(xi) (3)Xpeak=12(Xmax−Xmin) where xi is the signal for a period of time, and Xpeak is the peak value. The root-mean-square reflects the energy distribution of the signal in the time domain, expressed by Equation (4):(4)Xrms=1n∑i=1nxi2 where Xrms is the root-mean-square value. The kurtosis factor is sensitive to shock characteristics in the signal, as expressed by Equations (5) and (6):(5)Xk=∑i=1nxi4n (6)Xkf=XkXrms4 where Xk is kurtosis and Xkf is the kurtosis factor. The margin factor is more suitable for early impact faults, and its expression is as in Equations (7) and (8):(7)Xr=∑i=1nxi12n2 (8)Xcl=XpeakXr where Xr is the margin factor and Xcl is the root amplitude. In the context of anomaly detection, pressure signals exhibit a higher degree of discernibility compared to vibration signals. During normal operation of the fuel pump, pressure information tends to fluctuate around the expected normal value. However, in the presence of anomalies, such as instances of low or high pressure, the pressure information deviates significantly, reaching excessively high levels or persisting at abnormal values for a prolonged duration. Consequently, the analysis of pressure signals primarily revolves around assessing the amplitude and duration of such deviations. It has been observed that pressure signals exhibit a higher degree of discernibility compared to vibration signals. In normal operational scenarios, pressure information tends to fluctuate around the expected normal value. However, in the presence of anomalous events, such as instances of low or high pressure, the pressure information deviates significantly from the norm, reaching excessively high levels or persisting at abnormal values for prolonged periods. Therefore, the analysis of pressure signals primarily revolves around assessing the amplitude and duration of such deviations. 2.3.2. Threshold Detection For the abnormal judgment of a vibration signal, the 3σ threshold judgment method is used, which can be expressed as Equation (9):(9)f(x)=1σ2πexp[−12(x−μσ)2], −∞