An aircraft fault detection method based on random projection distance prediction

Abstract

The application discloses an airplane fault detection method based on random projection distance prediction and relates to the field of airplane intelligent fault detection.The method comprises the following steps: S1, collecting air flow, air pressure and temperature data through a pressure regulating stop valve sensor system installed on an airplane; S2, pre-processing the data collected in S1; S3, constructing an unsupervised anomaly detection model; S4, training and optimizing the unsupervised anomaly detection model; and S5, testing and optimizing the unsupervised anomaly detection model.The abnormal state of a pressure regulating stop valve (PRSOV) can be efficiently and accurately detected in an aviation health monitoring system (HUMS), so that the safety of airplane operation is improved, the maintenance cost is reduced, and the unplanned shutdown events are reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to intelligent fault detection of an aircraft, and in particular to an aircraft fault detection method based on random projection distance prediction. BACKGROUND

[0002] Air transportation is the most convenient mode of transportation in modern society, not only greatly promoting global economic activities, but also strengthening the connection between people. However, during flight, the aircraft may experience severe environmental changes, such as fluctuations in speed, temperature, humidity, and air pressure. These changes will have a negative impact on on-board equipment, especially the performance of core components may decline, or even abnormal, bringing serious challenges to the operation and maintenance of the aircraft. Among them, the pressure regulating shutoff valve (PRSOV, Pressure Regulated Shutoff Valve) is a key component of the aircraft environmental control system (ECS, Environmental Control System), which is mainly used to regulate gas flow and ensure that the cabin is provided with appropriate hot air supply under different flight stages and external environmental conditions. If the PRSOV fails, it may cause cabin pressure loss and temperature abnormalities, and in serious cases may cause aviation safety accidents. Therefore, a reliable and accurate prognostics and health management (PHM, Prognostics and Health Management) system is crucial for reducing unplanned maintenance, improving flight safety, and reducing operational risks. The current aircraft is equipped with a large number of sensors (possibly hundreds or even thousands), which are used to monitor different operating states, and these sensors generate a large amount of data. In order to extract effective information from these data and conduct health status monitoring, data-driven algorithms have become a key technology and are widely used in health monitoring and fault diagnosis tasks of key aviation components.

[0003] In early aircraft health monitoring research, researchers mainly relied on experience to identify and assess abnormal states by setting thresholds or logical rules. However, this method is limited in efficiency when facing complex systems and large-scale data, and it is difficult to accurately judge abnormal states. In order to deal with the increasing complexity and data size, feature engineering has become a key step in algorithm development. Researchers have proposed various feature extraction algorithms to extract representative fault features from high-dimensional data using domain knowledge and data characteristics. For example, Fourier transform and wavelet transform are widely used in signal processing to extract frequency features. However, these methods still have limitations, mainly in two aspects: on the one hand, manual feature engineering is heavily dependent on expert experience, which is prone to bias and cannot fully capture the dynamic characteristics of aircraft systems; on the other hand, the nonlinear modeling ability of traditional machine learning algorithms is limited, and they perform poorly when faced with high-dimensional and large-scale data, making it difficult to adapt to the complex operating environment of aircraft.

[0004] In recent years, deep learning technology has become an effective means of aircraft fault detection and diagnosis as a powerful data-driven method due to its excellent feature automatic learning ability. In particular, PRSOV as a key component has attracted more and more researchers' attention. However, the success of deep learning in this field mainly depends on a large amount of labeled abnormal data, but in actual operation, due to the rarity and diversity of aircraft failure events, it is difficult to collect large-scale labeled data covering all types of failures. In addition, accurately labeling all types of fault data is both time-consuming and expensive. Therefore, unsupervised anomaly detection (UAD, Unsupervised Anomaly Detection) has received widespread attention. The core idea of this method is to train only based on normal data, and automatically extract representative features through self-supervised proxy tasks. UAD methods learn the distribution characteristics and potential structure of normal data without relying on abnormal data, so they have great application potential in aviation application scenarios. However, our analysis found that in the PRSOV field, unsupervised anomaly detection still faces three key challenges.

[0005] Firstly, although the UAD method has been successful in multiple industrial applications, the current research on unsupervised anomaly detection for PRSOV is still very limited. Most existing methods focus on general industrial equipment, and there is little research on specialized optimization and adaptability for PRSOV. Secondly, there are also great challenges in directly applying existing UAD methods to PRSOV. For example, the latest unsupervised anomaly detection model based on Transformer usually requires a large amount of computing resources, which is far beyond the computing power of typical avionics processors. Therefore, it is particularly crucial to develop an efficient and low-computing-cost UAD method. Thirdly, most UAD methods lack theoretical guarantees, and the aviation system has very high requirements for reliability, which makes the methods lacking theoretical support face greater uncertainty in practical applications. Therefore, how to balance the constraints of computing resources, theoretical interpretability and high detection accuracy under the framework of unsupervised anomaly detection has become a problem that needs to be solved in current research. SUMMARY

[0006] The present application provides a kind of based on random projection distance prediction aircraft fault detection method, realizes real-time, high-precision anomaly detection.

[0007] According to an aspect of the present disclosure, a method for aircraft fault detection based on random projection distance prediction is provided, the method comprising:

[0008] S1, collecting air flow, air pressure and temperature data through the pressure regulating stop valve sensor system installed on the aircraft;

[0009] S2, pre-processing the data collected in S1;

[0010] In order to construct an efficient unsupervised anomaly detection model, the present research proposes RPDP-AD (unsupervised PRSOV multi-fault anomaly detection algorithm based on random projection distance prediction), the core technology S3, constructing an unsupervised anomaly detection model, including:

[0011] Random projection dimension reduction: using a random projection matrix, project high-dimensional input data into a low-dimensional space, and try to preserve the relative distance relationship between data points as much as possible; according to the Johnson-Lindenstrauss theorem, ensure that the projected data still retains the distribution information of the original data;

[0012] Supervised signal construction: by calculating the distance between data points in the projection space, construct a supervised signal, so that the model can learn the pattern of normal data in an unsupervised environment;

[0013] Lightweight MLP multi-layer perceptron network: due to the limited computing resources of avionics equipment, use MLP network instead of computationally intensive models to achieve efficient inference;

[0014] Distribution difference measure: By minimizing the distribution difference between the learnable mapping and the random mapping, the projection of the model in the low-dimensional space is ensured to maintain stability and interpretability;

[0015] The final RPDP-AD model can learn the normal PRSOV operating mode without labeled data and identify abnormal signals deviating from the normal mode during detection.

[0016] S4, training and optimization of unsupervised anomaly detection model;

[0017] S5, test and optimize the unsupervised anomaly detection model.

[0018] The data mainly comes from the pressure regulating stop valve (PRSOV) sensor system installed on the aircraft, which is used to monitor the operating state of PRSOV in different flight stages. Data collection involves multiple sensor locations and measurement parameters, covering: air flow and pressure parameters: monitoring the air flow, air pressure and temperature inside PRSOV to identify abnormal flow fluctuations or air tightness problems. Fault type data: if available, annotate the sensor data corresponding to known fault modes (such as leakage, blockage, mechanical wear) for subsequent performance evaluation.

[0019] The collected data is in time series format, with high-precision clock timestamps to ensure the synchronization of multi-sensor data for subsequent data fusion and analysis.

[0020] In one possible implementation, the method further comprises:

[0021] S6, aircraft actual verification and deployment, including:

[0022] After laboratory testing and optimization, the model is finally deployed to the actual aircraft health monitoring system to realize real-time fault detection:

[0023] Onboard verification: during flight missions, real-time acquisition of pressure regulating stop valve sensor data and online detection by the model; compare the model detection results with the artificial inspection results to verify its accuracy and stability;

[0024] Embedded deployment: due to the limited computing resources of avionics, the model is optimized for computing efficiency and deployed to the onboard computing platform or cloud monitoring system to realize real-time analysis and remote monitoring;

[0025] Real-time monitoring and maintenance optimization: integrate the model into the aircraft health monitoring system to realize all-weather online monitoring and provide early warning to maintenance personnel, optimize maintenance plans and reduce unscheduled maintenance.

[0026] The application can provide efficient, accurate and low-computational-overhead anomaly detection in actual aircraft operating environment, significantly improve PRSOV operating safety, and reduce aviation operating costs.

[0027] In a possible implementation, the pressure regulating stop valve sensor system monitors the air flow, air pressure and temperature inside the pressure regulating stop valve sensor to identify abnormal flow fluctuations or air tightness problems.

[0028] The collected data are in time series format, with time stamps recorded by high-precision clocks to ensure the synchronization of multi-sensor data for subsequent data fusion and analysis.

[0029] Since the actual collected data usually contain noise, missing values and irregular changes, data preprocessing is needed to improve data quality and enhance model learning effect. The main preprocessing steps include:

[0030] In a possible implementation, S2, the collected data in S1 are preprocessed, including:

[0031] De-noising: low-pass filter is used to remove high-frequency noise to ensure the smoothness of vibration signals; Kalman filter is used to process sensor data to make it more stable and reduce the influence of environmental noise; wavelet transform is used to extract main features and reduce interference caused by sensor errors;

[0032] Normalization: Z-score standardization or Min-Max normalization is used to ensure that all data remain within the same numerical range to reduce the influence of scale differences between features;

[0033] Time series slicing: sliding window method is used to divide data into fixed time windows for input into the model for analysis;

[0034] Data augmentation: data augmentation techniques are applied, including time warping, adding noise and signal mixing, to improve the generalization ability of the model.

[0035] After the model is constructed, it needs to be trained and optimized to ensure that the model can effectively learn the distribution characteristics of normal data.

[0036] In a possible implementation, S4, the unsupervised anomaly detection model is trained and optimized, including:

[0037] Unsupervised training: the model is trained only using normal data to minimize the distance prediction error in the random projection space, so that the model can learn the structural features of the normal pressure regulating stop valve PRSOV operating state;

[0038] Loss function optimization:

[0039] By using the distribution difference measure loss, it is ensured that the low-dimensional data after projection can still maintain the spatial structure of the high-dimensional data.

[0040] Hyperparameter adjustment: grid search is performed to adjust the learning rate, projection dimension, MLP layer number and other hyperparameters to improve the stability and performance of the model.

[0041] Regularization: L2 regularization is used to prevent overfitting, so that the model can be generalized to different PRSOV working conditions.

[0042] After the model training is completed, it needs to be tested to evaluate its detection ability under different fault modes and operating environments.

[0043] In one possible implementation, S5, the unsupervised anomaly detection model is tested and optimized, including:

[0044] Offline testing: test on real PRSOV sensor data set, calculate AUC, accuracy, recall rate, F1 score and other key indicators to evaluate the detection performance of the model; through ablation experiment to verify the contribution of different modules to the performance of the model, optimize the overall architecture;

[0045] Noise sensitivity analysis: add different levels of noise to the test data to evaluate the robustness of the model in actual application;

[0046] Multi-fault mode detection: use data of 7 known fault modes to test whether the model can effectively distinguish different fault types; according to the test results, fine-tune the model to ensure its stability under various flight conditions. Use data of 7 known fault modes to test whether the model can effectively distinguish different fault types.

[0047] According to the test results, fine-tune the model to ensure its stability under various flight conditions.

[0048] Compared with the prior art, the beneficial effects of the present application are:

[0049] An aircraft fault detection method based on random projection distance prediction is provided by the embodiments of the present disclosure. In order to overcome the above challenges, the random projection theory provides an effective solution. Random projection is a dimensionality reduction technique that projects high-dimensional data into a low-dimensional space using a random projection matrix while preserving the distance relationships between data points as much as possible. In high-dimensional space, the relative positions and distances between data points reflect their similarities and differences, and the random projection method preserves the structure of the original data by preserving these distance relationships. In recent years, researchers have attempted to combine random projection with deep learning to improve the feature extraction capabilities of models. However, current research on industrial unsupervised anomaly detection is still in the theoretical exploration stage and has not yet formed a systematic method system.

[0050] To solve the applicability problem of UAD methods in the PRSOV domain and further explore the potential of random projection theory based on deep learning in industrial anomaly detection, this study proposes an unsupervised PRSOV multi-fault anomaly detection algorithm based on random projection distance prediction (RPDP-AD). RPDP-AD constructs a supervision signal by utilizing distance information in the random projection space, enabling the model to automatically encode the data structure and distribution pattern of normal samples in an unsupervised environment. In addition, we introduce a distribution difference measurement method to minimize the difference between the learnable mapping and the random mapping in the low-dimensional space. Unlike existing deep learning methods, our method maintains theoretical interpretability through a fixed random mapping while achieving adaptive feature learning through a trainable projection head. To our knowledge, this study is the first to combine random projection theory with deep neural networks for industrial unsupervised anomaly detection, particularly in the field of aviation component monitoring. In addition, to provide theoretical support, we analyze the inner product preservation property based on the Johnson-Lindenstrauss theorem, further proving the feasibility of combining random projection with deep learning.

[0051] Notably, unlike other computationally intensive models, RPDP-AD adopts a simple MLP (Multi-Layer Perceptron) network structure, thus achieving real-time and high-precision anomaly detection under limited computing resources, effectively addressing the application requirements in the aviation field. Experiments conducted on real PRSOV vibration data sets (containing 7 different types of faults) show that RPDP-AD outperforms existing mainstream UAD methods on multiple fault types. In addition, we also conducted a sensitivity analysis to verify the stability of RPDP-AD under different random projection matrices and noise conditions. This research not only provides a new technical solution for aviation health monitoring, but also further expands the application of random projection theory in deep learning, laying an important foundation for future unsupervised anomaly detection research.

[0052] The application can efficiently and accurately detect the abnormal state of the pressure regulating and stop valve (PRSOV) in the aviation health monitoring system (HUMS), thereby improving the safety of the aircraft operation, reducing the maintenance cost, and reducing the unscheduled downtime. In practical applications, the traditional anomaly detection method often relies on manual threshold setting or expert-driven feature engineering, which cannot adapt to complex and high-dimensional sensor data, while the existing deep learning method relies heavily on a large number of labeled abnormal data, which is difficult to meet the needs of the aviation industry. In contrast, the RPDP-AD (unsupervised PRSOV multi-fault anomaly detection algorithm based on random projection distance prediction) proposed in the application adopts a completely unsupervised learning method, which only relies on normal data for training and does not need to collect expensive and scarce fault samples, which greatly improves the deployability and scalability of the method. In addition, the method also performs well in the environment with limited computing resources. Unlike the computationally intensive Transformer architecture, RPDP-AD adopts a lightweight MLP network that can run on the embedded platform of the avionics equipment to realize real-time detection without affecting the execution of other critical tasks. Experimental results show that the method outperforms existing unsupervised anomaly detection algorithms on the real PRSOV sensor dataset, achieving higher detection accuracy in AUC, F1 score and other key indicators, and maintaining stable performance in 7 different fault modes. In addition, the sensitivity analysis further verifies the robustness of RPDP-AD under different random projection matrices and noise conditions, which means that the method can adapt to different models and different working conditions while maintaining high-precision detection capability. In practical applications, RPDP-AD can be seamlessly integrated into existing aviation health monitoring systems and monitor the operating state of PRSOV in real time, identify potential faults in a timely manner, provide warning information to maintenance personnel, optimize maintenance strategies, and avoid flight interruptions caused by sudden failures. At the same time, the method can be extended to the health monitoring of other key aviation components (such as hydraulic systems, electronic control systems, etc.), providing reliable technical support for intelligent maintenance of the wider aviation industry. Therefore, the application not only theoretically promotes the combination of random projection theory and deep learning in industrial anomaly detection, but also provides an efficient, low-cost, interpretable and easy-to-deploy solution for the health monitoring of aircraft in practical applications, greatly improving aviation safety and operational efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0053] Figure 1 A pressure regulating and stop valve schematic diagram of an embodiment of the present disclosure is shown.

[0054] Figure 2 Different categories of PRSOV signal examples of an embodiment of the present disclosure are shown.

[0055] Figure 3 Workflow of the RPDP-AD framework of an embodiment of the present disclosure is shown.

[0056] Figure 4 An aircraft fault detection algorithm based on random projection distance prediction of an embodiment of the present disclosure is shown. DETAILED DESCRIPTION

[0057] Various exemplary embodiments, features, and aspects of the present disclosure will be described below in detail with reference to the accompanying drawings. Like reference numerals denote like elements throughout the drawings. Although various aspects of the embodiments are illustrated in the drawings, the drawings are not necessarily drawn to scale unless specifically indicated.

[0058] The term "exemplary" is used herein to mean "serving as an example, instance, or illustration." Any implementation described herein as "exemplary" is not necessarily to be construed as preferred or advantageous over other implementations.

[0059] In addition, for the purpose of convenience and brevity, detailed descriptions of well-known functions and structures incorporated in the present disclosure can be omitted. It will be appreciated by those skilled in the art that the present disclosure can be practiced without these specific details. In some instances, well-known methods, procedures, components, and circuits have not been described in detail since they can hinder the understanding of the present disclosure.

[0060] Theoretical basis: We provide a theoretical basis analysis for the proposed RPDP-AD anomaly detection model. Specifically, we explore the inner product preserving property based on the Johnson-Lindenstrauss lemma, which links the random projection theory to deep learning neural networks.

[0061] Random projection is an efficient dimensionality reduction technique. Given a high-dimensional data matrix where n is the number of data points and d is the original dimension of the data. Random projection utilizes a random matrix to project the data into a k-dimensional low-dimensional space (where ) to achieve dimensionality reduction. The Johnson-Lindenstrauss lemma provides a theoretical basis for this algorithm, which states that by using a random projection matrix, data in a high-dimensional space can be projected into a low-dimensional space while preserving the distance relationships between data points to some extent, which has also been mathematically rigorously proven. The theorem is expressed as follows:

[0062] Theorem 1. (Johnson-Lindenstrauss) Let . Let be a set containing n data points, . Then there exists a Lipshcitz mapping such that for all :

[0063] (1)

[0064] Theorem 2 then explicitly shows the norm preserving property of the random projection, which is stated as follows.

[0065] Theorem 2. (Norm preserving property) Let . Assume that all the elements in are independently sampled from N (0, 1), then

[0066] (2)

[0067] Theorem 2 requires that the random projection matrix A is independently sampled from a Gaussian distribution. However, Lemma 1 shows that it is not strictly necessary to use a Gaussian distribution. Many other distributions with unit variance and satisfying certain boundedness conditions (or high-order moment conditions) are also sufficient.

[0068] Lemma 1. Assume that for ， each element is uniformly distributed. Then for any vector

[0069] (3)

[0070] (4)

[0071] Based on the above theorem and lemma, a simple corollary can be obtained, which is that the inner product has a preserving property under random projection.

[0072] Corollary 1. Let and . Let where A is a matrix, each element in the matrix is independently sampled from a Gaussian distribution N(0, 1) (or from U(-1, 1)). Then

[0073] (5)

[0074] The above theory confirms that random projection can preserve the structure, features, and inner product of the original data. This further provides a solid theoretical basis for the RPDP-AD anomaly detection method proposed by us.

[0075] II. Random projection distance estimation

[0076] ​From Lemma 1, the inner product in the random projection space preserves the latent class results and feature information of the original data. To this end, we train a learnable model to predict the inner product in the random projection space to force it to represent the latent information in the data. This can provide a strong and effective supervision signal for the training of the model without any manual labels, thus enabling feature representation and training for large-scale data without labels. The proposed RPDP-AD model demonstrates as shown in Figure 3 . It is assumed that the PRSOV dataset is obtained from a real scene or an aircraft simulation model , , N is the total number of samples. A pair of samples is randomly selected from the dataset D , and then they are respectively sent into a weight-shared twin network architecture. The network is represented as , which maps the input samples from a d-dimensional space to a new low-dimensional space of k . is a learnable network parameter. Similarly, we also need to establish a random projection function , which also randomly projects the input samples from a d-dimensional space to a new low-dimensional space of k . To this end, RPDP-AD learns the implicit structural information in the original data space by establishing an inner product difference loss function, which is represented as follows:

[0077] (6)

[0078] The distance information (inner product) between data points contains important class structure information, for example, in the feature space, the relative position and distance between data points reflect their similarity and difference. This is crucial for encoding the distribution boundary between different classes. In other words, random projection preserves this distance information between data points in the dimension reduction process. For the anomaly detection task, the model only uses normal samples and aims to establish the distribution boundary of normal samples. RPDP-AD forces the model to represent the distance and similarity in the data space by predicting the inner product in the random space, thus accurately encoding the distribution of normal data patterns. Thus, the model effectively learns the internal structure of the normal data, making it easier to detect abnormal data that deviates from the learned distribution.

[0079] III. Distribution difference index

[0080] As shown in Figure 3 , the proposed RPDP-AD realizes the encoding of the class results and sample distance relationship by predicting the inner product in the random space. This section further introduces a distribution difference index to enhance the model's recognition performance for abnormal classes, which is represented as follows:

[0081] (7)

[0082] It can be seen The goal is to minimize the distribution difference in the two mapping spaces (the mapping space of the neural network and the mapping space of the random projection) so that the feature mapping of the neural network also has the characteristics of class structure and distance preservation. In addition, after the normal samples undergo the above optimization training, the model will obtain a lower distribution difference index. When a new abnormal sample is input, the feature structure of the normal sample encoded by the model will not be applicable, There will be a substantial increase. This can be a very good supplement to the anomaly monitoring signal.

[0083] Four, total loss of the model

[0084] The multi-fault anomaly monitoring model proposed in this study is based on the Johnson-Lindenstrauss lemma as the theoretical basis, and realizes the distribution encoding of the normal mode through the inner product prediction of the random projection space. The proposed RPDP-AD adopts the inner product difference and distribution difference loss function as the supervision signal to optimize the parameters of the model, which is expressed as:

[0085] (8)

[0086] where and are the weight hyperparameters of the two losses, respectively. Since the obtained PRSOV data samples are 201 pressure features, the weight-sharing twin network in RPDP-AD adopts a multi-layer perceptron to avoid overfitting. The Adam optimizer is used with a learning rate of 0.001 and a batch size of 64. RPDP-AD is implemented using Pytorch and Python 3.8, and is trained and tested on a workstation with a 3090 GPU.

[0087] Case data: The aircraft environmental control system (ECS) is a key subsystem of the aircraft, responsible for regulating and maintaining the temperature, humidity, and airflow inside the cabin to ensure passenger comfort and flight safety. ECS mainly includes cabin air supply and distribution, pressure control, temperature regulation, and humidity management functions, and its main air source is the pressurized bleed air extracted from the aircraft engine compressor part. The failure of the ECS system may cause cabin pressure loss or temperature instability, which seriously affects flight safety. In the ECS, the pressure regulating stop valve (PRSOV) is one of the key components, responsible for regulating and distributing high-temperature bleed air to different subsystems. Since the PRSOV needs to operate in a high-temperature and high-pressure environment, its internal components are prone to damage. Figure 1 shows the internal components of a PRSOV structural model.

[0088] Notably, the torque motor current is controlled by the system, which drives the pintle movement to regulate the airflow into the opening chamber and the discharge orifice to the outside environment. The pressure change in the opening chamber further drives the pintle movement, thereby adjusting the butterfly valve angle and downstream pressure. This dynamic adjustment mechanism effectively controls the airflow and pressure of the ECS. Malfunction of the PRSOV can lead to reduced efficiency of hot gas regulation or unstable pressure regulation. For example, charging or discharging faults of the torque motor can severely affect the PRSOV's ability to accurately regulate air pressure.

[0089] Table 1 contains PRSOV dataset description information for different abnormal categories

[0090] ;

[0091] Any malfunction of ECS components directly affects the safety of the aircraft. Since aviation regulations strictly prohibit the aircraft from continuing to fly in the case of ECS failure, experiments cannot be conducted in real flight environments. Therefore, this study uses a rigorously verified PRSOV Simulink simulation model to generate abnormal data under different working conditions. The generation of each PRSOV sample is guided by valve experts, who simulate different fault states by adjusting the inherent parameters of the PRSOV (such as friction coefficient, degree of charging and discharging blockage, etc.). For each sample, its regulated pressure data is recorded at each timestamp as training input. Each training sample contains 201 pressure-related features. The entire dataset contains 8,000 training samples, covering 8 different health states. Table 1 summarizes the fault types and their sample numbers, while Figure 2 shows examples of the collected signal samples. In the data division process, normal samples are further split into training and test sets, while all abnormal types are considered as test samples to evaluate the model's anomaly detection ability.

[0092] We conducted comparative experiments of the RPDP-AD model with other data-driven methods to evaluate its relative performance. In the experimental evaluation process, we used the following key indicators:

[0093] • AUC (Area Under the ROC Curve): measures the overall detection ability of the model;

[0094] • AUPRC (Area Under the Precision-Recall Curve): measures the performance of the model in the case of less abnormal data;

[0095] • Accuracy: evaluates the overall classification ability of the model;

[0096] • F1-score: a comprehensive indicator balancing precision and recall.

[0097] During the training process, we adopted the Adam optimizer with a learning rate of 0.001 and a batch size of 64. All models were implemented based on the PyTorch framework and Python 3.8, and trained for 500 epochs on an RTX 3090 GPU server before testing. In terms of data preprocessing, all data samples were normalized using Z-Score to ensure consistency across different feature dimensions. For hyperparameter selection and MLP network structure, we followed the MLP settings in the reference and used the LeakyReLU activation function. The random projection matrix was sampled from a standard Gaussian distribution, and the projection dimension was set to 128. However, as mentioned earlier, other probability distributions that satisfy unit variance and certain boundedness can also be used to construct random projection matrices. The selection of different random projection matrices will be further discussed in the sensitivity analysis section.

[0098] In this study, we benchmarked RPDP-AD and compared it with other representative data-driven anomaly detection methods. We selected a series of excellent unsupervised machine learning and deep learning anomaly detection models as comparison objects, including RCA, DeepSVDD, GOAD, REPEN, NeuTraL, SLAD, IForest, ROD, ECOD, and KDE. All these models were trained in an unsupervised setting, using only normal signals as training data, and the test results are shown in Table 2. From the experimental results, it can be seen that RPDP-AD significantly outperforms other methods in multiple key indicators, including AUC, AUPRC, Accuracy, and F1-score, demonstrating excellent anomaly detection capabilities.

[0099] Table 2 shows that RPDP-AD achieved the highest average AUC (0.990), AUPRC (0.953), accuracy (0.986), and F1-score (0.973) in all evaluation methods, indicating that the method has the best overall performance in the PRSOV anomaly detection task. The high AUC value reflects that RPDP-AD can effectively distinguish between normal and abnormal data under different types of anomaly patterns, and the highest AUPRC value further verifies that it can still maintain high precision and high recall rate in the class imbalance scenario. In addition, Table 2 further shows that RPDP-AD performs stably under all seven types of anomalies, with nearly perfect detection scores (AUC / AUPRC / Accuracy / F1 ≥ 0.990) on Type 5 (Charge & Friction) and Type 6 (Discharge & Friction), indicating that the method has strong robustness in handling multi-source anomaly interactions. At the same time, AUC = 0.996 in Type 1 (Charge) and F1-score = 0.979 in Type 2 (Discharge) show that RPDP-AD has good adaptability to independent fault patterns and mixed fault patterns.

[0100] In summary, RPDP-AD consistently maintains excellent detection performance on all types of PRSOV anomalies. The highest average scores in AUC, AUPRC, accuracy, and F1-score further verify its advantages in robustness and reliability. RPDP-AD still maintains high detection accuracy in complex anomaly combination scenarios, while some existing data-driven methods show obvious limitations in specific anomaly detection. Through this comparative experiment, we can determine that RPDP-AD has excellent adaptability, generalization ability, and industrial application value in the aviation health monitoring task.

[0101] Table 2 and the experimental results of other data-driven methods on the PRSOV dataset

[0102] ;

[0103] Aiming at the problem of anomaly detection of pressure regulating shut-off valve (PRSOV) in the aircraft environmental control system (ECS), an efficient, unsupervised and computationally friendly intelligent diagnosis method is proposed to improve the safety and maintenance efficiency of the aircraft. PRSOV regulates and controls the airflow during flight to ensure the stability of the temperature and pressure in the cabin. However, under the influence of long-term operation and complex environmental changes, the component may wear, leak or fail, affecting the normal operation of the aircraft and even causing safety accidents. Due to the strict reliability requirements of the aviation industry, ensuring the health status of PRSOV is crucial for flight safety. Traditional threshold monitoring and rule-based anomaly detection methods have limited performance when dealing with high-dimensional, nonlinear, and multi-fault mode complex data. Existing machine learning methods, on the other hand, heavily rely on expert-driven feature engineering and cannot fully capture the dynamic characteristics of the system. At the same time, although deep learning methods have performed well in other industrial fields, their training often requires a large amount of labeled abnormal data. However, in practical aviation applications, it is difficult to obtain comprehensive labeled data due to the scarcity, diversity, and high labeling cost of fault data. Therefore, the goal of this study is to design an unsupervised anomaly detection method that does not require abnormal data, has theoretical explainability, and has low computational overhead to overcome the limitations of existing methods in PRSOV monitoring tasks. To this end, we propose RPDP-AD (Random Projection Distance Prediction-based Unsupervised PRSOV Multi-fault Anomaly Detection Algorithm), which uses random projection theory to map high-dimensional data to low-dimensional space and constructs a supervision signal by calculating the distance information after projection, so that it can learn the distribution pattern of normal data without relying on abnormal data. At the same time, we introduce a distribution difference measurement method to minimize the deviation between the mapping learned by the model and the random mapping, ensuring that the model can preserve the structural information of the data. In addition, to adapt to the computational resource limitations in the aviation field, RPDP-AD adopts a lightweight MLP (Multi-Layer Perceptron) network, avoiding the high computational demand of computationally intensive methods such as Transformers, thus achieving real-time and high-precision anomaly detection. More importantly, we provide theoretical analysis based on the Johnson-Lindenstrauss theorem, proving the feasibility and effectiveness of the random projection method in the deep learning framework. Finally, this method can efficiently detect multiple fault modes at low computational cost, not only filling the gap in unsupervised anomaly detection research in the PRSOV field, but also expanding the application of random projection theory in industrial deep learning, providing a new solution for aviation health monitoring systems that is precise, reliable and explainable.

[0104] Having described above several embodiments of the disclosure, any modifications and variations that fall within the scope of the described embodiments are also intended to be within the scope of the disclosure. As will be apparent to those skilled in the art, some modifications and variations to the embodiments described above can be practiced while staying within the scope and spirit of the described embodiments. The foregoing description of the described embodiments has been presented for the purposes of illustration and description. It is not intended to be exhaustive or to limit the described embodiments to the precise form disclosed. Many modifications and variations are possible in light of the above teachings. It is intended that the disclosed embodiments be limited only by the claims.

Claims

1. A method for aircraft fault detection based on random projection distance prediction, characterized in that, The method comprises: S1, collecting air flow, air pressure and temperature data through the pressure regulating stop valve sensor system installed on the aircraft; S2, preprocessing the data collected in S1; S3, constructing an unsupervised anomaly detection model, comprising: Random projection dimension reduction: using a random projection matrix, projecting high-dimensional input data into a low-dimensional space while preserving the relative distance relationship between data points as much as possible; according to the Johnson-Lindenstrauss theorem, it is ensured that the projected data still retains the distribution information of the original data; Supervised signal construction: by calculating the distance between data points in the projection space, a supervised signal is constructed to enable the model to learn the pattern of normal data in an unsupervised environment; Lightweight MLP multi-layer perceptron network: due to the limited computing resources of avionics, an MLP network is used to replace computationally intensive models to achieve efficient inference; Distribution difference measurement: by minimizing the distribution difference between the learnable mapping and the random mapping, the stability and interpretability of the model in the low-dimensional space are ensured; S4, training and optimizing the unsupervised anomaly detection model; S5, testing and optimizing the unsupervised anomaly detection model.

2. The method of claim 1, wherein, The method further comprises: S6, aircraft actual verification and deployment, comprising: After laboratory testing and optimization, the model is finally deployed to the actual aircraft health monitoring system to realize real-time fault detection: Onboard verification: during flight tasks, real-time pressure regulating stop valve sensor data is collected and detected online through the model; compare the model detection results with the artificial inspection results to verify its accuracy and stability; Embedded deployment: due to the limited computing resources of avionics, the optimized model is deployed to the onboard computing platform or cloud monitoring system to realize real-time analysis and remote monitoring; Real-time monitoring and maintenance optimization: integrate the model in the aircraft health monitoring system to realize all-weather online monitoring and provide early warning to maintenance personnel to optimize maintenance plans and reduce unscheduled maintenance.

3. The method of claim 1, wherein, The pressure regulating stop valve sensor system monitors the air flow, air pressure and temperature inside the pressure regulating stop valve sensor to identify abnormal flow fluctuations or air tightness problems; The collected data is in time series format, with high-precision clock recording timestamps to ensure the synchronization of multi-sensor data for subsequent data fusion and analysis.

4. The method of claim 1, wherein, S2, preprocessing the data collected in S1, comprising: Denoising: using a low-pass filter to remove high-frequency noise and ensure the smoothness of the vibration signal; through Kalman filtering of sensor data, it is more stable and reduces the influence of environmental noise; using wavelet transform denoising to extract main features and reduce interference caused by sensor errors; Normalization: using Z-score standardization or Min-Max normalization to ensure that all data remains within the same numerical range to reduce the impact of feature scale differences; Time series slicing: using a sliding window method to divide the data into fixed time windows for input into the model for analysis; Data augmentation: Apply data augmentation techniques, including time warping, adding noise, signal mixing, to improve the generalization ability of the model.

5. The method of claim 1, wherein, S4, training and optimizing unsupervised anomaly detection model, including: Unsupervised training: Only use normal data to train the model to minimize the distance prediction error in the random projection space, so that the model can learn the structural characteristics of the normal pressure regulating stop valve PRSOV operating state; Loss function optimization: Use distribution difference measure loss to ensure that the low-dimensional data after projection can still maintain the spatial structure of the high-dimensional data; Hyperparameter adjustment: Grid search is performed to adjust learning rate, projection dimension, MLP layer number and other hyperparameters to improve the stability and performance of the model; Regularization: Use L2 regularization to prevent overfitting, so that the model can generalize to different PRSOV working conditions.

6. The method of claim 1, wherein, S5, test and optimize the unsupervised anomaly detection model, including: Offline test: Test on real PRSOV sensor data set, calculate AUC, accuracy, recall rate, F1 score and other key indicators to evaluate the detection performance of the model; Through ablation experiment to verify the contribution of different modules to the performance of the model, optimize the overall architecture; Noise sensitivity analysis: Add different levels of noise to the test data to evaluate the robustness of the model in actual application; Multi-fault mode detection: Use 7 kinds of known fault mode data to test whether the model can effectively distinguish different fault types; According to the test result, the model is fine-tuned to ensure its stability in various flight conditions.

Images