Turbofan engine fault detection method based on variational neural anomaly detection architecture

Abstract

The invention discloses a turbofan engine fault detection method based on a variational neural anomaly detection architecture, and relates to the technical field of aero-engine fault detection and diagnose.The method comprises the steps that multi-element sensor data of a turbofan engine in the flight state is obtained, data preprocessing is carried out, and the data of the multi-element sensor data is obtained; obtaining a turbofan engine sensor fault diagnosis data set; the method comprises the following steps: constructing a turbofan engine fault diagnosis model based on variational neural anomaly detection and a unified multi-task joint optimization objective function, training the fault diagnosis model by using a turbofan engine sensor fault diagnosis data set, and realizing iterative updating of parameters through the unified multi-task joint optimization objective function; and collecting real-time multi-sensor data of the turbofan engine in a flight state and inputting the real-time multi-sensor data into the trained fault diagnosis model, and evaluating the operation state of the engine in real time. The method effectively improves the fault diagnosis accuracy of the turbofan engine, and guarantees the operation safety of the aero-engine.

Description

Technical Field

[0001] The present invention relates to the technical field of aeroengine fault detection and diagnosis, and particularly relates to a method for detecting faults of a turbofan engine based on a variational neural anomaly detection architecture. Background Art

[0002] In recent years, the rapid development of the aviation industry has greatly improved the application range and performance requirements of turbofan engines. As the core power device of modern aircraft, the reliability of the operating state of turbofan engines is directly related to flight safety and the successful implementation of missions. With the increasing complexity of the structure of aeroengines, for example, the design of key components such as high-pressure compressors, high-pressure turbines, and combustion chambers is more refined, and the coupling relationship between components is closer, resulting in a more demanding operating environment, which significantly increases the possibility of various faults such as blade cracks, bearing wear, combustion instability, and compressor surges. Once a fault occurs in an aeroengine, if it cannot be detected in time and effective measures are not taken, it may cause serious flight safety accidents. Existing traditional fault detection methods usually rely on manual experience or diagnostic models based on fixed thresholds, and it is difficult to effectively cope with the complex and changeable operating conditions and unknown fault modes of the engine, resulting in insufficient diagnostic accuracy and reliability. Therefore, there is an urgent need for a more intelligent, efficient, and accurate method for detecting faults of turbofan engines to achieve real-time monitoring and early warning of the abnormal state of the engine and ensure the flight safety and reliability of aircraft.

[0003] In the past few decades, researchers have proposed a variety of anomaly detection algorithms, including methods based on statistical thresholds, principal component analysis (PCA), and classical support vector machines (SVM). These methods usually rely on setting clear thresholds or analyzing linear features of historical data to achieve fault detection. However, these traditional anomaly detection algorithms show obvious deficiencies in dealing with complex non-linear feature relationships. Especially, the actual flight data of turbofan engines is usually scarce, and many unknown fault modes are not fully reflected in the existing datasets. For example, traditional threshold-based methods are prone to false alarms or missed detections when faced with blade cracks or new types of combustion instability that have not appeared before; while classical principal component analysis (PCA) has insufficient dimensionality reduction ability in the case of strong non-linear features in the data and is difficult to effectively capture fault symptoms; support vector machines (SVM) are highly dependent on training data, and if the training data lacks effective descriptions of unknown fault modes, the accuracy of detecting new faults will be significantly reduced. Therefore, traditional methods cannot efficiently and accurately identify unknown fault modes in turbofan engines under complex and data-deficient conditions, and there is an urgent need for a new method that still has high-precision detection ability under data-sparse conditions.

[0004] Specifically, an ideal anomaly detection model should possess the following capabilities: First, it should be able to effectively generalize and identify unknown anomaly patterns that do not exist in the dataset; second, it should be able to comprehensively analyze data from multiple sensors and have the ability to make global anomaly judgments on the operating state of the engine; in addition, it should also have the ability to accurately identify and locate subtle anomalies under local or specific operating conditions of the engine. Currently, there are not many models that possess all of the above capabilities simultaneously, but some studies have proposed the NeuTraL-AD method, which improves the semantic consistency and data diversity of the model by jointly learning data transformation and an encoder and using deterministic contrastive loss (DCL), and has shown outstanding performance in multiple time series and tabular datasets.

[0005] However, to achieve accurate fault detection of turbofan engines, the following problems still need to be faced.

[0006] 1) The NeuTraL-AD algorithm can only output anomaly scores at the full sequence level and cannot locate the specific anomaly moments in the sequence, making it difficult to locate local faults. In addition, since the anomaly scores are continuous values, it is difficult to automatically determine the optimal threshold in the case of no labels.

[0007] 2) Aeroengine fault data usually has non-linear and strongly coupled characteristics, which makes it difficult and costly to obtain anomalies, especially data on unknown or rare fault patterns is even more scarce, limiting the model's learning and generalization ability for unknown fault patterns.

[0008] 3) Engines in actual operation are often accompanied by complex changes in operating conditions, such as frequent changes in temperature, pressure, and speed, making the boundaries of anomaly patterns fuzzy and variable, increasing the difficulty for the anomaly detection model to accurately identify fault patterns. Summary of the Invention

[0009] Based on the technical problems existing in the above background technology, the purpose of the present invention is to propose a turbofan engine fault detection method based on a variational neural anomaly detection architecture (Variational Neural-AD) to solve the deficiencies of traditional anomaly detection methods in dealing with unknown fault patterns, global judgment, and accurate local anomaly positioning. Specifically, the present invention combines NeuTraL-AD with a variational autoencoder (VAE) to improve the robustness and generalization of anomaly detection through a multi-task joint training mechanism; innovatively introduces the reconstruction error as a supplementary index for anomaly detection to help the model more carefully capture and locate local anomalies of the engine; at the same time, through KL divergence regularization in the latent space, the stability and performance of the model in high-dimensional data scenarios are improved. The technical solution of the present invention not only effectively improves the accuracy of turbofan engine fault diagnosis, but also significantly enhances the engineering applicability of the model, ensuring the operation safety of aeroengines.

[0010] Specifically, the present invention provides a fault detection method for a turbofan engine based on a variational neural anomaly detection architecture, including: Obtain multivariate sensor data of the turbofan engine in flight; Perform data preprocessing on the multivariate sensor data to obtain a turbofan engine sensor fault diagnosis dataset; Construct a fault diagnosis model architecture for a turbofan engine based on variational neural anomaly detection, where the fault diagnosis model includes a convolutional feature encoding module, a variational auto-encoding reconstruction module, and a neural transformation contrast learning module; Construct a unified multi-task joint optimization objective function to jointly train and optimize the convolutional feature encoding module, the variational auto-encoding reconstruction module, and the neural transformation contrast learning module; Use the turbofan engine sensor fault diagnosis dataset to train the fault diagnosis model, and realize iterative update of parameters through the unified multi-task joint optimization objective function; Collect real-time multivariate sensor data of the turbofan engine in flight, input the collected multivariate sensor data into the trained fault diagnosis model, and evaluate the engine operating state in real time.

[0011] Further, the multivariate sensor data includes: flight altitude, flight Mach number, power control angle, fuel flow rate, and high-pressure compressor speed.

[0012] Further, the data preprocessing includes data cleaning, standardization, and abnormal sample construction.

[0013] Further, the turbofan engine sensor fault diagnosis dataset is: , where N is the total number of samples, represents the th multivariate sensor signal sample, is the corresponding fault category label, and each fault sample in this dataset corresponds to an ideal fault-free sample .

[0014] Further, the convolutional feature encoding module consists of several convolutional neural network layers and fully connected layers, which are used to map the multivariate sensor data to a high-dimensional latent feature space; Further, a fault classification module is arranged at the back end of the convolutional feature encoding module, which consists of a convolutional layer and a fully connected layer, and is used to map the high-dimensional feature information obtained by the convolutional feature encoding module to the fault category space to realize health state recognition.

[0015] Further, the variational auto-encoder reconstruction module is used to reconstruct the original sensor signal of the turbofan engine from the latent features generated by the convolutional feature encoding module through a decoder, and calculate the reconstruction error of the signal to determine the position of the abnormal signal.

[0016] Further, the neural transformation contrastive learning module consists of multiple learnable neural transformation functions and is used to perform diverse semantic transformations on the input data; the neural transformation contrastive learning module introduces a deterministic contrast loss function to optimize the latent feature representation of the convolutional feature encoding module, so that the consistency and difference between different transformed views and the original view in the semantic space can be effectively distinguished.

[0017] Further, when training the fault diagnosis model, the joint optimization objective function used is: ; where: is the deterministic contrast loss, which is used to optimize the semantic transformation ability of the neural transformation contrastive learning module; is the reconstruction error loss of the variational auto-encoder reconstruction module, which is used to ensure the accurate reconstruction of the input data by the model and the precise positioning of abnormalities; is the KL divergence loss, which is used to constrain the latent feature distribution to conform to the Gaussian distribution characteristics; λ is a hyperparameter, which is used to control the optimization balance degree between the deterministic contrast loss and the variational reconstruction error loss.

[0018] One or more technical solutions provided in the embodiments of the present application have at least the following technical effects or advantages: The present invention proposes a deep learning architecture based on variational neural anomaly detection (Variational Neural-AD) for the application scenarios of turbofan engine sensor fault diagnosis and abnormal location. This architecture combines the advantages of variational autoencoder (VAE) and self-supervised neural transform contrast learning module (NeuTraL-AD), which can not only accurately identify various fault types of sensors, but also achieve high-precision positioning of abnormal signals to ensure flight safety. In particular, in order to solve the problem of scarce turbofan engine data and difficulty in identifying unknown fault modes, the present invention innovatively proposes an adaptive neural transform enhancement strategy, which effectively improves the model's generalization recognition ability for unknown faults by generating more potential abnormal samples. At the same time, through the variational autoencoder reconstruction mechanism (VAE Reconstruction), fine-grained semantic representation in the latent space and accurate backtracking and positioning of local abnormal signals are achieved. Most importantly, the encoder and decoder of the present invention achieve collaborative optimization of deep features through multi-task joint training, thereby effectively improving the fault diagnosis accuracy and robustness of the model under complex flight conditions.

[0019] Compared with the prior art, the present invention significantly improves the accuracy of turbofan engine sensor fault diagnosis and anomaly detection, can efficiently identify and locate various unknown sensor failure conditions, and ensure the safe and reliable operation of the turbofan engine under various flight conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] The drawings described herein are used to provide a further understanding of the embodiments of the present invention and constitute a part of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings: Figure 1 This is a flow chart of the turbofan engine fault detection method based on the variational neural anomaly detection architecture proposed in the present invention. DETAILED DESCRIPTION

[0021] In order to make the technical problems, technical solutions and beneficial effects to be solved by the present invention more clearly understood, the present invention is further described in detail below. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0022] In recent years, turbofan engines have achieved remarkable breakthroughs in performance optimization and reliability improvement. Their high-efficiency operation ability makes them the core of modern aviation power. This breakthrough benefits from the leapfrog development of materials science, fluid dynamics, and intelligent control technology. Turbofan engines use a high-precision sensor network to monitor key parameters such as combustion chamber temperature, turbine speed, fuel pressure, and vibration spectrum in real time, and combine digital twin models to achieve dynamic assessment of the operating state. Among them, high-temperature thermocouples, dynamic pressure probes, and fiber optic vibration sensors form the basis of the health monitoring system, while special sensors such as lubricating oil metal debris detection and gas path electrostatic monitoring are directly related to the engine's fault warning ability. However, under continuous high-load operating conditions, the extreme internal environment of the engine (high temperature, high pressure, high speed) will cause sensor performance drift, and progressive damages such as turbine blade creep and combustion chamber thermal fatigue will continue to accumulate, which may lead to sensor failure or component function degradation. When abnormal deviations occur in key parameters, if not identified and corrected in time, it may induce chain failures such as compressor stall and turbine overheating, seriously threatening flight safety. Constructing an accurate fault detection and health management system (PHM) is decisive for preventing unplanned engine removals and avoiding catastrophic accidents. Therefore, developing intelligent fault diagnosis and remaining life prediction technologies is a necessary guarantee for achieving high reliability and high maintainability operation of aeroengines.

[0023] Based on the above considerations, the present invention proposes a fault detection method for turbofan engines based on a variational neural anomaly detection architecture to solve the deficiencies of traditional anomaly detection methods in dealing with unknown fault modes, global judgment, and accurate localization of local anomalies. Specifically, the present invention combines NeuTraL-AD with a variational autoencoder (VAE), and improves the robustness and generalization of anomaly detection through a multi-task joint training mechanism; innovatively introduces the reconstruction error as a supplementary index for anomaly detection to help the model more carefully capture and locate local anomalies in the engine; at the same time, through the KL divergence regularization of the latent space, the stability and performance of the model in high-dimensional data scenarios are improved. The technical solution of the present invention not only effectively improves the accuracy of turbofan engine fault diagnosis, but also significantly enhances the engineering applicability of the model, ensuring the operating safety of aeroengines.

[0024] Specifically, as Figure 1 shown, the present invention provides a fault detection method for turbofan engines based on a variational neural anomaly detection architecture, and the method includes the following steps: Step 1: Obtain multi-sensor data of the turbofan engine in flight state.

[0025] The turbofan engine uses multiple sensors deployed at key parts of the engine to collect operating state data in real time to monitor the health status and performance indicators of the engine. Therefore, the multi-sensor data samples of the turbofan engine collected can be expressed as: Among them, M represents the number of sensor signals, represents the - th time - series data collected by the

[0026] Step 2: Perform data pre - processing on the multi - sensor data to obtain a turbofan engine sensor fault diagnosis data set for training the subsequent model.

[0027] Specifically, the data pre - processing includes data cleaning, standardization, and abnormal sample construction.

[0028] The turbofan engine operates under different flight conditions (such as take - off, cruise, landing, etc.), and the status data of key sensors during the flight are collected in real - time, including information such as flight altitude, flight Mach number, power control angle, fuel flow, high - pressure compressor speed, etc. However, the original multi - sensor data collected during the flight of the turbofan engine usually has problems such as noise interference, signal drift, and data missing. In order to ensure that the subsequent diagnosis model can efficiently extract accurate and robust features from the data and address the problem of large differences in numerical ranges between different sensors, the present invention proposes an adaptive robust standardization method to improve the robustness of data processing. Let the original data of the sensor sequence be , and its standardization calculation formula is defined as: ; Among them, represents the median of the sequence data, represents the median absolute deviation (Median Absolute Deviation) of the sequence data, is a very small constant to prevent division - by - zero errors. Using this method can effectively suppress the abnormal peaks in the data and improve the stability of the subsequent anomaly detection model.

[0029] In the actual flight environment, due to the degradation, contamination of the sensors themselves or environmental interference, these sensor signals may present various abnormal patterns, such as drift, mutation, or noise pollution, etc. In order to construct a high - quality engine fault diagnosis model, through a large number of actual flight - state data acquisitions and simulation fault injection experiments, the present invention constructs a turbofan engine sensor fault diagnosis data set: Among them, N is the total number of samples, represents the - th multi - sensor signal sample, is the corresponding fault - class label. In addition, each fault sample in this data set corresponds to an ideal fault - free sample for the reconstruction target of the model.

[0030] Step 3: Construct a fault diagnosis model architecture for turbofan engines based on Variational Neural-AD, where the fault diagnosis model includes a convolutional feature encoding module, a variational autoencoder reconstruction module, and a neural transformation contrastive learning module.

[0031] To effectively achieve real-time fault monitoring and unknown anomaly pattern recognition of turbofan engines under complex flight conditions, the present invention proposes an innovative fault diagnosis model based on Variational Neural-AD. Specifically, the present invention creatively integrates the probabilistic generation characteristics and feature extraction capabilities of the variational autoencoder (VAE), as well as the self-supervised anomaly enhancement strategy provided by neural transformation contrastive learning (NeuTraL-AD). It realizes the anomaly capture and precise positioning of turbofan engine multi-sensor data at both the global and local levels.

[0032] Based on the method of NeuTraL-AD for generating anomaly samples through adaptive data transformation, this method can simulate more potential fault patterns in the case of scarce data, and uses contrastive loss to ensure that the transformed samples maintain a high similarity in the original space while having a greater difference from the original samples in the semantic space. Subsequently, by replacing the original encoder in NeuTraL-AD with the encoder-decoder reconstruction mechanism of VAE, deeper semantic representations in the latent space are further explored. The feature extraction of VAE can not only trace and locate local anomalies, but also reconstruct or generate more anomaly samples by adjusting the latent representation, continuously enriching the fault pattern learning ability of the model. The combination of the two enables the present invention to have excellent anomaly recognition performance in both the global and local dimensions, especially achieving better results in unknown fault scenarios.

[0033] Specifically, the model designed by the present invention includes a convolutional feature encoding module (Convolutional Encoder), a variational autoencoder reconstruction module (VAE Reconstruction), and a neural transformation contrastive learning module (NeuralTransformation Contrastive Learning). Through an innovative multi-task joint training mechanism, deep collaborative optimization is achieved among the modules. The detailed derivation and description of each module and the corresponding mechanism are as follows: The convolutional feature encoding module (Convolutional Encoder) consists of several convolutional neural network layers and fully connected layers, which are used to map multi-sensor data to a high-dimensional latent feature space; the latent features output by this module fully retain the key semantic information of the data, providing a reliable basis for subsequent anomaly detection.

[0034] The core function of the convolutional feature encoding module is to perform deep feature extraction and encoding of the multi-sensor signals of the turbofan engine to generate high-dimensional and discriminative potential feature expressions. Specifically, the input standardized sensor data is: ; in, represents the number of sensors, Indicates the time sequence length. Convolutional encoder Through a multi-layer one-dimensional convolutional neural network and a fully connected network structure, the input data is mapped to a high-dimensional and high-semantic feature space, and the distribution parameters μ and σ of the latent space are obtained: .

[0035] A fault classification module is set at the back end of the convolutional feature encoding module, which consists of a convolutional layer and a fully connected layer. It is used to map the high-dimensional feature information obtained by the convolutional feature encoding module to the fault category space to realize health status identification.

[0036] The variational autoencoder reconstruction module (VAE Reconstruction) uses the latent features generated by the convolutional feature encoding module to reconstruct the original sensor signal of the turbofan engine through the decoder, and calculates the reconstruction error of the signal to determine the location of the abnormal signal. Specifically, during the training phase, the module constrains the latent features to obey the Gaussian distribution to obtain a stable feature representation; during the inference phase, it performs fine-grained anomaly detection on the input data based on the reconstruction error to identify faults caused by small deviations.

[0037] In order to further achieve accurate positioning of abnormal patterns and local anomaly detection, this paper creatively introduces the structure of variational autoencoder (VAE) and uses its reconstruction mechanism to achieve fine-grained anomaly capture. Specifically, the latent variable Reparameterize sampling based on the distribution parameters generated by the encoder: ; Then, through the decoder The potential characteristics Reconstructed into raw sensor data: ; By calculating the reconstruction error, we can accurately locate the abnormal points where local deviations occur in the signal. The error is expressed as: ; In addition, by introducing the KL-divergence constraint, the feature distribution in the latent space can be made more regular and smooth, enhancing the generalization ability of the feature space. The KL-divergence is defined as: ; On this basis, this module not only has the ability to accurately locate anomalies but also can generate or reconstruct potential anomaly samples through the smooth characteristics of the latent space, expanding the diversity of the data.

[0038] The Neural Transformation contrastive learning module consists of multiple learnable neural transformation functions for performing diverse semantic transformations on the input data; the Neural Transformation contrastive learning module introduces the Deterministic Contrastive Loss (DCL) function to optimize the latent feature representation of the convolutional feature encoding module, enabling the effective distinction of the consistency and difference between different transformed views and the original view in the semantic space. In this way, the Neural Transformation module enhances the model's sensitivity to global anomaly patterns and improves the detection ability for various types of faults.

[0039] Given the difficulty in obtaining fault samples of turbofan engines and the large number of unknown anomaly patterns in the actual dataset, the present invention creatively proposes the Neural Transformation contrastive learning module to adaptively learn data transformations to generate rich anomaly semantic views. Specifically, the module defines multiple learnable neural transformation functions , which act on the input data to generate diverse transformed views: ; To strengthen the model's global sensitivity to anomaly features, the Deterministic Contrastive Loss (DCL) is defined: ; where is the cosine similarity metric function, and is the temperature parameter. By optimizing the DCL loss, the encoder is driven in the latent space to make the transformed samples significantly different from the original samples in the semantic space while having a high similarity in the original data space, achieving the purpose of more efficiently identifying unknown anomaly patterns.

[0040] Step 4: Construct a unified multi-task joint optimization objective function to jointly train and optimize the convolutional feature encoding module, the variational auto-encoder reconstruction module, and the Neural Transformation contrastive learning module.

[0041] To integrate the advantages of the above three modules, the present invention further proposes a unified multi-task joint optimization objective function: ; Among them, λ is a trade-off hyperparameter that controls the optimization balance between contrastive learning and the VAE reconstruction task.

[0042] Specifically: DCL loss: Through contrastive learning in the semantic space, it enhances the encoder's ability to capture global anomaly patterns, especially having better generalization performance for unknown anomalies; VAE reconstruction loss and KL divergence: Through a refined reconstruction error measurement, it achieves precise localization of local anomalies, and the KL divergence ensures the smoothness of the latent space distribution, improving the model's generalization and stability.

[0043] On this basis, the present invention innovatively integrates the variational autoencoder with NeuTraL-AD. The encoder of the VAE enhances the semantic space expression ability of the encoder in NeuTraL-AD, while the decoder of the VAE provides a way to generate more anomaly samples, thereby further enhancing the model's learning and generalization ability for unknown faults. Specifically, by actively adjusting the latent feature distribution or performing latent space sampling, new anomaly samples can be reconstructed or generated, expanding the diversity of anomaly patterns in the training data.

[0044] Step 5: Use the turbofan engine sensor fault diagnosis dataset to train the fault diagnosis model, and achieve iterative update of parameters through a unified multi-task joint optimization objective function. The training objective of the model is to maximize the discrimination ability of latent semantic information while accurately reconstructing and precisely locating abnormal signals.

[0045] Specifically, when training the fault diagnosis model, the joint optimization objective function used is: ; Where: is the deterministic contrast loss, used to optimize the semantic transformation ability of the neural transformation contrast learning module; is the reconstruction error loss of the variational autoencoding reconstruction module, used to ensure the accurate reconstruction of the input data by the model and precise anomaly localization; is the KL divergence loss, used to constrain the latent feature distribution to conform to the Gaussian distribution characteristics; λ is a hyperparameter used to control the degree of optimization balance between the deterministic contrast loss and the variational reconstruction error loss.

[0046] Step 6: Collect real-time multi-sensor data of the turbofan engine in flight, input the collected multi-sensor data into the trained fault diagnosis model, and evaluate the engine operating state in real time.

[0047] The above fault diagnosis model extracts the potential feature representation of data through convolutional feature encoding, identifies global unknown abnormal patterns through neural transformation contrastive learning (NeuTraL-AD), and further accurately locates local abnormal positions by combining the reconstruction error of the variational autoencoder (VAE), so as to realize real-time fault diagnosis, abnormal location and health status monitoring of the turbofan engine under complex flight conditions.

[0048] To better illustrate the technical effects of the present invention, a specific embodiment is used to test and verify the present invention.

[0049] This study uses multivariate sensor data generated by the simulation platform of the advanced geared turbofan engine (AGTF30) of NASA in the United States to construct multiple fault modes covering steady-state / dynamic conditions and the full flight envelope. The dataset contains 18 health states, which include both long-term healthy conditions and simulated flight data under 17 different fault modes, with high complexity and representativeness. Each fault mode corresponds to the performance degradation (such as efficiency reduction or capacity attenuation) of key components such as the fan, low-pressure compressor (LPC), high-pressure compressor (HPC), high-pressure turbine (HPT), and low-pressure turbine (LPT). Each fault occurs separately during a flight and lasts for about 200 seconds, thus forming time-series data with clear fault labels. In addition, the dataset also contains about 10,000 seconds of healthy cruise steady-state operation records, which can be used as normal control data for model training and evaluation.

[0050] The core of this study is to realize the fault diagnosis and fault correction of the turbofan engine. Fault diagnosis is defined as a multi-classification problem. Therefore, we use three indicators, namely the area under the curve (AUC, Area Under Curve), average precision (AP, Average Precision), and F1 score (F1-Score), to measure the performance of the diagnosis model. The higher the value, the better the model performance. They are defined as: ; ; ; In addition, to further evaluate the reliability and practical application performance of the model, we also introduce two indicators, namely the missed detection rate (MDR, Missed Detection Rate) and the false alarm rate (FAR, False Alarm Rate). The missed detection rate represents the proportion of faults that actually exist but are identified as normal states by the model. Its calculation formula is: ; The false alarm rate represents the proportion of normal states that are misidentified as fault states by the model. Its calculation formula is: ; wherein ; . In the formula: TP (True Positive) is the number of true positives; FP (False Positive) is the number of false positives; FN (False Negative) is the number of false negatives; TN (True Negative) is the number of true negatives. The above indicators comprehensively reflect the overall classification performance of the proposed model in the multi-class fault diagnosis scenario.

[0051] To verify the effectiveness of the proposed fault diagnosis method for turbofan engines based on Variational Neural-AD and each module, this study set up ablation experiments and designed multiple groups of comparative experiments to evaluate the performance advantages of the proposed method. Specifically, the ablation experiment comparison models designed include: (1) Variational Neural-AD-F: Remove the Variational Autoencoder (VAE) module, only retain the convolutional encoder and the neural transformation contrast learning module, and are dedicated to the fault diagnosis task; (2) Variational Neural-AD-V: Remove the neural transformation contrast learning module, only retain the convolutional encoder and the Variational Autoencoder (VAE) reconstruction module, focusing on local anomaly localization and reconstruction; (3) Variational Neural-AD-E: Remove the convolutional feature encoding module, directly send the original data into the variational autoencoder and the neural transformation module to analyze the contribution of the convolutional encoder in feature extraction; (4) Variational Neural-AD-Base: Remove both the VAE module and the contrast learning module at the same time, only retain the convolutional encoder and the fault classifier, and only use it as a basic classification model for comparison.

[0052] Fault diagnosis of turbofan engine sensors: To demonstrate the effectiveness of the proposed method, Table 1 shows the fault diagnosis performance of Variational Neural-AD, Variational Neural-AD-F, Variational Neural-AD-V, Variational Neural-AD-E, and Variational Neural-AD-Base on the turbofan engine sensor dataset. As can be seen from Table 1, the effectiveness of each module has been fully verified. For example, the AUCs of fault diagnosis for Variational Neural-AD-V and Variational Neural-AD-F reach 0.978 and 0.971 respectively, which are significantly better than 0.946 of the basic model Variational Neural-AD-Base. This indicates that both the VAE reconstruction module and the neural transformation module have significantly improved the model performance. In addition, the AP of Variational Neural-AD-V reaches 0.963, which is 4 percentage points higher than 0.922 of the basic model. This shows that local anomaly reconstruction through VAE reconstruction can effectively improve the fineness of anomaly detection. The overall performance of Variational Neural-AD is the most prominent, with its AUC, AP, and F1-Score reaching 0.986, 0.975, and 97.68% respectively, indicating the excellent performance improvement brought by the collaborative work of each module of the model. In addition, this study further analyzes two key indicators: the miss detection rate (MDR) and the false alarm rate (FAR). The miss detection rate and false alarm rate of Variational Neural-AD are 1.52% and 1.37% respectively, which are significantly lower than 3.85% and 3.42% of the basic model Variational Neural-AD-Base. This result further verifies the collaborative optimization among the modules proposed in the present invention, effectively reducing the risks of misjudgment and missed detection, and significantly improving the reliability and stability of fault diagnosis.

[0053] Table 1 Ablation Experiments of the Proposed Method

[0054] Those skilled in the art should understand that the embodiments of the present invention can be provided as a method, a system, or a computer program product. Therefore, the present invention can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk memory, CD-ROM, optical memory, etc.) containing computer-usable program code.

[0055] The present invention is described with reference to the flowcharts and / or block diagrams of systems, apparatus (systems), and computer program products according to embodiments of the invention. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and combinations of flows and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing device to produce a machine, such that the instructions executed by the processor of the computer or other programmable data processing device produce means for implementing the functions specified in the flow Figure 1 one or more flows and / or blocks Figure 1 or means for implementing the functions specified in one or more blocks.

[0056] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to operate in a particular manner, such that the instructions stored in the computer-readable memory produce a manufacture including instruction means for implementing the functions specified in the flow Figure 1 one or more flows and / or blocks Figure 1 or means for implementing the functions specified in one or more blocks.

[0057] These computer program instructions can also be loaded onto a computer or other programmable data processing device, such that a series of operational steps are performed on the computer or other programmable device to produce a computer-implemented process, and thus the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in the flow Figure 1 one or more flows and / or blocks Figure 1 or means for implementing the functions specified in one or more blocks.

[0058] Although the preferred embodiments of the present invention have been described, additional changes and modifications can be made by those skilled in the art once they learn of the basic inventive concept. Therefore, the appended claims are intended to be construed to include the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0059] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalent technologies, the present invention is also intended to include these modifications and variations.

Claims

1. A fault detection method for a turbofan engine based on a variational neural anomaly detection architecture, characterized in that Including: Obtain multi - sensor data of a turbofan engine in flight state; Perform data pre - processing on the multi - sensor data to obtain a turbofan engine sensor fault diagnosis data set; Construct a turbofan engine fault diagnosis model architecture based on variational neural anomaly detection, where the fault diagnosis model includes a convolutional feature encoding module, a variational auto - encoding reconstruction module, and a neural transformation contrast learning module; Construct a unified multi - task joint optimization objective function to jointly train and optimize the convolutional feature encoding module, the variational auto - encoding reconstruction module, and the neural transformation contrast learning module; Use the turbofan engine sensor fault diagnosis data set to train the fault diagnosis model, and realize the iterative update of parameters through the unified multi - task joint optimization objective function; Collect real - time multi - sensor data of a turbofan engine in flight state, input the collected multi - sensor data into the trained fault diagnosis model, and evaluate the engine operation state in real time.

2. The method for detecting faults in a turbofan engine based on a variational neural anomaly detection architecture according to claim 1, wherein: The multi - sensor data includes: flight altitude, flight Mach number, power control angle, fuel flow rate, and high - pressure compressor speed.

3. The fault detection method for a turbofan engine based on the variational neural anomaly detection architecture according to claim 1, wherein: The data pre - processing includes data cleaning, standardization, and abnormal sample construction.

4. The fault detection method for a turbofan engine based on a variational neural anomaly detection architecture according to claim 1, characterized in that: The sensor fault diagnosis dataset of the turbofan engine is as follows: , where N is the total number of samples, denotes the th multi-sensor signal sample, is the corresponding fault class label, and each fault sample in this dataset corresponds to an ideal fault-free sample .

5. The fault detection method for a turbofan engine based on the variational neural anomaly detection architecture according to claim 1, characterized in that: The convolutional feature encoding module consists of several convolutional neural network layers and fully - connected layers, which is used to map the multi - sensor data to a high - dimensional latent feature space.

6. The fault detection method for a turbofan engine based on a variational neural anomaly detection architecture according to claim 1, wherein: A fault classification module is arranged at the backend of the convolutional feature encoding module, which consists of a convolutional layer and a fully - connected layer, and is used to map the high - dimensional feature information obtained by the convolutional feature encoding module to the fault category space to realize health state recognition.

7. The method for detecting faults in a turbofan engine based on a variational neural anomaly detection architecture according to claim 1, characterized in that: The variational auto - encoding reconstruction module is used to reconstruct the original sensor signal of the turbofan engine through a decoder from the latent features generated by the convolutional feature encoding module, and calculate the reconstruction error of the signal to determine the position of the abnormal signal.

8. The method for detecting faults in a turbofan engine based on a variational neural anomaly detection architecture according to claim 1, characterized in that: The neural transformation contrast learning module consists of multiple learnable neural transformation functions, which is used to perform diverse semantic transformations on the input data; the neural transformation contrast learning module introduces a deterministic contrast loss function to optimize the latent feature representation of the convolutional feature encoding module, so that the consistency and difference between different transformed views and the original view in the semantic space can be effectively distinguished.

9. The method for detecting faults in a turbofan engine based on a variational neural anomaly detection architecture according to claim 1, wherein: When training the fault diagnosis model, the joint optimization objective function used is: ； Where: is a deterministic contrastive loss for optimizing the semantic transformation ability of the neural transformation contrastive learning module; It is the reconstruction error loss of the variational auto - encoding reconstruction module, which is used to ensure the accurate reconstruction of the input data by the model and the precise positioning of anomalies; is the KL divergence loss, which is used to constrain the latent feature distribution to conform to the Gaussian distribution characteristics; λ is a hyperparameter used to control the optimization balance degree between the deterministic contrast loss and the variational reconstruction error loss.

Images