A Multi-Data Fusion and Failure Mode Recognition Method Based on Branch Adversarial Networks

Abstract

This invention relates to a multi-data fusion and failure mode identification method based on branch adversarial networks, comprising the following steps: acquiring sensor signal data, failure mode labels, and total operating time of historical aircraft engines and performing data preprocessing to construct a dataset; loading the preprocessed dataset into a constructed multi-branch deep learning model and training the model using an adversarial training mechanism, wherein the multi-branch deep learning model is used to construct failure mode-specific Hiers (HIs), with each branch layer constructing an HI for a specific failure mode; acquiring sensor signal data of in-service engines and inputting it into the trained multi-branch deep learning model to obtain candidate HIs corresponding to different failure modes; extracting features from each candidate HI; and performing failure mode identification based on the extracted features to achieve Restricted Low-Use (RUL) prediction. Compared with existing technologies, this invention has advantages such as accurate description of different degradation modes, good interpretability, and high flexibility.

Description

Technical Field

[0001] This invention relates to the field of aircraft engine performance health management and remaining life prediction technology, and in particular to a multi-data fusion and failure mode identification method based on branch adversarial networks. Background Technology

[0002] Predictive health management (PHM) of manufacturing systems helps detect potential degradation modes early, enabling the development of appropriate maintenance and safety inspection strategies to effectively prevent unexpected equipment failures and extend their service life. Over time, systems may degrade through various failure modes, ultimately leading to unexpected failures and significant losses. Therefore, utilizing sensor signals for machine failure mode identification and remaining useful life (RUL) prediction is a core research area in PHM.

[0003] Data-driven RUL prediction methods can be divided into direct methods and health index (HI)-based methods. Existing direct RUL prediction methods use sensor signals as model input and RUL directly as model output, without considering potential degradation mechanisms. This "black box" approach not only limits the interpretability of feature extraction but also fails to effectively visualize the degradation process of units, thus limiting a deep understanding of the system's degradation process, which may hinder engineers from accurately grasping system performance. In contrast, HI-based methods offer better interpretability and visualization capabilities. HI is a one-dimensional time series constructed by fusing multiple sensor signals, quantifying the health status based on the system's inherent degradation patterns. However, in multi-failure-mode scenarios, failure mode identification and RUL prediction are often treated as completely independent tasks, or the failure modes and RUL are directly output through a joint learning model. The former ignores the complex relationships between tasks, while the latter suffers from the same drawback as the direct RUL method, failing to provide interpretable visualization of the health status.

[0004] In summary, the challenges and research gaps in aircraft engine degradation modeling and RUL prediction under multiple failure modes are as follows: First, existing research uses joint learning networks to treat fault diagnosis and RUL prediction as collaborative tasks, sharing information related to different tasks. However, this method cannot characterize the complete degradation evolution process of the equipment. In engineering practice, clear visualization of the equipment degradation state is highly needed, which remains a challenging task in multi-failure mode scenarios. Second, different failure modes have their own degradation patterns, and existing methods struggle to construct a unified framework of different degradation characteristics (HI) for multiple failure modes to characterize these degradation processes. Third, existing methods struggle to simultaneously address the RUL prediction task while ensuring that the constructed HI has clearly distinguishable features across different failure modes, thus failing to assist in failure mode identification. Summary of the Invention

[0005] The purpose of this invention is to provide a multi-data fusion and failure mode identification method based on branch adversarial networks, which utilizes multi-sensor signal data reflecting the health status of aircraft engines collected during operation to identify failure modes and predict RUL of in-service aircraft engines.

[0006] The objective of this invention can be achieved through the following technical solutions:

[0007] A method for multi-data fusion and failure mode identification based on branch adversarial networks includes the following steps:

[0008] Acquire historical aircraft engine sensor signal data, failure mode labels, and total runtime, and perform data preprocessing to construct a dataset;

[0009] The preprocessed dataset is loaded into the constructed multi-branch deep learning model, and the model is trained using an adversarial training mechanism. The multi-branch deep learning model is used to construct HIs specific to failure modes, with each branch layer constructing an HI for a specific failure mode.

[0010] The sensor signal data of the in-service engine is acquired and input into the trained multi-branch deep learning model to obtain candidate failure modes (HIs) for different failure modes. Features of each candidate HI are extracted, and failure mode identification is performed based on the extracted features to achieve RUL prediction.

[0011] In this invention, aircraft engines are divided into two types: historical aircraft engines, for which all sensor signals up to their failure point can be acquired and the failure mode is known; and in-service aircraft engines, for which only sensor signals at a specific point in time before failure can be acquired. The multidimensional sensor signals of a usable aircraft engine are represented as x. i (t)=(x i,1 (t),x i,2 (t),…,x i,L (t)), where x i,l (t) represents the l-th sensor signal of the i-th engine, where l = 1, ..., L, and t = t1, ..., T. i t1 represents the first observable time point, T i This represents the last observable time point. There are K failure modes in total, and each engine degrades in only one failure mode, i.e., there is a set of indices i∈φ. k , φ k ∈φ={φ k |k=1,…,K}.

[0012] The multi-branch deep learning model is represented as follows:

[0013]

[0014] in, θ represents the shared layer function of a neural network that extracts features extensively from multidimensional sensors. s For the shared layer parameter set, This represents the extracted features obtained from the shared layer. Represents a branched neural network, θ k =[θ (1) ,…,θ (k) ,…,θ (K) [ ] represents the parameter set for each branch. This represents the HI of engine i belonging to the k-th failure mode at time t; the neural network first uses... The system learns the overall degradation characteristics, and then each branch layer learns degradation characteristics specific to different failure modes.

[0015] The Health Index (HI) is defined as the sum of potential failure states and white noise, as shown in the following formula:

[0016]

[0017] Wherein, the basis function ψ at time t i (t)=[1,t,t 2 The basis function parameter vector for describing the failure process of an aircraft engine i, belonging to the k-th failure mode, is: Estimate using the least squares method, i.e. Among them Ψ i =[ψ i (t1),…,ψ i (T i )] T Thus, the fitted potential degenerate state is obtained: This is the noise term corresponding to the k-th failure mode, assumed to have a mean of 0 and a variance of . normal distribution

[0018] To ensure that HI accurately estimates failure states, HI must satisfy the following properties:

[0019] Attribute 1: Minimum Variance (MVF) at the Failure Threshold: HI should satisfy the condition that it reaches the threshold of its failure mode at the time of failure. This invention uses the variance of the last observation point relative to the threshold as a measure, and the calculation formula is as follows:

[0020]

[0021] Among them, M iThe failure threshold is used only to distinguish between the engine's operating condition and failure state, and to indicate different degradation trends (increasing or decreasing) under different failure modes. Its absolute value has no specific meaning, so it is set to have two values: 1 or -1.

[0022] Attribute 2: Degradation Range (DR): HI should satisfy the requirement of having a clear degradation trend to characterize the degradation from a relatively healthy state to a failure state. Therefore, potential failure states should cover the interval between 0 and the failure threshold as much as possible. The calculation formula is as follows:

[0023]

[0024] Attribute 3: Minimum Fit Error (MFE): HI should satisfy the condition of having the minimum fit error relative to the potential failure state at all observable time points to ensure that HI can clearly reflect the equipment health status. The calculation formula is as follows:

[0025]

[0026] Attribute 4: Monotonicity (MON): The health degradation of a device is an irreversible process, therefore HI should exhibit a clear monotonic trend. The calculation formula is as follows:

[0027]

[0028] Where α is a slack variable, ReLU(x) = max(0,x); if M i =1, Expected value The degradation trend should be upward; otherwise, the degradation trend should be downward.

[0029] Property 5: Convexity (CON): HI should satisfy the condition that the rate of degradation accelerates with the deterioration of health status, using the degradation coefficient. The convexity, reflecting the degenerate state, is calculated using the following formula:

[0030]

[0031] If M i =1, then the expected degradation trend should be an accelerating upward trend, that is If M i =-1, then the expected degradation trend should be an accelerated decline, that is

[0032] The aforementioned attributes are incorporated into the joint loss function to achieve unsupervised learning of the neural network branch layers, thereby enabling the construction of a High-Intensity Scale (HI) that accurately reflects the failure process under multiple failure modes. Specifically, this invention utilizes historically collected aircraft engine crew data for model training and uses the five attributes that the HI should satisfy to formulate the loss function for each branch used to construct the HI. The calculation formula is as follows:

[0033]

[0034] Wherein, λ1, λ2, λ3, and λ4 are the tuning parameters determined through cross-validation. The differences between the loss functions of different branches are attributed only to the failure mode labels, thus allowing HI to be constructed under multiple failure modes within a unified framework.

[0035] To avoid branch models ignoring the differences between different failure modes and posing challenges to subsequent failure modes, this invention proposes an adversarial training mechanism. This mechanism enables branch models to learn the differences between different failure modes, thereby establishing a failure mode-specific Hierarchy (HI). Specifically, in each iteration cycle, K branch models are trained sequentially. For the k-th branch model, all sensor data are used as model input, where the sensor data belonging to the k-th failure mode label is represented as X. i =[x i,1 ,x i,2 ,…,x i,L ], i∈φ k Used to establish "true HI" Sensor data belonging to other failure modes were used as adversarial examples to establish a "pseudo-HI" h. i '(t); True HI constructs the loss function by minimizing HI. Ensure that the constructed HI satisfies the aforementioned HI properties; conversely, a pseudo HI should belong to other failure modes and therefore should not satisfy the properties of the k-th failure mode. The difference between true and pseudo HIs becomes more apparent through iterative updates of the model parameters.

[0036] This invention uses Euclidean distance to quantify the differences in features between genuine and counterfeit HIs, and designs three features—failure threshold, degradation range, and convexity—to comprehensively characterize these differences. The specific calculation formulas are as follows:

[0037]

[0038] in, The function for calculating Euclidean distance, fea T This represents the feature vector extracted from the true HI, fea F This represents the feature vector extracted from the pseudo-HI. Both contain three components: failure threshold, degradation range, and convexity. The specific calculation formula is as follows:

[0039]

[0040] Where, N k This represents the number of engine units in the training dataset that belong to the k-th failure mode;

[0041] Thus, the adversarial loss function is constructed. The calculation formula is as follows:

[0042]

[0043] Countermeasure loss value The number of layers should be gradually increased during the iteration process; this invention employs a gradient inversion layer (GRL). Specifically, during the forward propagation process, The output is the same as the input. However, during backpropagation, The output will be multiplied by the negative gradient scaling factor, causing the gradient direction to reverse, as shown below:

[0044]

[0045] Therefore, the total loss function of the k-th branch of the multi-branch deep learning model is:

[0046]

[0047] Specifically, ρ = 1 when the input engine unit belongs to the k-th failure mode, otherwise ρ = 0; the model output passes through the gradient inversion layer only when ρ = 0. Towards The gradient direction is optimized. Therefore, the construction loss is minimized by minimizing HI. and maximizing the fight against loss Update the branch neural network. Due to the adversarial training mechanism, the branch model can learn the differences between degradation data and other failure modes to extract more reliable degradation features specific to the failure mode, thereby further improving the failure mode recognition performance.

[0048] The failure mode identification specifically includes:

[0049] For in-service aircraft engine packs, the sensor signal is represented as X. j j = 1, ..., U, where U represents the total number of generating units, using τ j This indicates the last observation time of the sensor signal of the unit.

[0050] The branch model obtained from offline adversarial training will construct the true HI for the corresponding failure mode and the pseudo HI for the other failure modes. Therefore, the sensor signals of the in-service aircraft engine with unknown failure modes are input into the multi-branch deep learning model, and the candidate HI obtained by the k-th branch model is denoted as h. j ′ ,kOf these, only one is the true HI of the unit under its actual failure mode. Therefore, the failure mode identification task is transformed into selecting the true HI that best matches the unit's degradation characteristics from the K candidate HIs.

[0051] Considering the true HI established by each branch for the historical engine unit It can accurately represent the degradation characteristics of each failure mode. Therefore, the failure mode of an in-service unit can be determined by measuring the similarity of degradation characteristics between candidate HIs and true HIs. The candidate HI with the highest similarity is determined as the true HI of that in-service unit. Specifically, features are extracted from all true HIs of historical units to construct a prototype representation P for each failure mode. k Feature vectors are obtained by extracting features from each candidate HI of the in-service aircraft crew. Euclidean distance is used to measure the feature vector of in-service aircraft units. With the prototype P in the corresponding failure mode k The failure mode k with the smallest distance based on the similarity between the two is selected as the failure mode of the in-service unit, as shown in the following formula:

[0052]

[0053] P k =[p FT, ,p DI, ,p FIT,k ,p CON,k ],

[0054]

[0055] Among them, P k) Indicates all The degenerate feature vector, This represents the degenerate feature vector of each candidate HI. This represents the function for calculating Euclidean distance.

[0056] As shown above, the degradation feature vector includes four features: the final observed value, the degradation range, the fitting error, and the convexity, specifically:

[0057] Feature 1) Final observation

[0058] Units operating under the same failure mode should have the same failure threshold; even for in-service units that have not yet experienced a failure, they will degrade to the failure threshold of their respective failure modes. Therefore, the failure threshold of historical units under each failure mode is used as a prototype feature, estimated by the average HI value at the last observation time; for in-service units, their last observation is used as the feature.

[0059]

[0060]

[0061] Where, p FT, This represents the prototype characteristic of the last observed true HI for all historical units belonging to the k-th failure mode. This represents the final observation feature of the candidate HI constructed by the k-th branch model for the in-service aircraft.

[0062] Feature 2) Degradation range

[0063] Based on attribute 2 that HI must satisfy, units with the same failure mode should have the same degradation range. This invention extracts this feature using the following formula:

[0064]

[0065] Where, p DI, This represents the true HI degradation range prototype characteristic of all historical units belonging to the k-th failure mode. The degradation range characteristics of candidate HIs constructed by the k-th branch model for in-service aircraft units.

[0066] Feature 3) Fitting error

[0067] Based on the aforementioned attribute 3 that HI must satisfy and the definition of HI, the noise items of units under the same failure mode should conform to the same distribution, and the noise items of units under different failure modes should conform to different distributions. This invention extracts this feature using the following formula:

[0068]

[0069] Where, p FIT,k This represents the prototype feature of the fitting error of the true HI for all historical units belonging to the k-th failure mode. This represents the fitting error characteristics of the candidate HI constructed for the k-th branch model for in-service aircraft units.

[0070] Feature 4) Convexity

[0071] According to attribute 5 that HI must satisfy, units under the same failure mode will experience degradation at the same rate, while different failure modes will exhibit different degradation rates. This invention extracts this feature using the following formula:

[0072]

[0073] Where, p CON,k This represents the convexity prototype feature of the true HI of all historical units belonging to the k-th failure mode. This represents the convexity feature of the candidate HI constructed for the k-th branch model for in-service aircraft units.

[0074] In the above formula, Let Hi and Ti represent the HI and Ti values ​​of engine i belonging to the k-th failure mode at time t. i Indicates the last observable time point. h represents the fitted potential degradation state of historical engine i belonging to the k-th failure mode at time t. j ′ ,k (t) represents the candidate HI value at time t obtained by the k-th branch model for in-service engine j. N represents the fitted potential degradation state at time t obtained by the k-th branch model for in-service engine j. k τ represents the number of engine units in the training dataset that belong to the k-th failure mode, t1 represents the first observable time point, and τ j This indicates the last observation time of the sensor signal of the in-service engine j. This is the degradation coefficient.

[0075] This invention, after identifying the failure modes of in-service units and obtaining the HI (Hypothesis) with their characteristics, uses historical unit failure data to update the degradation parameters of in-service units under the corresponding failure modes. This overcomes the difficulty of limited observable data for in-service units, thereby accurately predicting the RUL (Recovery Limitation Rate) of in-service units. Since the in-service units have not yet experienced a failure, T... j Let represent the failure time of unit j, and its cumulative distribution function is updated as follows:

[0076]

[0077] Where, τ j This indicates the last observation time of the sensor signal of the in-service engine j. Let HI represent the failure mode of in-service engine j as determined by the model, P(x) be the probability density function of random variable x, and M be the probability density function of random variable x. j This represents the failure threshold for the j-th unit, defined in the same way as the failure thresholds for historical units: M j ∈Ω k , ψ j (τ j ), Let be the basis functions and their parameters for the j-th unit HI, respectively. The median of this probability density function is the predicted failure time. That is, the predicted for

[0078] By adopting the technical solution provided by this invention, it is helpful to accurately describe and visualize the health status and failure process of aircraft engines, and to realize the diagnosis of aircraft engine failure modes and prediction of remaining life, which can effectively reduce the economic and social losses caused by damage to aircraft engines.

[0079] Compared with the prior art, the present invention has the following beneficial effects:

[0080] (1) The present invention can clearly visualize the degradation and evolution of the health status of aircraft engines through the constructed HI, thereby gaining a comprehensive understanding of the health status of the unit under various failure mode scenarios.

[0081] (2) Based on engineering experience, this invention has developed a unified HI structural framework with different properties tailored for multiple failure modes. The unified framework has excellent interpretability and flexibility, while ensuring that the model can accurately describe different degradation modes.

[0082] (3) The HI constructed by the adversarial training mechanism in this invention exhibits different characteristics under different failure modes, which facilitates fault diagnosis and further improves the accuracy of RUL prediction. Attached Figure Description

[0083] Figure 1 This is a flowchart of the method of the present invention;

[0084] Figure 2 The failure mode identification and RUL prediction results of the present invention are shown in one embodiment, wherein (2a) shows the failure mode identification results and (2b) shows the remaining lifetime estimation results;

[0085] Figure 3 The diagram shows the failure mode identification accuracy and RUL prediction error of HI for in-service aircraft engines, based on the HI establishment scheme and the scheme of eliminating each HI attribute respectively in one embodiment of the present invention.

[0086] Figure 4 In one embodiment, the present invention compares the failure mode diagnosis of in-service aircraft engines under conditions with and without adversarial training mechanisms. (4a) shows the failure mode identification results under the condition with adversarial training mechanism, and (4b) shows the failure mode identification results under the condition without adversarial training mechanism. Detailed Implementation

[0087] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0088] To evaluate the model's performance, this embodiment presents a case study of an aircraft gas turbine engine failure dataset. This dataset, generated by NASA's Commercial Modular Aero-Propulsion System Simulation (C-MAPSS), simulates the failure process of turbofan engines. The dataset comprises four subsets; this embodiment focuses on the FD003 subset. In this dataset, each in-service aircraft engine enters one of two potential failure modes: high-pressure compressor failure (HPC) or engine fan failure (FAN). Fourteen sensors are considered for monitoring aircraft engine failure signals: T24, T30, T50, P30, Nf, Nc, Ps30, phi, NRf, NRc, BPR, htBleed, W31, and W32. FD003 includes a training set and a test set. The training set contains sensor signals for each aircraft engine up to its failure stage, with the failure mode known; while the test set interrupts sensor signals at a point in time before failure, providing the actual remaining lifetime data after that point, with the failure mode unknown.

[0089] The following describes how the model of this invention is used to construct a High-Intensity Attribute (HI) for aircraft engines, thereby achieving failure mode identification (FMO) and remaining service life (RUL) prediction. This invention includes two aspects: offline model training and online model testing. During offline training, a multi-branch neural network is used to construct HIs specific to each failure mode for the aircraft engine. Each branch model learns the degradation characteristics of the aircraft engine under various failure modes unsupervised through a unified HI attribute framework. An adversarial training mechanism is used to enable each branch model to capture the differences between different failure modes. During online measurement, the input to the multi-branch neural network is the sensor of the in-service aircraft engine. The model generates multiple candidate HIs for the engine, and a similarity-based selection method is used to determine the unique HI that accurately describes the degradation process, thus achieving failure mode identification. Finally, the constructed HI is used to complete RUL prediction. Specifically, as shown... Figure 1 As shown, it includes the following steps:

[0090] Acquire historical aircraft engine sensor signal data, failure mode labels, and total runtime, and perform data preprocessing to construct a dataset;

[0091] The preprocessed dataset is loaded into the constructed multi-branch deep learning model, and the model is trained using an adversarial training mechanism. The multi-branch deep learning model is used to construct HIs specific to failure modes, with each branch layer constructing an HI for a specific failure mode.

[0092] The sensor signal data of the in-service engine is acquired and input into the trained multi-branch deep learning model to obtain candidate failure modes (HIs) for different failure modes. Features of each candidate HI are extracted, and failure mode identification is performed based on the extracted features to achieve RUL prediction.

[0093] The specific implementation method is as follows:

[0094] 1. Establish a neural network model

[0095] In this embodiment, aircraft engines are divided into two types: one is a historical aircraft engine, for which all sensor signals up to its failure point can be acquired and the failure mode is known; the other is an in-service aircraft engine, for which only sensor signals at a specific point in time before its failure can be acquired. The multidimensional sensor signal of a usable aircraft engine is represented as x. i (t)=(x i,1 (t),x i,2 (t),…,x i,L (t)), where x i,l (t) represents the l-th sensor signal of the i-th engine, where l = 1, ..., L, and t = t1, ..., T. i t1 represents the first observable time point, T i This represents the last observable time point. There are K failure modes in total, and each engine degrades in only one failure mode, i.e., there is a set of indices i∈φ. k , φ k ∈φ={φ k |k=1,…,K}.

[0096] This embodiment utilizes a multi-branch neural network to construct failure mode-specific hierarchies (HIs). Each branch layer focuses on constructing an HI for a single failure mode, as shown in the following formula:

[0097]

[0098] in, θ represents the shared layer function of a neural network that extracts features extensively from multidimensional sensors. S For the shared layer parameter set, This represents the extracted features obtained from the shared layer. Represents a branched neural network, θ k =[θ (1) ,…,θ (k) ,…,θ (K) [ ] represents the parameter set for each branch. This represents the HI of engine i belonging to the k-th failure mode at time t; the neural network first uses... The system learns the overall degradation characteristics, and then each branch layer learns degradation characteristics specific to different failure modes.

[0099] The Health Index (HI) is defined as the sum of potential failure states and white noise, as shown in the following formula:

[0100]

[0101] Wherein, the basis function ψ at time t i (t)=[1,t,t 2 The basis function parameter vector for describing the failure process of an aircraft engine i, belonging to the k-th failure mode, is: Estimate using the least squares method, i.e. Among them Ψ i =[ψ i (t1),…,ψ i (T i )] T Thus, the fitted potential degenerate state is obtained: This is the noise term corresponding to the k-th failure mode, assumed to have a mean of 0 and a variance of . normal distribution

[0102] To ensure that HI accurately estimates failure states, HI must satisfy the following properties:

[0103] Attribute 1: Minimum Variance (MVF) at the Failure Threshold: HI should satisfy the condition that it reaches the threshold of its failure mode at the time of failure. This invention uses the variance of the last observation point relative to the threshold as a measure, and the calculation formula is as follows:

[0104]

[0105] Among them, M i This is the failure threshold, used only to distinguish between the engine's operating condition and failure state, and to represent different degradation trends (increasing or decreasing) under different failure modes. Its absolute value has no specific meaning, so it is set to have two values: 1 or -1.

[0106] Attribute 2: Degradation Range (DR): HI should satisfy the requirement of having a clear degradation trend to characterize the degradation from a relatively healthy state to a failure state. Therefore, potential failure states should cover the interval between 0 and the failure threshold as much as possible. The calculation formula is as follows:

[0107]

[0108] Attribute 3: Minimum Fit Error (MFE): HI should satisfy the condition of having the minimum fit error relative to the potential failure state at all observable time points to ensure that HI can clearly reflect the equipment health status. The calculation formula is as follows:

[0109]

[0110] Attribute 4: Monotonicity (MON): The health degradation of a device is an irreversible process, therefore HI should exhibit a clear monotonic trend. The calculation formula is as follows:

[0111]

[0112] Where α is a slack variable, ReLU(x) = max(0,x); if M i =1, Expected value The degradation trend should be upward; otherwise, the degradation trend should be downward.

[0113] Property 5: Convexity (CON): HI should satisfy the condition that the rate of degradation accelerates with the deterioration of health status, using the degradation coefficient. The convexity, reflecting the degenerate state, is calculated using the following formula:

[0114]

[0115] If M i =1, then the expected degradation trend should be an accelerating upward trend, that is If M i =-1, then the expected degradation trend should be an accelerated decline, that is

[0116] 2. Offline Training: Constructing an adversarial training system specific to failure mode HI

[0117] The aforementioned attributes are incorporated into the joint loss function to achieve unsupervised learning of the neural network branch layers, thereby enabling the construction of a High-Intensity Scale (HI) that accurately reflects the failure process under multiple failure modes. Specifically, this invention utilizes historically collected aircraft engine crew data for model training and uses the five attributes that the HI should satisfy to formulate the loss function for each branch used to construct the HI. The calculation formula is as follows:

[0118]

[0119] Wherein, λ1, λ2, λ3, and λ4 are the tuning parameters determined through cross-validation. The differences between the loss functions of different branches are attributed only to the failure mode labels, thus allowing HI to be constructed under multiple failure modes within a unified framework.

[0120] To avoid branch models ignoring the differences between different failure modes and posing challenges to subsequent failure modes, this invention proposes an adversarial training mechanism. This mechanism enables branch models to learn the differences between different failure modes, thereby establishing a failure mode-specific Hierarchy (HI). Specifically, in each iteration cycle, K branch models are trained sequentially. For the k-th branch model, all sensor data are used as model input, where the sensor data belonging to the k-th failure mode label is represented as X. i =[x i,1 ,x i,2 ,…,x i,L ], i∈φ k Used to establish "true HI" Sensor data belonging to other failure modes were used as adversarial examples to establish a "pseudo-HI" h. i i (t); True HI constructs the loss function by minimizing HI. (i.e., formula (9)) ensures that the constructed HI satisfies the above HI properties; conversely, the pseudo HI should belong to other failure modes and therefore should not satisfy the properties of the k-th failure mode. The difference between true HI and pseudo HI becomes more apparent through iterative updates of model parameters.

[0121] This invention uses Euclidean distance to quantify the differences in features between genuine and counterfeit HIs, and designs three features—failure threshold, degradation range, and convexity—to comprehensively characterize these differences. The specific calculation formulas are as follows:

[0122]

[0123] in, The function for calculating Euclidean distance, fea T This represents the feature vector extracted from the true HI, fea F This represents the feature vector extracted from the pseudo-HI. Both contain three components: failure threshold, degradation range, and convexity. The specific calculation formula is as follows:

[0124]

[0125] Where, N k This represents the number of engine units in the training dataset that belong to the k-th failure mode.

[0126] Thus, the adversarial loss function is constructed. The calculation formula is as follows:

[0127]

[0128] Countermeasure loss value The number of layers should be gradually increased during the iteration process; this invention employs a gradient inversion layer (GRL). Specifically, during the forward propagation process, The output is the same as the input. However, during backpropagation, The output will be multiplied by the negative gradient scaling factor, causing the gradient direction to reverse, as shown below:

[0129]

[0130] Therefore, the total loss function of the k-th branch of the multi-branch deep learning model is:

[0131]

[0132] Specifically, ρ = 1 when the input engine unit belongs to the k-th failure mode, otherwise ρ = 0; the model output passes through the gradient inversion layer only when ρ = 0. Towards The gradient direction is optimized. Therefore, the construction loss is minimized by minimizing HI. and maximizing the fight against loss Update the branch neural network. Due to the adversarial training mechanism, the branch model can learn the differences between degradation data and other failure modes to extract more reliable degradation features specific to the failure mode, thereby further improving the failure mode recognition performance.

[0133] 3. Online Testing: Failure Mode Identification and RUL Prediction of In-Service Engines Based on Similarity

[0134] For in-service aircraft engine packs, the sensor signal is represented as X. j j = 1, ..., U, where U represents the total number of generating units, using τ j This indicates the last observation time of the sensor signal of the unit.

[0135] The branch model obtained from offline adversarial training will construct the true HI for the corresponding failure mode and the pseudo HI for the other failure modes. Therefore, the sensor signals of the in-service aircraft engine with unknown failure modes are input into the multi-branch deep learning model, and the candidate HI obtained by the k-th branch model is denoted as h. j j ,k Of these, only one is the true HI of the unit under its actual failure mode. Therefore, the failure mode identification task is transformed into selecting the true HI that best matches the unit's degradation characteristics from the K candidate HIs.

[0136] Considering the true HI established by each branch for the historical engine unit It can accurately represent the degradation characteristics of each failure mode. Therefore, the failure mode of an in-service unit can be determined by measuring the similarity of degradation characteristics between candidate HIs and true HIs. That is, the candidate HI with the highest similarity is determined as the true HI of that in-service unit. Specifically, features are extracted from all true HIs of historical units to construct a prototype representation P for each failure mode. k Feature vectors are obtained by extracting features from each candidate HI of the in-service aircraft crew. Euclidean distance is used to measure the feature vector of in-service aircraft units. With the prototype P in the corresponding failure mode k The failure mode k with the smallest distance based on the similarity between the two is selected as the failure mode of the in-service unit, as shown in the following formula:

[0137]

[0138] P k =[p FT, ,p DI, ,p FIT,k ,p CON,k ],(17)

[0139]

[0140] Among them, P k Indicates all The degenerate feature vector, This represents the degenerate feature vector of each candidate HI. This represents the function for calculating Euclidean distance.

[0141] As shown above, the degradation feature vector includes four features: the final observed value, the degradation range, the fitting error, and the convexity, specifically:

[0142] Feature 1) Final observation

[0143] Units operating under the same failure mode should have the same failure threshold; even for in-service units that have not yet experienced a failure, they will degrade to the failure threshold of their respective failure modes. Therefore, the failure threshold of historical units under each failure mode is used as a prototype feature, estimated by the average HI value at the last observation time; for in-service units, their last observation is used as the feature.

[0144]

[0145] Where, p FT,k This represents the prototype characteristic of the last observed true HI for all historical units belonging to the k-th failure mode. This represents the final observation feature of the candidate HI constructed by the k-th branch model for the in-service aircraft.

[0146] Feature 2) Degradation range

[0147] Based on attribute 2 that HI must satisfy, units with the same failure mode should have the same degradation range. This invention extracts this feature using the following formula:

[0148]

[0149] Where, p DI, This represents the true HI degradation range prototype characteristic of all historical units belonging to the k-th failure mode. The degradation range characteristics of candidate HIs constructed by the k-th branch model for in-service aircraft units.

[0150] Feature 3) Fitting error

[0151] Based on the aforementioned attribute 3 that HI must satisfy and the definition of HI, the noise items of units under the same failure mode should conform to the same distribution, and the noise items of units under different failure modes should conform to different distributions. This invention extracts this feature using the following formula:

[0152]

[0153] Where, p FIT,k This represents the prototype feature of the fitting error of the true HI for all historical units belonging to the k-th failure mode. This represents the fitting error characteristics of the candidate HI constructed for the k-th branch model for in-service aircraft units.

[0154] Feature 4) Convexity

[0155] According to attribute 5 that HI must satisfy, units under the same failure mode will experience degradation at the same rate, while different failure modes will exhibit different degradation rates. This invention extracts this feature using the following formula:

[0156]

[0157] Where, p CON,k This represents the convexity prototype feature of the true HI of all historical units belonging to the k-th failure mode. This represents the convexity feature of the candidate HI constructed for the k-th branch model for in-service aircraft units.

[0158] In the above formula, Let Hi and Ti represent the HI and Ti values ​​of engine i belonging to the k-th failure mode at time t. i Indicates the last observable time point. h represents the fitted potential degradation state of historical engine i belonging to the k-th failure mode at time t. j ′,k (t) represents the candidate HI value at time t obtained by the k-th branch model for in-service engine j. N represents the fitted potential degradation state at time t obtained by the k-th branch model for in-service engine j. k τ represents the number of engine units in the training dataset that belong to the k-th failure mode, t1 represents the first observable time point, and τ j This indicates the last observation time of the sensor signal of the in-service engine j. This is the degradation coefficient.

[0159] This invention, after identifying the failure modes of in-service units and obtaining the HI (Hypothesis) with their characteristics, uses historical unit failure data to update the degradation parameters of in-service units under the corresponding failure modes. This overcomes the difficulty of limited observable data for in-service units, thereby accurately predicting the RUL (Recovery Limitation Rate) of in-service units. Since the in-service units have not yet experienced a failure, T... j Let represent the failure time of unit j, and its cumulative distribution function is updated as follows:

[0160]

[0161] Where, τ j This indicates the last observation time of the sensor signal of the in-service engine j. Let HI represent the failure mode of in-service engine j as determined by the model, P(x) be the probability density function of random variable x, and M be the probability density function of random variable x. j This represents the failure threshold for the j-th unit, defined in the same way as the failure thresholds for historical units: M j ∈Ω k , ψ j (τ j ), Let be the basis functions and their parameters for the j-th unit HI, respectively. The median of this probability density function is the predicted failure time. That is, the predicted for

[0162] 4. Failure Mode Identification and RUL Prediction Result Analysis

[0163] This invention sets two metrics to evaluate model performance.

[0164] 1) For failure mode identification of in-service aircraft engines, accuracy is used as a metric to evaluate failure mode identification performance. Accuracy is defined as the ratio of engine units that correctly identify failure modes.

[0165] 2) For RUL prediction of in-service aircraft engines, the remaining life prediction error is used as a metric to evaluate RUL prediction performance. The remaining life prediction error δ for the j-th in-service aircraft engine is... j Defined as predicted remaining lifespan With the actual remaining lifespan ζ j The absolute difference between them divided by the actual failure time T j ,Right now:

[0166]

[0167] Figure 2 Figure 2a shows the failure mode identification and remaining lifetime prediction results of the present invention in one embodiment. The horizontal axis of the image represents different RUL states. "20, 40, 60, 80, 100, 125, 150" represent the failure mode identification and remaining lifetime estimation results for all in-service aircraft engines with RUL values ​​less than or equal to 20, 40, 60, 80, 100, 125, and 150. Figure 2b shows the failure mode identification results, where the vertical axis represents the average failure mode identification accuracy for each RUL state. When the in-service engine RUL is less than or equal to 100, the present invention can completely and accurately identify the failure mode. When the in-service engine RUL is greater than 100, the present invention can still guarantee 99% accuracy in identifying the failure mode. Figure 2b shows the remaining lifetime estimation results, where the bars represent the average prediction error, and the error bins represent the standard deviation of the prediction error for the corresponding RUL state. The model of the present invention can achieve good RUL prediction for aircraft engines under various RUL states, while possessing high stability and low uncertainty (i.e., a small standard deviation). Especially when the RUL is small, the stability of the model in this invention for RUL estimation is more significant. This has important practical significance for preventing sudden failures of aircraft engines.

[0168] Figure 3 The figure illustrates the failure mode identification (FMD) accuracy and range-limited error (RUL) prediction error of in-service aircraft engines, based on the present invention's scheme for establishing the HI attribute and the scheme for eliminating each HI attribute individually. As shown in the figure, eliminating any HI attribute constraint negatively impacts the model's FMD and RUL prediction accuracy. Therefore, the present invention's scheme, which retains all five HI attributes, is crucial for achieving optimal FMD and RUL prediction performance.

[0169] Figure 4This paper compares the performance of the present invention in failure mode diagnosis of in-service aircraft engines under adversarial training and non-adversarial training conditions. The horizontal and vertical axes are plotted as the relative distance between the health index similarity feature and the training prototypes for the two failure modes. Figure (4a) shows the failure mode identification results under the adversarial training mechanism, and Figure (4b) shows the failure mode identification results under the non-adversarial training mechanism. The in-service aircraft engine was diagnosed with the failure mode corresponding to its closest prototype. As shown in the figures, the adversarial training mechanism proposed in this invention can more accurately distinguish failure modes compared to the non-adversarial training mechanism.

[0170] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A method for multi-data fusion and failure mode identification based on branch adversarial networks, characterized in that, Includes the following steps: Acquire historical aircraft engine sensor signal data, failure mode labels, and total runtime, and perform data preprocessing to construct a dataset; The preprocessed dataset is loaded into the constructed multi-branch deep learning model, and the model is trained using an adversarial training mechanism. The multi-branch deep learning model is used to construct HIs specific to failure modes, with each branch layer constructing an HI for a specific failure mode. The sensor signal data of the in-service engine is acquired and input into the trained multi-branch deep learning model to obtain candidate HIs corresponding to different failure modes. The features of each candidate HI are extracted, and failure mode identification is performed based on the extracted features, thereby realizing RUL prediction. In the dataset, the multidimensional sensor signals of a certain available aircraft engine are represented as follows: ,in Indicates the first The first engine One sensor signal, , , Indicates the first observable point in time. Indicates the last observable time point; total There are multiple failure modes, and each engine degrades in only one failure mode, i.e., there is a set of subscripts. ; The multi-branch deep learning model is represented as follows: in, This represents a shared layer function of a neural network that extracts features extensively from multidimensional sensors. For the shared layer parameter set, This represents the extracted features obtained from the shared layer. This represents a branched neural network. For each branch parameter set, Indicates belonging to the first Engines with various failure modes exist The HI time; the neural network first through Learn the overall degradation characteristics, and then each branch layer learns degradation characteristics specific to different failure modes. The HI is defined as a potential failure state plus a white noise term, as shown in the following formula: Among them, time basis functions Used to describe the failure process of an aircraft engine, and belongs to the first... Aircraft engines with various failure modes The basis function parameter vector is The least squares method is used to estimate that... in Thus, the fitted potential degenerate state is obtained: ; It is the first The noise term corresponding to each failure mode is assumed to have a mean of 0 and a variance of . normal distribution .

2. The method for multi-data fusion and failure mode identification based on branch adversarial networks according to claim 1, characterized in that, To ensure that HI accurately estimates failure states, HI must satisfy the following properties: 1) Minimum variance at the failure threshold: HI satisfies the condition of reaching the threshold of its failure mode at the time of failure. The variance between the last observed point and the threshold is used as the measure, and the calculation formula is as follows: in, This is the failure threshold, used only to distinguish between the engine's operating condition and failure state, and to represent different degradation trends under different failure modes. It has two possible values: 1 or -1. 2) Degradation Scope: The HI should meet the requirement of having a clear degradation trend to characterize the degradation from a relatively healthy state to a failure state. Therefore, it should cover as many potential failure states as possible. The interval between the failure threshold and the failure rate is calculated using the following formula: 3) Minimum Fitting Error: HI should satisfy the condition of having the minimum fitting error relative to the potential failure state at all observable time points, to ensure that HI can clearly reflect the health status of the equipment. The calculation formula is as follows: 4) Monotonicity: Equipment health degradation is an irreversible process, therefore HI should exhibit a clear monotonic trend. The calculation formula is as follows: in, It is a slack variable. ;if ,expect The degradation trend is upward; otherwise, the degradation trend should be downward. 5) Convexity: HI should satisfy the condition that the rate of degradation accelerates with the deterioration of health status, utilizing the degradation coefficient. The convexity, reflecting the degenerate state, is calculated using the following formula: if Therefore, the expected degradation trend should be an accelerating upward trend, that is... ;if Therefore, the expected degradation trend should be an accelerated decline, i.e. .

3. The method for multi-data fusion and failure mode identification based on branch adversarial networks according to claim 2, characterized in that, The loss function used to construct HI is determined by using the five properties that HI should satisfy. The calculation formula is as follows: in, These are the tuning parameters determined through cross-validation.

4. The method for multi-data fusion and failure mode identification based on branch adversarial networks according to claim 3, characterized in that, The adversarial training mechanism is specifically as follows: in each iteration cycle... The branch model is trained sequentially, for the first branch... The model has several branches, with all sensor data used as input. The data belonging to the first branch... Sensor data for each failure mode label is represented as follows: = , Used to establish "true HI" Sensor data belonging to other failure modes are used as adversarial examples to establish "fake HI" (hypothetical high-incidence) models. True HI constructs the loss function by minimizing HI. Ensure that the constructed HI satisfies the above HI properties; The differences in features between genuine and counterfeit HIs are quantified using Euclidean distance, and three features—failure threshold, degradation range, and convexity—are designed to comprehensively characterize these differences. The specific calculation formulas are as follows: in, This represents the feature vector extracted from the true HI. This represents the feature vector extracted from the pseudo-HI. Both contain three components: failure threshold, degradation range, and convexity. The specific calculation formula is as follows: in, Indicates that the data in the training dataset belongs to the first... Number of engine units with each failure mode Indicates the first observable point in time; Thus, the adversarial loss function is constructed. The calculation formula is as follows: Then, the first branch of the multi-branch deep learning model The total loss function for the branch model is: Among them, when the input engine unit belongs to the first When there are multiple failure modes, ,otherwise Only when The model output is passed through a gradient reversal layer ℵ(∙) to... Gradient direction optimization; where, during forward propagation, The output is the same as the input. During backpropagation, The output will be multiplied by the negative gradient scaling factor, causing the gradient direction to reverse, as shown below: By minimizing the HI construction loss and maximizing the fight against loss Update the branch neural network.

5. The method for multi-data fusion and failure mode identification based on branch adversarial networks according to claim 1, characterized in that, The failure mode identification specifically includes: For in-service aircraft engine packs, sensor signals are represented as: , ,in To indicate the total number of generating units, use This indicates the last observation time of the unit's sensor signals; The sensor signals from in-service aircraft engines with unknown failure modes are input into a multi-branch deep learning model, which is then processed by the first branch... The candidate HI obtained from the branch model are denoted as... ; The failure modes of in-service units are determined by measuring the degradation feature similarity between candidate failure modes (HIs) and true failure modes (HIs) established from historical units under each failure mode. The candidate HI with the highest similarity is identified as the true failure mode (HI) of that in-service unit. Specifically, features are extracted from all true failure modes of historical units to construct a prototype representation for each failure mode. ; Feature vectors are obtained by extracting features from each candidate HI of the in-service crew. The Euclidean distance is used to measure the feature vector of in-service units. Prototypes in the corresponding failure modes The similarity between them, the failure modes with the minimum distance The formula for selecting the failure mode for this in-service unit is as follows: in, Indicates all The degenerate feature vector, This represents the degenerate feature vector of each candidate HI. .

6. The method for multi-data fusion and failure mode identification based on branch adversarial networks according to claim 5, characterized in that, The degradation feature vector comprises four features: the final observed value, the degradation range, the fitting error, and the convexity, specifically represented as follows: in, Indicates belonging to the first The prototype characteristics of the last observed true HI of all historical units for each failure mode. Indicates by the first Each branch model represents the final observed features of candidate HIs constructed for in-service aircraft units. Indicates belonging to the first The true HI degradation range prototype characteristics of all historical units for each failure mode. By the The branch model constructs the degradation range characteristics of candidate HIs for in-service units. Indicates belonging to the first The prototype characteristics of the fitting error of the true HI of all historical units for each failure mode. Indicates the first The fitting error characteristics of the candidate HI constructed by the branch model for in-service units. Indicates belonging to the first The true convexity prototype characteristics of the HI of all historical units for each failure mode. Indicates the first The convexity features of candidate HIs constructed by the branch model for in-service units. in, Indicates belonging to the first Engines with various failure modes exist HI at all times Indicates the last observable time point. Indicates belonging to the first Historical Engines with Various Failure Modes exist The potential degenerate state of the fit at time t, This indicates the status of engines currently in service. By the The branch model yields the following: Candidate HI value at time 1, This indicates the status of in-service engines. By the The branch model yields the following: The potential degenerate state of the fit at time t, Indicates that the data in the training dataset belongs to the first... Number of engine units with each failure mode Indicates the first observable point in time. Indicates the engine in service The last observation time of the sensor signal, This is the degradation coefficient.

7. The method for multi-data fusion and failure mode identification based on branch adversarial networks according to claim 1, characterized in that, The RUL prediction specifically involves: based on the identified failure modes of in-service units and the HI (Hypothesis) characteristic of those failure modes, updating the degradation parameters of in-service units under the corresponding failure modes using historical unit failure data, and predicting the RUL of the in-service units. Indicates the unit The failure time, and its cumulative distribution function update are as follows: in, Indicates the engine in service The last observation time of the sensor signal, This indicates that the model determines it belongs to the first... In-service engines with various failure modes HI, For random variables The probability density function, Indicates the first The failure threshold for each unit is defined in the same way as the historical failure threshold for units: , The first The basis functions and parameters of the individual unit HI are used, and the median of this probability density function is the predicted failure time. That is, the predicted RUL for - 。

Images