Aero-engine early fault detection method and device based on multi-resolution consensus and storage medium

Abstract

The application discloses an aero-engine early fault detection method and device based on multi-resolution consensus and a storage medium. The method comprises the following steps: acquiring multi-channel sensor time series data and preprocessing; generating multiple views with different resolutions through time domain downsampling and / or frequency domain masking; inputting each view into a shared encoder to extract features, and then outputting fault prediction probabilities of each view by a classifier; in the detection model training stage, a multi-layer consensus loss function containing a first consensus loss and a second consensus loss is constructed, and the model is jointly trained in combination with a supervised loss; in the detection stage, integrated probabilities are obtained by adaptively fusing the prediction results of each view, and accurate determination of the fault starting point is realized based on a persistent threshold mechanism. Through multi-level consensus constraints and adaptive fusion, the stability and accuracy of early fault detection under long health periods and complex working conditions are significantly improved.

Description

Technical Field

[0001] This invention relates to the field of aero-engine health management technology, specifically to an aero-engine early fault detection method, device, and storage medium based on multi-resolution learning and consensus mechanisms. Background Technology

[0002] As a high-value, high-reliability complex system, early fault detection of aero-engines is crucial for flight safety and maintenance costs. Existing deep learning-based fault detection methods, such as LSTM or single-scale CNN, face severe challenges when processing long-term health data of aero-engines: over hundreds or even thousands of operating cycles, fault signs are extremely weak and easily masked by complex variations in operating conditions (such as changes in engine speed and load), resulting in long detection delays and high false positive rates.

[0003] In recent years, multi-view learning has been introduced into the field of fault diagnosis, aiming to improve model robustness by analyzing multiple perspectives (views) of data. For example, existing techniques generate multiple views by performing different transformations (such as wavelet transform and Fourier transform) on the original time-series data, train classifiers independently, and finally integrate the results through voting or other methods. However, such methods have significant drawbacks: First, each view model is trained independently, lacking active constraints on consistency between views during the learning process. When variations in operating conditions lead to increased differences in the distribution of data across different views, the independent models are prone to making contradictory judgments, and even after later integration, decision stability cannot be guaranteed. Second, existing methods typically perform simple result fusion at the model output layer, failing to delve into the feature representation level to force the model to learn essential features that are robust to changes in operating conditions and sensitive to faults. Finally, traditional methods treat all views as equally important, failing to dynamically adjust the weights of views based on data quality or fault modes.

[0004] Therefore, how to design a method that can proactively promote multi-view consensus at the feature learning and decision-making levels, thereby achieving stable and accurate early fault detection over long health periods and under various operating conditions, has become an urgent technical problem to be solved. Summary of the Invention

[0005] The purpose of this invention is to address the problems of detection delay and poor stability in existing aero-engine fault detection methods under long health periods and varying operating conditions. It provides an aero-engine early fault detection method, device, and storage medium based on multi-resolution consensus that can promote multi-view consensus from the deep feature level and achieve adaptive fusion, so as to significantly improve the accuracy and robustness of fault initiation point detection.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] In a first aspect, the present invention provides a method for early fault detection of aero-engines based on multi-resolution consensus, comprising:

[0008] Acquire and preprocess multi-channel sensor timing data of aero-engines;

[0009] For the preprocessed time-series data segment, K views with different resolutions are generated by at least one of time-domain downsampling and frequency-domain masking, where K is an integer greater than 1;

[0010] The K views are input into a pre-trained detection model, which includes a shared encoder and a classifier; wherein the shared encoder is used to extract features from the input views, and the classifier is used to output a fault prediction probability based on the features extracted by the shared encoder.

[0011] The detection model is trained through the following steps:

[0012] A multi-layer consensus loss function is constructed, which includes a first consensus loss and a second consensus loss. The first consensus loss is calculated based on the features output by the intermediate network layer of the shared encoder for K views of the same sample, and is used to constrain the consistency of different views in the feature space. The second consensus loss is calculated based on the fault prediction probability output by the classifier for K views of the same sample, and is used to constrain the consistency of different views in the decision space.

[0013] The detection model is jointly trained based on the multi-layer consensus loss function and the supervision loss for fault classification of each view; the supervision loss is used to measure the difference between the fault prediction results of each view and the true label.

[0014] Based on the fault prediction probabilities of the K views output by the detection model, the detection results related to the fault state are obtained.

[0015] Preferably, generating K views with different resolutions includes: dynamically determining the sampling rate of the time-domain downsampling and / or the mask range of the frequency-domain mask based on the statistical characteristics of the currently input time-series data segment through a view policy network.

[0016] Preferably, the first consensus loss is obtained by calculating the cosine distance or mean square error between the feature vectors output by the K views of the same sample at a specified intermediate network layer; the second consensus loss is obtained by calculating the KL divergence or JS divergence between the fault prediction probability distributions of the K views of the same sample.

[0017] Preferably, obtaining the detection results related to the fault state includes:

[0018] Based on the detection model, the view fault prediction probability of each of the K views is obtained;

[0019] The fusion weight of each view is dynamically calculated based on the confidence level of the failure prediction probability of the K views.

[0020] The integrated fault prediction probability is obtained by weighting the fault prediction probabilities of the K views according to the fusion weight.

[0021] Early fault status determination is based on the integrated fault prediction probability.

[0022] Preferably, the confidence level is measured by the entropy or variance of the probability distribution of the view fault prediction; the fusion weight is obtained by Softmax normalization of the reciprocal of the confidence level.

[0023] Preferably, the early fault state determination based on the integrated fault prediction probability includes fault initiation point detection:

[0024] The continuous time-series data of the sensor to be detected are input into the detection model in sequence to obtain an integrated fault prediction probability set arranged in time order, forming an integrated fault prediction probability sequence.

[0025] A persistent threshold mechanism is applied to determine the integrated fault prediction probability sequence;

[0026] When the integrated fault prediction probability value of M consecutive time points in the integrated fault prediction probability sequence exceeds a set threshold, the starting point of the consecutive time period is determined as the fault initiation point.

[0027] Preferably, the method further includes: after determining the fault initiation point, predicting the remaining service life of the operating data after the fault initiation point.

[0028] Preferably, the shared encoder is a Transformer encoder, a temporal convolutional network, or a recurrent neural network.

[0029] In a second aspect, the present invention provides an early fault detection device for aero-engines based on multi-resolution consensus, comprising a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the method described in the first aspect.

[0030] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the method described in the first aspect.

[0031] Beneficial effects:

[0032] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0033] 1. This invention applies consensus constraints simultaneously to the intermediate network layer (feature space) of the shared encoder and the output layer (decision space) of the classifier, forcing the model to learn essential features and decision boundaries that are insensitive to multi-resolution views of simulated operating conditions but sensitive only to faults, thereby fundamentally improving the stability and robustness of the model under complex and variable operating conditions.

[0034] 2. This invention introduces dynamic fusion weights based on prediction confidence, enabling the model to autonomously assess the reliability of each view during inference, assigning higher weights to views with high confidence, achieving more intelligent and accurate decision fusion, and further improving detection accuracy.

[0035] 3. This invention combines an integrated fault prediction probability sequence with a persistent threshold mechanism, which can effectively smooth out probability fluctuations caused by noise, accurately and stably locate the fault initiation time, and provide high-quality input for subsequent remaining service life prediction.

[0036] 4. The entire framework (view generation, shared encoding, consensus constraints, classification and fusion) can be jointly trained end-to-end, ensuring consistent optimization goals and avoiding suboptimal problems caused by staged training, thus fully leveraging the advantages of deep learning. Attached Figure Description

[0037] Figure 1 A flowchart of the early fault detection method for aero-engines based on multi-resolution consensus provided in this embodiment of the invention;

[0038] Figure 2 This is an exemplary schematic diagram of the layout of aero-engine sensors in an embodiment of the present invention;

[0039] Figure 3 This is a schematic diagram illustrating the generation of a coarse-grained view through time-domain downsampling in an embodiment of the present invention;

[0040] Figure 4 This is a schematic diagram of the frequency domain mask generation filtering view in an embodiment of the present invention. Detailed Implementation

[0041] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0042] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, in the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0043] It should be noted that, unless otherwise defined, the technical or scientific terms used in this invention should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar words used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components.

[0044] This invention provides a method, apparatus, and storage medium for early fault detection in aero-engines based on multi-resolution consensus. Taking the FD001 subset of the NASA C-MAPSS (Commercial Modular Aero-Propulsion System Simulation) dataset, a publicly available benchmark dataset in the field of aero-engine health monitoring, as an example, the specific implementation of this invention is described in detail. The FD001 subset contains multivariate sensor data from 100 engines operating under a single fault mode and a single operating condition until failure, suitable for validating early fault detection algorithms.

[0045] Method Implementation Examples

[0046] This embodiment details the specific implementation steps of the early fault detection method for aero-engines based on multi-resolution consensus provided by the present invention. For example... Figure 1 As shown, it includes the following steps:

[0047] S101: Data Acquisition and Preprocessing

[0048] Specifically, the first step is to acquire multi-channel sensor time-series data of the aero-engine. In this embodiment, 21 sensor channels (such as total temperature, pressure, speed, fuel flow, etc.) related to the degradation of core mechanical components of the engine are selected from the C-MAPSSFD001 dataset. Each data record represents an operating cycle, and the data for each engine constitutes a time series from healthy operation to failure. For example... Figure 2As shown, it illustrates a specific example of an aircraft engine sensor layout, in which sensors S1, S2, S3 and S4 are arranged at corresponding locations of the aircraft engine's fan, compressor, combustion chamber, turbine and exhaust nozzle to detect fan speed, compressor temperature, turbine pressure and exhaust temperature, respectively.

[0049] Then, the acquired raw time-series data is preprocessed, including:

[0050] Missing value handling: Linear interpolation is used to fill in individual missing sensor readings.

[0051] Standardization: Z-score standardization is performed on the data from each sensor channel to ensure a mean of 0 and a standard deviation of 1. For a time series of length n, x = (x1, x2, ..., xn), the standardization formula is:

[0052]

[0053] in, This is the mean of the sensor sequence. Standard deviation, This is the standardized value.

[0054] Sliding window slicing: To construct training samples, a sliding window is used to cut long sequences into segments of fixed length. The detection window length L = 30 (cycles) and the sliding step size S = 1. For each window segment, a binary label is assigned based on whether its ending cycle exceeds a preset fault start cycle (which needs to be estimated through RUL labels or expert annotation; in this embodiment, the RUL decay segment start point provided by the dataset is used): healthy (0) or faulty (1). Finally, a training dataset containing N samples is obtained. ,in, To represent a sample, Label them as follows: 0 for healthy, 1 for faulty.

[0055] S102: Adaptive Multi-Resolution View Generation

[0056] This step generates K views with different resolutions from the preprocessed time-series data segments by using at least one of time-domain downsampling and frequency-domain masking, where K is an integer greater than 1.

[0057] Specifically, for each time-series data segment after preprocessing K views with different resolutions are generated. In this embodiment, K=3. The view generation combines temporal downsampling and frequency domain masking, and the parameters are adaptively determined by a lightweight view policy network.

[0058] The three views with different resolutions generated in this embodiment are defined as follows:

[0059] See Figure 1 (V1): Original resolution view. That is, the original time-series data segment. .

[0060] See Figure 2 (V2): Coarse-grained time-domain view. (This is achieved through...) It is obtained by temporal downsampling in the time dimension. The downsampling rate r is dynamically determined by the policy network.

[0061] See Figure 3 (V3): Frequency domain filtered view. For Each sensor channel undergoes a Discrete Fourier Transform (DFT) to mask out the high-frequency components, followed by an inverse DFT to reconstruct the signal. The proportion of low-frequency components retained is then preserved. The frequency domain mask (i.e., the frequency domain mask) is dynamically determined by the policy network.

[0062] The view policy network described in this embodiment is a two-layer fully connected neural network, with the current time-series data segment as its input. The statistical characteristics (such as mean, variance, and dominant frequency of each channel) are output as two scalar parameters. and After mapping, the actual parameters are obtained: downsampling rate. (i.e., r can be 2, 3, or 4), retaining the proportion of low frequencies. (Right now (Within the range of [0.3, 0.9]). By introducing a policy network, the sampling rate of the time-domain downsampling and / or the mask range of the frequency-domain mask are dynamically determined based on the statistical characteristics of the currently input time-series data segment.

[0063] The generation process is as follows: For V2: By sampling rate in the time dimension Perform average pooling to obtain For V3: Each column (sensor channel) Performing the DFT (Discrete Fourier Transform) yields the frequency domain representation. Preserve low-frequency components (previous) ( ), set the high-frequency components to zero, and obtain Then perform an inverse DFT (IDFT) to reconstruct the signal back to the time domain. The result is obtained by splicing together all channels after processing. .

[0064] To more intuitively illustrate the generation mechanism of multi-resolution views in the embodiments of the present invention, the following detailed explanation is provided in conjunction with the accompanying drawings.

[0065] Temporal downsampling generates coarse-grained views:

[0066] like Figure 3 As shown, the horizontal axis represents time (operational cycle), and the vertical axis represents the normalized amplitude of the sensor data. The original signal curve represents a preprocessed, standardized time-series data segment acquired from a certain sensor channel, which has high resolution in the time dimension. To simulate low-temporal-resolution observation scenarios caused by reduced data acquisition frequency or gradual changes in operating conditions, this embodiment of the invention employs temporal downsampling to generate a coarse-grained view. Specifically, the original signal is resampled or averaged at a fixed sampling rate r (e.g., r=2). As shown in the downsampled signal curve in the figure, the processed signal retains the basic shape and trend of the original signal while significantly reducing the number of data points and lowering the temporal resolution. This coarse-grained view helps the model learn features that are insensitive to high-frequency noise and focus more on long-term trends, making it an effective means of simulating the slow changes in signal characteristics over a long healthy period.

[0067] Frequency domain mask generation filter view:

[0068] like Figure 4 As shown, the horizontal axis represents frequency (Hz), and the vertical axis represents normalized amplitude. The full-spectrum curve represents the amplitude spectrum obtained after performing a discrete Fourier transform on the original signal. It can be seen that the signal energy is distributed within a certain frequency band. To simulate frequency domain observation scenarios caused by limitations in sensor frequency response characteristics or by focusing on specific fault frequency bands, this embodiment of the invention employs frequency domain masking. Specifically, a low-pass mask function is defined, the effect of which is shown in the frequency mask (low-pass) curve in the figure: high-frequency components above the preset cutoff frequency (as shown in the transition region in the figure) are significantly suppressed or zeroed out, while low-frequency components are completely preserved. Performing an inverse Fourier transform on the masked spectrum reconstructs a time-domain signal that filters out high-frequency noise and irrelevant fluctuations, focusing on low-frequency and main harmonic components—the filtered view. This view effectively enhances the low-frequency characteristics associated with the slow degradation of engine mechanical components and suppresses high-frequency random interference.

[0069] The coarse-grained view and the filtered view generated through the two methods described above, together with the original resolution view, constitute the K views with different resolutions defined in this embodiment of the invention. These views differentiate the same operating state from different physical dimensions (temporal domain resolution, frequency domain components), providing rich and complementary information sources for subsequent shared encoders and multi-layer consensus learning. This is an important foundation for the model to achieve stability and early fault detection under varying operating conditions.

[0070] S103: Construction and Training of Multi-Level Consensus Detection Model

[0071] First, a detection model is constructed, with a shared encoder and classifier at its core, and a multi-layer consensus loss is introduced during training.

[0072] Shared encoder: It consists of multiple sequentially connected encoding layers. This embodiment uses a 4-layer Transformer encoder as the shared encoder, with each layer containing a multi-head self-attention mechanism and a feedforward neural network. It should be noted that the use of a multi-layer Transformer structure as the shared encoder in this embodiment is only an example; the core function of the shared encoder is to extract effective features from multi-resolution views, and any deep learning model that can achieve this function (such as Temporal Convolutional Network (TCN), Gated Recurrent Unit (GRU), etc.) is applicable.

[0073] This shared encoder is used to extract features from the K views generated in step S102 that are input to it. Specifically, for views of different lengths (e.g., V2), they are first adjusted to a uniform feature dimension (e.g., 128 dimensions) through linear projection. After adding positional encoding, the data is input into the shared encoder, which ultimately outputs a context-aware feature representation sequence for each view. To obtain a fixed-length feature vector, global average pooling (GAP) is performed on the output sequence along the time dimension to obtain a feature representation that aggregates information from the entire sequence. .

[0074] The classifier consists of a linear layer followed by a softmax activation function. It represents the features. Mapped to a 2D vector, the output represents the fault prediction probability for each view. ,in Let be the probability that the sample belongs to the fault category (label 1). This represents the probability that the sample belongs to the health category (label 0).

[0075] During the training phase of the detection model, a multi-layer consensus loss function is constructed. In this embodiment, the multi-layer consensus loss function includes a first consensus loss and a second consensus loss. Specifically:

[0076] First consensus loss (feature layer consensus, The calculation of features from the output of an intermediate network layer in a shared encoder for K views based on the same sample is described. The intermediate network layer refers to a layer located between the encoder's input layer and the final output layer, capable of learning high-level abstract semantic features of the data. By imposing consistency constraints on this layer, the model can be forced to learn essential features insensitive to view differences. This embodiment specifies calculation at the output of the second layer of the Transformer encoder. This layer is the intermediate network layer selected in this embodiment because it can better balance the level of feature abstraction and task relevance, making it an effective location for imposing feature consistency constraints between views. Of course, other suitable intermediate network layers can be selected based on model depth and specific tasks. Indicates the first The view in the first layer( =2) output feature vectors. The first consensus loss measures the inconsistency of the feature space by calculating the sum of the cosine distances between the feature vectors output by the K views of the same sample at a specified intermediate network layer (layer 2):

[0077]

[0078] in, Represents cosine similarity. Minimize This can be used to constrain the consistency of different views in the feature space. The first consensus loss can also be obtained by calculating the mean square error between feature vectors.

[0079] Second consensus loss (decision-making consensus, The second consensus loss is calculated based on the fault prediction probabilities of K views from the same sample in the classifier output. The second consensus loss can be obtained by calculating the KL divergence or JS divergence between the fault prediction probability distributions of the K views from the same sample. This embodiment uses KL divergence to measure the fault prediction probability distribution of each view. Differences between them:

[0080]

[0081] in, The mean distribution of the predicted probability of failure for each view. for arrive The KL divergence. Minimize It can be used to constrain the consistency of different views in the decision space.

[0082] Supervised Loss: A supervised loss is set for fault classification, which measures the difference between the fault prediction results and the true labels for each view; this embodiment uses cross-entropy loss. The supervised loss is used for fault classification. Assuming a training batch contains B samples, the loss is calculated as follows:

[0083]

[0084] in, It is the first The true label (0 or 1) of each sample. Model for the first The first sample The probability of failure predicted by each view.

[0085] It should be noted that the use of cross-entropy loss as the supervised loss for fault classification in this embodiment is only a preferred implementation. It should be understood that any loss function that can measure the difference between the model's predicted probability and the true label can be used to implement the above-mentioned supervised loss, such as Focal Loss (suitable for class imbalance scenarios) and Hinge Loss (suitable for maximum margin classification). These can be selected and used according to the specific application scenario, as long as they can effectively drive the model to learn the fault classification task.

[0086] Joint Training: The detection model is jointly trained based on a multi-layer consensus loss function and the supervised loss for fault classification of each view. The total loss of joint training is a weighted sum of the cross-entropy loss, the first consensus loss, and the second consensus loss.

[0087]

[0088] In this embodiment, hyperparameters are set. , .

[0089] Then use the Adam optimizer with an initial learning rate. Batch size 32, in the training dataset Train the model for 50 epochs until the loss converges. After training, you will obtain the trained detection model.

[0090] S104: Adaptive View Fusion and Fault Detection

[0091] This step inputs the time-series data of the aircraft engine sensors to be detected into the trained detection model. Based on the fault prediction probabilities of the K views output by the detection model, the detection results related to the fault state are obtained. Specifically,

[0092] First, the time-series data segment to be detected undergoes the same preprocessing and view generation as in steps S101 and S102. Then, it is fed into the detection model trained in step S103. Based on the detection model, the view fault prediction probability of each of the K views is obtained. .

[0093] Then, based on the confidence levels of the fault prediction probabilities of the K views, the fusion weight for each view is dynamically calculated. This embodiment uses the entropy of the fault prediction probability distribution. To measure uncertainty, the lower the entropy, the higher the confidence level. Additionally, confidence can also be measured by the variance of the failure probability distribution of the view. If variance is used, confidence can be defined as the reciprocal of the variance of the prediction probability; the smaller the variance, the higher the confidence. The fusion weight of each view is obtained by Softmax normalization of the reciprocal of the confidence level. Fusion weights The calculation is as follows:

[0094]

[0095] in, The temperature parameter is used to adjust the sharpness of the weight distribution. In this embodiment, it is set as follows: .

[0096] Furthermore, the integrated fault prediction probability is obtained by weighting the fault prediction probabilities of the K views according to the fusion weights. Specifically, the integrated fault prediction probability... .

[0097] Finally, an early fault state determination is made based on the integrated fault prediction probability. Specifically, take... Failure probability in As the integrated fault prediction probability of the current time series data segment. If If so, it can be determined that the segment is in a faulty state.

[0098] S105: Fault Initiation Point Detection

[0099] To achieve more accurate early warning, early fault status determination is based on integrated fault prediction probability, and further includes fault initiation point detection, which specifically includes the following steps:

[0100] Generate probability sequence: For complete operating data (from healthy to faulty) of an engine to be monitored, repeat step S104 sequentially. Specifically, input continuous time-series data of the sensors to be monitored into the detection model in sequence, and slide the detection window (length L) defined in step S101 on the time axis at fixed time intervals (e.g., each operating cycle). For each window position, execute step S104 to calculate the integrated fault prediction probability corresponding to that window. Arranging all these probability values ​​in order of their corresponding window center time point (or start time point) yields the integrated fault prediction probability set arranged in chronological order, forming the integrated fault prediction probability sequence. , The sequence length depends on the total number of cycles, the window length L, and the sliding step size.

[0101] A persistent threshold mechanism is applied: the above integrated fault prediction probability sequence is evaluated using a persistent threshold mechanism. A probability threshold is set. The duration M = 5 (cycles).

[0102] Determining the starting point: When the integrated fault prediction probability value at M consecutive time points in the integrated fault prediction probability sequence exceeds a set threshold, the starting point of that consecutive time period is determined as the fault initiation point. That is, finding the point that satisfies... The earliest time point ,but This refers to the fault start point of the detected fault start cycle.

[0103] S106: Remaining useful life prediction

[0104] After determining the fault initiation point, the remaining service life (RUL) is predicted based on the operating data after the fault initiation point. Specifically, an independent RUL prediction model (such as another regression neural network) can be used, taking the detected fault initiation point as the time origin and using the sensor data sequence after the initiation point as input, to predict the number of remaining operating cycles of the engine from the current moment until failure.

[0105] The proposed method was validated on the C-MAPSS FD001 test set. Compared to the baseline LSTM model and the traditional multi-view ensemble model, the ensemble fault prediction probability sequence output by the proposed method is more stable during the healthy period (low false alarm rate), and rises more sharply and exceeds the threshold earlier in the fault initiation stage, exhibiting shorter detection latency and higher stability. Quantitative results show that the proposed method achieves an F1 score of 94.5% for fault initiation point detection, with an average detection latency of 9.8 cycles, significantly outperforming traditional methods.

[0106] Device and Medium Examples

[0107] This invention also provides an early fault detection device for aero-engines based on multi-resolution consensus. This device can be a computer, server, or dedicated hardware deployed in an airborne health management unit or ground monitoring center. The device includes a memory and a processor. The memory (such as ROM, RAM, or hard disk) stores a computer program capable of implementing the above-described method steps. When the processor (such as CPU or GPU) executes the computer program, it implements the early fault detection method described in S101 to S106.

[0108] The present invention also provides a computer-readable storage medium, such as a USB flash drive, optical disc, solid-state drive, or server storage space. The medium stores a computer program that, when read and executed by one or more processors, causes the processors to perform the early fault detection method described in S101 to S106 above.

[0109] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, various modifications and variations can be made to the embodiments of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for early fault detection in aero-engines based on multi-resolution consensus, characterized in that, include: Acquire and preprocess multi-channel sensor timing data of aero-engines; For the preprocessed time-series data segment, K views with different resolutions are generated by at least one of time-domain downsampling and frequency-domain masking, where K is an integer greater than 1; The K views are input into a pre-trained detection model, which includes a shared encoder and a classifier; wherein the shared encoder is used to extract features from the input views, and the classifier is used to output a fault prediction probability based on the features extracted by the shared encoder. The detection model is trained through the following steps: A multi-layer consensus loss function is constructed, which includes a first consensus loss and a second consensus loss. The first consensus loss is calculated based on the features output by the intermediate network layer of the shared encoder for K views of the same sample, and is used to constrain the consistency of different views in the feature space. The second consensus loss is calculated based on the fault prediction probability output by the classifier for K views of the same sample, and is used to constrain the consistency of different views in the decision space. The detection model is jointly trained based on the multi-layer consensus loss function and the supervision loss for fault classification of each view; the supervision loss is used to measure the difference between the fault prediction results of each view and the true label. Based on the fault prediction probabilities of the K views output by the detection model, the detection results related to the fault state are obtained.

2. The method according to claim 1, characterized in that, The generation of K views with different resolutions includes: dynamically determining the sampling rate of the time-domain downsampling and / or the mask range of the frequency-domain mask based on the statistical characteristics of the currently input time-series data segment through a view policy network.

3. The method according to claim 1, characterized in that, The first consensus loss is obtained by calculating the cosine distance or mean square error between the feature vectors output by the K views of the same sample at a specified intermediate network layer; the second consensus loss is obtained by calculating the KL divergence or JS divergence between the fault prediction probability distributions of the K views of the same sample.

4. The method according to claim 1, characterized in that, The obtained detection results related to the fault state include: Based on the detection model, the view fault prediction probability of each of the K views is obtained; The fusion weight of each view is dynamically calculated based on the confidence level of the failure prediction probability of the K views. The integrated fault prediction probability is obtained by weighting the fault prediction probabilities of the K views according to the fusion weight. Early fault status determination is based on the integrated fault prediction probability.

5. The method according to claim 4, characterized in that, The confidence level is measured by the entropy or variance of the probability distribution of the failure prediction of the view; the fusion weight is obtained by Softmax normalization of the inverse of the confidence level.

6. The method according to claim 4 or 5, characterized in that, The early fault state determination based on the integrated fault prediction probability includes fault initiation point detection: The continuous time-series data of the sensor to be detected are input into the detection model in sequence to obtain an integrated fault prediction probability set arranged in time order, forming an integrated fault prediction probability sequence. A persistent threshold mechanism is applied to determine the integrated fault prediction probability sequence; When the integrated fault prediction probability value of M consecutive time points in the integrated fault prediction probability sequence exceeds a set threshold, the starting point of the consecutive time period is determined as the fault initiation point.

7. The method according to claim 6, characterized in that, The method further includes: after determining the fault initiation point, predicting the remaining service life of the operating data after the fault initiation point.

8. The method according to claim 1, characterized in that, The shared encoder is a Transformer encoder, a temporal convolutional network, or a recurrent neural network.

9. An early fault detection device for aero-engines based on multi-resolution consensus, characterized in that, It includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps of the method according to any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 8.

Images