Aero-engine performance degradation path evolution, model training method and device

Abstract

The application provides an aero-engine performance degradation path evolution and model training method and device, which is applied to the technical field of aero-engine health management. The method comprises the following steps: obtaining full-life data of an aero-engine; inputting the full-life data of the aero-engine into a pre-trained aero-engine performance degradation path evolution model, wherein the aero-engine performance degradation path evolution model comprises a shared encoder and a Koopman transfer matrix; extracting an initial latent variable through the shared encoder, and generating a latent variable sequence of all time steps within a sliding window through linear recursion through the Koopman transfer matrix; extracting the latent variable of the last time step in the latent variable sequence as a health state representation corresponding to the sliding window, and splicing the health state representations of all sliding windows in chronological order to obtain an evolution path of aero-engine performance degradation. The method realizes high-precision and high-stable continuous evolution of the aero-engine performance degradation path.

Description

Technical Field

[0001] This application relates to the field of aero-engine health management technology, and in particular to an aero-engine performance degradation path evolution, model training method and device. Background Technology

[0002] Aero engines operate under extreme conditions of high temperature, high pressure, and high speed for extended periods. Rotating components (such as fans, compressors, and turbines) gradually experience efficiency decline and reduced flow capacity. This performance degradation directly leads to increased fuel consumption, higher exhaust temperatures, and reduced thrust, and in severe cases, may cause safety accidents. Accurately estimating the evolution path of aero engine performance degradation is a crucial prerequisite for achieving remaining service life prediction, condition-based maintenance decisions, and optimized fleet scheduling.

[0003] Currently, the mainstream methods for estimating the performance degradation path of aero-engines mainly rely on the construction of data-driven health indicators. Existing data-driven methods typically utilize traditional dimensionality reduction techniques such as principal component analysis, or employ deep generative models (such as variational autoencoders) to extract features from multi-sensor monitoring data in order to construct a latent variable space that can reflect the engine's health status.

[0004] However, existing performance degradation path estimation methods cannot provide high-precision and stable continuous degradation path estimation results for aero-engines in practical applications. Summary of the Invention

[0005] This application provides a method and apparatus for the evolution of aero-engine performance degradation paths and model training, which addresses the technical deficiency in the prior art that cannot provide high-precision and stable continuous estimation results of aero-engine performance degradation paths, and realizes deep decoupling characterization of multi-source sensor features and high-precision, highly stable, and continuous evolution prediction of long-cycle performance degradation paths.

[0006] In a first aspect, embodiments of this application provide a method for the evolution of performance degradation paths in aero-engines, including: Obtain full life-cycle data of aero engines; The full life-cycle data of the aero-engine is input into a pre-trained aero-engine performance degradation path evolution model, wherein the aero-engine performance degradation path evolution model includes a shared encoder and a Koopman transition matrix; the shared encoder extracts initial latent variables characterizing the health state of the aero-engine, and the Koopman transition matrix generates a sequence of latent variables for all time steps within a sliding window through linear recursion of the initial latent variables; The latent variable at the last time step in the latent variable sequence is extracted as the health status representation of the corresponding sliding window, and the health status representations of all sliding windows are spliced ​​together in chronological order to obtain the evolution path of the aero-engine performance degradation.

[0007] Secondly, embodiments of this application also provide a training method for an aero-engine performance degradation path evolution model. The aero-engine performance degradation path evolution model includes a cascaded shared encoder, a Koopman transition matrix, and independent decoders corresponding to each sensor. The method includes: Obtain historical sample data of the entire life cycle of aero engines; Using the shared encoder, initial latent variables characterizing the health status of the aero-engine are extracted based on the full-lifetime historical sample data; Using the Koopman transition matrix, the initial latent variables are linearly recursively processed to generate a sequence of latent variables; The latent variable sequences are reconstructed using the independent decoders described above to obtain multi-sensor reconstructed data. Based on the full-lifetime historical sample data, the multi-sensor reconstructed data, and the hyperparameters... The weighted Kullback-Leibler divergence is used to construct a joint loss function. Based on the joint loss function, the network parameters of the shared encoder, the Koopman transition matrix, and each of the independent decoders are updated to obtain the trained aero-engine performance degradation path evolution model.

[0008] Thirdly, embodiments of this application also provide an aero-engine performance degradation path evolution device, comprising: The data acquisition module is used to acquire full-lifecycle data of the aero-engine; The feature extraction module is used to input the full life-cycle data of the aero-engine into a pre-trained aero-engine performance degradation path evolution model, wherein the aero-engine performance degradation path evolution model includes a shared encoder and a Koopman transition matrix; the shared encoder extracts initial latent variables characterizing the health state of the aero-engine, and the Koopman transition matrix generates a sequence of latent variables for all time steps within a sliding window through linear recursion of the initial latent variables; The path generation module is used to extract the latent variable of the last time step in the latent variable sequence as the health status representation of the corresponding sliding window, and to splice the health status representations of all sliding windows in chronological order to obtain the evolution path of the aero-engine performance degradation.

[0009] Fourthly, embodiments of this application also provide a training device for an aero-engine performance degradation path evolution model, comprising: The acquisition module is used to acquire historical sample data of the entire life cycle of the aero-engine; The calculation module is used to extract initial latent variables characterizing the health status of the aero-engine based on the full-lifetime historical sample data using the shared encoder; to linearly recursively calculate the initial latent variables using the Koopman transition matrix to generate a sequence of latent variables; and to reconstruct the sequence of latent variables using each of the independent decoders to obtain multi-sensor reconstructed data. The training module is used to train data based on the full-lifetime historical sample data, the multi-sensor reconstructed data, and the hyperparameters. The weighted Kullback-Leibler divergence is used to construct a joint loss function. Based on the joint loss function, the network parameters of the shared encoder, the Koopman transition matrix, and each of the independent decoders are updated to obtain the trained aero-engine performance degradation path evolution model.

[0010] Fifthly, embodiments of this application also provide an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the aero-engine performance degradation path evolution model training as described in the first aspect or the aero-engine performance degradation path evolution model training as described in the second aspect.

[0011] In a sixth aspect, embodiments of this application also provide a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the aero-engine performance degradation path evolution model training as described in the first aspect or the aero-engine performance degradation path evolution model training as described in the second aspect.

[0012] This application also provides a computer program product, including a computer program, which, when executed by a processor, implements the training of the aero-engine performance degradation path evolution model as described in the first aspect or the aero-engine performance degradation path evolution model as described in the second aspect.

[0013] The aero-engine performance degradation path evolution and model training method and apparatus provided in this application overcome the prediction deviation defects caused by neglecting the physical correlation between parameters in traditional methods by using a shared encoder to fuse multi-source sensor data into a unified low-dimensional initial latent variable, thus improving the accuracy and robustness of performance degradation path modeling. By introducing a Koopman transition matrix into the latent space to perform linear recursive evolution of the initial latent variable, the strong nonlinear degradation in the observation space is transformed into linear recursion in the latent space, effectively avoiding the error accumulation caused by frame-by-frame iteration of traditional recursive networks, and significantly improving long-term evolution prediction. The accuracy and stability of the model are improved; at the same time, the Koopman transition matrix is ​​used to construct a structured and analyzable degenerate manifold in the latent space, providing a clear mathematical reference for the performance degradation path, overcoming the technical defect of the traditional deep black box model that is not interpretable, and providing a reliable quantitative basis for remaining lifetime prediction; in the online real-time inference stage, only a single forward propagation is required by the shared encoder, and the subsequent evolution states are all linearly inferred by the Koopman transition matrix through matrix multiplication, without the need for frame-by-frame decoding by the recursive network, which greatly reduces the consumption of computing resources and inference latency, and meets the real-time health management requirements of airborne or near-airborne scenarios. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 This is a flowchart illustrating the evolution method of the performance degradation path of an aero-engine provided in the embodiments of this application; Figure 2 This is a schematic diagram of the structural connection of the aero-engine performance degradation path evolution model provided in the embodiments of this application; Figure 3 This is a flowchart illustrating the training method for the aero-engine performance degradation path evolution model provided in this application embodiment; Figure 4 This is a schematic diagram of the visual experimental verification results provided in the embodiments of this application; Figure 5 This is a schematic diagram of the structure of the aero-engine performance degradation path evolution device provided in the embodiments of this application; Figure 6 This is a schematic diagram of the structure of the training device for the evolution model of the performance degradation path of an aero-engine provided in the embodiments of this application; Figure 7 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0017] To better understand the embodiments of this application, the prior art will first be described in detail: Despite the progress made in feature extraction by existing deep generative models, current performance degradation path estimation methods still suffer from the following technical shortcomings when faced with heterogeneous monitoring data throughout the entire life cycle of aero-engines: Disadvantage 1: Incomplete decoupling of latent space features, resulting in a mixture of operating noise and degradation information. Aero engines experience various complex operating conditions during actual flight, including takeoff, cruise, and climb, and the sensors themselves contain measurement noise. In standard variational autoencoders (VAEs), latent variables lack explicit independence decoupling constraints, causing the engine's true internal degradation information to be deeply mixed with external operating noise and fluctuations in operating conditions. Although β-VAEs introduce a hyperparameter β to force latent variable decoupling, in actual training, setting β to a fixed constant often presents a dilemma: too large a β makes it difficult to optimize the reconstruction loss and loses key degradation features; too small a β results in insufficient decoupling capability. Furthermore, if the KL divergence constraint is too strong in the early stages of training, it can easily lead to posterior collapse, causing encoder failure.

[0018] Disadvantage 2: The latent space temporal evolution lacks dynamic constraints, making accurate long-term predictions impossible. Although deep learning models are often used for modeling latent variables in time series, they are essentially black-box nonlinear evolution models. The transfer and diffusion of latent variables over time lack explicit mathematical or physical dynamic constraints, resulting in an unanalyzable evolutionary trajectory. When predicting long-term decline paths, black-box models are prone to accumulating prediction errors at each step, causing the predicted trajectory to diverge rapidly and failing to provide stable and robust long-term recursive estimates.

[0019] Disadvantage 3: Heterogeneous characteristics of multiple sensors lead to gradient conflicts, limiting reconstruction accuracy. Aero-engine airflow monitoring data includes various physical quantities such as temperature, pressure, speed, and flow rate. The dimensions of the signals from different sensors differ significantly, and their dynamic ranges are completely different. Existing technologies typically use a unified decoder for all sensors to reconstruct the network. During backpropagation training, the reconstruction error gradients generated by sensors with different dimensions will cause gradient conflicts and parameter dilution in the shared network layers. This "one-size-fits-all" shared decoding architecture severely limits the accurate fitting of specific degradation modes from various heterogeneous sensors, leading to decreased reconstruction accuracy and consequently compromising the purity of latent space features.

[0020] In summary, there is an urgent need for a method for estimating the performance degradation path of aero-engines that can completely decouple degradation information, construct linearly predictable latent space evolution laws, and adapt to the heterogeneous characteristics of multiple sensors.

[0021] To address the aforementioned technical issues, this application proposes a method and apparatus for analyzing the performance degradation path evolution and model training of aero-engines. Through a cascaded design of a channel attention module, a residual convolution module, and a bidirectional long short-term memory network module, it achieves comprehensive extraction of multi-level degradation features—including spatial coupling between channels, short-term local correlation, and long-term temporal dependence—from multi-sensor data of aero-engines. This effectively solves the technical problem of traditional methods neglecting the inherent physical relationships between parameters, leading to degradation trajectories deviating from the true evolution trend, and significantly improves the accuracy and robustness of health status characterization. By introducing a Koopman transition matrix to impose linear dynamic constraints on the latent variable sequence, the aero-engine's originally strongly nonlinear degradation process in the original observation space is transformed into a more dynamic process in the latent space. The process transforms the trajectory into a global linear evolution, enabling long-term state deduction through single-step matrix multiplication. This effectively avoids the error accumulation problem caused by frame-by-frame iteration in recurrent neural networks, significantly improving the accuracy and stability of long-term recursive prediction. It provides a health state reference with clear physical meaning and analytical capability for remaining lifetime estimation. By configuring independent decoders for each sensor channel with decoupled parameters, the technical bottleneck of mutual interference between different physical dimension features in traditional multi-channel joint decoding is overcome. Each independent decoder focuses on the reconstruction of the physical dimension signal of a single sensor channel, completely eliminating crosstalk between multi-source sensor features during the reconstruction stage, significantly improving the fidelity of multi-sensor reconstruction data. Furthermore, by fusing reconstruction loss and hyperparameters in the joint loss function... The weighted KL divergence loss enables end-to-end collaborative training and joint optimization of the shared encoder, Koopman transition matrix, and each independent decoder. This allows the latent variable space to achieve an optimal balance while satisfying the triple objectives of probability regularization, linear dynamic constraints, and reconstruction fidelity. After training, the independent decoders can be removed, and only the shared encoder and Koopman transition matrix are retained to independently generate online inference for the performance degradation path of aero-engines. This significantly reduces the computational resource requirements and storage overhead during the online deployment phase.

[0022] To better understand the embodiments of this application, the application scenarios are described in detail below: The aero-engine performance degradation path evolution method and model training method provided in this application can be widely applied to the health management and fault prediction of various gas turbine engines. This method is applicable to continuous degradation data modeling, low-dimensional health feature extraction, latent space linear dynamic constraints, and remaining life prediction of gas turbine engines such as turbofan engines, turbojet engines, and turboshaft engines throughout their entire life cycle. It can be deployed in aero-engine maintenance, health management, and flight safety assessment systems.

[0023] It should be noted that the execution subject of the aero-engine performance degradation path evolution method provided in this application embodiment can be an aero-engine performance degradation path evolution device or an electronic device, and no specific limitation is made here.

[0024] Optionally, the electronic device includes: a computer, a mobile terminal, an industrial control computer, a data acquisition workstation, an airborne data processing unit, a server, or an embedded controller, etc.

[0025] The following section uses electronic devices as an example to elaborate on the aero-engine performance degradation path evolution method provided in this application: Figure 1 This is a flowchart illustrating the evolution method of the performance degradation path of an aero-engine provided in an embodiment of this application. Figure 1 As shown, the method includes the following steps 101-103.

[0026] Step 101: Obtain the full lifespan data of the aero-engine.

[0027] Among them, full life-cycle data refers to the time-series monitoring data continuously collected by multiple types of sensor channels at equal time intervals throughout the entire life cycle of an aero-engine, from its first operation to when its performance indicators reach the failure threshold.

[0028] Optionally, the full lifecycle data may include key physical parameters acquired by multiple sensor channels, such as exhaust temperature, fuel flow rate, vibration amplitude, compressor outlet pressure, high-pressure speed, low-pressure speed, and core engine vibration phase.

[0029] It should be noted that, during the decay process of aero-engines, the aforementioned key physical parameters exhibit non-stationary, strongly coupled, and time-varying nonlinear degradation characteristics, and the physical constraint relationships between various sensor channels drift with the degree of decay. If a general data-driven model is directly used to fit them, it is very easy to cause the degradation trajectory to deviate from the true evolution trend by ignoring the inherent physical correlation between parameters.

[0030] In this embodiment, the electronic device acquires the full life-cycle data of the aero-engine as input to the subsequent aero-engine performance degradation path evolution model. The aim is to utilize the multi-source sensing information and physical coupling prior contained in the full life-cycle data to provide a reliable data foundation for the model to learn the real degradation law, so as to drive the model to capture the implicit physical evolution law between parameters, thereby supporting the accurate modeling and evolution inference of the aero-engine performance degradation path.

[0031] Step 102: Input the full life-cycle data of the aero-engine into the pre-trained aero-engine performance degradation path evolution model, which includes a shared encoder and a Koopman transition matrix; extract the initial latent variables representing the health status of the aero-engine through the shared encoder, and generate a sequence of latent variables for all time steps within the sliding window through linear recursion using the Koopman transition matrix.

[0032] Among them, the evolution model of the performance degradation path of aero-engines includes a shared encoder and a Koopman transition matrix.

[0033] A shared encoder is used to compress and map the sensor monitoring parameters of the first time step within the current sliding window into an initial latent variable, which represents the embedding of the aircraft engine's health indicators at the current moment.

[0034] The Koopman transition matrix is ​​a trainable linear transformation matrix used to generate the latent variables for all remaining time steps within a sliding window through a linear recursive method from the initial latent variables.

[0035] In this embodiment, the electronic device divides the lifecycle data obtained in step 101 into preset time windows and inputs the time-series data within the current sliding window into a shared encoder. The shared encoder extracts the implicit degradation state behind the sensor monitoring parameters through nonlinear mapping, which serves as the initial latent variable at the starting point of the sliding window. This initial latent variable serves as the degradation state benchmark at the starting point of the sliding window. Then, the electronic device inputs this initial latent variable into a Koopman transition matrix. The Koopman transition matrix generates latent variables for all future times within the sliding window step by step through a linear recursive relationship, forming a continuous sequence of latent variables. This sequence of latent variables depicts the continuous evolution trajectory of the aero-engine degradation state along the time axis in the latent space.

[0036] It should be noted that, due to the strong nonlinear characteristics of the aero-engine performance degradation process in the original observation space, and within the framework of Koopman operator theory, there always exists a suitable latent space in which the evolution of this nonlinear system behaves as an approximately linear dynamic system. Based on this, this step constructs a high-dimensional latent degradation state of the aero-engine through a shared encoder, and then applies linear dynamic constraints to this state via the Koopman transition matrix, so that the generated latent variable sequence is distributed along a global linear evolution trajectory in the latent space. This linear recursive mechanism does not rely on frame-by-frame iteration of a recurrent neural network and can achieve long-term state derivation with single-step matrix multiplication, effectively avoiding the problem of error accumulation, thus possessing long-term recursive prediction capabilities. At the same time, the linearly evolving latent variable sequence constitutes a structured analytically degradable manifold in the latent space, providing a health indicator reference benchmark with clear physical meaning for subsequent remaining lifetime estimation.

[0037] Step 103: Extract the latent variable of the last time step in the latent variable sequence as the health status representation of the corresponding sliding window, and splice the health status representations of all sliding windows in chronological order to obtain the evolution path of the aero-engine performance degradation.

[0038] In this embodiment, the electronic device acquires the sequence of latent variables generated in step 102 within the current sliding window and extracts the latent variable at the last time step of the sequence as the health status representation of the current sliding window. Since the latent variable sequence is constrained by Koopman linear dynamics along the time axis, the latent variable at the last time step centrally reflects the comprehensive degradation state under the coupled effects of multiple failure modes within the engine at the end of the sliding window, thus more accurately reflecting the engine's health level at the end of the sliding window. Subsequently, the electronic device splices the health status representations corresponding to each sliding window according to the chronological order of their original acquisition times, forming a high-dimensional state trajectory that evolves over time. This trajectory depicts the continuous degradation manifold of the aero-engine from healthy operation to performance degradation, i.e., the evolution path of aero-engine performance degradation.

[0039] In this embodiment, the technical solutions of steps 101-103 use a shared encoder to fuse multi-source sensor data into a unified low-dimensional initial latent variable, overcoming the prediction deviation caused by neglecting the physical correlation between parameters in traditional methods, and improving the accuracy and robustness of performance degradation path modeling. By introducing a Koopman transition matrix in the latent space to linearly recursively evolve the initial latent variable, the strong nonlinear degradation in the observation space is transformed into linear recursion in the latent space, effectively avoiding the error accumulation caused by frame-by-frame iteration of traditional recursive networks, and significantly improving the accuracy and stability of long-term evolution prediction. At the same time, the Koopman transition matrix is ​​used to construct a structured analyzable degenerate manifold in the latent space, providing a clear mathematical reference benchmark for the performance degradation path, overcoming the technical defect of the uninterpretable nature of traditional deep black box models, and providing a reliable quantitative basis for remaining lifetime prediction. In the online real-time inference stage, only a single forward propagation is required by the shared encoder, and the subsequent evolution states are all linearly recursively derived by the Koopman transition matrix through matrix multiplication, without the need for frame-by-frame decoding by a recursive network, which greatly reduces the consumption of computing resources and inference latency, and meets the real-time health management requirements of airborne or near-airborne scenarios.

[0040] In some embodiments, the shared encoder includes a channel attention module. The step of the electronic device extracting initial latent variables characterizing the health status of the aero-engine through the shared encoder may include: the electronic device using the current sliding window to extract full-lifetime data to obtain the time-series monitoring data corresponding to the current sliding window, and inputting the time-series monitoring data into the channel attention module to perform adaptive channel weighting on the time-series monitoring data of each sensor channel to obtain weighted sensor data; the electronic device determining the initial latent variables characterizing the health status of the aero-engine based on the weighted sensor data.

[0041] The channel attention module is used to establish explicit interdependencies between different sensor channels. By learning the dynamic weighting coefficients corresponding to each sensor channel, it adaptively adjusts the response level of each sensor feature.

[0042] In this embodiment, the electronic device first reads the acquired full life data of the aero-engine, and performs sliding window addressing and truncation on the full life data according to the preset sliding window length L (e.g., L=16) and the number of sensor channels C (e.g., C=7) to obtain the timing monitoring data corresponding to the current sliding window. The timing monitoring data of the current sliding window includes the timing monitoring parameters continuously collected by C sensor channels within the current sliding window length L.

[0043] Secondly, the electronic device performs dimensionality reduction and global feature compression on the time-series monitoring data of the current sliding window along the time dimension. The electronic device uses global average pooling to calculate the mean of the time-series monitoring parameters of each sensor channel on the time axis, compressing the time-series monitoring data of dimension L×C into a channel description vector that represents the global background degradation information of each sensor channel. Each element in the channel description vector represents the global degradation background information of a sensor channel in the current time window.

[0044] Subsequently, the electronic device constructs explicit interdependencies between different sensor channels based on the channel description vector. The electronic device then inputs this channel description vector into a nonlinear structure constructed from two linear fully connected layers. In this process, the electronic device first uses the first fully connected layer to reduce the dimensionality of the channel description vector and employs a nonlinear activation function for feature excitation to fully capture the complex nonlinear aerodynamic-thermodynamic coupling relationships between different sensor channels. Then, the electronic device uses the second fully connected layer to restore the feature dimension to the original number of sensor channels and uses a normalized activation function (e.g., the Sigmoid function) to map the values ​​to a preset weight range (e.g., [0,1]), thereby adaptively generating dynamic weighting coefficients for each sensor channel. These dynamic weighting coefficients reflect the importance of each sensor channel in representing the overall health status of the engine at the current degradation stage.

[0045] Next, the electronic device uses the calculated dynamic weighting coefficients to adaptively weight the original time-series monitoring data. The electronic device multiplies the original input time-series monitoring data with the corresponding dynamic weighting coefficients element-wise for each channel to obtain the weighted sensor data.

[0046] Finally, based on the weighted sensor data and combined with the subsequent cascaded modules of the shared encoder, the electronic equipment continues to perform spatial local nonlinear decoupling and spatiotemporal degradation feature extraction, and finally obtains the initial latent variables characterizing the health state of the aero-engine. This provides a highly pure and physically consistent health state benchmark for subsequent high-precision, iterative-free long-term linear recursive evolution in the latent space using the Koopman transition matrix.

[0047] It should be noted that, through the above channel weighting operation, the characteristics of the sensor channels that are most sensitive to degradation information and have the best characterization ability can be automatically identified and enhanced according to the current actual degradation state of the aero-engine. At the same time, the characteristic responses of noise channels or faulty sensor channels can be adaptively suppressed, thereby realizing autonomous collaborative filtering of heterogeneous multi-source monitoring parameters at the source of feature input.

[0048] In some embodiments, the shared encoder further includes a residual convolution module cascaded with the channel attention module. The step of the electronic device determining the initial latent variable characterizing the health status of the aero-engine based on the weighted sensor data may include: the electronic device inputting the weighted sensor data into the residual convolution module and extracting a short-time degradation feature sequence characterizing the local coupling characteristics between the various sensor channels from the weighted sensor data; the electronic device determining the initial latent variable characterizing the health status of the aero-engine based on the short-time degradation feature sequence.

[0049] The residual convolution module is used to capture the nonlinear dynamic interaction and local degradation fluctuation characteristics of different heterogeneous sensor parameters in a short time scale by multi-scale convolution mapping across channels, while keeping the time series length unchanged.

[0050] In this embodiment of the application, taking a three-layer stacked one-dimensional residual convolutional layer as an example, the electronic device specifically extracts short-term degradation feature sequences using the residual convolutional module through the following steps: First, the electronic device inputs the weighted sensor data into the first one-dimensional residual convolutional layer of the residual convolution module. This first one-dimensional residual convolutional layer uses a one-dimensional convolutional kernel of a preset size to perform sliding convolution feature extraction along the time dimension on the adaptively weighted multi-sensor data, in order to initially establish the spatial and temporal cross-coupling of features from different sensor channels within locally adjacent time steps. After convolution calculation, the electronic device sequentially performs batch normalization and nonlinear activation (e.g., GELU activation function) mapping on the output features of the first one-dimensional residual convolutional layer to stabilize the feature distribution and introduce nonlinear expressive power. During this process, the electronic device expands the number of feature channels from the original number of channels to a first intermediate number of channels (e.g., from 7 dimensions to 32 dimensions) and introduces a first-level residual connection. The features output from the first one-dimensional residual convolutional layer are element-wise added to the original input features. If the number of channels does not match, the electronic device uses a one-dimensional channel-adjusting convolution (e.g., 1×1 convolution) to map and align the original input features before adding them, obtaining the first-level added features to ensure lossless information flow during deep transmission.

[0051] Secondly, the electronic device inputs the features obtained from the first-level summation into a second one-dimensional residual convolutional layer. This second one-dimensional residual convolutional layer further extracts deep coupling features within a wider spatiotemporal receptive field between the sensors through convolutional kernels. After batch normalization and nonlinear activation processing, the number of feature channels is expanded from the first intermediate channel number to the second intermediate channel number (e.g., from 32 dimensions to 64 dimensions). Similarly, the electronic device introduces a second-level residual connection, adding the features calculated and activated by this convolutional layer element-wise with the features input to the second one-dimensional residual convolutional layer to obtain the features obtained from the second-level summation.

[0052] Next, the electronic device feeds the features obtained from the second-level summation into a third second-one-dimensional residual convolutional layer. This third second-one-dimensional residual convolutional layer performs the final integration of local degradation patterns on the deep features and, after batch normalization and nonlinear activation processing, expands the number of channels to the target high-dimensional number of channels (e.g., from 64 dimensions to 128 dimensions). Through the third-level residual connection, the electronic device performs element-wise summation of the output after activation of the third layer with the features input to this third second-one-dimensional residual convolutional layer, ultimately obtaining a short-term degradation feature sequence representing the local coupling features between the various sensor channels.

[0053] Finally, based on the short-term degradation feature sequence and combined with the subsequent cascaded modules of the shared encoder, the electronic equipment ultimately determines the initial latent variables characterizing the health status of the aero-engine.

[0054] Optionally, in the above one-dimensional residual convolution calculation process, the convolution kernel size is set to 3, the convolution stride is set to 1, and the corresponding zero padding parameter Padding=1 is set so that the length of the output feature matrix on the time axis always remains consistent with the length of the sliding window before and after the convolution operation.

[0055] It should be noted that by using the cross-channel multi-scale residual mapping of the residual convolution module, it is possible to capture the local fluctuations and thermodynamic interaction features of multiple heterogeneous sensors under microscopic transients while maintaining temporal spatial resolution, effectively avoiding gradient vanishing and information loss in the deep feature extraction process.

[0056] In some embodiments, the shared encoder further includes a bidirectional long short-term memory network module cascaded with the channel attention module and the residual convolution module. The step of the electronic device determining the initial latent variables characterizing the health status of the aero-engine based on the short-term degradation feature sequence may include: the electronic device inputting the short-term degradation feature sequence into the bidirectional long short-term memory network module, extracting the bidirectional long-term dependencies of the short-term degradation feature sequence to obtain the hidden state sequence; and the electronic device determining the initial latent variables characterizing the health status of the aero-engine based on the hidden state sequence.

[0057] Among them, the bidirectional long short-term memory network module simultaneously captures the causal relationships along the forward and backward time directions in time-series data, capturing the long-range dependencies between time steps before and after the aero-engine degradation process.

[0058] In this embodiment, the electronic device first inputs the short-term degradation feature sequence output by the residual convolution module into the forward long short-term memory network along the time axis. The forward long short-term memory network uses its internal forget gate, input gate and output gate to control the flow of information. In the process of gradually updating the cell state, it extracts the cumulative evolution trend of engine performance degradation and outputs the positive hidden state corresponding to each time step.

[0059] Meanwhile, the electronic device inputs the short-term degradation feature sequence in reverse time axis into the backward long short-term memory network. The reverse long short-term memory network also uses its internal control gating mechanism to extract the reverse physical constraints of the degradation trajectory and the correlation feedback between the future state and the current state, and outputs the reverse hidden state corresponding to each time step.

[0060] Next, at each identical time step, the electronic device concatenates the corresponding forward hidden state with the reverse hidden state. Through this bidirectional fusion operation, the electronic device obtains a fused hidden feature at each time step that integrates bidirectional causal dependency information from the past and future, thereby arranging and generating a complete sequence of hidden states in the time dimension.

[0061] Finally, based on the hidden state sequence and in conjunction with the subsequent cascaded modules of the shared encoder, the electronic equipment ultimately determines the initial latent variables characterizing the health status of the aero-engine.

[0062] It should be noted that by introducing bidirectional time-dependent modeling, it is possible to simultaneously extract the cumulative time trend of engine performance degradation and future causal constraints, thus fully capturing the degradation trajectory with highly nonlinear and long time-delay characteristics during the entire life cycle of aero-engines.

[0063] In some embodiments, the step of determining initial latent variables characterizing the health state of the aero-engine based on the hidden state sequence may include: the electronic device performing average pooling and layer normalization on the hidden state sequence to obtain a feature vector; the electronic device inputting the feature vector into two independent linear fully connected layers for mapping, and outputting the mean vector and log-variance vector of the latent variables respectively; the electronic device reparameterizing the mean vector and log-variance vector to obtain the initial latent variables characterizing the health state of the aero-engine.

[0064] The mean vector refers to the expected center location in the latent space probability distribution of the deterministic, mainstream degradation state of the aero-engine reflected by multi-source sensor parameters within the current sliding window.

[0065] The logarithmic variance vector is used to characterize the range of uncertainty in state estimation caused by sensor measurement noise, transient environmental interference, and drastic fluctuations in engine operating conditions. It is used to impose statistical boundary constraints on the latent space probability manifold by mapping the logarithmic domain.

[0066] In this embodiment, the electronic device acquires a hidden state sequence output by a bidirectional long short-term memory network module. The length of this hidden state sequence in the time step dimension is consistent with the sliding window length, and it has a specific feature channel dimension. To extract stationary degradation features within the sliding window and eliminate feature redundancy along the time axis, the electronic device performs global average pooling on the hidden state sequence along the time step dimension. This involves calculating the average value of each feature channel at each time step, extracting a fixed-length intermediate feature vector. Subsequently, the electronic device performs layer normalization on this intermediate feature vector. By centering and standardizing the features of this intermediate feature vector, the distribution range of the feature data is stabilized to suppress internal covariate shifts during deep network training, thereby stabilizing model training and accelerating convergence, resulting in a normalized feature vector.

[0067] Secondly, the electronic device inputs the normalized feature vectors in parallel to two independent linear fully connected layers. During this process, the electronic device uses the first linear fully connected layer to perform a dimension reduction linear mapping on the normalized feature vectors and outputs the mean vector of the corresponding latent variables. At the same time, the electronic device uses the second linear fully connected layer to perform an independent dimension reduction linear mapping on the feature vectors and outputs the log-variance vector of the corresponding latent variables.

[0068] Finally, in order to construct a continuous and generalizable latent space probability manifold during the model training phase, the electronic equipment performs probabilistic reparameterization sampling based on the mean vector and the log-variance vector to obtain initial latent variables characterizing the health status of the aero-engine.

[0069] It should be noted that by mapping deterministic degradation features to a probabilistic latent variable distribution containing a mean vector and a log-variance vector, the initial latent variables can not only accurately characterize the mainstream degradation state of the aero-engine at the current moment, but also quantitatively characterize the uncertainty in state estimation caused by sensor measurement noise, environmental condition fluctuations, and individual engine differences. This provides a numerically stable and physically consistent health state benchmark for subsequent Koopman linear recursion from the source, effectively improving the reliability and robustness of long-term degradation prediction.

[0070] Optionally, the electronic device performs probabilistic reparameterization sampling based on the mean vector and the log-variance vector. Specifically, the electronic device first randomly samples a random noise vector with the same dimension as the mean vector from a standard normal distribution. Second, the electronic device performs an exponential operation on the log-variance vector to calculate the standard deviation vector corresponding to the feature scale. Next, the electronic device performs element-wise multiplication of the standard deviation vector and the random noise vector in the corresponding dimension to obtain a dynamic disturbance vector after uncertainty scaling. Finally, the electronic device performs element-wise addition of the dynamic disturbance vector and the mean vector to obtain the initial latent variable characterizing the health status of the aero-engine.

[0071] In some embodiments, the Koopman transition matrix is ​​used to establish a linear mapping relationship between adjacent time steps for latent variables characterizing the health status of the aero-engine, and the size of the Koopman transition matrix matches the dimension of the initial latent variables.

[0072] The Koopman transition matrix is ​​a learnable Koopman transition matrix used to linearly map the latent variables of the current time step to the latent variables of the next time step in the latent space constructed by the shared encoder through single-step matrix multiplication, thereby approximately describing the global linear evolution law of the aero-engine degradation state in the latent space.

[0073] Optionally, when the initial latent variables are 64-dimensional, the size of the Koopman transition matrix is ​​64×64.

[0074] For example, the initial latent variable output by the shared encoder is dimensional vectors (e.g.) This represents the implicit degradation state of the aero-engine at the current sliding window initiation point. The Koopman transition matrix... For one A learnable square matrix. The introduction of the Koopman transition matrix into electronic devices. The initial latent variables are transformed using the Koopman transition matrix. Linear recursion generates latent variables for all future times within the sliding window step by step, forming a continuous sequence of latent variables.

[0075] It should be noted that the Koopman transition matrix is ​​optimized by applying linear dynamic constraints during training, ensuring that the generated latent variable sequence is distributed along a global linear evolution trajectory in the latent space. Since the linear recursive mechanism does not rely on frame-by-frame iteration of the recurrent neural network, it can achieve long-term state derivation with single-step matrix multiplication, effectively avoiding the error accumulation problem caused by gradient vanishing or exploding in long-term predictions of recurrent networks, thus possessing long-term recursive prediction capabilities.

[0076] In some embodiments, the aero-engine performance degradation path evolution model further includes an independent decoder corresponding to each sensor of the aero-engine. The electronic device independently reconstructs and maps the latent variable sequence through the independent decoder corresponding to each sensor to obtain the sensor time-series signal corresponding to each independent decoder. The electronic device splices the sensor time-series signals output by each independent decoder to obtain multi-sensor reconstruction data.

[0077] Among them, the independent decoder corresponding to each sensor refers to the configuration of a reconstruction mapping channel with independent parameters for each of the various sensors with different key physical parameters deployed in the aero-engine.

[0078] Understandably, given that different sensors have different dimensions, physical characteristics, and varying sensitivities and response patterns to engine performance degradation, adopting a one-to-one independent decoding design can achieve physical decoupling of the reconstructed features of different channels. This effectively prevents mutual interference and feature entanglement of features from multiple sources during the unified decoding process, thereby ensuring that each independent decoder can accurately reproduce the physical scale and local fluctuation details of the corresponding sensor.

[0079] In this embodiment, the electronic device acquires a sequence of latent variables obtained by evolution and recursion of the Koopman transition matrix in the latent space. Then, the electronic device distributes the sequence of latent variables in parallel and in a split manner to independent decoders corresponding to each sensor. Within each independent decoder, the sequence of latent variables is reverse-mapped in the time domain to restore the single-channel time-series signal of the sensor corresponding to a specific measurement point. Subsequently, the electronic device collects the single-channel sensor time-series signals output by all independent decoders and maps them. Under the premise of ensuring the alignment of the time step dimension, these independent single-channel sensor time-series signals are horizontally concatenated to recombine and restore a complete multi-sensor reconstruction data containing the original number of channels, which is used to calculate the reconstruction loss with the original input data.

[0080] It should be noted that by configuring independent decoding channels for each sensor, each independent decoder can adaptively focus and learn the specific physical degradation mode of its corresponding sensor. This effectively avoids reconstruction interference between heterogeneous sensors, while making the overall system parameter count far lower than the computational load required to train a complete feature extraction and reconstruction model for each sensor individually. This achieves a synergy between high-precision decoupled reconstruction and lightweight deployment.

[0081] In some embodiments, the independent decoder includes a pointwise fully connected network and a residual convolutional network. The step of the electronic device independently reconstructing and mapping the latent variable sequence through the independent decoder corresponding to each sensor to obtain the sensor time-series signal corresponding to each independent decoder may include: for each independent decoder, the electronic device inputs the latent variable sequence into the pointwise fully connected network and performs independent fully connected mapping on the latent variables at each time step to obtain a first reconstructed signal; the electronic device inputs the latent variable sequence into the residual convolutional network and performs one-dimensional convolutional feature extraction and smoothing filtering using information from adjacent time steps to obtain a second reconstructed signal; the electronic device superimposes and adds the first reconstructed signal and the second reconstructed signal to obtain the sensor time-series signal corresponding to the independent decoder.

[0082] Among them, the point-to-point fully connected network refers to a nonlinear network that performs point-to-point feature remapping on low-dimensional latent variables at each independent time step in the latent variable sequence without crossing time steps. This point-to-point fully connected network is used to establish an absolute mapping relationship between the latent space degradation state and physical observations.

[0083] Residual convolutional networks are convolutional neural networks that use one-dimensional convolutional kernels with a set window length to perform sliding feature extraction and local context association modeling on time series. These residual convolutional networks are used to capture and constrain the local smoothness and dynamic evolution characteristics of time-series reconstructed signals.

[0084] The first reconstructed signal refers to the sensor transient output signal predicted individually by the point-to-point fully connected network at each isolated time step. This first reconstructed signal is used to characterize the transient fluctuation characteristics of a specific channel that are independent of preceding and following time constraints and respond extremely quickly to engine state transitions.

[0085] The second reconstructed signal refers to the temporally smoothed reconstructed signal output by the residual convolutional network after capturing the neighborhood context features of the time step. This second reconstructed signal represents the trend features that integrate the continuity of time series dynamics and physical evolution constraints, and is used to filter and correct the local high-frequency noise in the first reconstructed signal.

[0086] In this embodiment, the electronic device acquires the latent variable sequence obtained by evolutionary recursion, and inputs the latent variable sequence in parallel into a point-wise fully connected network and a residual convolutional network for dual-path feature restoration.

[0087] In the branch of point-to-point fully connected networks, for each time step in the sequence of latent variables, the electronic device will... The latent variable vector is input into the first fully connected layer, and a linear transformation is performed to... The feature dimension of the latent variable vector is compressed to a preset intermediate dimension (e.g., the intermediate dimension is set to 16 dimensions), and nonlinear excitation is performed using the GELU activation function to capture the nonlinear thermodynamic state characteristics of the aero-engine at a single time step. Subsequently, the electronic device inputs the nonlinearly excited features into the second fully connected layer, further mapping them to the single-channel target dimension, and outputs the single-channel reconstruction value of that time step. After the electronic device performs the above point-by-point mapping operation sequentially for all time steps, it summarizes the outputs of each time step to obtain the first reconstruction signal consistent with the sliding window length. In this calculation process, the calculations of each time step are completely independent and uncoupled, enabling this branch to sensitively and without delay capture channel-specific local transient fluctuation characteristics such as aero-engine surge, transient overheating, or valve sudden changes.

[0088] Subsequently, in the residual convolutional network branch, the electronic device first performs a matrix transpose operation on the latent variable sequence, converting it into a matrix format with the channel dimension first and the time step dimension last, resulting in a transposed matrix that conforms to the input specification of one-dimensional convolution. The electronic device then inputs the transposed matrix into a first one-dimensional convolutional layer. This first one-dimensional convolutional layer uses a pre-defined kernel size (e.g., kernel size of 3, stride of 1, and including zero-padding of 1) and performs sliding convolution calculations along the time dimension to capture the local state evolution trend between adjacent time steps. After convolution calculation, the GELU activation function is used for non-linear feature projection, initially condensing the high-dimensional spatial features into a lower-dimensional space. Then, the electronic device inputs the output features of the first one-dimensional convolutional layer into a second one-dimensional convolutional layer. This second one-dimensional convolutional layer has the same kernel size, stride, and padding as the first one-dimensional convolutional layer. The second one-dimensional convolutional layer further performs temporal dimensionality reduction, compressing the channel dimension to a single-channel target dimension and outputting reconstructed features in the transposed matrix format. Finally, the electronic device performs a transpose recovery operation on the reconstructed feature again, restoring it to a second reconstructed signal with the same length as the sliding window. In this calculation process, because the one-dimensional convolution kernel incorporates the degenerate state information of adjacent time steps when sliding to extract features, the second reconstructed signal exhibits extremely strong stationarity and continuity in the time domain, acting as a time-domain low-pass smoothing filter.

[0089] Next, the electronic device performs element-wise addition of the first reconstructed signal and the second reconstructed signal step by step to obtain the single-channel sensor timing signal corresponding to the independent decoder. Through the superposition and fusion of dual-path features, the single-channel sensor timing signal simultaneously takes into account the high-frequency transient response capability of the point-by-point fully connected branch and the global temporal smoothness and continuity of the residual convolution branch.

[0090] Finally, the electronic device splices together the single-channel sensor timing signals mapped and output by all the independent decoders, which correspond one-to-one with the channels of each physical sensor, to reconstruct complete multi-sensor reconstruction data.

[0091] It should be noted that the latent variables of each time step are independently mapped through a point-by-point fully connected network, enabling the decoder to keenly capture the transient fluctuation characteristics of the aero-engine in a single time step. By using a residual convolutional network to perform sliding filtering and smoothing constraints on the information of adjacent time steps, high-frequency measurement noise in the first reconstructed signal is effectively filtered out and local abnormal jumps are corrected. Finally, the two reconstructed signals are superimposed and added together, taking into account both transient high-frequency response sensitivity and global continuity in the time domain. This achieves high-fidelity and low-noise reconstruction and restoration of the sensor time-series signal, provides accurate supervision signals for reconstruction loss, and thus constrains the latent variables extracted by the encoder to retain key degradation information in the original sensor data.

[0092] It should be noted that the execution subject of the training method for the evolution model of the performance degradation path of the aero-engine provided in another embodiment of this application can be an aero-engine performance degradation path evolution model training method device or an electronic device, which is not specifically limited here.

[0093] Optionally, the electronic device includes: a computer, a mobile terminal, an industrial control computer, a data acquisition workstation, an airborne data processing unit, a server, or an embedded controller, etc.

[0094] The following section uses electronic devices as an example to illustrate in detail the training method for the aero-engine performance degradation path evolution model provided in this application: Figure 2 This is a schematic diagram of the structural connection of the aero-engine performance degradation path evolution model provided in this application embodiment. Figure 2 The model for the performance degradation path of this aero-engine includes a cascaded shared encoder, a Koopman transition matrix, and independent decoders corresponding to each sensor. From... Figure 2 It can be seen that the input of the shared encoder is historical sample data from multiple sensors captured by a sliding window, including the sliding window itself. Various measuring points, such as fan speed Core machine speed Exhaust temperature Total pressure Physical parameters, etc. During the encoding stage, initial latent variables characterizing the current health state of the aero-engine are extracted sequentially through a channel attention module, a residual convolution module, and a bidirectional long short-term memory network module. In the latent space, the Koopman transition matrix... The system receives the initial latent variables from the shared encoder and generates a sequence of latent variables for all time steps within the current sliding window through a linear recursive relationship. During the decoding phase, each independent decoder receives the latent variables from each time step in the latent variable sequence and independently reconstructs the temporal reconstruction signal for the corresponding sensor channel. Finally, the original input data, latent space probability features, and the reconstructed temporal reconstruction signal are used to construct a joint loss function. Based on this joint loss function, a parameter gradient update path established through the error backpropagation algorithm is connected to the shared encoder, the Koopman transition matrix, and the trainable network weights in each independent decoder. By minimizing the joint loss during iteration, the collaborative training and joint optimization of the network parameters are ultimately achieved.

[0095] Understandable Figure 2 The structure of the aero-engine performance degradation path evolution model shown is only an example provided by the embodiments of this application. The specific implementation of each module in the aero-engine performance degradation path evolution model is not limited to this.

[0096] Combination Figure 2 , Figure 3 This is a flowchart illustrating the training method for the aero-engine performance degradation path evolution model provided in this application embodiment. Figure 3 In this method, steps 301-305 are included.

[0097] Step 301: Obtain historical sample data of the entire life cycle of the aero-engine.

[0098] It is understandable that the full-lifetime historical sample data obtained in step 301 is completely consistent with the full-lifetime data for inference evolution obtained in step 101 in terms of feature dimensions and sensor channel structure. The only difference between the two is that the full-lifetime historical sample data is an offline training dataset containing complete historical decay trajectories, used to support the parameter optimization and convergence of the model; while the full-lifetime data is the time-series signal to be evaluated collected online during the current actual operation of the aero-engine, used to support the online health state evolution inference of the model after training.

[0099] Step 302: Using a shared encoder, extract initial latent variables characterizing the health status of the aero-engine based on the full lifecycle historical sample data.

[0100] Among them, by Figure 2 It is known that the shared encoder includes a cascaded channel attention module, a residual convolution module, and a bidirectional long short-term memory network module.

[0101] Optionally, the channel attention module includes a global average pooling unit, a two-layer fully connected network and an activation unit, and a weighted multiplication unit. The global average pooling unit performs global average pooling on the input temporal features along the time dimension to extract global features at each time step, obtaining a feature vector. The two-layer fully connected network generates channel weight vectors that correspond one-to-one with the number of sensor channels. The weighted multiplication unit multiplies the channel weight vectors with the original input temporal features channel by channel, achieving adaptive weighting for the multi-channel sensor dimension.

[0102] This channel attention module enables the model to automatically focus on the sensor channel with the richest degradation feature information in the current time window during training, while effectively suppressing interference from irrelevant noise channels or faulty sensors.

[0103] Optionally, the residual convolution module includes stacked multi-layer one-dimensional residual convolutional layers, batch normalization layers, nonlinear activation units, and residual connection paths. The multi-layer one-dimensional residual convolutional layers are used to extract one-dimensional convolutional features from the adaptively weighted multi-channel sensor time-series signals to capture local spatial correlation features and short-term degradation features between different sensor channels. The batch normalization layers and nonlinear activation units are cascaded at the output of each one-dimensional residual convolutional layer, respectively used to perform batch normalization constraints on the convolutionally extracted features and to execute GELU nonlinear activation transformations to improve the network's fitting ability and accelerate network convergence. The residual connection paths are established between the input and output of each one-dimensional residual convolutional layer, used to directly connect and superimpose the input features of the current layer onto the output features after nonlinear activation transformation. Furthermore, when the number of channels is mismatched, one-dimensional kernel convolution is used to adjust the number of input channels for dimensional alignment, achieving cross-layer feature fusion.

[0104] Optionally, the number of stacked one-dimensional residual convolutional layers can be set according to the actual situation, such as stacking three one-dimensional residual convolutional layers, without specific limitations here.

[0105] Optionally, the kernel size of each one-dimensional residual convolutional layer is set to 3, the stride to 1, and the padding parameter to 1 to maintain the same length of the input and output temporal feature sequences. The number of channels in the three one-dimensional residual convolutional layers is expanded layer by layer. For example, the number of channels in the first one-dimensional residual convolutional layer is expanded from 7 to 32, in the second layer from 32 to 64, and in the third layer from 64 to 128.

[0106] This residual convolution module, by introducing residual connection paths, can effectively alleviate the gradient vanishing problem in the forward propagation and error backpropagation processes of multi-layer deep convolutional networks, thereby improving the network's fitting robustness while ensuring the ability to extract deep features.

[0107] Optionally, the bidirectional long short-term memory network module includes a bidirectional long short-term memory unit, which receives the feature sequence output by the residual convolution module and captures the past and future context information of the feature sequence in the time dimension through a cyclic long short-term structure in both the forward and reverse directions, thereby effectively modeling the long-term evolutionary dependency in multi-sensor time-series signals and outputting a bidirectional hidden state sequence covering all times.

[0108] Optionally, the shared encoder also includes an average pooling and layer normalization module cascaded at the back end of the bidirectional long short-term memory network module, as well as a fully connected mapping layer.

[0109] In step 302, the electronic device employs a shared encoder, using the entire lifetime historical sample data as initial input temporal features to feed into the channel attention module. Through global average pooling and a two-layer fully connected network, the weights of each sensor channel are adaptively learned. The channel weights are then multiplied by the initial input temporal features channel by channel to obtain the adaptively weighted multi-channel sensor temporal signal. Next, the electronic device feeds the adaptively weighted multi-channel sensor temporal signal into the residual convolution module. It sequentially processes the signal through stacked multi-layer one-dimensional residual convolutional layers, batch normalization layers, and the GELU activation function, and uses residual connections for feature fusion to extract the local spatial correlations and short-term degradation features between sensors. Subsequently, the electronic device feeds the feature sequence output from the residual convolution module into the bidirectional long short-term memory network module. The bidirectional long short-term memory units model the temporal dependencies and output the hidden state sequences corresponding to all time points. Next, the electronic device utilizes average pooling and layer normalization modules to calculate the average value of the hidden state sequence over time, aggregating the temporal features into a fixed-length feature vector (e.g., a 128-dimensional feature vector). Layer normalization is then used to normalize the feature vector, stabilizing model training and accelerating gradient convergence. Finally, the electronic device inputs the normalized feature vector into a fully connected mapping layer, mapping it through two independent linear fully connected layers to calculate the mean and log-variance parameters of the output latent space probability distribution. During model training iterations, the electronic device introduces a reparameterization technique, randomly sampling a noise vector from a standard normal distribution and combining it with the mean and log-variance parameters to calculate the initial latent variables characterizing the current health state of the aero-engine.

[0110] This step, by cascading and fusing attention-adaptive weighting, local short-term feature convolution extraction, bidirectional global temporal modeling, and probability-regularized latent space mapping, can accurately and robustly construct low-dimensional, noise-resistant, and highly generalizable latent variables of health status.

[0111] Step 303: Using the Koopman transition matrix, the initial latent variables are linearly recursively transformed to generate a sequence of latent variables.

[0112] Among them, the Koopman transition matrix is ​​used to linearly model the temporal dynamic evolution behavior of aero-engines in a low-dimensional latent space.

[0113] In step 303, the electronic device acquires the initial latent variables output in step 302. The initial latent variable for dimensional vectors (e.g.) This characterizes the implicit degradation state of the aero-engine at the current sliding window initiation point. Subsequently, the electronic equipment uses the Koopman transition matrix to transform the initial latent variables... Based on the following linear recursive relationship, generate the latent variables for all future times within the current sliding window step by step: ,in, Indicates the length of the sliding window. Indicates the first A vector of latent variables at each time step. Represents the Koopman transition matrix. This represents matrix multiplication.

[0114] For example, such as Figure 2 As shown, the electronic device uses an initial latent variable Starting from the recursive starting point, first calculate , and then with Given the current state, calculate And so on, using the same Koopman transition matrix Repeatedly applied across all time steps, generating the data recursively step by step. This ultimately forms a continuous sequence of latent variables. This latent variable sequence characterizes the continuous evolution trajectory of the aero-engine degradation state along the time axis in the latent space.

[0115] It should be noted that the Koopman transition matrix shares parameters across all time steps; that is, the linear mapping at different time steps uses the same matrix, rather than using a different transition matrix for each time step. Through this parameter-sharing mechanism, the number of recursive steps for the Koopman transition matrix is ​​not limited by network depth and can extend to any future time, thus providing reliable technical support for the long-term prediction of aero-engine performance degradation paths. Alternatively, in another optional implementation, multi-level evolution calculations can be performed using transition matrices with independent parameters for different time steps to adapt to strongly non-stationary degradation processes; this is not the only possible approach.

[0116] Step 304: Using each independent decoder, the latent variable sequence is reconstructed to obtain multi-sensor reconstructed data.

[0117] Each independent decoder employs a dual-path parallel reconstruction architecture, specifically comprising a pointwise fully connected network and a residual convolutional network. The pointwise fully connected network is used to independently map each time step in the latent variable sequence to extract specific transient change features of each sensor channel; the residual convolutional network is used to perform sliding window filtering on the latent variable sequence using temporal context information from adjacent time steps to suppress high-frequency noise introduced by pointwise prediction.

[0118] Optionally, the point-to-point fully connected network includes a first fully connected layer, a GELU activation function, and a second fully connected layer.

[0119] Optionally, the residual convolutional network includes a first one-dimensional convolutional layer, a GELU activation function, and a second one-dimensional convolutional layer, which are cascaded in sequence.

[0120] Alternatively, in other embodiments, the residual convolutional network may also employ three or more stacked one-dimensional convolutional layers, with the number of channels in each convolutional layer first expanded and then compressed to a single-channel target dimension, in order to enhance the feature extraction capability for complex temporal patterns.

[0121] Optionally, each independent decoder also includes a feature fusion layer, which adaptively fuses the first reconstructed signal output by the pointwise fully connected network with the second reconstructed signal output by the residual convolutional network, so as to comprehensively utilize the pointwise mapping's sensitive response to transient high-frequency changes and the sliding window filtering's ability to smoothly maintain the global continuity in the time domain, thereby obtaining a single-channel sensor timing signal that balances local accuracy and global stability.

[0122] Understandably, there is a one-to-one mapping between each independent decoder and the sensor channel that acquires historical sample data. Each independent decoder is used to reconstruct the sequence of latent variables in the low-dimensional latent space in parallel and independently into the physical dimension signal of the corresponding sensor channel, so as to eliminate the interference of the coupling of physical characteristics of different sensors on the reconstruction accuracy.

[0123] In step 304, the electronic device inputs the latent variable sequence generated in step 303 into each independent decoder. Within each independent decoder, the latent variable sequence is fed in parallel into a pointwise fully connected network and a residual convolutional network for reconstruction.

[0124] Subsequently, the electronic device, for each time step in the latent variable sequence, will... The latent variable vector is input into the first fully connected layer, and a linear transformation is performed to convert it into a linear vector. The latent variable vector is compressed to a preset intermediate dimension and nonlinearly excited using the GELU activation function. Subsequently, the electronic device inputs the nonlinearly excited features into the second fully connected layer, which is further mapped to the single-channel target dimension, and outputs the single-channel reconstruction value of that time step. After performing the above point-by-point mapping operation on all time steps in sequence, the outputs of each time step are summarized to obtain the first reconstruction signal with the same length as the sliding window.

[0125] The pointwise fully connected network performs time-independent dimensionality reduction and feature recovery on the latent variables at each isolated time step in the latent variable sequence, outputting a first reconstructed signal. The values ​​of each element in the first reconstructed signal are the pointwise predicted values, which can sensitively and without delay capture and reflect the transient high-frequency response and rapid abrupt degradation characteristics of the aero-engine caused by sudden anomalies.

[0126] Simultaneously, the electronic device performs matrix transpose on the latent variable sequence, converting it into a matrix format with the channel dimension first and the time step dimension last. The electronic device inputs the transposed matrix into the first one-dimensional convolutional layer, performs sliding convolution calculation along the time dimension to capture the local state evolution trend between adjacent time steps, and performs nonlinear feature projection through the GELU activation function. The output features of the first one-dimensional convolutional layer are input into the second one-dimensional convolutional layer for further temporal feature aggregation, compressing the channel dimension to a single-channel target dimension, and outputting the reconstructed features in transposed format. The reconstructed features are transposed and restored again to a second reconstructed signal with the same length as the sliding window.

[0127] Next, the electronic device uses a feature fusion layer to add the first reconstructed signal and the second reconstructed signal element-wise, step-by-step, to obtain the single-channel sensor timing signal corresponding to the independent decoder. Through the superposition and fusion of dual-path features, the single-channel sensor timing signal simultaneously takes into account the sensitive response capability of the point-by-point fully connected branch to transient high-frequency changes and the smoothing filtering capability of the residual convolution branch for global continuity in the time domain.

[0128] Finally, the electronic device splices together the single-channel sensor timing signals output by each independent decoder, which correspond one-to-one with each sensor channel, according to the sensor channel dimension, to reconstruct complete multi-sensor reconstruction data. This multi-sensor reconstruction data is consistent in shape with the original timing monitoring data input to the shared encoder.

[0129] It should be noted that this step breaks away from the conventional approach of multi-channel joint decoding in the decoding stage, configuring a completely decoupled and one-to-one independent decoder for each sensor channel. This channel-independent parallel decoding mechanism completely eliminates crosstalk and coupling interference between different sensors during the reconstruction stage. Simultaneously, within each independent decoder, a dual-path parallel and heterogeneous reconstruction design balances the ability to quickly capture high-frequency occasional anomalies with the ability to smooth low-frequency degradation trends, achieving extremely high-precision time-series reconstruction. This provides crucial data support for subsequently constructing the joint loss function and faithfully reconstructing the engine performance degradation path.

[0130] Step 305: Based on the full-lifetime historical sample data, multi-sensor reconstructed data, and hyperparameters... We construct a joint loss function using weighted Kullback-Leibler divergence, and update the network parameters of the shared encoder, Koopman transition matrix, and each independent decoder based on the joint loss function to obtain the trained aero-engine performance degradation path evolution model.

[0131] Kullback-Leibler divergence (KL divergence) is used to measure the difference between the latent variable distribution of the shared encoder output and the preset standard normal prior distribution, thereby achieving regularization constraints on the latent space.

[0132] Hyperparameters It refers to the balance coefficients, or hyperparameters, used in the joint loss function to control the weights of the KL divergence term. The value of determines the trade-off between the degree of decoupling of latent variables and the accuracy of reconstruction.

[0133] By hyperparameters Weighted Kullback-Leibler divergence refers to the sum of the KL divergence values ​​and the hyperparameters. The regularization loss term obtained by multiplication forces the latent space features to tend to follow a standard normal distribution, and is determined according to the hyperparameters. The decoupling performance of the configuration strength adjustment feature.

[0134] In step 305, the electronic device uses full-lifetime historical sample data, multi-sensor reconstructed data, and hyperparameters to... Weighted Kullback-Leibler divergences are jointly calculated to construct a joint loss function with multi-objective collaborative constraints. Subsequently, based on this joint loss function, the network parameters of the shared encoder, Koopman transition matrix, and each independent decoder are iteratively updated using the error backpropagation algorithm and gradient descent optimizer until the network converges. During the parameter update process, after each forward propagation calculation and obtaining the joint loss function value, the electronic device automatically calculates the parameter gradients of the joint loss function with respect to the shared encoder (including the network weights of the channel attention module, residual convolution module, bidirectional LSTM module, layer normalization module, and fully connected mapping layer), the Koopman transition matrix, and each independent decoder (including the branch weights of the pointwise fully connected network and residual convolution network) using a chain rule. The electronic device performs backward parameter correction along the gradient update path to minimize the joint loss in multiple iterations. When the joint loss function reaches a preset convergence threshold, or the number of iterations reaches a preset maximum iteration limit, the update stops, the currently saved network weight parameters are locked, and the trained aero-engine performance degradation path evolution model is output.

[0135] In this embodiment, the technical solutions of steps 301-305, through the cascaded design of the channel attention module, residual convolution module, and bidirectional long short-term memory network module, achieve comprehensive extraction of multi-level degradation features of inter-channel spatial coupling, short-term local correlation, and long-term temporal dependence in multi-sensor data of aero-engines. This effectively solves the technical problem that traditional methods cause degradation trajectories to deviate from the true evolution trend due to neglecting the inherent physical correlation between parameters, and significantly improves the accuracy and robustness of health status characterization. By introducing the Koopman transition matrix to apply linear dynamic constraints to the latent variable sequence, the degradation process of aero-engines, which originally exhibited a strongly nonlinear degradation process in the original observation space, is transformed into a globally linear degradation process in the latent space. The trajectory is simplified, enabling long-term state deduction via single-step matrix multiplication. This effectively avoids the error accumulation problem caused by frame-by-frame iteration in recurrent neural networks, significantly improving the accuracy and stability of long-term recursive prediction. It provides a health state reference with clear physical meaning and analytical capability for remaining lifetime estimation. By configuring independent decoders for each sensor channel with decoupled parameters, the technical bottleneck of mutual interference between different physical dimension features in traditional multi-channel joint decoding is broken. Each independent decoder focuses on the reconstruction of the physical dimension signal of a single sensor channel, completely eliminating crosstalk between multi-source sensor features during the reconstruction stage, significantly improving the fidelity of multi-sensor reconstruction data. Furthermore, by fusing reconstruction loss and hyperparameters in the joint loss function... The weighted KL divergence loss enables end-to-end collaborative training and joint optimization of the shared encoder, Koopman transition matrix, and each independent decoder. This allows the latent variable space to achieve an optimal balance while satisfying the triple objectives of probability regularization, linear dynamic constraints, and reconstruction fidelity. After training, the independent decoders can be removed, and only the shared encoder and Koopman transition matrix are retained to independently generate online inference for the performance degradation path of aero-engines. This significantly reduces the computational resource requirements and storage overhead during the online deployment phase.

[0136] In some embodiments, the electronic device uses full-lifetime historical sample data, multi-sensor reconstructed data, and hyperparameters to... The steps for constructing the joint loss function using the weighted Kullback-Leibler divergence may include: the electronic device calculating the mean of the sum of squared errors between the lifetime historical sample data and the multi-sensor reconstructed data, as the reconstruction loss term; the electronic device calculating the Kullback-Leibler divergence between the latent variable distribution of the shared encoder output and the preset standard normal prior distribution, and incorporating the hyperparameters... The value is multiplied by the calculated Kullback-Leibler divergence to obtain the result from the hyperparameters. The weighted Kullback-Leibler divergence is used as a decoupling constraint term; the electronic device performs a weighted sum of the reconstruction loss term and the decoupling constraint loss term to obtain the joint loss function.

[0137] In this embodiment, the electronic device first acquires the multi-sensor reconstruction data output in step 304 and the original full-lifetime historical sample data acquired in step 301, and calculates the mean of the sum of squared errors between the two as a reconstruction loss term used to constrain reconstruction fidelity. Simultaneously, the electronic device acquires the latent variable distribution parameters output by the shared encoder in step 302, namely the mean vector and log-variance vector, calculates the Kullback-Leibler divergence between this latent variable distribution and a preset standard normal prior distribution, and multiplies the hyperparameter β with the Kullback-Leibler divergence value to obtain a decoupling constraint loss term used to constrain latent variable space regularization and feature decoupling. Finally, the electronic device performs a weighted summation of the reconstruction loss term and the decoupling constraint loss term to obtain a joint loss function.

[0138] It should be noted that by constructing the joint loss function, the latent variables extracted by the shared encoder can not only reconstruct the original sensor signal with high fidelity, but also form a structured, decoupled and continuously smooth probability distribution in the latent space, providing a high-quality initial state reference for the subsequent linear recursion of the Koopman transition matrix.

[0139] Optionally, the joint loss function also includes a Koopman linear constraint loss term, which is used to constrain the sequence of latent variables to satisfy a linear recursive relationship.

[0140] Optionally, the joint loss function may also include a health indicator supervision loss term, used to constrain the mapping relationship between the latent variable sequence and the health indicator labels.

[0141] In some embodiments, the electronic device will use hyperparameters The value is multiplied by the calculated Kullback-Leibler divergence to obtain the result from the hyperparameters. The steps for weighted Kullback-Leibler divergence may include: during the iterative training of the aero-engine performance degradation path evolution model, the electronic device records the current training round of the aero-engine performance degradation path evolution model; when the current training round is less than or equal to a preset transition round threshold, the electronic device controls the hyperparameters. The value starts from 0 and increases linearly with the current training round until it reaches the preset maximum target weight value; if the current training round is greater than the preset transition round threshold, the electronic device will adjust the hyperparameter. The value of the hyperparameter remains constant at the preset maximum target weight value; the electronic device will maintain the hyperparameter corresponding to the current training round. The value is multiplied by the calculated Kullback-Leibler divergence to obtain the result from the hyperparameters. The weighted Kullback-Leibler divergence is used as a decoupling constraint term in the joint loss function.

[0142] In this embodiment, during each training iteration, the electronic device first reads the number of completed training rounds as the current training round. The electronic device then compares the current training round with a preset transition round threshold. If the current training round is less than or equal to the preset transition round threshold (e.g., the first 30 training rounds), the electronic device calculates the value of the hyperparameter β corresponding to the current training round in a linearly increasing manner. At the start of training, i.e., round 0, the value of hyperparameter β is 0; as the training rounds increase, the value of hyperparameter β increases linearly; when the training round reaches the preset transition round threshold, the value of hyperparameter β reaches the preset maximum target weight value. If the current training round is greater than the preset transition round threshold, the electronic device fixes the value of hyperparameter β to the preset maximum target weight value and keeps it unchanged in subsequent training. Finally, the electronic device multiplies the value of hyperparameter β corresponding to the current training round with the Kullback-Leibler divergence calculated in step 302 to obtain a Kullback-Leibler divergence weighted by hyperparameter β, which serves as the decoupling constraint term in the joint loss function.

[0143] It should be noted that by employing the aforementioned warm-up strategy to dynamically adjust the hyperparameter β, it is gradually increased from 0 in the early stages of training rather than being directly set to a large target value. In the initial training phase, the shared encoder's network parameters are in a randomly initialized state and have not yet learned effective degenerate feature representations. If a strong KL divergence constraint is applied directly at this time, it will force the latent variable distribution to align prematurely with the standard normal prior distribution, causing the shared encoder to abandon encoding effective features in the input data and map all samples to similar latent variable distributions, thus triggering posterior collapse and causing the shared encoder to lose its feature extraction capability. Through the aforementioned warm-up strategy, the shared encoder can freely learn key degenerate features in the input data under weak constraints or without obstacles in the early stages of training, obtaining a relatively accurate latent variable representation capability. When training progresses to the later stages and the shared encoder has a stable feature extraction capability, the KL divergence constraint is gradually strengthened, forcing the latent variable space to undergo structured regularization and feature decoupling. This avoids posterior collapse while ensuring that the final latent variable space has both good decoupling characteristics and retains complete degenerate information, significantly improving the model's training stability and final performance.

[0144] In some embodiments, the process of updating the parameters of the Koopman transition matrix by the electronic device may further include: extracting the Koopman transition matrix corresponding to the current training round, calculating the maximum singular value of the Koopman transition matrix; and, if the maximum singular value is greater than 1, using the maximum singular value to perform spectral normalization scaling on the Koopman transition matrix to constrain the spectral radius of the Koopman transition matrix to not exceed 1.

[0145] In this embodiment, after updating the parameters of the Koopman transition matrix in each iteration, the electronic device extracts the Koopman transition matrix corresponding to the current training round, performs singular value decomposition on the Koopman transition matrix, and calculates the maximum singular value of the Koopman transition matrix. The electronic device then calculates the maximum singular value. If the maximum singular value is compared with a preset threshold of 1, and if the maximum singular value is greater than 1, the electronic device follows the formula. For the Koopman transition matrix Scaling is performed to obtain the scaled Koopman transition matrix. The maximum singular value is equal to 1. If the maximum singular value is less than or equal to 1, the electronic device keeps the Koopman transition matrix unchanged.

[0146] It should be noted that by applying spectral normalization scaling to the Koopman transition matrix, constraining its spectral radius to no more than 1, the amplitude of the latent variable sequence remains stable during long-term recursion. This prevents prediction failure due to a spectral radius greater than 1, and also avoids attenuation and loss of degradation information due to a spectral radius less than 1. This ensures the numerical stability and physical rationality of the aero-engine performance degradation path evolution model in long-term recursive prediction. Furthermore, spectral normalization scaling directly normalizes the matrix after each parameter update, without involving additional loss function terms or gradient calculations. This results in low computational cost, ease of implementation, and suitability for real-time application during model training.

[0147] To better understand the technical effects of the aero-engine performance degradation path evolution model provided in the embodiments of this application, this application also evaluated the effectiveness of the above evolution model through the following verification experiments.

[0148] This application employs the aforementioned trained aero-engine performance degradation path evolution model to perform low-dimensional health status characterization of the time-series monitoring signals of the aero-engine (taking Engine Unit 36 ​​as an example) during its entire life cycle operation. To intuitively demonstrate the modeling capability of high-dimensional latent space features for aero-engine degradation and to visualize and analyze the degradation trajectory of the model in low-dimensional space, this application uses Principal Component Analysis (PCA) dimensionality reduction to project the 64-dimensional latent space feature vector output by the shared encoder. The visualization experimental verification results are as follows: Figure 4 As shown.

[0149] For example, Figure 4 This is a schematic diagram illustrating the visual experimental verification results provided in the embodiments of this application. For example... Figure 4 As shown, the left subfigure illustrates the two-dimensional latent trajectory of engine unit 36 ​​on the low-dimensional principal component plane (PC1-PC2). In the figure, "Start" indicates the initial degradation state of the aero-engine during its initial operation (PC1-PC2 coordinates corresponding to the initial latent variables), and "End" indicates the final degradation state when the aero-engine fails or reaches the end of its service life. The blue broken line connecting each sampling point depicts the continuous evolution path of the health state in the low-dimensional latent space as the operating cycle progresses. With the increase in the number of operating cycles, the health state of the aero-engine gradually and continuously evolves from the starting region in the lower right to the ending region in the upper left. The entire latent trajectory exhibits strong spatial manifold continuity in the PC1-PC2 space. This result demonstrates that the shared encoder and Koopman transition matrix constructed in this application can successfully couple and map high-dimensional, noisy multi-sensor physical signals into a latent degradation manifold with a clear structure, continuous evolution, and physical dynamic laws in the latent space, achieving temporal continuity modeling of the health state evolution of complex dynamic systems.

[0150] To further quantitatively evaluate the accuracy of low-dimensional features in representing engine degradation and remaining useful life (RUL), this application extracted the first principal component PC1, which has the highest contribution rate after dimensionality reduction, and performed a correlation analysis between PC1 and the engine's actual operating cycles and true remaining useful life. The evolution curve is shown below. Figure 5 The right-hand sub-figure is shown below.

[0151] exist Figure 4In the right-hand subplot, the solid green line represents the evolution trajectory of the first principal component PC1 as the engine cycle number changes, while the dashed red line represents the actual remaining lifetime (RUL) degradation curve of the engine unit. As can be seen from the right-hand subplot, the first principal component PC1 exhibits a highly significant monotonically decreasing trend throughout the overall lifespan (from cycle 0 to over 160 cycles), and its overall fluctuation trend shows a very high degree of parallelism with the linear decreasing trajectory of the actual RUL. This proves that the latent variable extracted through the model in this application, its main energy distribution PC1, can serve as a golden health indicator for assessing the degree of engine degradation. The slope of the PC1 evolution curve has a high numerical correlation with the deceleration rate of the remaining lifetime. In the early, middle, and late stages of engine operation, the deceleration rate of PC1 perfectly matches the gradual physical degradation law of system performance, and near the end of its life, the PC1 index rapidly drops to its lowest point, demonstrating extremely high sensitivity to sudden accelerated degradation characteristics. Furthermore, although the original multi-channel sensor data contains a large amount of high-frequency noise due to operating condition fluctuations, measurement errors, and occasional system disturbances, the PC1 evolution trajectory obtained after dimensionality reduction still maintains a very clear and smooth monotonically decaying path on a macroscopic level, without any violent disordered oscillations. This fully verifies that the cascaded encoding mechanism of channel attention, residual convolution, and bidirectional long short-term memory networks, as well as the Koopman linear recursion mechanism, demonstrates powerful denoising and temporal filtering performance in the feature extraction and latent space modeling stages, greatly preserving the deep dynamic characteristics related to the essence of performance degradation.

[0152] In summary, Figure 4 The experimental results fully and intuitively verify that the potential health status characterization extracted in the embodiments of this application has extremely strong nonlinear characterization ability and evolution tracking accuracy for the remaining life of aero-engines, and can provide decisive and high-confidence data support for the medium and long-term health management, condition-based maintenance and ultra-precise prediction of the remaining life of engines.

[0153] The following describes the aero-engine performance degradation path evolution device provided in the embodiments of this application. The aero-engine performance degradation path evolution device described below and the aero-engine performance degradation path evolution method described above can be referred to in correspondence.

[0154] Figure 5 This is a schematic diagram of the structure of the aero-engine performance degradation path evolution device provided in the embodiments of this application. Figure 5 As shown, the device includes: a data acquisition module 501, a feature extraction module 502, and a path generation module 503.

[0155] The data acquisition module 501 is used to acquire full-life data of the aero-engine.

[0156] The feature extraction module 502 is used to input the full life data of the aero-engine into a pre-trained aero-engine performance degradation path evolution model, wherein the aero-engine performance degradation path evolution model includes a shared encoder and a Koopman transition matrix; the shared encoder extracts the initial latent variables characterizing the health state of the aero-engine, and the Koopman transition matrix generates a sequence of latent variables for all time steps within a sliding window through linear recursion of the initial latent variables.

[0157] The path generation module 503 is used to extract the latent variable of the last time step in the latent variable sequence as the health status representation of the corresponding sliding window, and to splice the health status representations of all sliding windows in chronological order to obtain the evolution path of the performance degradation of the aero-engine.

[0158] Optionally, the shared encoder includes a channel attention module.

[0159] Optionally, the feature extraction module 502 is specifically used to extract the full-lifetime data using the current sliding window to obtain the time-series monitoring data corresponding to the current sliding window, and input the time-series monitoring data into the channel attention module to perform adaptive channel weighting on the time-series monitoring data of each sensor channel to obtain weighted sensor data; based on the weighted sensor data, the initial latent variables characterizing the health status of the aero-engine are determined.

[0160] Optionally, the shared encoder also includes a residual convolution module cascaded with the channel attention module.

[0161] Optionally, the feature extraction module 502 is specifically used to input the weighted sensor data into the residual convolution module, extract the short-term degradation feature sequence that characterizes the local coupling characteristics between each sensor channel in the weighted sensor data, and determine the initial latent variable characterizing the health status of the aero-engine based on the short-term degradation feature sequence.

[0162] Optionally, the shared encoder also includes a bidirectional long short-term memory network module cascaded with the channel attention module and the residual convolution module.

[0163] Optionally, the feature extraction module 502 is specifically used to input the short-term degradation feature sequence into the bidirectional long short-term memory network module, extract the bidirectional long-term dependency of the short-term degradation feature sequence to obtain the hidden state sequence, and determine the initial latent variables characterizing the health state of the aero-engine based on the hidden state sequence.

[0164] Optionally, the aero-engine performance degradation path evolution model also includes an independent decoder corresponding to each sensor of the aero-engine.

[0165] Optionally, the feature extraction module 502 is specifically used to independently reconstruct and map the latent variable sequence through the independent decoder corresponding to each sensor to obtain the sensor time series signal corresponding to each independent decoder; and to splice the sensor time series signals output by each independent decoder to obtain multi-sensor reconstruction data.

[0166] Optionally, the standalone decoder includes a pointwise fully connected network and a residual convolutional network.

[0167] Optionally, the feature extraction module 502 is specifically used to input the latent variable sequence into the pointwise fully connected network for each independent decoder, perform independent fully connected mapping on the latent variables at each time step to obtain a first reconstructed signal; input the latent variable sequence into the residual convolutional network, perform one-dimensional convolutional feature extraction and smoothing filtering using information from adjacent time steps to obtain a second reconstructed signal; and superimpose the first reconstructed signal and the second reconstructed signal to obtain the sensor time-series signal corresponding to the independent decoder.

[0168] Optionally, the Koopman transition matrix is ​​used to establish a linear mapping relationship between adjacent time steps for the latent variables characterizing the health status of the aero-engine, and the size of the Koopman transition matrix matches the dimension of the initial latent variables.

[0169] The following describes the training device for the aero-engine performance degradation path evolution model provided in the embodiments of this application. The aero-engine performance degradation path evolution model training device described below and the aero-engine performance degradation path evolution model training method described above can be referred to in correspondence.

[0170] Figure 6 This is a schematic diagram of the structure of the training device for the aero-engine performance degradation path evolution model provided in this application embodiment. (See attached diagram.) Figure 6 As shown, the device may include: an acquisition module 601, a calculation module 602, and a training module 603.

[0171] The acquisition module 601 is used to acquire historical sample data of the entire life cycle of the aero-engine.

[0172] The calculation module 602 is used to extract initial latent variables representing the health status of the aero-engine based on the full-lifetime historical sample data using the shared encoder; to linearly recursively derive the initial latent variables using the Koopman transition matrix to generate a sequence of latent variables; and to reconstruct the sequence of latent variables using each of the independent decoders to obtain multi-sensor reconstructed data.

[0173] Training module 603 is used to train the data based on the lifetime historical sample data, the multi-sensor reconstructed data, and the hyperparameters. The weighted Kullback-Leibler divergence is used to construct a joint loss function. Based on this joint loss function, the network parameters of the shared encoder, the Koopman transition matrix, and each independent decoder are updated to obtain the trained aero-engine performance degradation path evolution model.

[0174] Optionally, the training module 603 is specifically used to calculate the mean of the sum of squared errors between the full-lifetime historical sample data and the multi-sensor reconstructed data, as the reconstruction loss term; and to calculate the Kullback-Leibler divergence between the latent variable distribution of the shared encoder output and the preset standard normal prior distribution, and to incorporate the hyperparameters... Multiplying the value of the hyperparameter by the calculated Kullback-Leibler divergence yields the result obtained from the hyperparameter. The weighted Kullback-Leibler divergence is used as a decoupling constraint term; the reconstruction loss term and the decoupling constraint loss term are weighted and summed to obtain the joint loss function.

[0175] Optionally, the training module 603 is specifically used to record the current training round of the aero-engine performance degradation path evolution model during iterative training; and to control the hyperparameter when the current training round is less than or equal to a preset transition round threshold. The value starts from 0 and increases linearly with the current training epoch until it reaches the preset maximum target weight value; if the current training epoch is greater than the preset transition epoch threshold, the hyperparameter is... The value of this hyperparameter remains constant at the preset maximum target weight value; the hyperparameter corresponding to the current training round is... The value is multiplied by the calculated Kullback-Leibler divergence to obtain the result from the hyperparameter. The weighted Kullback-Leibler divergence is used as a decoupling constraint term in the joint loss function.

[0176] Optionally, the training module 603 is specifically used to extract the Koopman transition matrix corresponding to the current training round, calculate the maximum singular value of the Koopman transition matrix, and if the maximum singular value is greater than 1, use the maximum singular value to perform spectral normalization scaling on the Koopman transition matrix to constrain the spectral radius of the Koopman transition matrix to not exceed 1.

[0177] Figure 7 This is a schematic diagram of the structure of the electronic device provided in an embodiment of this application. For example... Figure 7As shown, the electronic device may include a processor 710, a communications interface 720, a memory 730, and a communication bus 740. The processor 710, communications interface 720, and memory 730 communicate with each other via the communication bus 740. The processor 710 can call logical instructions from the memory 730 to execute methods for analyzing the evolution of aero-engine performance degradation paths or training methods for aero-engine performance degradation path evolution models.

[0178] Furthermore, the logical instructions in the aforementioned memory 730 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0179] On the other hand, this application also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the aero-engine performance degradation path evolution method or the aero-engine performance degradation path evolution model training method provided by the above methods.

[0180] In another aspect, embodiments of this application also provide a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, is implemented to perform the aero-engine performance degradation path evolution method or the aero-engine performance degradation path evolution model training method provided by the above methods.

[0181] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0182] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0183] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for the evolution of performance degradation paths in aero-engines, characterized in that, The method includes: Obtain full life-cycle data of aero engines; The full life-cycle data of the aero-engine is input into a pre-trained aero-engine performance degradation path evolution model, wherein the aero-engine performance degradation path evolution model includes a shared encoder and a Koopman transition matrix; the shared encoder extracts initial latent variables characterizing the health state of the aero-engine, and the Koopman transition matrix generates a sequence of latent variables for all time steps within a sliding window through linear recursion of the initial latent variables; The latent variable at the last time step in the latent variable sequence is extracted as the health status representation of the corresponding sliding window, and the health status representations of all sliding windows are spliced ​​together in chronological order to obtain the evolution path of the aero-engine performance degradation.

2. The method for predicting the performance degradation path of an aero-engine according to claim 1, characterized in that, The shared encoder includes a channel attention module, and the step of extracting initial latent variables characterizing the health state of the aero-engine through the shared encoder includes: The full-lifetime data is extracted using the current sliding window to obtain the time-series monitoring data corresponding to the current sliding window. The time-series monitoring data is then input into the channel attention module to perform adaptive channel weighting on the time-series monitoring data of each sensor channel, resulting in weighted sensor data. Based on the weighted sensor data, initial latent variables characterizing the health status of the aero-engine are determined.

3. The method for predicting the performance degradation path of an aero-engine according to claim 2, characterized in that, The shared encoder further includes a residual convolution module cascaded with the channel attention module. The step of determining the initial latent variables characterizing the health status of the aero-engine based on the weighted sensor data includes: The weighted sensor data is input into the residual convolution module to extract short-term degradation feature sequences that characterize the local coupling features between each sensor channel from the weighted sensor data. Based on the short-term degradation feature sequence, initial latent variables characterizing the health status of the aero-engine are determined.

4. The method for predicting the performance degradation path of an aero-engine according to claim 3, characterized in that, The shared encoder further includes a bidirectional long short-term memory network module cascaded with the channel attention module and the residual convolution module. The step of determining the initial latent variables characterizing the health state of the aero-engine based on the short-term degradation feature sequence includes: The short-term degradation feature sequence is input into the bidirectional long short-term memory network module to extract the bidirectional long-term dependency of the short-term degradation feature sequence and obtain the hidden state sequence. Based on the hidden state sequence, initial latent variables characterizing the health state of the aero-engine are determined.

5. A training method for an aero-engine performance degradation path evolution model, characterized in that, The aero-engine performance degradation path evolution model includes a cascaded shared encoder, a Koopman transition matrix, and independent decoders corresponding to each sensor; the method includes: Obtain historical sample data of the entire life cycle of aero engines; Using the shared encoder, initial latent variables characterizing the health status of the aero-engine are extracted based on the full-lifetime historical sample data; Using the Koopman transition matrix, the initial latent variables are linearly recursively processed to generate a sequence of latent variables; The latent variable sequences are reconstructed using the independent decoders described above to obtain multi-sensor reconstructed data. Based on the full-lifetime historical sample data, the multi-sensor reconstructed data, and the hyperparameters... The weighted Kullback-Leibler divergence is used to construct a joint loss function. Based on the joint loss function, the network parameters of the shared encoder, the Koopman transition matrix, and each of the independent decoders are updated to obtain the trained aero-engine performance degradation path evolution model.

6. The training method for the aero-engine performance degradation path evolution model according to claim 5, characterized in that, The data is based on the full-lifetime historical sample data, the multi-sensor reconstructed data, and the hyperparameters. The steps for constructing the joint loss function using weighted Kullback-Leibler divergence include: Calculate the mean of the sum of squared errors between the full-lifetime historical sample data and the multi-sensor reconstructed data, and use it as the reconstruction loss term; Calculate the Kullback-Leibler divergence between the latent variable distribution of the shared encoder output and the preset standard normal prior distribution, and then use the hyperparameters... Multiplying the value of the hyperparameter by the calculated Kullback-Leibler divergence yields the result obtained from the hyperparameter. The weighted Kullback-Leibler divergence is used as a decoupling constraint term; The joint loss function is obtained by weighted summation of the reconstruction loss term and the decoupling constraint loss term.

7. The training method for the aero-engine performance degradation path evolution model according to claim 6, characterized in that, The hyperparameters Multiplying the value of the hyperparameter by the calculated Kullback-Leibler divergence yields the result obtained from the hyperparameter. The steps for weighted Kullback-Leibler divergence include: During the iterative training of the aero-engine performance degradation path evolution model, the current training round of the aero-engine performance degradation path evolution model is recorded. When the current training round is less than or equal to a preset transition round threshold, the hyperparameters are controlled. The value starts from 0 and increases linearly with the current training round until the preset maximum target weight value is reached. If the current training round is greater than the preset transition round threshold, the hyperparameters will be adjusted. The value remains constant at the preset maximum target weight value; The hyperparameters corresponding to the current training round The value is multiplied by the calculated Kullback-Leibler divergence to obtain the value derived from the hyperparameter. The weighted Kullback-Leibler divergence is used as a decoupling constraint term in the joint loss function.

8. The training method for the aero-engine performance degradation path evolution model according to claim 5, characterized in that, The process of updating the parameters of the Koopman transition matrix also includes: Extract the Koopman transition matrix corresponding to the current training round, and calculate the maximum singular value of the Koopman transition matrix; When the maximum singular value is greater than 1, the Koopman transition matrix is ​​spectrally normalized and scaled using the maximum singular value, constraining the spectral radius of the Koopman transition matrix to not exceed 1.

9. A device for analyzing the performance degradation path evolution of an aero-engine, characterized in that, The apparatus is used to implement the aero-engine performance degradation path evolution method as described in any one of claims 1-4, comprising: The data acquisition module is used to acquire full-lifecycle data of the aero-engine; The feature extraction module is used to input the full life-cycle data of the aero-engine into a pre-trained aero-engine performance degradation path evolution model, wherein the aero-engine performance degradation path evolution model includes a shared encoder and a Koopman transition matrix; the shared encoder extracts initial latent variables characterizing the health state of the aero-engine, and the Koopman transition matrix generates a sequence of latent variables for all time steps within a sliding window through linear recursion of the initial latent variables; The path generation module is used to extract the latent variable of the last time step in the latent variable sequence as the health status representation of the corresponding sliding window, and to splice the health status representations of all sliding windows in chronological order to obtain the evolution path of the aero-engine performance degradation.

10. A training device for an aero-engine performance degradation path evolution model, characterized in that, The apparatus is used to implement the training method for the aero-engine performance degradation path evolution model as described in any one of claims 5-8, including: The acquisition module is used to acquire historical sample data of the entire life cycle of the aero-engine; The calculation module is used to extract initial latent variables characterizing the health status of the aero-engine based on the full-lifetime historical sample data using the shared encoder; to linearly recursively calculate the initial latent variables using the Koopman transition matrix to generate a sequence of latent variables; and to reconstruct the sequence of latent variables using each of the independent decoders to obtain multi-sensor reconstructed data. The training module is used to train data based on the full-lifetime historical sample data, the multi-sensor reconstructed data, and the hyperparameters. The weighted Kullback-Leibler divergence is used to construct a joint loss function. Based on the joint loss function, the network parameters of the shared encoder, the Koopman transition matrix, and each of the independent decoders are updated to obtain the trained aero-engine performance degradation path evolution model.

Images