Turboshaft engine operation monitoring method and system based on digital twinning

Abstract

The application discloses a turbofan engine operation monitoring method and system based on digital twinning, belongs to the technical field of turbofan engine monitoring, and aims to solve the problems of difficult extraction of early crack, wear and other weak fault signals and low prediction accuracy of multi-source fault propagation path in a strong noise background. Original operation signals are collected through a sensing array, and demodulation signals are generated through adaptive resonance demodulation chain processing. Time-frequency transformation is performed on the demodulation signals, the measured characteristic frequency is matched with the theoretical characteristic frequency, an initial candidate feature set is constructed, independent feature sets are generated by fusing a multi-scale decoupling network constrained by digital twinning, and effective causal pairs are screened out for the independent features by adopting Granger causality analysis combined with physical coupling relationship simulation of digital twinning. A dynamic Bayesian network is constructed to simulate fault propagation, the post-probability is calculated through digital twinning verification and Monte Carlo simulation, early warning is triggered, and maintenance decisions are generated, so that weak signal extraction and accurate fault prediction are realized in a strong noise background.

Description

Technical Field

[0001] The present invention relates to the technical field of turbofan engine monitoring, and more particularly to a turbofan engine operation monitoring method and system based on digital twin. Background Art

[0002] As a core component of aviation propulsion, monitoring the operating status of turbofan engines is crucial for their safety and reliability. In actual operating conditions, engines often face complex conditions characterized by high temperatures, high pressures, and strong vibration and noise. Transient impact signals generated by early-stage faults, such as bearing microwear and initial blade cracks, are weak and suffer from frequency aliasing, making them easily masked by strong background noise. This makes it difficult for traditional signal processing techniques to effectively capture fault characteristics, leading to delayed early fault identification and a high risk of missed detection.

[0003] On the other hand, turbofan engine faults often exhibit multi-component coupled propagation characteristics. For example, bearing failure may cause rotor imbalance and further lead to blade rubbing. Existing fault prediction models rely on single-dimensional features or simple causal associations, and have difficulty handling the strong coupling relationship between time domain, frequency domain, and time-frequency joint features, resulting in fuzzy fault propagation path analysis and insufficient prediction accuracy. In addition, traditional methods lack effective constraints on feature independence, are easily affected by noise interference, and affect feature matching accuracy, and cannot meet the needs of accurate tracing of multi-source faults under complex working conditions. Therefore, in order to overcome these limitations, the present invention proposes a turbofan engine operation monitoring method and system based on digital twins. Summary of the Invention

[0004] To address the shortcomings of existing technologies, the present invention aims to provide a turbofan engine operation monitoring method and system based on digital twins. Through multi-dimensional signal processing and dynamic causal modeling, this method efficiently isolates subtle fault signatures and predicts complex fault propagation paths. This improves the detectability of early-stage faults in high-noise environments, overcomes the technical bottleneck of multi-source coupled fault propagation modeling, and significantly enhances the reliability and prediction accuracy of turbofan engine operation monitoring under complex operating conditions.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] The turbofan engine operation monitoring system based on digital twin includes:

[0007] Build a digital twin of the turbofan engine to map the physical operating status of the turbofan engine in real time;

[0008] Collect the raw operating signals of the turbofan engine, reduce noise and enhance features of the raw operating signals through an adaptive resonance demodulation chain, output a pre-processed and standardized demodulated signal set, and dynamically map and visualize it in the digital twin;

[0009] The adaptive resonance demodulation chain initializes the variational mode decomposition parameters based on the main energy frequency bands of the original operating signal. By constructing a variational constrained optimization model, the original operating signal is structurally decomposed in the frequency domain. The effective intrinsic mode function components obtained by decomposition are subjected to nonlinear envelope demodulation using the discrete Teager energy operator, and a standardized demodulated signal set is generated through normalization processing.

[0010] The standardized demodulated signal set is transformed into a time-frequency matrix. The theoretical characteristic frequencies of turbofan engine component faults are quantified by constructing a priori calculation rules for fault characteristic frequencies. An initial candidate feature set is constructed. The time domain features, frequency domain features, and time-frequency joint features of the initial candidate feature set are separated through the hierarchical architecture of a multi-scale decoupling network to generate an independent feature set for each fault type of turbofan engine component.

[0011] For the independent feature set of each fault type of turbofan engine components, Granger causality analysis is used in combination with the physical coupling relationship of digital twin simulation to screen effective causal pairs, and a dynamic Bayesian network is constructed to predict the fault propagation path and component failure probability, trigger fault warning and generate maintenance decision recommendations.

[0012] Specifically, the steps for constructing the initial candidate feature set include:

[0013] Performing time-frequency transformation on the standardized demodulated signal set to generate a time-frequency matrix. In the time-frequency matrix, the frequency of the frequency peak point in the energy concentration area is identified as the measured frequency;

[0014] Construct a priori calculation rule system of fault characteristic frequency consisting of component classification rules and theoretical calculation formulas to quantify the theoretical characteristic frequency of turbofan engine component failures;

[0015] Match the measured frequency extracted from the time-frequency matrix with the theoretical characteristic frequency obtained by the fault characteristic frequency prior calculation rule system, and calculate the relative error;

[0016] Set a component error tolerance threshold. If the relative error between the measured frequency and the theoretical characteristic frequency of the component is less than the corresponding component error tolerance threshold, obtain the frequency peak point corresponding to the measured frequency and construct the initial candidate feature set.

[0017] An effective threshold is set and the number of frequency peak points in the initial candidate feature set is counted. If the number is less than the effective threshold, the missing warning mechanism is triggered and the component error tolerance threshold is dynamically adjusted. The component error tolerance threshold is relaxed through a dynamic threshold adjustment strategy based on operating condition parameters, and the initial candidate feature set is updated.

[0018] Specifically, the fault characteristic frequency prior calculation rule system includes:

[0019] The turbofan engine is divided into components, including bearings, blades, transmissions, and air paths;

[0020] Define the determinants and physical mechanisms of the characteristic frequencies of turbofan engine components; the determinants are used to quantify the influence of component structural and operating parameters on the characteristic frequencies, and the physical mechanisms are used to establish a mapping relationship between characteristic frequencies and failure modes;

[0021] According to the determining factors and physical mechanisms of the characteristic frequency of turbofan engine components, a characteristic frequency calculation formula is constructed for different component categories, and the theoretical characteristic frequency is quantitatively deduced and calculated to obtain the theoretical characteristic frequency of turbofan engine components.

[0022] Specifically, the multi-scale decoupling network includes:

[0023] A bidirectional long short-term memory network architecture is used to construct a time-domain decoupling layer. By configuring an attention mechanism, the time series characteristics of the standardized demodulated signal are analyzed to construct a time-domain feature vector.

[0024] A convolutional neural network with a bandpass convolution kernel is configured as the frequency domain decoupling layer to perform hierarchical filtering on the spectrum of the standardized demodulated signal, extract the amplitude spectrum and harmonic structure, and construct the frequency domain feature vector.

[0025] The Transformer architecture is used to construct a joint time-frequency decoupling layer, configure a multi-head self-attention mechanism, set a mutual information minimization loss function, and construct a joint time-frequency feature vector.

[0026] The physical constraint loss terms of the multi-scale decoupling network based on the digital twin physical model are set, including frequency constraint and amplitude constraint. The frequency constraint is used to measure the deviation between the characteristic frequency after decoupling and the theoretical characteristic frequency of the digital twin simulation. The amplitude constraint is used to measure the fitting error between the characteristic amplitude after decoupling and the nonlinear mapping relationship between fault degree and amplitude established by the digital twin.

[0027] Specifically, the specific steps of generating an independent feature set for each fault type of a turbofan engine component include:

[0028] The initial candidate feature set of turbofan engine components and the standardized demodulated signal are processed by a multi-scale decoupling network. The multi-scale decoupling network separates the time domain features, frequency domain features and time-frequency joint features of the initial candidate feature set through a hierarchical architecture to output multi-scale feature vectors, including time domain feature vectors, frequency domain feature vectors and time-frequency joint feature vectors.

[0029] The multi-scale feature vector output by the multi-scale decoupling network is similar to the pre-built standard fault sample feature vector library of turbofan engine components. The fault type is matched by configuring the similarity threshold to determine whether the multi-scale feature vector output by the multi-scale decoupling network and the corresponding standard fault sample feature vector belong to the same fault type.

[0030] The multi-scale feature vectors output by the multi-scale decoupling network are classified and reconstructed according to the fault types of turbofan engine components, and multi-dimensional fault features are constructed, including time domain features, frequency domain features, and time-frequency domain features. The multi-dimensional fault features are standardized to generate an independent feature set for each fault type of turbofan engine components.

[0031] Specifically, the steps of performing noise reduction and feature enhancement on the original running signal through the adaptive resonant demodulation chain and outputting a set of pre-processed standardized demodulated signals include:

[0032] Calculate the frequency range of the original running signal by short-time Fourier transform and define the main energy frequency band of the original running signal;

[0033] Based on the main energy frequency band, the initial value of the number of decomposition layers is set, and a variational constrained optimization model of variational mode decomposition is constructed. The objective function of the variational constrained optimization model is to minimize the bandwidth sum of each intrinsic mode function component. The constraint condition of the variational constrained optimization model is that the sum of all intrinsic mode function components is equal to the input original operating signal; and the initial value of the center frequency of each intrinsic mode function component is evenly distributed within the main energy frequency band;

[0034] Iteratively solve the variational constrained optimization model of variational mode decomposition to obtain independent frequency subbands corresponding to different intrinsic mode function components; and calculate the kurtosis value for each intrinsic mode function component during the iterative process;

[0035] Configure the kurtosis threshold, filter the intrinsic mode function components according to the kurtosis threshold, obtain the valid intrinsic mode function components, and construct the valid component set.

[0036] Specifically, the steps of performing noise reduction and feature enhancement on the original running signal through the adaptive resonance demodulation chain and outputting a pre-processed standardized demodulated signal set also include:

[0037] Perform nonlinear envelope demodulation on the effective intrinsic mode function components in the effective component set, calculate the instantaneous energy of the effective intrinsic mode function components point by point through the discrete Teager energy operator, and generate a demodulated signal set through envelope detection;

[0038] A bandpass filter is set, and the passband range of the bandpass filter is constrained within the main energy frequency band. The demodulated signal set is filtered, and the demodulated signal is standardized by calculating the effective value, peak value and peak factor of the demodulated signal to generate a standardized demodulated signal set.

[0039] Specifically, the steps for predicting fault propagation paths and component failure probabilities, triggering fault warnings, and generating maintenance decision recommendations include:

[0040] For each independent feature set of turbofan engine component fault type, a vector autoregression model was constructed using Granger causality analysis. The time lag order was configured to capture the temporal dependencies between the fault type features. The causal test statistics between the independent feature sets were calculated based on the time lag order. Significant causal pairs of component fault types were screened by setting a significance level threshold.

[0041] Construct a pseudo-causal relationship library, which contains pseudo-causal relationships that violate physical laws. Compare the selected significant causal pairs with the pseudo-causal relationship library, eliminate the significant causal pairs with pseudo-causal relationships, and obtain valid causal pairs.

[0042] A feature correlation matrix is ​​constructed based on effective causal pairs. The elements of the feature correlation matrix represent the intensity of the leading and lagging influence between independent feature sets through standardized Granger causality statistics.

[0043] Specifically, the steps for predicting fault propagation paths and component failure probabilities, triggering fault warnings, and generating maintenance decision recommendations include:

[0044] A dynamic Bayesian network is constructed based on the feature correlation matrix. The number of network time slices is set. The edge weights of each component node are determined by combining the standardized Granger causal statistics and the causal strength of the digital twin simulation, and the conditional probability table is initialized.

[0045] Based on a dynamic Bayesian network, Monte Carlo simulation and digital twin virtual verification are combined to generate fault propagation paths. For each component node, the digital twin performs multi-physics simulation of the fault propagation path, quantifying the frequency of fault states to calculate the posterior probability of component node failure. A fault probability warning threshold is set, and warnings are triggered for component nodes with posterior probabilities greater than the fault probability warning threshold. The fault propagation path is also visualized.

[0046] For the fault propagation path, the dynamic Bayesian network is used to simulate the inhibitory effect of the intervention plans in the causal intervention plan library on the failure probability of each component node. The virtual maintenance cost of the fusion digital twin is introduced, including working hours, virtual loss of component replacement, and the comprehensive cost function of actual operation and maintenance data. The failure probability reduction rate and input-output ratio of different intervention plans are calculated to generate maintenance decision recommendations ranked by comprehensive benefits.

[0047] The turbofan engine operation monitoring method based on digital twinning includes the following steps:

[0048] Step S1: Construct a turbofan engine digital twin, which maps the real-time operation state of the turbofan engine entity;

[0049] Step S2: Collect the original operation signal of the turbofan engine, and perform noise reduction and feature enhancement on the original operation signal through an adaptive resonance demodulation chain, output a set of preprocessed standardized demodulation signals, and dynamically map and visualize in the digital twin;

[0050] The adaptive resonance demodulation chain initializes the variational modal decomposition parameters based on the main energy band of the original operation signal, and performs structured decomposition on the original operation signal in the frequency domain by constructing a variational constraint optimization model; the effective intrinsic modal function components obtained by decomposition are nonlinearly envelope demodulated by a discrete Teager energy operator, and a set of standardized demodulation signals is generated by standardization processing;

[0051] Step S3: Perform time-frequency transformation on the set of standardized demodulation signals to generate a time-frequency matrix, and construct an initial candidate feature set by quantifying the theoretical characteristic frequency of the turbofan engine component fault through the construction of a fault characteristic frequency prior calculation rule system, separate the time domain features, frequency domain features and time-frequency joint features of the initial candidate feature set through the hierarchical architecture of the multi-scale decoupling network, and generate an independent feature set for each type of turbofan engine component fault;

[0052] Step S4: For each type of independent feature set of the turbofan engine component fault, effective causal pairs are screened by using Granger causality analysis combined with the physical coupling relationship of digital twin simulation, and a dynamic Bayesian network is constructed to predict the fault propagation path and component fault probability, trigger fault warning and generate maintenance decision suggestions.

[0053] The beneficial effects of the present application are:

[0054] The present invention addresses the problem that early weak fault signals are easily drowned out under strong noise, resulting in a high missed detection rate. By decomposing the frequency domain and demodulating through an adaptive resonance demodulation chain linked to the digital twin, the decomposition parameters are initialized based on the main energy frequency band simulated by the digital twin, thereby enhancing the fault characteristics such as bearing wear and blade cracks, suppressing noise interference, improving signal detectability, and solving the missed detection problem of traditional methods. In order to address the problems of difficult processing of multi-dimensional coupling features and fuzzy propagation path analysis, the present invention separates the time domain, frequency domain and time-frequency joint features through a multi-scale decoupling network, combines the physical constraint loss term of the digital twin to generate an independent feature set, and uses the virtual fault samples of the digital twin to expand the training data. , solve the problem of separation of strongly coupled features and provide highly recognizable features for accurate tracing; in response to the problem of insufficient prediction accuracy of multi-component coupled faults, Granger causality analysis is used and the physical coupling relationship of digital twins is integrated to screen causal pairs, construct a dynamic Bayesian network, and use the multi-physical field simulation data of digital twins to initialize network parameters. The propagation path is analyzed by combining Monte Carlo simulation and digital twin virtual verification, and a comprehensive cost function that integrates digital twin maintenance costs is introduced to generate maintenance decisions, realizing the integration of decoupling, prediction and intervention, improving the accuracy of fault propagation analysis and prediction under complex working conditions, and solving the bottlenecks of existing technologies in strong noise feature extraction and multi-source coupling modeling. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 Schematic diagram of the structure of the turbofan engine operation monitoring system based on digital twin of the present invention;

[0056] Figure 2 A flowchart of the specific steps for obtaining a standardized demodulated signal set according to the present invention;

[0057] Figure 3 Flowchart for obtaining the initial candidate feature set of the present invention;

[0058] Figure 4 A flow chart for generating an independent feature set for each fault type of a turbofan engine component for the present invention;

[0059] Figure 5 This is a flow chart of the turbofan engine operation monitoring method based on digital twins of the present invention. DETAILED DESCRIPTION

[0060] Example 1:

[0061] See also Figure 1 ,This embodiment introduces a turbofan engine operation monitoring system based on digital twin,,including a digital twin mapping module, an anti-noise perception module, a feature decoupling module, and a dynamic causal module;

[0062] The digital twin mapping module is used to construct a digital twin that includes the turbofan engine's geometry, material characteristics, and power parameters based on the turbofan engine's physical model and historical operating data, and to map the turbofan engine's physical operating status in real time. The historical operating data includes a full life cycle operating parameter database, a fault case feature library, a material performance degradation library, a repair and maintenance log, and an operating environment dataset.

[0063] In this example, the digital twin mapping module builds a multi-physics coupling model containing components such as the compressor, combustion chamber, and turbine based on the three-dimensional CAD model of the turbofan engine. It uses the steady-state data such as the engine speed, pressure ratio, temperature field distribution, and vibration amplitude under various operating conditions in the full life cycle operating parameter database, covering rated operating conditions, variable load conditions, and start-stop transient processes. It iteratively optimizes dynamic parameters such as bearing stiffness and blade damping through finite element analysis; and integrates the characteristic frequency, vibration waveform, oil abrasive concentration, and other diagnostic features of typical faults such as bearing spalling and blade cracks in historical fault events in the integrated fault case feature library. Data is used to generate virtual fault samples with different degrees of damage, drive multi-scale decoupling network training, and improve the accuracy of feature decoupling; relying on the fatigue life data, thermal expansion coefficient change curve and corrosion rate history of high-temperature alloys, composite materials and other components in the material performance degradation library, the remaining life loss rate is calculated in real time; according to the operation and maintenance data such as component replacement records, calibration parameters, lubrication cycles in the maintenance log, it is used to correct the historical records of the initial damage state of the digital twin; combined with the performance degradation laws under different altitudes, temperatures and humidity conditions in the working environment data set, it is used for parameter calibration of the environmental compensation model.

[0064] The anti-noise sensing module is used to collect raw operating signals, including mechanical vibration signals, acoustic characteristic signals, thermal state signals, fluid dynamics signals, oil state signals, and environmental compensation signals, through a sensor array configured on the turbofan engine. The sensor array uses double-shielded cables and optoelectronic isolation technology to ensure the reliability and integrity of raw signal acquisition in high-noise environments. The adaptive resonant demodulation chain reduces noise and enhances features of the raw operating signals to obtain a pre-processed, standardized demodulated signal set containing clear fault characteristics. This provides a highly reliable input data source for subsequent feature decoupling and enables dynamic mapping and visualization of multiple components in the digital twin, creating an integrated solution for signal acquisition and preprocessing in high-noise environments. This significantly improves the detectability of early-stage fault characteristics and lays a data foundation for fault diagnosis and prediction of turbofan engines.

[0065] In this embodiment, the original operating signal is collected by the sensor array of the turbofan engine to fully characterize the operating status of the turbofan engine and capture early signs of faults. Among them, the mechanical vibration signal is used to monitor mechanical faults such as rotor imbalance, bearing peeling, and blade rubbing; the acoustic characteristic signal is used to identify faults such as abnormal wear of bearings and gears, and unstable combustion in the combustion chamber; the thermal state signal is used to evaluate combustion stability, thermal fatigue of hot components and cooling system efficiency; the fluid mechanics signal is used to analyze the risk of compressor surge, combustion oscillation and damage to gas path components; the oil state signal is used for bearing and gear wear trend analysis and lubrication system fault diagnosis; the environmental compensation signal is used to correct the influence of environmental factors such as altitude, temperature, and humidity on signal characteristics; the original operating signal is transmitted through double-shielded cables and optoelectronic isolation technology to ensure transmission reliability, and then the adaptive resonance demodulation chain is used to reduce noise and enhance features: the noise and fault frequency bands are adaptively separated by variational mode decomposition, combined with Teag The er energy operator enhances the transient impact characteristics and ultimately outputs a standardized demodulated signal set containing clear fault characteristics, such as the bearing fault characteristic frequency, blade crack peak factor, and abrasive concentration mutation gradient. Dynamic mapping and visualization of multiple components are realized in the digital twin. Through the coupled rendering of the digital twin 3D model and signal characteristics, key parameters such as the bearing fault characteristic frequency, blade crack peak factor, and abrasive concentration mutation gradient in the standardized demodulated signal set are converted into interactive virtual representations. The spatial distribution of the fault characteristics of each component is presented in the form of a dynamic heat map, and the time series evolution of the characteristic frequency is displayed through a spectrum waterfall diagram. Combined with vibration mode rendering, the physical mapping of signal characteristics and mechanical behavior is realized; it provides a high-reliability input data source for the subsequent feature decoupling engine, realizes an integrated solution for signal acquisition and preprocessing in a strong noise environment, significantly improves the detectability of early faults, and lays a data foundation for fault diagnosis and remaining life prediction of turbofan engines.

[0066] See also Figure 2 Preferably, the specific steps of obtaining the standardized demodulated signal set include:

[0067] In the high-noise environment of a turbofan engine, characteristic fault frequency bands can be overwhelmed. Direct full-band decomposition introduces significant noise computation and can miss weak fault signals. A short-time Fourier transform is used to calculate the frequency range of the original operating signal and define its primary energy band. For example, the range from the lowest to the highest frequency where 95% of the energy is accumulated is considered the primary energy band, ensuring coverage of the signal's primary energy components.

[0068] The penalty factor is set to balance the decomposition accuracy and noise suppression effect in the variational mode decomposition process; the bandwidth constraint strength of each intrinsic mode function can be controlled by adjusting the penalty factor value.

[0069] The initial value of the decomposition layer number is adaptively set based on the energy density spectrum distribution characteristics of the main energy frequency bands, which is used to perform preliminary frequency band division on the original operating signal and provide an initial decomposition structure for subsequent variational constrained optimization; the preliminary frequency band division is made to cover the core energy area to avoid over-decomposition or under-decomposition.

[0070] Specifically, the specific steps of setting the initial value of the decomposition layer number include:

[0071] The short-time Fourier transform results in the main energy frequency band are smoothed, the spectral energy density function is calculated, the first-order derivative of the spectral energy density function is solved, and the inflection point of the energy change of the spectral energy density function is identified. The inflection point corresponds to the frequency band boundary of different vibration sources in the original operating signal.

[0072] Based on the statistical characteristics of the spectrum energy density spectrum, including the mean and variance, the gradient threshold is dynamically set, and the inflection point where the absolute value of the gradient is greater than the gradient threshold is marked as a potential frequency band splitting point;

[0073] The number of potential frequency band segmentation points is counted, and combined with the characteristics of variational mode decomposition, the initial value of the decomposition layer is set to the number of potential frequency band segmentation points plus one.

[0074] Construct a variational constrained optimization model for variational mode decomposition. Its objective function is to minimize the bandwidth sum of each intrinsic mode function component. The constraint condition is that the sum of all intrinsic mode function components is equal to the input original operating signal. This ensures that the optimization starting point is located in the effective frequency band, accelerates convergence, and prevents the center frequency from being initialized to the noise frequency band, resulting in decomposition failure. For example, the variational constrained optimization model for variational mode decomposition is:

[0075]

[0076] in, is the set of intrinsic mode function components to be decomposed, For the eigenmode function components, Indicates the The intrinsic mode function component signals are used to express the separated independent frequency band signals such as bearing vibration and blade flutter; is the set of center frequencies of each eigenmode function component, It is The center frequency of the eigenmode function component, , is the frequency value in Hz, which is used to represent the corresponding fault characteristic frequency, such as the fault frequency of the bearing outer ring; is the unit impulse function, which is used to construct frequency domain constraints; It is an imaginary unit, representing the complex frequency domain characteristics of the signal; is the circular constant pi, is a time variable, which is the time axis data of signal acquisition; is a convolution operator, which is used to construct the frequency domain bandwidth constraint of the signal; is a time differential operator, which is used to extract the instantaneous frequency and bandwidth characteristics of the signal; is a complex exponential modulation factor, which realizes the translation of the signal from the time domain to the frequency domain; is is the square of the norm, which is used to calculate the energy norm of the signal; is the number of decomposition layers, which is used to represent the number of frequency bands to be decomposed, is a minimum function; is the input original running signal, which is used to represent the original monitoring data containing noise.

[0077] The initial values of the center frequencies of the principal modal function components are uniformly distributed in the main energy frequency band, that is, the center frequencies of the principal modal function components are initialized with equal intervals in the main energy frequency band. The alternating direction multiplier method is used to iteratively solve the variational constraint optimization model of the variational modal decomposition. By optimizing the center frequencies and bandwidths of the principal modal function components in the frequency domain, the original running signal is structurally decomposed in the frequency domain, and the frequency centers of the principal modal function components are stabilized in the corresponding fault characteristic frequency band.

[0078] In the iteration process, the kurtosis value of each principal modal function component is calculated. The kurtosis value is an index for reflecting the impact characteristics of the signal, and the greater the kurtosis value, the more obvious the impact characteristics.

[0079] A kurtosis threshold is configured to screen the principal modal function components with significant impact characteristics. According to the kurtosis threshold, the principal modal function components with kurtosis values greater than the kurtosis threshold are screened and marked as effective principal modal function components. An effective component set is constructed, and noise-dominant components with low kurtosis are removed, thereby retaining high-frequency impact characteristic components related to early faults.

[0080] The engine speed change rate is monitored, and the penalty factors of the objective function and the constraint conditions of the variational constraint optimization model are dynamically adjusted based on the speed change ratio to avoid frequency band drift caused by speed change and ensure effective separation of high-frequency fault characteristics under variable speed conditions. Finally, an effective component set containing fault frequency bands is obtained. Each effective principal modal function component in the effective component set corresponds to a specific fault-related frequency band, and the frequency domain separation of the original running signal in a strong noise environment is realized.

[0081] Nonlinear envelope demodulation is performed on the effective intrinsic mode function components in the effective component set. The instantaneous energy of the effective intrinsic mode function components is calculated point by point using the discrete Teager energy operator. The instantaneous energy characteristics of the effective intrinsic mode function components are extracted through nonlinear operations. The transient impact signal is amplified and the sinusoidal steady-state noise is suppressed to form an energy distribution sequence. Then, a demodulated signal set is generated through envelope detection. Each demodulated signal highlights the impact component in the original operating signal, such as the transient vibration response caused by cracks.

[0082] A bandpass filter is set according to the fault characteristic frequency of the engine component. The passband range of the bandpass filter is constrained within the main energy frequency band. The demodulated signal set is filtered and standardized by calculating the effective value, peak value and peak factor of the demodulated signal to eliminate amplitude differences and generate a standardized demodulated signal set.

[0083] The feature decoupling module is used to separate the multi-source fault characteristics of the standardized demodulated signal set output by the anti-noise perception module based on the physical constraints of the digital twin and the data-driven dual-engine architecture. The digital twin is based on the component dynamics model of real-time working condition simulation, and dynamically updates the fault feature frequency prior calculation rule system. It first reconstructs the time-frequency space of the standardized demodulated signal and embeds the engine component fault feature frequency prior to screen the initial candidate feature set; based on physical prior and time-frequency analysis, the fault features of components such as bearings and blades that are mixed in a wide frequency band are separated, solving the problem that traditional methods cannot distinguish coupling features with close frequencies; then, by building a multi-scale decoupling network and generating virtual fault samples with different damage degrees, the training data set of the multi-scale decoupling network is expanded, and the strong coupling features are separated by combining the mutual information minimization constraint and the physical law regularization loss term; the physical constraint loss of the digital twin is deeply embedded in the feature decoupling process through the loss function of the multi-scale decoupling network. The frequency constraint term introduces the component natural frequency fluctuation threshold calibrated by the digital twin, and the amplitude constraint term is associated with the nonlinear mapping relationship between the fault degree and amplitude established by the digital twin. After optimization by the Lagrange multiplier method, the fit between the decoupling feature and the digital twin physical model is significantly improved. In addition, the real-time material degradation parameters output by the digital twin dynamically adjust the attention mechanism weights of the multi-scale decoupling network, enhance the ability to extract fault-sensitive features, and ultimately reconstruct and output an independent feature set covering time domain, frequency domain, and time-frequency joint features according to the fault type, providing the dynamic causal module with uncoupled fault input that integrates physical priors, greatly shortening the fault tracing time and significantly enhancing the fault analysis capability under complex working conditions.

[0084] Preferably, the specific steps of separating the multi-source fault features of the standardized demodulated signal set include:

[0085] See also Figure 3, perform time-frequency transformation on the standardized demodulated signal set, select algorithms such as synchronous compression wavelet transform, convert the standardized demodulated signal from time domain to time-frequency domain, generate a time-frequency matrix, and achieve coordinated optimization of time resolution and frequency resolution; effectively improve the resolution of the standardized demodulated signal in the time and frequency dimensions, and can clearly present the energy distribution of transient impact signals such as high-frequency vibrations caused by component cracks in the time-frequency plane, where the time axis records the occurrence moment and the frequency axis represents the frequency component, providing multi-dimensional visual expression for subsequent feature extraction.

[0086] In the time-frequency matrix, the frequency peak point of the energy concentration area is identified by the local peak detection algorithm. This frequency is used as the measured frequency to characterize the frequency component with the strongest energy in the standardized demodulated signal, which corresponds to potential fault characteristics such as turbofan engine bearing spalling, blade cracks, and gear wear.

[0087] Based on the physical principles of turbofan engine mechanical dynamics, thermodynamics, and real-time operating simulation data of digital twins, a fault characteristic frequency prior calculation rule system consisting of component classification rules and theoretical calculation formulas is constructed to quantify the theoretical characteristic frequencies of faults of different turbofan engine components. Through multi-source parameter fusion calculation, the component operating status parameters output by the digital twin are integrated to achieve effective matching between the measured frequency and the theoretical characteristic frequency.

[0088] Specifically, the fault characteristic frequency prior calculation rule system includes:

[0089] Classify turbofan engines into components, including bearings, blades, transmissions, and gas paths;

[0090] Define the determinants and physical mechanisms of the characteristic frequencies of turbofan engine components. The determinants are used to quantitatively characterize the degree of influence of component structural parameters such as the number of rolling elements and the number of blades, and operating parameters such as speed and pressure ratio on the characteristic frequencies, and provide parameter input for constructing the characteristic frequency calculation formula. The physical mechanism is used to establish a mapping relationship between characteristic frequencies and fault modes, such as the periodic disturbance caused by the rolling elements impacting the outer ring of the bearing and the blades passing through the sensor, to give physical meaning to the characteristic frequencies and establish a mapping relationship with the fault modes. Specifically, for bearing components, the determinants of their characteristic frequencies include structural and motion parameters such as the number of rolling elements, bearing pitch circle diameter, rolling element diameter, contact angle and speed. The physical mechanism originates from the contact between the rolling elements and the inner and outer rings of the bearings due to faults such as peeling and cracking. Periodic impact excitation, the frequency of which is directly related to bearing geometry and speed. For blade-type components, the characteristic frequency is primarily determined by the number of blades, speed, and blade profile parameters. The physical mechanism manifests itself as periodic perturbations as the blades pass through the sensor, namely the blade pass frequency, or the blade's own aeroelastic vibrations, such as flutter frequency. Factors determining the characteristic frequency of transmission components include the number of teeth in the gear mesh, the transmission ratio, and the input speed. The physical mechanism is closely related to rotational excitation caused by periodic loads in the gear mesh or coupling imbalance. The characteristic frequency of gas path components is related to aerodynamic thermodynamic parameters such as speed, pressure ratio, and flow path geometry, such as blade pitch and flow path diameter. The physical mechanism involves the frequency characteristics of processes such as compressor surge, turbine blade stall, or unstable combustion oscillations within the combustion chamber. The digital twin dynamically corrects the theoretically calculated parameters of the characteristic frequency based on real-time simulation of the component dynamics model, such as the current speed and temperature field distribution, significantly improving the matching accuracy between the theoretical characteristic frequency and the measured frequency.

[0091] According to the determinants and physical mechanisms of the characteristic frequencies of turbofan engine components, characteristic frequency calculation formulas based on mechanical dynamics and thermodynamics principles are constructed for different component categories to perform quantitative deduction and calculation of theoretical characteristic frequencies. These quantitative calculation formulas express the determinants of the characteristic frequencies of each component through mathematical relationships, providing a calculable theoretical basis for matching the measured frequencies with the theoretical characteristic frequencies, ensuring the physical interpretability of the fault characteristic frequencies and the accuracy of the quantitative analysis. For example,

[0092] For bearing components, the characteristic frequency of outer ring fault The calculation formula is:

[0093]

[0094] in, is the number of rolling elements, is the rotation speed (unit: revolutions per minute), is the rolling element diameter, is the bearing pitch diameter, is the contact angle, and the periodic frequency of the rolling elements impacting the outer ring when the outer ring fails is quantitatively described through the coupling relationship between the number of rolling elements, the rotational speed and the bearing geometric parameters; the blade characteristic frequency calculation formula of blade components is the product of the number of blades and the rotational speed divided by sixty, reflecting the direct influence of the number of blades and the rotational speed on the frequency; the gear meshing characteristic frequency of transmission components is calculated by the number of gear meshing teeth, the input speed and the transmission ratio, and the formula is the product of the number of gear meshing teeth and the input speed divided by sixty and then multiplied by the transmission ratio, reflecting the joint effect of the number of gear meshing teeth, the rotational speed and the transmission ratio; the compressor surge characteristic frequency of gas path components is related to the speed of sound, the compressor impeller diameter and the pressure ratio. It is calculated through the ratio of the speed of sound to the impeller diameter, combined with the difference between the pressure ratio and one and the square root of the pressure ratio, and the influence of aerodynamic thermodynamic parameters on the surge frequency is quantified. These quantitative calculation formulas express the determining factors of the characteristic frequencies of each component through mathematical relationships, providing a calculable theoretical basis for matching the measured frequencies with the theoretical characteristic frequencies, ensuring the physical interpretability of the fault characteristic frequencies and the accuracy of the quantitative analysis.

[0095] The measured frequency extracted from the time-frequency matrix is ​​matched with the theoretical characteristic frequency obtained from the characteristic frequency calculation formula, the relative error is calculated, and differentiated component error tolerance thresholds are set according to the sensitivity of component failure. If the relative error of the component is less than the corresponding component error tolerance threshold, the frequency peak point corresponding to the measured frequency is obtained to construct the initial candidate feature set; this matching process filters out false frequencies through the physical prior constraints provided by the digital twin, such as excluding irregular frequency fluctuations caused by sensor noise, to ensure that each frequency peak in the initial candidate feature set corresponds to a potential failure mode of a specific component.

[0096] An effective threshold is set to represent the minimum feature set integrity threshold corresponding to the number of component categories. For example, it is configured to cover the feature quantity baseline value of four major component categories: bearings, blades, transmissions, and air paths. The number of frequency peak points in the initial candidate feature set is counted. If it is less than the effective threshold, it indicates that the key component features are missing, triggering the missing warning mechanism, dynamically adjusting the component error tolerance threshold, and relaxing the component error tolerance threshold through a dynamic threshold adjustment strategy based on operating condition parameters. The initial candidate feature set is updated to ensure the integrity and reliability of the initial candidate feature set. The dynamic threshold adjustment strategy dynamically corrects the error tolerance threshold of each component based on predefined parameter mapping rules by correlating operating parameter data such as engine speed fluctuation rate, load status, and environmental parameters in real time. At the same time, it queries the historical database for failure cases with the same operating parameters, and uses the statistically obtained average error value of feature matching as the threshold relaxation benchmark, and ensures reliability through abnormal data filtering. In addition, by defining the physical correlation rules of component characteristic frequencies, the frequency peak points newly added after the threshold is relaxed are cross-component collaboratively verified, and their coupling relationship with other component features is verified, and false features with no reasonable physical correlation are eliminated, so that the threshold can be adaptively adjusted when features are missing, balancing the integrity of the initial candidate feature set and diagnostic reliability.

[0097] See also Figure 4 , the initial candidate feature set of turbofan engine components and the standardized demodulated signal are processed by multi-scale decoupling network respectively. The multi-scale decoupling network realizes feature separation in time domain, frequency domain and time-frequency joint dimensions through a hierarchical architecture, namely:

[0098] Through timing analysis, we focus on the fault-sensitive transient impact patterns and provide timing rationality verification for the frequency peaks of the initial candidate feature set. We use a bidirectional long short-term memory network architecture to construct a time domain decoupling layer, and configure the attention mechanism to analyze the time series characteristics of the standardized demodulated signal. We use bidirectional neurons to capture the forward and backward timing dependencies of the signal, focusing on fault-sensitive features such as the impact interval period and energy mutation points. We output a time domain feature vector, which includes timing statistics such as the impact interval standard deviation, energy rising edge slope, and pulse repetition frequency, to verify the timing rationality of the initial candidate features.

[0099] A convolutional neural network with a bandpass convolution kernel is configured as the frequency domain decoupling layer. The convolution kernel parameters are preset according to the characteristic frequency range of the component simulated by the digital twin. The spectrum of the standardized demodulated signal is subjected to layered filtering. Features such as the amplitude spectrum, harmonic amplitude ratio, and sideband energy distribution within each frequency band are extracted. A frequency domain feature vector is constructed, which includes the energy proportion of the target frequency band, harmonic distortion, and octave energy distribution. This is used to verify the harmonic rationality of the initial candidate frequency.

[0100] A Transformer architecture is used to construct a joint time-frequency decoupling layer, equipped with a multi-head self-attention mechanism. By setting a mutual information minimization loss function, the mutual information value of strongly coupled features in the time-frequency domain is forced to be reduced. A mutual information threshold is set to ensure independent feature separation. A joint time-frequency feature vector is constructed, including time-frequency energy concentration and cross-time frequency migration entropy, to decouple aliasing features in the time and frequency dimensions.

[0101] Physical constraint loss terms are set for the multi-scale decoupling network based on the digital twin physical model, including frequency and amplitude constraints. The frequency constraint measures the deviation between the decoupled characteristic frequency and the theoretical characteristic frequency simulated by the digital twin. It is constrained by the component's natural frequency fluctuation threshold calibrated by the digital twin. The decoupled characteristic frequency refers to the single frequency component directly corresponding to the specific component fault, separated from the standardized demodulated signal through layered processing of the multi-scale decoupling network. Mutual information minimization eliminates coupling interference from other components, and dynamic threshold adjustment is used to adapt to the current operating conditions, ultimately converging to the standard range of the theoretical characteristic frequency.

[0102] The amplitude constraint measures the fitting error between the decoupled characteristic amplitude and the nonlinear mapping relationship between fault severity and amplitude established by the digital twin, quantified by the root mean square error. The decoupled characteristic amplitude refers to the signal amplitude corresponding to the decoupled characteristic frequency. The nonlinear mapping between fault severity and amplitude refers to the proportional relationship between the fault severity of a turbofan engine component and the vibration signal amplitude, obtained by fitting test data and historical failure cases.

[0103] Weight coefficients are set for frequency and amplitude constraints to balance their importance in the constraint process. The physical constraint loss term is integrated into the multiscale decoupling network using the Lagrange multiplier method. The decoupled features must satisfy the physical laws related to engine dynamics, such as the correlation between blade crack depth and impact amplitude. The degree of fit between the features and the physical laws must meet a set standard.

[0104] The multiscale feature vectors output by the multiscale decoupling network are compared to a pre-built library of standard turbofan engine fault sample feature vectors to quantify their similarity in feature space. The fault type is then matched by configuring a similarity threshold. The multiscale feature vectors contain fault-sensitive features derived from time-domain feature vectors, frequency-domain feature vectors, and a combined time-frequency feature vector. The standard fault sample feature vector library is constructed from typical fault modes calibrated with historical fault data and integrated with virtual fault sample features generated by the digital twin. When the similarity exceeds the similarity threshold, the multiscale feature vector output by the multiscale decoupling network and the corresponding standard fault sample feature vector are determined to belong to the same fault type, providing a reliable basis for type matching for subsequent diagnosis.

[0105] The decoupled multi-scale feature vectors are classified and reconstructed according to the fault type of turbofan engine components to construct multi-dimensional fault signatures, including time domain features, frequency domain features, and time-frequency domain features. These multi-dimensional fault signatures are then standardized to generate independent feature sets for each turbofan engine component fault type.

[0106] The dynamic causal module uses the independent feature set for each turbofan engine component fault type output by the feature decoupling module as input to construct a fault tracing network that integrates temporal evolution and dynamic propagation mechanisms. It calculates and predicts fault propagation paths and turbofan engine component failure probabilities, providing early warning of turbofan engine component failures. Granger causality analysis is used to mine time-varying causal relationships between features. Combined with the physical coupling relationships between components simulated by the digital twin, a pseudo-causal relationship library is constructed to eliminate false causal pairs that violate physical mechanisms. Real-time material degradation parameters output by the digital twin, such as bearing raceway wear and blade crack growth rate, are used to dynamically adjust the time lag order of the vector autoregressive model, significantly improving the computational accuracy of the causal test statistics. Algorithmic learning is used to construct a dynamic Bayesian network with multiple component nodes. Network parameters are initialized based on multi-physics simulation data from the digital twin. Edge weights for each component node are determined jointly by standardized Granger causality statistics and the causal influence strength from the digital twin simulation. Conditional probability tables are trained by fusing virtual fault propagation samples generated by the digital twin with historical data, enabling adaptive updating of network parameters to complex operating conditions. By integrating Monte Carlo simulation with digital twin virtual verification to predict fault propagation paths, the digital twin performs multi-physics simulation verification on high-probability propagation paths predicted by a dynamic Bayesian network, such as those from bearing failure to rotor eccentricity to blade rubbing, quantifying the physical thresholds and timescales of fault propagation. For each component node, the remaining life loss rate calculated by the digital twin is combined with the Monte Carlo simulation results to generate a posterior probability of fault that includes physical explainability. Causal intervention analysis also provides a quantitative basis for maintenance decisions. By using the digital twin to simulate the implementation effects of different maintenance strategies in the causal intervention solution library, a comprehensive cost function that integrates the digital twin's virtual maintenance costs with actual operation and maintenance data is introduced to calculate the failure probability reduction rate and input-output ratio of the intervention solution. Maintenance decision recommendations are generated that combine virtual verification with physical constraints to support proactive prevention and control of engine health management.

[0107] Preferably, a fault tracing network integrating time series evolution and dynamic propagation mechanism is constructed to calculate and predict the fault propagation path and the failure probability of turbofan engine components. The specific steps of early warning of turbofan engine component failure include:

[0108] For the independent feature set of each fault type of turbofan engine components, Granger causality analysis is used to construct a vector autoregressive model, and the time lag order is configured to capture the temporal dependency between the fault type features. The time lag order is dynamically set according to the physical evolution characteristics of the fault features simulated by the digital twin, such as the time delay from bearing wear to rotor imbalance, to improve the accuracy of capturing temporal dependency.

[0109] Based on the time lag order, the causal test statistics between the independent feature sets of each fault type, such as the F statistic, are calculated. By setting the significance level threshold, the significance level threshold is set according to the confidence requirement of engine fault diagnosis, and the significant causal pairs of component fault types are screened according to the significance level threshold. For example, the feature variable pairs whose F statistics exceed the significance level threshold are screened as significant causal pairs. The feature variables refer to the specific feature parameters in the independent feature set that integrates the digital twin physical prior, that is, the time domain features, frequency domain features and time-frequency joint features directly corresponding to the specific component faults generated by the multi-scale decoupling network.

[0110] Combining the mechanical dynamics principles of turbofan engines with the component coupling relationships of digital twin simulations, a pseudo-causal relationship library is constructed. The pseudo-causal relationship library contains characteristic variable combinations of pseudo-causal relationships that violate physical laws, which is used to eliminate pseudo-causal relationships that violate physical laws in significant causal pairs. The selected significant causal pairs are compared with the pseudo-causal relationship library. By traversing the characteristic variable combinations in the library and matching the characteristic variable types, significant causal pairs with pseudo-causal relationships are eliminated to obtain valid causal pairs, ensuring that the retained causal pairs conform to the actual physical coupling relationships between engine components.

[0111] Based on the effective causal pairs after eliminating spurious causal relationships, a feature correlation matrix is ​​constructed by normalizing the Granger causal statistics and mapping them to the [0, 1] interval. This matrix clearly defines the leading and lagging influence relationships between independent feature sets of effective causal pairs. Elements of the feature correlation matrix characterize the strength of the leading and lagging influence between independent feature sets, such as the Granger causal influence weight of a sudden change in bearing vibration amplitude on gear mesh frequency fluctuations. The feature correlation matrix clearly defines the direction of influence and degree of correlation of each effective causal pair in the time series, providing physical constraints for subsequent structural learning of dynamic Bayesian networks.

[0112] Based on the feature correlation matrix of effective causal pairs, the K2 algorithm is used for dynamic Bayesian network structure learning. A dynamic Bayesian network is configured, including nodes for components such as bearings, blades, transmissions, and air paths. The number of network time slices is dynamically set based on the timescale characteristics of fault propagation to capture the temporal evolution of fault characteristics. The edge weights of each component node are determined by combining the standardized Granger causal statistics with the causal strength of the digital twin simulation. The value range is normalized to the interval [0, 1] to represent the strength of the causal influence between the characteristic variables. The conditional probability table is initialized to form a dynamic Bayesian network model framework that reflects the dynamic propagation of fault characteristics. The conditional probability distribution is initialized by fusing virtual fault samples generated by the digital twin with historical data, improving the network's adaptability to complex operating conditions.

[0113] Based on a dynamic Bayesian network, Monte Carlo simulation and digital twin virtual verification are integrated to generate a fault propagation path. For each component node, the digital twin performs multi-physics field simulation of the fault propagation path, quantifies the frequency of occurrence of the fault state, calculates the posterior probability of component node failure, sets a fault probability warning threshold, and triggers a warning for component nodes whose posterior probability is greater than the fault probability warning threshold. The fault propagation path is visualized through a directed acyclic graph, such as bearing failure to rotor imbalance to blade rubbing. The dynamic rendering of component damage from the digital twin is superimposed when visualizing the fault propagation path.

[0114] A causal intervention plan library is constructed to store preventive maintenance intervention plans for different fault types. A set of intervention plans is selected from the library based on the fault propagation path. The digital twin simulates the effect of the intervention plans on component failure probability. A comprehensive cost function integrating virtual maintenance costs with digital twins is introduced, including labor hours, virtual component replacement losses, and actual operation and maintenance data. The failure probability reduction rate and input-output ratio of different intervention plans are calculated, and maintenance decision recommendations are generated, ranked by comprehensive benefits. These recommendations include priority intervention nodes, operation sequences, and quantitative indicators of expected effects. For example, bearing replacement can reduce the probability of transmission component failure, providing data-driven decision support for operation and maintenance personnel.

[0115] A closed-loop verification mechanism for early warning results is established, combining virtual verification data from the digital twin with actual monitoring data to calculate performance indicators such as prediction accuracy, missed alarm rate, and false alarm rate. If indicators fail to meet preset standards over multiple consecutive monitoring cycles, the dynamic Bayesian network's structural parameters, including the number of time slices, edge weight thresholds, or the lag order of Granger causality analysis, are automatically adjusted retroactively. Simultaneously, the pseudo-causal relationship library is updated to incorporate newly identified physical constraints, continuously improving the reliability of fault tracing and the generalization capability of the early warning model under complex operating conditions.

[0116] Example 2:

[0117] See also Figure 5 This embodiment introduces a turbofan engine operation monitoring method based on digital twin, which includes the following steps:

[0118] Step S1: Build a digital twin of the turbofan engine to map the physical operating status of the turbofan engine in real time;

[0119] Step S2: The sensor array configured on the turbofan engine collects the original operating signal including mechanical vibration, acoustic characteristics, thermal state and other multi-dimensional signals, and uses the adaptive resonance demodulation chain to reduce noise and enhance features of the original operating signal:

[0120] The main energy frequency band of the signal is determined based on short-time Fourier transform, and the signal is structurally decomposed in the frequency domain through variational mode decomposition to screen out effective intrinsic mode function components with kurtosis values ​​greater than a preset kurtosis threshold;

[0121] The discrete Teager energy operator is used to perform nonlinear envelope demodulation on the effective intrinsic mode function components. After bandpass filtering and normalization, a standardized demodulated signal set containing clear fault characteristics is generated.

[0122] Step S3: Perform time-frequency transformation on the standardized demodulated signal to generate a time-frequency matrix, and extract the measured frequency through local peak detection; construct a fault characteristic frequency prior calculation rule system based on the engine mechanical dynamics principle, and calculate the theoretical characteristic frequency of each component; match the measured frequency with the theoretical characteristic frequency through physical prior constraints to construct an initial candidate feature set; use a multi-scale decoupling network to separate and obtain multi-scale feature vectors, and after similarity matching and normalization processing, generate an independent feature set for each fault type.

[0123] Step S4: Granger causality analysis is used to identify effective causal pairs for the independent feature set, integrating the physical coupling relationships of the digital twin simulation. A feature correlation matrix is ​​constructed to characterize the lead-lag relationship between the feature variables. A dynamic Bayesian network is constructed based on the feature correlation matrix, and a network architecture is configured that includes key component nodes such as bearings and blades. Fault propagation paths are generated through Monte Carlo simulation and virtual verification of the digital twin. The posterior probability of failure of the component nodes is calculated and high-risk fault warnings are triggered. Dynamic rendering of component damage in the digital twin is superimposed when visualizing the fault propagation paths. The dynamic Bayesian network is used to simulate the effect of measures in the causal intervention solution library on the suppression of failure probability. A comprehensive cost function is introduced that integrates the virtual maintenance cost of the digital twin and the actual downtime loss. The fault suppression effect and input-output ratio of the intervention solution are quantified, and maintenance decision recommendations are generated based on the overall benefits. The quantitative indicators of the expected effect of the priority intervention nodes, operation sequence, and digital twin verification are clearly defined.

[0124] Preferably, the specific steps of constructing the initial candidate feature set include:

[0125] Performing time-frequency transformation on the standardized demodulated signal set to generate a time-frequency matrix. In the time-frequency matrix, the frequency of the frequency peak point in the energy concentration area is identified as the measured frequency;

[0126] Construct a priori calculation rule system for fault characteristic frequency, which consists of component classification rules and theoretical calculation formulas, to quantify the theoretical characteristic frequency of turbofan engine component failures;

[0127] Match the measured frequency extracted from the time-frequency matrix with the theoretical characteristic frequency obtained by the fault characteristic frequency prior calculation rule system, and calculate the relative error;

[0128] Set a component error tolerance threshold. If the relative error between the measured frequency and the theoretical characteristic frequency of the component is less than the corresponding component error tolerance threshold, obtain the frequency peak point corresponding to the measured frequency and construct the initial candidate feature set.

[0129] An effective threshold is set and the number of frequency peak points in the initial candidate feature set is counted. If the number is less than the effective threshold, the missing warning mechanism is triggered and the component error tolerance threshold is dynamically adjusted. The component error tolerance threshold is relaxed through a dynamic threshold adjustment strategy based on operating condition parameters, and the initial candidate feature set is updated.

[0130] Preferably, the fault characteristic frequency prior calculation rule system includes:

[0131] The turbofan engine is divided into components, including bearings, blades, transmissions, and air paths;

[0132] Define the determinants and physical mechanisms of the characteristic frequencies of turbofan engine components; the determinants are used to quantify the degree to which component structural and operating parameters affect the characteristic frequencies, and the physical mechanisms are used to establish a mapping relationship between characteristic frequencies and failure modes;

[0133] According to the determining factors and physical mechanisms of the characteristic frequency of turbofan engine components, a characteristic frequency calculation formula is constructed for different component categories, and the theoretical characteristic frequency is quantitatively deduced and calculated to obtain the theoretical characteristic frequency of turbofan engine components.

[0134] Preferably, the multi-scale decoupling network includes:

[0135] A bidirectional long short-term memory network architecture is used to construct a time-domain decoupling layer. By configuring an attention mechanism, the time series characteristics of the standardized demodulated signal are analyzed to construct a time-domain feature vector.

[0136] A convolutional neural network with a bandpass convolution kernel is configured as the frequency domain decoupling layer to perform hierarchical filtering on the spectrum of the standardized demodulated signal, extract the amplitude spectrum and harmonic structure, and construct the frequency domain feature vector.

[0137] The Transformer architecture is used to construct a joint time-frequency decoupling layer, configure a multi-head self-attention mechanism, set a mutual information minimization loss function, and construct a joint time-frequency feature vector.

[0138] The physical constraint loss terms of the multi-scale decoupling network based on the digital twin physical model are set, including frequency constraints and amplitude constraints. The frequency constraint is used to measure the deviation between the characteristic frequency after decoupling and the theoretical characteristic frequency of the digital twin simulation. The amplitude constraint is used to measure the fitting error between the characteristic amplitude after decoupling and the nonlinear mapping relationship between fault degree and amplitude established by the digital twin.

[0139] Preferably, the specific steps of generating an independent feature set for each fault type of a turbofan engine component include:

[0140] The initial candidate feature set of turbofan engine components and the standardized demodulated signal are processed by a multi-scale decoupling network. The multi-scale decoupling network separates the time domain features, frequency domain features and time-frequency joint features of the initial candidate feature set through a hierarchical architecture to output multi-scale feature vectors, including time domain feature vectors, frequency domain feature vectors and time-frequency joint feature vectors.

[0141] The multi-scale feature vector output by the multi-scale decoupling network is similar to the pre-built standard fault sample feature vector library of turbofan engine components. The fault type is matched by configuring the similarity threshold to determine whether the multi-scale feature vector output by the multi-scale decoupling network and the corresponding standard fault sample feature vector belong to the same fault type.

[0142] The multi-scale feature vectors output by the multi-scale decoupling network are classified and reconstructed according to the fault types of turbofan engine components, and multi-dimensional fault features are constructed, including time domain features, frequency domain features, and time-frequency domain features. The multi-dimensional fault features are standardized to generate an independent feature set for each fault type of turbofan engine components.

[0143] Working principle and its effect:

[0144] The present invention collects the multi-dimensional original operating signals of the turbofan engine through the sensor array, uses the adaptive resonance demodulation chain linked with the digital twin to perform variational modal decomposition based on the main energy frequency band of the signal, and structurally separates the effective intrinsic modal function components containing weak fault characteristics in the frequency domain. The nonlinear demodulation of the Teager energy operator enhances the transient impact signals of early faults such as bearing wear and blade cracks, suppresses strong background noise, and generates a high signal-to-noise ratio demodulation signal, which solves the problem of missed early fault detection in traditional methods and significantly improves the detectability of weak signals. After the demodulated signal is transformed into time and frequency, the multi-scale decoupling network is used to fuse the bidirectional long short-term memory network, the bandpass convolutional neural network and the Transformer architecture, and the physical constraint loss term of the digital twin is combined to separate the time domain, frequency domain and time-frequency joint features to generate independent features for each fault type. The proposed method breaks through the limitation of traditional methods in their ability to separate strongly coupled features. For independent features, Granger causality analysis is used and the physical coupling relationship of digital twins is integrated to screen effective causal pairs. A dynamic Bayesian network is constructed and the posterior probability of component node failure is calculated in real time through Monte Carlo simulation and digital twin virtual verification. The fault time series evolution and cross-component propagation path under complex working conditions are analyzed. At the same time, a comprehensive cost function integrating digital twin maintenance costs is introduced to generate maintenance decision recommendations, realizing an integrated process from fault feature enhancement and decoupling to propagation prediction and active intervention, significantly improving the prediction accuracy of multi-source coupled faults, combining dynamic updating of working conditions with a closed-loop verification mechanism, improving adaptability to extreme working conditions, meeting the needs of efficient monitoring of the entire working envelope, and comprehensively solving key technical problems such as the difficulty in extracting strong noise features and the low accuracy of multi-source coupling modeling.

[0145] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiment. All technical solutions based on the concept of the present invention are within the scope of protection of the present invention. It should be noted that for those skilled in the art, various improvements and modifications that do not depart from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. The turbofan engine operation monitoring system based on digital twin is characterized by: Including digital twin mapping module, anti-noise perception module, feature decoupling module and dynamic causal module: The digital twin mapping module is used to construct a digital twin of the turbofan engine and map the physical operating status of the turbofan engine in real time; The anti-noise sensing module is used to collect the original operating signal of the turbofan engine, reduce noise and enhance features of the original operating signal through an adaptive resonance demodulation chain, output a set of pre-processed standardized demodulated signals, and dynamically map and visualize them in the digital twin; The adaptive resonance demodulation chain initializes variational mode decomposition parameters based on the main energy frequency band of the original operating signal, and performs structured decomposition of the original operating signal in the frequency domain by constructing a variational constrained optimization model; nonlinear envelope demodulation is performed on the effective intrinsic mode function components obtained by decomposition using a discrete Teager energy operator, and a standardized demodulation signal set is generated through normalization processing; The feature decoupling module is used to perform time-frequency transformation on the standardized demodulated signal set to generate a time-frequency matrix, and quantify the theoretical characteristic frequency of the turbofan engine component fault by constructing a priori calculation rule system for the fault characteristic frequency, construct an initial candidate feature set, and separate the time domain features, frequency domain features and time-frequency joint features of the initial candidate feature set through the hierarchical architecture of the multi-scale decoupling network to generate an independent feature set for each fault type of the turbofan engine component; The dynamic causal module is used to identify valid causal pairs based on the independent feature set of each fault type of turbofan engine component using Granger causality analysis combined with the physical coupling relationship of digital twin simulation, and to construct a dynamic Bayesian network to predict the fault propagation path and component failure probability, trigger fault warnings, and generate maintenance decision recommendations; The specific steps of constructing the initial candidate feature set include: Performing a time-frequency transformation on the standardized demodulated signal set to generate a time-frequency matrix, wherein the frequency of the frequency peak point of the energy concentration area is identified in the time-frequency matrix as the measured frequency; Construct a priori calculation rule system for fault characteristic frequency, which consists of component classification rules and theoretical calculation formulas, to quantify the theoretical characteristic frequency of turbofan engine component failures; Match the measured frequency extracted from the time-frequency matrix with the theoretical characteristic frequency obtained by the fault characteristic frequency prior calculation rule system, and calculate the relative error; Set a component error tolerance threshold. If the relative error between the measured frequency and the theoretical characteristic frequency of the component is less than the corresponding component error tolerance threshold, obtain the frequency peak point corresponding to the measured frequency and construct the initial candidate feature set. Set an effective threshold and count the number of frequency peak points in the initial candidate feature set. If the number is less than the effective threshold, trigger the missing warning mechanism and dynamically adjust the component error tolerance threshold. The component error tolerance threshold is relaxed through a dynamic threshold adjustment strategy based on operating condition parameters, and the initial candidate feature set is updated. The specific steps of performing noise reduction and feature enhancement on the original running signal through the adaptive resonance demodulation chain and outputting a pre-processed standardized demodulated signal set include: Calculate the frequency range of the original operating signal by short-time Fourier transform, and define the main energy frequency band of the original operating signal; Based on the main energy frequency band, an initial value of the number of decomposition layers is set, and a variational constrained optimization model of variational mode decomposition is constructed. The objective function of the variational constrained optimization model is to minimize the bandwidth sum of each intrinsic mode function component. The constraint condition of the variational constrained optimization model is that the sum of all intrinsic mode function components is equal to the input original operating signal; and the initial value of the center frequency of each intrinsic mode function component is evenly distributed within the main energy frequency band; Iteratively solve the variational constrained optimization model of variational mode decomposition to obtain independent frequency subbands corresponding to different intrinsic mode function components; and calculate the kurtosis value for each intrinsic mode function component during the iterative process; Configure the kurtosis threshold, filter the intrinsic mode function components according to the kurtosis threshold, obtain the valid intrinsic mode function components, and construct the valid component set.

2. The turbofan engine operation monitoring system based on digital twin according to claim 1, characterized in that: The fault characteristic frequency prior calculation rule system includes: Classify the turbofan engine into components, including bearings, blades, transmissions, and air paths; Determinants and physical mechanisms of characteristic frequencies of turbofan engine components are defined; the determinants are used to quantify the degree to which component structural and operating parameters affect characteristic frequencies, and the physical mechanisms are used to establish a mapping relationship between characteristic frequencies and failure modes; According to the determining factors and physical mechanisms of the characteristic frequency of turbofan engine components, a characteristic frequency calculation formula is constructed for different component categories, and the theoretical characteristic frequency is quantitatively deduced and calculated to obtain the theoretical characteristic frequency of turbofan engine components.

3. The turbofan engine operation monitoring system based on digital twin according to claim 1, characterized in that: The multi-scale decoupling network includes: A bidirectional long short-term memory network architecture is used to construct a time-domain decoupling layer. By configuring an attention mechanism, the time series characteristics of the standardized demodulated signal are analyzed to construct a time-domain feature vector. A convolutional neural network with a bandpass convolution kernel is configured as the frequency domain decoupling layer to perform hierarchical filtering on the spectrum of the standardized demodulated signal, extract the amplitude spectrum and harmonic structure, and construct the frequency domain feature vector. The Transformer architecture is used to construct a joint time-frequency decoupling layer, configure a multi-head self-attention mechanism, set a mutual information minimization loss function, and construct a joint time-frequency feature vector. The physical constraint loss terms of the multi-scale decoupling network based on the digital twin physical model are set, including frequency constraints and amplitude constraints. The frequency constraint is used to measure the deviation between the characteristic frequency after decoupling and the theoretical characteristic frequency of the digital twin simulation. The amplitude constraint is used to measure the fitting error between the characteristic amplitude after decoupling and the nonlinear mapping relationship between fault degree and amplitude established by the digital twin.

4. The turbofan engine operation monitoring system based on digital twin according to claim 1, characterized in that: The specific steps of generating an independent feature set for each fault type of a turbofan engine component include: performing multi-scale decoupling network processing on the initial candidate feature set and the standardized demodulated signal of the turbofan engine component, respectively. The multi-scale decoupling network separates the time domain features, frequency domain features, and time-frequency joint features of the initial candidate feature set through a hierarchical architecture to output a multi-scale feature vector, including a time domain feature vector, a frequency domain feature vector, and a time-frequency joint feature vector; The multi-scale feature vector output by the multi-scale decoupling network is similar to the pre-built standard fault sample feature vector library of turbofan engine components. The fault type is matched by configuring the similarity threshold to determine whether the multi-scale feature vector output by the multi-scale decoupling network and the corresponding standard fault sample feature vector belong to the same fault type. The multi-scale feature vectors output by the multi-scale decoupling network are classified and reconstructed according to the fault types of turbofan engine components, and multi-dimensional fault features are constructed, including time domain features, frequency domain features, and time-frequency domain features. The multi-dimensional fault features are standardized to generate an independent feature set for each fault type of turbofan engine components.

5. The turbofan engine operation monitoring system based on digital twin according to claim 1, characterized in that: The specific steps of performing noise reduction and feature enhancement on the original running signal through the adaptive resonance demodulation chain and outputting a pre-processed standardized demodulated signal set also include: Perform nonlinear envelope demodulation on the effective intrinsic mode function components in the effective component set, calculate the instantaneous energy of the effective intrinsic mode function components point by point through the discrete Teager energy operator, and generate a demodulated signal set through envelope detection; A bandpass filter is set, and the passband range of the bandpass filter is constrained within the main energy frequency band. The demodulated signal set is filtered, and the demodulated signal is standardized by calculating the effective value, peak value and peak factor of the demodulated signal to generate a standardized demodulated signal set.

6. The turbofan engine operation monitoring system based on digital twin according to claim 1, characterized in that: The specific steps of predicting the fault propagation path and component failure probability, triggering fault warning and generating maintenance decision suggestions include: For each independent feature set of turbofan engine component fault type, a vector autoregression model is constructed using Granger causality analysis. Time lag orders are configured to capture the temporal dependencies between fault type features. Causal test statistics between independent feature sets are calculated based on the time lag orders. Significant causal pairs of component fault types are screened by setting a significance level threshold. Constructing a pseudo-causal relationship library, wherein the pseudo-causal relationship library contains pseudo-causal relationships that violate physical laws, comparing the screened significant causal pairs with the pseudo-causal relationship library, eliminating the significant causal pairs with pseudo-causal relationships, and obtaining valid causal pairs; A feature correlation matrix is ​​constructed based on the effective causal pairs, and the elements of the feature correlation matrix represent the leading and lagging influence strengths between independent feature sets through standardized Granger causal statistics.

7. The turbofan engine operation monitoring system based on digital twin according to claim 6, characterized in that: The specific steps of predicting the fault propagation path and component failure probability, triggering fault warning and generating maintenance decision suggestions also include: A dynamic Bayesian network is constructed based on the characteristic association matrix. The number of network time slices is set. The edge weights of each component node are determined by combining the standardized Granger causal statistics and the causal strength of the digital twin simulation, and the conditional probability table is initialized. Based on the dynamic Bayesian network, Monte Carlo simulation and digital twin virtual verification are integrated to generate fault propagation paths. For each component node, the digital twin performs multi-physics field simulation of the fault propagation path, quantifies the frequency of fault states, and calculates the posterior probability of component node failure. A fault probability warning threshold is set, and warnings are triggered for component nodes whose posterior probability is greater than the fault probability warning threshold. The fault propagation path is also visualized. For the fault propagation path, the dynamic Bayesian network is used to simulate the inhibitory effect of the intervention plans in the causal intervention plan library on the failure probability of each component node. The virtual maintenance cost of the fusion digital twin is introduced, including working hours, virtual loss of component replacement, and the comprehensive cost function of actual operation and maintenance data. The failure probability reduction rate and input-output ratio of different intervention plans are calculated to generate maintenance decision recommendations ranked by comprehensive benefits.

8. A turbofan engine operation monitoring method based on digital twin, which is implemented based on the turbofan engine operation monitoring system based on digital twin according to any one of claims 1 to 7, characterized in that: The following steps are involved: Step S1: Build a digital twin of the turbofan engine to map the physical operating status of the turbofan engine in real time; Step S2: collecting the original operating signal of the turbofan engine, performing noise reduction and feature enhancement on the original operating signal through an adaptive resonance demodulation chain, outputting a set of pre-processed standardized demodulated signals, and dynamically mapping and visualizing them in the digital twin; The adaptive resonance demodulation chain initializes variational mode decomposition parameters based on the main energy frequency band of the original operating signal, and performs structured decomposition of the original operating signal in the frequency domain by constructing a variational constrained optimization model; nonlinear envelope demodulation is performed on the effective intrinsic mode function components obtained by decomposition using a discrete Teager energy operator, and a standardized demodulation signal set is generated through normalization processing; Step S3: performing time-frequency transformation on the standardized demodulated signal set to generate a time-frequency matrix, and quantifying the theoretical characteristic frequencies of turbofan engine component faults by constructing a priori calculation rule system for fault characteristic frequencies, constructing an initial candidate feature set, and separating the time domain features, frequency domain features, and time-frequency joint features of the initial candidate feature set through a hierarchical architecture of a multi-scale decoupling network to generate an independent feature set for each fault type of the turbofan engine component; Step S4: For the independent feature set of each fault type of turbofan engine components, Granger causality analysis is used in combination with the physical coupling relationship of digital twin simulation to screen effective causal pairs, and a dynamic Bayesian network is constructed to predict the fault propagation path and component failure probability, trigger fault warning and generate maintenance decision recommendations.

Images