A flow-guiding sealing intelligent diagnosis method and system

Abstract

The application relates to the technical field of rotating machinery health monitoring and fault diagnosis, and discloses a flow guide sealing intelligent diagnosis method and system, wherein the flow guide sealing intelligent diagnosis method comprises the following steps: collecting an original monitoring data set; performing working condition self-adaptive pretreatment on the original monitoring data set; extracting multi-scale time-frequency-space collaborative features, and performing dimension reduction by using a feature selection method; training by using a physical consistency constraint and a small sample learning method to obtain a small sample diagnosis model; inputting an optimized feature vector into the diagnosis model to perform abnormality detection, failure mode identification and severity assessment; establishing a sealing performance degradation model, estimating degradation model parameters, and predicting a remaining service life; integrating the diagnosis model and the remaining life prediction model into an intelligent diagnosis system, adopting online real-time diagnosis, and obtaining a diagnosis report; and the application effectively solves the problems of incomplete monitoring information, early fault detection difficulty and lack of life prediction capability in the flow guide sealing diagnosis.

Description

Technical Field

[0001] This invention relates to the field of rotating machinery health monitoring and fault diagnosis technology, and more specifically, to a method and system for intelligent diagnosis of flow guide seals. Background Technology

[0002] Flow guide seals are critical sealing components in high-end rotating machinery such as aero-engines, gas turbines, and compressors. They are primarily used for interstage sealing in high-pressure compressors, sealing structures between turbine disks and stators, and clearance control between rotors and casings. Flow guide seals achieve sealing of high-pressure gases through a labyrinthine flow channel structure, reducing interstage leakage and improving thermodynamic efficiency. Simultaneously, flow guide seals have a significant impact on the rotor's dynamic characteristics; the aerodynamic forces generated by the airflow within the sealing gap provide stiffness and damping to the rotor, affecting its critical speed and stability.

[0003] During long-term operation, flow guide seals can fail in various ways due to factors such as high temperature, high pressure, high-speed rotation, and rotor vibration. These failures include uniform wear of the sealing teeth, non-uniform wear, fatigue cracks at the root of the sealing teeth, crushing of the honeycomb sealing surface, thermal deformation, and localized rubbing. These failures lead to degraded sealing performance, increased leakage, and decreased engine efficiency. In severe cases, they can even cause rotor dynamic instability, threatening the safe operation of the equipment. Traditional flow guide seal condition monitoring mainly relies on periodic shutdown inspections, which has problems such as downtime affecting equipment availability, inability to detect sudden failures, and difficulty in quantitatively assessing health status and predicting remaining life.

[0004] With the development of sensor technology, signal processing technology, and artificial intelligence technology, predictive maintenance based on condition monitoring has become an important direction for the health management of rotating machinery. However, existing fault diagnosis technologies face many challenges when applied to the diagnosis of flow guide seals: a single monitoring signal cannot fully reflect the multi-physics coupling state of the seal; early weak fault features are submerged in strong background noise and difficult to extract; the fault features of multiple failure modes are coupled and cross-linked, making accurate identification difficult; purely data-driven methods lack physical interpretability and are highly dependent on training data; changes in operating conditions cause feature drift, resulting in poor adaptability of the diagnostic model; and there is a lack of ability to predict the trend of seal performance degradation and estimate remaining life. Therefore, there is an urgent need to study an intelligent diagnostic method for flow guide seals that can fully utilize multi-source monitoring information, integrate physical mechanisms and data-driven methods, accurately identify multiple failure modes, achieve early fault warning and trend prediction, and has good engineering practicality. Summary of the Invention

[0005] This invention provides a method and system for intelligent diagnosis of flow-guiding seals, which solves the technical problems in related technologies such as downtime affecting equipment availability, inability to detect sudden faults, and difficulty in quantitatively assessing health status and predicting remaining life.

[0006] This invention provides a method for intelligent diagnosis of flow-guiding seals, comprising the following steps: Collect raw monitoring datasets, which include vibration signals, pressure signals, temperature signals, and rotational speed signals; The original monitoring dataset is subjected to working condition adaptive preprocessing to obtain a preprocessed dataset. Multi-scale time-frequency-space collaborative features are extracted from the preprocessed dataset, and dimensionality reduction is performed using feature selection methods to obtain optimized feature vectors. Based on optimized feature vectors, a physical-guided deep learning diagnostic algorithm is adopted, which is trained using physical consistency constraints and few-shot learning methods to obtain a physical-guided few-shot diagnostic model. The optimized feature vector is input into the diagnostic model for anomaly detection, failure mode identification, and severity assessment. The health score and confidence score are calculated to obtain the hierarchical diagnostic results. A sealing performance degradation model is established based on the health sequence and operating parameters in the stratified diagnostic results. The degradation model parameters are estimated using the Bayesian method, and the remaining service life is predicted using Monte Carlo simulation to obtain the remaining service life prediction results. The diagnostic model and the remaining life prediction model are integrated into an intelligent diagnostic system, which uses online real-time diagnosis to obtain a diagnostic report.

[0007] In a preferred embodiment, the process of collecting the raw monitoring dataset includes: Based on the classic leakage calculation model, the leakage mass flow rate is calculated according to the inlet and outlet pressures, temperatures, and rotational speeds. The current sealing gap is determined by inverting the measured pressure difference and temperature. The aerodynamic stiffness coefficient and aerodynamic damping coefficient of the seal are inverted using the parameter identification method; By merging the measured dataset and the parameters inferred from the virtual sensing, a complete monitoring dataset is obtained.

[0008] In a preferred embodiment, the condition-adaptive preprocessing of the original monitoring dataset includes: Calculate the instantaneous rotational speed and interpolate the discrete rotational speed signal into a continuous instantaneous rotational speed curve; Angular domain resampling is performed on the time-domain vibration signal to convert equal time interval sampling into equal angular interval sampling; Order analysis is performed to extract the amplitude of key orders as characteristic parameters, thereby achieving normalization of operating conditions.

[0009] In a preferred embodiment, the extraction of multi-scale time-frequency-space coordinated features includes: Calculate the time-domain statistical characteristics of the order-domain signals for each channel; Time-frequency analysis is performed on the vibration signal to extract time-frequency domain features; Calculate spatial correlation characteristics based on vibration signals from different measuring points; Dimensionality reduction is performed using feature selection methods to obtain optimized feature vectors.

[0010] In a preferred embodiment, the physically guided small-sample diagnostic model includes: A dual-channel feature extraction network is constructed, where the physical channel extracts features with physical meaning, and the data channel performs a nonlinear transformation on the optimized feature vector. The contrastive learning method is used to pre-train feature representations on normal operating data; Physical consistency constraints are used to ensure that network predictions are consistent with the results calculated by the physical model. A physical-guided small-sample diagnostic model was obtained by training a few-sample classifier.

[0011] In a preferred embodiment, the anomaly detection, failure mode identification, and severity assessment include: Perform anomaly detection to determine if the seal is in an abnormal state; Identify failure modes and calculate the distance between the sample to be classified and various prototypes; Health score is calculated based on the degree of feature deviation, and the severity is divided into different levels according to the health score; The overall confidence level is calculated based on the consistency of the indicators to obtain the stratified diagnostic results.

[0012] In a preferred embodiment, obtaining the remaining lifetime prediction result includes: A sealing performance degradation model considering the influence of operating conditions is established, and a proportional risk model is used to describe the degradation law of health over time. The parameters of the degradation model are estimated using the Bayesian method, the prior distribution of the parameters is established, and the posterior sampling is performed using the Markov chain Monte Carlo method. The remaining service life is predicted by Monte Carlo simulation, and a failure threshold is set. When the health status first crosses the failure threshold, it is determined to be a sealing failure.

[0013] In a preferred embodiment, the online real-time diagnostics includes: Based on the edge computing architecture, edge computing nodes perform data acquisition, preprocessing, and feature extraction in real time close to the data acquisition end; The cloud server receives feature vectors transmitted from edge nodes and performs diagnostic model inference and hierarchical diagnostic decisions. The cloud regularly collects new monitoring data and diagnostic results to incrementally update the model, and uses online learning methods to update the model parameters.

[0014] In a preferred embodiment, the anomaly detection, failure mode identification, and severity assessment further include: Health status is calculated based on the degree of deviation of features and the amount of change of key physical parameters, and the distance between real-time features and normal state prototype is calculated. The confidence assessment of the diagnostic results includes confidence in anomaly detection, confidence in failure mode identification, and confidence in severity. Based on the diagnostic results and the inversion of key physical parameters using the physical model, the gap estimate is updated using Bayesian inference.

[0015] This invention provides an intelligent diagnostic system for flow guiding seals, used to perform the aforementioned intelligent diagnostic method for flow guiding seals, comprising: The information acquisition module is used to collect raw monitoring datasets, which include vibration signals, pressure signals, temperature signals, and rotational speed signals. The preprocessing module is used to perform condition-adaptive preprocessing on the original monitoring dataset to obtain a preprocessed dataset. The feature extraction module extracts multi-scale time-frequency-space collaborative features based on the preprocessed dataset, and uses feature selection methods to reduce dimensionality to obtain optimized feature vectors. The diagnostic model building module, based on optimized feature vectors, adopts a physical-guided deep learning diagnostic algorithm, and is trained using physical consistency constraints and few-shot learning methods to obtain a physical-guided few-shot diagnostic model. The diagnostic analysis module is used to input optimized feature vectors into the diagnostic model for anomaly detection, failure mode identification, and severity assessment, calculate health status and confidence level, and obtain stratified diagnostic results. The life prediction module establishes a sealing performance degradation model based on the health sequence and operating parameters in the hierarchical diagnosis results, estimates the degradation model parameters using the Bayesian method, and uses Monte Carlo simulation to predict the remaining service life, thus obtaining the remaining service life prediction results. The results output module is used to integrate the diagnostic model and the remaining life prediction model into an intelligent diagnostic system, which uses online real-time diagnosis to obtain a diagnostic report.

[0016] The beneficial effects of this invention are as follows: By employing multi-level monitoring information acquisition and virtual sensing technology, sensors are deployed in accessible locations to acquire external measurable information. Simultaneously, a physical model of the flow-guiding seal is used to calculate key internal parameters that are difficult to measure directly, enabling multi-physics collaborative monitoring of the sealed state. Virtual sensing technology overcomes the difficulty of sensor deployment under high temperature, high pressure, and high speed environments. By fusing measured data and parameters calculated from the physical model, more complete and abundant monitoring information is obtained, providing a reliable data foundation for accurate diagnosis.

[0017] An adaptive preprocessing method based on order analysis is adopted. Through angular domain resampling and order domain analysis, speed-related features are transformed into speed-independent order features, effectively eliminating the influence of operating condition changes on diagnostic features. The order analysis method makes the features extracted under various operating conditions such as different speeds and loads comparable, improving the operating condition adaptability and robustness of the diagnostic model, and solving the problems of feature drift and model failure caused by operating condition changes.

[0018] A multi-scale time-frequency-space collaborative feature extraction method is proposed, extracting sealing state features from multiple dimensions including the time domain, order domain, time-frequency domain, and spatial domain, fully exploiting the fault information contained in the monitoring signals. In particular, the use of synchronous compressed wavelet transform to extract time-frequency ridges clearly shows the evolution trajectory of fault feature frequencies, effectively extracting features of early, weak faults. The synergistic effect of multi-domain features enhances the ability to distinguish different failure modes, providing high-quality feature input for subsequent accurate classification.

[0019] Fourth, this invention constructs a physically guided few-shot deep learning diagnostic model, innovatively integrating the fluid dynamics and rotor dynamics mechanisms of the flow-guiding seal into a deep neural network. Physical consistency constraints ensure that the features learned by the network conform to physical laws. This method effectively solves the problems of lack of interpretability and poor generalization ability in purely data-driven methods. Furthermore, through contrastive learning, transfer learning, and few-shot learning techniques, it overcomes the challenge of difficult model training under conditions of scarce fault samples, achieving accurate failure mode identification even with a very small number of fault samples.

[0020] A hierarchical, multi-granularity diagnostic architecture was designed. The first layer uses anomaly detection based on normal data to quickly determine whether the seal deviates from the normal state. The second layer uses a small-sample classifier to accurately identify specific failure mode types. The third layer assesses the severity and health of the failure. This hierarchical diagnostic strategy balances accuracy and real-time performance, while providing a confidence assessment for each layer's diagnostic results, quantifying the reliability of the diagnosis and providing maintenance personnel with a more credible basis for decision-making.

[0021] A remaining life prediction method based on a condition-dependent degradation model was established. By considering the influence of operating parameters such as speed, pressure, and temperature on the degradation rate, a stochastic degradation model consistent with the actual degradation law of seals was established. A Bayesian method was used to fuse prior knowledge and monitoring data for parameter estimation and model updating. Monte Carlo simulation was used to give the probability distribution and confidence interval of the remaining life. This method realizes the transformation from passive fault diagnosis to proactive health management, providing reliable technical support for formulating scientific preventive maintenance strategies. Attached Figure Description

[0022] Figure 1This is a flowchart of a flow-guiding seal intelligent diagnostic method according to the present invention; Figure 2 This is a block diagram of a flow-guiding and sealing intelligent diagnostic system according to the present invention. Detailed Implementation

[0023] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0024] At least one embodiment of the present invention discloses a method for intelligent diagnosis of flow-guiding seals, such as... Figure 1 As shown, it includes: Step 1: Collect the raw monitoring dataset, which includes vibration signals, pressure signals, temperature signals, and rotational speed signals; Specifically, it includes the following: Based on the structural characteristics and accessibility constraints of the high-pressure compressor's flow guide seal, a triaxial vibration accelerometer was placed at two measuring points, A and B, before and after the bearing housing, respectively. Vibration acceleration signals in three orthogonal directions were collected, resulting in a total of six vibration signals. A piezoelectric accelerometer with a sensitivity of 100 mV / g and a frequency response range of 0.5 Hz to 10 kHz was selected. The sensor was installed at a rigid support location on the bearing housing and secured with bolts to ensure the rigidity of the signal transmission path.

[0025] Based on the layout space of the compressor inlet and seal outlet pipelines, an absolute pressure sensor is installed at a distance of 2 times the pipe diameter from the front end of the seal on the inlet pipeline to measure the inlet pressure, and an absolute pressure sensor is installed at a distance of 3 times the pipe diameter from the rear end of the seal on the outlet pipeline to measure the outlet pressure. Piezoresistive pressure sensors are used, with a measurement range of 0-5MPa and an accuracy of 0.1%FS.

[0026] K-type thermocouple temperature sensors are installed on the inlet and outlet pipes respectively to measure the inlet and outlet temperatures. The response time of the temperature sensors is less than 0.5 seconds. A non-contact infrared temperature sensor is installed on the outer wall of the casing, closest to the sealed cavity, to measure the temperature of the outer wall of the casing. The temperature of the sealed cavity is then calculated using a thermal conduction model.

[0027] Based on the compressor measurement and control system, rotor speed, rotor axial displacement, and rotor radial displacement signals are acquired synchronously. The speed signal is obtained by a Hall effect speed sensor or a magnetoelectric speed sensor, outputting a pulse signal; one pulse is generated per revolution for speed calculation and phase reference. The displacement signal is acquired by an eddy current displacement sensor, with a measurement range of 0-2 mm and an accuracy of 0.01 mm.

[0028] A multi-channel data acquisition system was used to synchronously acquire signals from all the aforementioned sensors. A 24-bit high-precision A / D converter was employed, with the sampling frequency set to 20 times the rotational frequency corresponding to the rotor's maximum speed. The rotational frequency was calculated as follows: obtain the rotor's maximum speed value (in rpm), divide this speed value by 60 to obtain the revolutions per second, which is the maximum rotational frequency (in Hz). For a rotor maximum speed of 18,000 rpm, the maximum rotational frequency is 300 Hz, and the sampling frequency is set to 6,000 Hz. All channels were synchronously triggered for acquisition, with each acquisition lasting 10 seconds, resulting in a raw data matrix containing multiple channels and 60,000 sampling points per channel.

[0029] Based on the one-dimensional flow theory model of the flow guide seal, the leakage and sealing gap are calculated according to the measured inlet pressure, outlet pressure, inlet temperature and rotation speed.

[0030] Using the classic leakage calculation model for labyrinth seals, the leakage mass flow rate calculation process includes: calculating the seal circumference (seal diameter multiplied by pi), calculating the flow area (seal circumference multiplied by seal radial clearance), calculating the inlet gas dynamic head (twice the inlet pressure divided by the inlet gas density, then taking the square root), and multiplying the flow coefficient, flow area, dynamic head, and flow correction function to obtain the leakage mass flow rate. The inlet gas density is calculated using the ideal gas law based on the inlet pressure and temperature; the flow correction function considers the influence of the multi-tooth labyrinth effect and is a function of the pressure ratio and the number of seal teeth.

[0031] The flow correction function is calculated using the Martin formula. The calculation process includes: calculating the pressure ratio square, calculating the pressure difference effect term (1 minus the square of the pressure ratio), calculating the throttling effect term (the negative logarithm of the pressure ratio, reflecting the throttling effect of a single-stage seal), calculating the average flow coefficient per tooth (the pressure difference effect term divided by the product of the number of seal teeth and the throttling effect term), and taking the square root of this coefficient to obtain the correction coefficient. This coefficient takes into account the cascade throttling effect and pressure recovery effect of multi-tooth labyrinth seals.

[0032] When seal wear leads to increased clearance, leakage increases and pressure differential decreases. The calculation process for retrieving the current seal clearance using measured pressure differential and inlet / outlet temperatures includes estimating the actual leakage based on the pressure differential and temperature using the energy equation, calculating the seal circumference, calculating the inlet gas dynamic head, and dividing the actual leakage by the product of the flow coefficient, seal circumference, dynamic head, and flow correction function to obtain the current seal clearance.

[0033] Based on rotor dynamics theory, the aerodynamic force of the flow-guiding seal will produce stiffness and damping effects on the rotor. Based on the measured rotor vibration response and rotational speed, the aerodynamic stiffness coefficient and aerodynamic damping coefficient of the seal are inverted using the parameter identification method, and a simplified dynamic model of the rotor-seal system is established. This model describes the second-order differential equation of the rotor radial displacement, including mass terms, damping terms (bearing damping and seal damping), stiffness terms (bearing stiffness and seal stiffness), and unbalanced force excitation.

[0034] By performing amplitude-frequency response analysis on the measured vibration signal, the natural frequency and damping ratio of the system are identified. Combined with the known bearing parameters, the aerodynamic stiffness coefficient and aerodynamic damping coefficient of the seal are obtained by inversion and used as virtual sensing parameters.

[0035] The measured dataset and the parameters (sealing gap, aerodynamic stiffness coefficient, aerodynamic damping coefficient) calculated by virtual sensing are combined to obtain a complete monitoring dataset.

[0036] Step 2: Perform adaptive preprocessing on the original monitoring dataset to obtain the preprocessed dataset; Specifically, it includes the following: Based on the rotational speed pulse signal acquired in step 1, the instantaneous rotational speed at each sampling moment is calculated. The instantaneous rotational speed is calculated as follows: measure the time interval between two adjacent pulses, divide 60 by the time interval to obtain the revolutions per minute, which is the instantaneous rotational speed.

[0037] The discrete rotational speed pulse signal is interpolated into a continuous instantaneous rotational speed curve using cubic spline interpolation. The instantaneous angular position is then calculated using time integration. The calculation method for the angular position is as follows: the instantaneous rotational speed is converted from revolutions per minute to radians per second (i.e., the rotational speed is divided by 60 and then multiplied by 2π). This angular velocity is then integrated over time (from the initial time 0 to the current time) to obtain the cumulative rotational angle of the rotor relative to the initial position.

[0038] Based on the calculated angular positions, the time-domain vibration signal is resampled in the angular domain, converting equal-time interval sampling into equal-angle interval sampling. The angular sampling interval is set by dividing a complete circle (2π radians) by the number of sampling points per revolution, with 1024 sampling points per revolution. The angular domain signal is calculated by determining the position of each angular sampling point (sampling point number multiplied by the angular sampling interval), finding the time corresponding to that angle using the inverse interpolation angular position function, and obtaining the vibration signal value at that time point using linear interpolation as the angular domain signal value. The pressure pulsation signal and other vibration channel signals are resampled in the same way to obtain the angular domain signal set.

[0039] Based on angular domain signals, order analysis is performed using Fast Fourier Transform (FFT). The calculation process for order analysis includes: determining the analysis parameters, setting the number of sampling points per revolution to 1024, and determining the range of orders to be analyzed, including key orders such as 0th, 0.5th, 1st, 1.5th, and 2nd; executing a Discrete Fourier Transform, traversing all angular sampling points for each order n, and calculating the complex contribution value of each sampling point. Specifically, the vibration signal value of the point is multiplied by a complex exponential function, where the exponent part of the complex exponential function is the negative imaginary unit, multiplied by 2π, multiplied by the order n, multiplied by the sampling point index k, and then divided by the number of sampling points per revolution; accumulating the contributions of all sampling points, summing the complex contribution values ​​of all 1024 sampling points to obtain the total complex amplitude of that order; performing normalization processing, dividing the accumulated result by the number of sampling points per revolution to obtain the final complex amplitude of that order. The physical meaning of the order is the ratio of the vibration frequency to the rotational frequency. Therefore, the order spectrum is independent of the rotational speed, thus achieving normalization of the operating conditions.

[0040] The amplitudes of key orders are extracted as characteristic parameters. For flow-guiding seal diagnosis, the key orders of interest include: 0.5 order (half an order, reflecting rotor eccentricity and asymmetry), 1st order (first order, reflecting rotor imbalance), 2nd order (second order, reflecting rotor misalignment or circumferential two-point rubbing of the seal), and higher orders (3rd order and above, reflecting seal tooth damage or localized rubbing). The method for calculating the amplitude of each order is as follows: for the complex amplitude of each key order, calculate its modulus, which is the square root of the sum of the squares of the real and imaginary parts of the complex number, to obtain the magnitude of the amplitude of that order.

[0041] The method for calculating the energy percentage of the nth order is as follows: calculate the energy percentage of each order (by squaring the amplitude of each order to obtain the energy value of that order, summing the energies of all orders from order 0 to the highest analysis order to obtain the total energy, and finally dividing the energy of the nth order by the total energy to obtain the energy percentage of the nth order).

[0042] The dimensionless differential pressure coefficient is calculated as follows: Obtain the measured values ​​of the inlet and outlet pressures (both in Pa), subtract the outlet pressure from the inlet pressure to obtain the pressure difference, and divide the pressure difference by the inlet pressure to obtain the dimensionless differential pressure coefficient. This coefficient reflects the sealing effect of the seal. When the seal wears, the pressure difference decreases, and the dimensionless differential pressure coefficient decreases.

[0043] The dimensionless temperature rise coefficient is calculated as follows: Obtain the measured values ​​of the outlet and inlet temperatures (both in K), subtract the inlet temperature from the outlet temperature to obtain the temperature rise value, and divide the temperature rise value by the inlet temperature to obtain the dimensionless temperature rise coefficient. This coefficient reflects the temperature rise caused by sealing friction and flow loss; the temperature rise increases when the seal rubs against the surface.

[0044] The method for calculating the corrected Reynolds number is as follows: Obtain the necessary physical parameters (inlet gas density in kg / m³). 3 The rotor speed is in rad / s, the seal diameter is in m, the seal radial clearance is in m, and the gas dynamic viscosity is in Pa·s. The characteristic velocity is obtained by multiplying the rotor speed by the seal diameter, and the characteristic length is obtained by using the seal radial clearance as the characteristic length. The corrected Reynolds number is obtained by dividing the product of the inlet gas density, characteristic velocity, and characteristic length by the gas dynamic viscosity. This parameter varies with operating conditions and requires normalization.

[0045] Based on the angular domain signal and the normalized physical parameter signal, wavelet packet decomposition is used for adaptive denoising. The Daubechies 4 (db4) wavelet is selected as the mother wavelet, and a 3-level wavelet packet decomposition is performed to obtain the decomposition coefficients of 8 frequency bands.

[0046] The adaptive threshold is calculated as follows: Noise standard deviation estimation: For the wavelet coefficients of the i-th frequency band, a robust median estimation method is used to calculate the noise standard deviation. Specifically, the median of the absolute values ​​of the coefficients in that frequency band is calculated and divided by 0.6745. Threshold calculation: The noise standard deviation is multiplied by the square root of the product of twice the natural logarithm and the number of coefficients in that frequency band to obtain the adaptive threshold for that frequency band. Coefficients smaller than the threshold are soft-thresholded, while coefficients larger than the threshold are retained. Then, wavelet packet reconstruction is performed to obtain the denoised signal.

[0047] Through the above preprocessing, a normalized and denoised dataset is obtained, including order domain features, dimensionless physical parameters, and denoised signals, providing high-quality input for subsequent feature extraction.

[0048] Step 3: Extract multi-scale time-frequency-space collaborative features based on the preprocessed dataset, and use feature selection methods to reduce dimensionality to obtain optimized feature vectors; Specifically, it includes the following: Based on the preprocessed signal obtained in step 2, time-domain statistical characteristics are calculated for the order domain signals of each channel. For vibration signals, the following characteristics are calculated: the mean and standard deviation of the signal are calculated as basic descriptions of the amplitude level; the peak factor is further calculated, defined as the ratio of the signal peak value to the standard deviation, reflecting the impact of the signal; skewness and kurtosis are calculated, where skewness reflects the symmetry of the signal amplitude distribution, positive skewness indicates the presence of a positive amplitude pulse, and negative skewness indicates the presence of a negative amplitude pulse; kurtosis reflects the sharpness of the signal amplitude distribution, and increased kurtosis usually corresponds to the occurrence of impact faults; the above statistical characteristics are calculated for all vibration and pressure channel signals, including mean, standard deviation, peak value, peak-to-peak value, peak factor, skewness, kurtosis, waveform factor, and impulse factor, forming a time-domain feature vector, where the time-domain feature dimension is 72 (for 6 vibration channels and 2 pressure channels, 9 statistical features are extracted for each channel).

[0049] The calculation method for the energy proportions of low-order energies (0-2), medium-order energies (2-10), and high-order energies (above 10) is as follows: for low-order energies, the energy proportions of all orders from 0 to 2 are summed; for medium-order energies, the energy proportions of all orders from 2 to 10 are summed; and for high-order energies, the energy proportions of all orders from 10 to the highest analysis order are summed.

[0050] The method for calculating the order energy entropy is as follows: calculate the natural logarithm of the energy proportion of each order, multiply the energy proportion of each order by its corresponding natural logarithm, sum the above products of all orders, and add a negative sign before the summation result to obtain the order energy entropy. The larger the energy entropy, the more dispersed the energy distribution, indicating the existence of multiple frequency components, which may correspond to the coexistence of multiple fault modes.

[0051] The amplitude and phase of the characteristic orders (0.5, 1, and 2) are extracted. In particular, the appearance and enhancement of the half-order components often correspond to rotor eccentricity or asymmetric seal wear. The order energy features are combined into a vector.

[0052] High-resolution time-frequency analysis of vibration signals is performed based on Synchrosqueezed Wavelet Transform (SSWT). SSWT improves time-frequency resolution and displays the trajectory of instantaneous frequency changes more clearly by synchronously compressing the results of Continuous Wavelet Transform (CWT).

[0053] The steps for calculating wavelet coefficients by performing CWT on the vibration signal are as follows: Calculate the wavelet coefficients, selecting the Morlet wavelet as the mother wavelet, which has good time-frequency localization characteristics and can simultaneously provide time and frequency information; set a series of scale parameters according to the frequency range to be analyzed, with the scale parameters inversely proportional to the frequency; for each scale parameter, convolve the vibration signal with the translated and scaled mother wavelet function, and implement the integration operation through discretization, converting continuous integration into finite summation, with the integration interval determined according to the effective support length of the wavelet; the result of the convolution operation is the wavelet coefficient corresponding to that scale and position.

[0054] The specific implementation of the mother wavelet function is as follows: Morlet wavelet is used as the mother wavelet, and the center frequency parameter is usually taken as 5; the integration interval is usually the sampling points within 4 times the scale parameter range.

[0055] The steps for calculating the instantaneous frequency are as follows: Calculate the instantaneous frequency by taking the complex argument of each wavelet coefficient to obtain phase information; perform numerical differentiation of the phase information in the angular direction to obtain the phase gradient; divide the phase gradient by -2π to obtain the instantaneous frequency corresponding to that position and scale; compress and rearrange the wavelet coefficients according to the instantaneous frequency to obtain the SSWT coefficients, which have higher energy concentration in the time-frequency plane; extract the time-frequency ridge, which is the trajectory with the maximum energy in the time-frequency plane, and this trajectory corresponds to the evolution of the fault characteristic frequency; use the ridge extraction algorithm to find the frequency corresponding to the maximum energy at each time moment to form the time-frequency ridge; calculate the time-frequency domain features: time-frequency energy entropy, time-frequency energy concentration, average frequency of the ridge, frequency variance of the ridge, etc., to form the time-frequency feature vector.

[0056] Furthermore, Variational Mode Decomposition (VMD) can be used to replace synchronous compressed wavelet transform for time-frequency feature extraction, thereby further improving the decomposition capability and noise resistance of multiple oscillating modes in non-stationary signals. The VMD method decomposes the signal into several intrinsic mode functions, each with a specific center frequency and finite bandwidth. Based on the vibration signal, the objective optimization problem of VMD is to minimize the sum of the bandwidths of each mode, while constraining the sum of the modes to equal the original signal.

[0057] The solution method for the VMD optimization problem is as follows: Set an optimization objective of minimizing the sum of the bandwidths of all modal functions, where the bandwidth of each mode is measured using the L2 norm of the squared gradient of its analytic signal; set a constraint that the sum of all modal functions must equal the original input signal; construct a Lagrange function and introduce the constraint into the objective function using the Lagrange multiplier method; use the alternating direction multiplier method to iteratively update each modal function and its corresponding center frequency until convergence. The optimization problem is solved using the alternating direction multiplier method, iteratively updating each modal function and its center frequency until convergence. The signal is decomposed into K modal components, each corresponding to a center frequency. The instantaneous amplitude and instantaneous frequency of each mode are calculated, and modal energy features, center frequency features, and frequency variance features are extracted.

[0058] Compared to wavelet transform, the VMD method is more robust to noise and can achieve clearer separation of multi-frequency components generated by sealing faults, making it particularly suitable for complex situations where multiple failure modes coexist in the seal. The modal features extracted by VMD are fused with or replaced with the original time-frequency features to form a new time-frequency feature vector, while subsequent steps remain unchanged. The purpose of this alternative embodiment is to enhance the feature extraction capability of complex multimodal signals and improve the accuracy of fault identification.

[0059] Spatial correlation characteristics are calculated based on vibration signals from multiple measurement points (measurement point A and measurement point B).

[0060] The method for calculating the cross-correlation function is as follows: acquire vibration signal data of measuring point A and measuring point B, calculate the correlation between the two signals under each lag angle, multiply the signal of measuring point A with the signal of measuring point B lagging by a certain angle point by point and then calculate the average value; traverse all possible lag angles to obtain the complete cross-correlation function; extract the peak value of the cross-correlation function and the corresponding lag angle to reflect the phase relationship between the vibration signals of the two measuring points; when the seal is locally damaged, the phase difference between the signals of the two measuring points will change.

[0061] The coherence function is calculated as follows: calculate the self-power spectral density of measuring point A and measuring point B respectively, and calculate the cross-power spectral density between measuring point A and measuring point B; calculate the coherence function value for each frequency point by dividing the square of the modulus of the cross-power spectral density by the product of the two self-power spectral densities; the value of the coherence function at key frequencies (such as the rotational frequency and its harmonics) reflects the degree of correlation between the vibrations of the two measuring points, and the coherence will decrease when a local fault occurs in the seal.

[0062] Spatial correlation features are combined into a vector, which contains correlation features between multiple measurement points.

[0063] Based on the extracted multi-domain features, time-domain features, order-domain features, time-frequency features, and spatial correlation features are concatenated into a high-dimensional comprehensive feature vector.

[0064] Before concatenating features from different domains, standardization preprocessing is required to eliminate differences in dimensions. Specifically, this involves performing Z-score standardization on each feature domain, calculating the mean and standard deviation of the features for that domain, and then subtracting the mean from the feature value and dividing by the standard deviation. Standardized features have zero mean and unit variance, eliminating differences in dimensions and numerical ranges between features from different domains. This ensures that each feature in the concatenated feature vector has equal importance, preventing feature domains with larger numerical values ​​from dominating the entire feature vector.

[0065] High-dimensional feature vectors can reach hundreds of dimensions, exhibiting redundancy and correlation. A feature selection method based on Max-Relevance and Min-Redundancy (mRMR) is used for dimensionality reduction. The goal of the mRMR method is to select a subset of features that are most correlated with the target variable and have the least redundancy among features. The specific steps are as follows: The mRMR method calculates the mutual information between each feature and the target variable (sealed state category) to measure correlation; calculates the mutual information among features to measure redundancy; constructs an objective function that maximizes the average mutual information between the selected features and the target variable while minimizing the average mutual information among the selected features; uses a forward search algorithm to progressively select features that maximize the objective function; and finally selects M optimal features (M is typically 20-50) to form an optimized feature vector.

[0066] Step 4: Based on the optimized feature vector, a physical-guided deep learning diagnostic algorithm is used, which is trained using physical consistency constraints and few-shot learning methods to obtain a physical-guided few-shot diagnostic model. Specifically, it includes the following: Based on the physical mechanism of the flow-guiding seal and the optimized features obtained in step 3, a dual-channel feature extraction network is constructed, which includes two parallel branches: a physical channel and a data channel.

[0067] Based on the flow theory of the guide seal and the rotor dynamics theory, the physical channel extracts features with clear physical meaning. The input consists of physical parameters (sealing gap, aerodynamic stiffness coefficient, aerodynamic damping coefficient) and dimensionless parameters (dimensional pressure difference coefficient, dimensionless temperature rise coefficient, corrected Reynolds number) obtained by virtual sensing in step 1. These parameters are mapped to the physical feature space using a two-layer fully connected neural network, with a feature dimension of 32.

[0068] The data channel uses a multilayer perceptron (MLP) structure to perform a nonlinear transformation on the optimized feature vector obtained in step 3. The network contains three hidden layers with 128, 64, and 32 neurons respectively, and uses ReLU as the activation function to output data-driven features.

[0069] Physical features and data features are fused, and an attention mechanism is used to adaptively assign weights to the two types of features. The attention weight calculation steps are as follows: the attention fusion mechanism concatenates the physical features and data features, calculates the attention weight vector through a fully connected layer and a sigmoid activation function, and uses the attention weights to perform weighted fusion of the two features. The fused feature is equal to the element-wise product of the attention weights and the physical features, plus 1 and minus the element-wise product of the attention weights and the data features.

[0070] Due to the scarcity of fault samples, a contrastive learning approach is used to pre-train feature representations on a large amount of unlabeled normal operation data. The goal of contrastive learning is to ensure that similar samples are close in the feature space, while dissimilar samples are far apart. The implementation steps of contrastive learning are as follows: Contrastive learning uses data augmentation on normal operation data samples to generate two views. Data augmentation methods include adding Gaussian noise, time shifting, amplitude scaling, and frequency band filtering. The two views are treated as positive sample pairs, and the views of different samples are treated as negative sample pairs. The two views are mapped to the feature space through a feature extraction network to obtain the corresponding feature representations. The contrastive learning loss is calculated. For each sample, the cosine similarity between the two views is calculated, and the similarity between the sample and all other samples in the batch is also calculated. The normalized temperature cross-entropy loss is used as the loss function for contrastive learning. The numerator is the exponent of the similarity between positive sample pairs, and the denominator is the sum of the similarity exponents between the sample and all samples in the batch. The temperature parameter is set to 0.1. By minimizing the contrastive learning loss, the network learns feature representations that are sensitive to changes in the sealing state.

[0071] To ensure that the features learned by the network conform to physical laws, a physical consistency constraint is introduced during training. Based on the relationship between leakage and gap in the flow-guiding seal, the constraint ensures that the sealing state predicted by the network is consistent with the calculation results of the physical model. The implementation method of the physical consistency constraint is as follows: a regression branch is added using fused features to predict the sealing gap; the virtual sensing gap is calculated using the physical model in step one based on the measured pressure and temperature; a physical consistency loss is defined, and the square of the difference between the gap predicted by the network and the gap calculated by the physical model is calculated; this loss term constrains the network's prediction to be consistent with the physical model, preventing the network from learning mappings that violate physical laws.

[0072] A classifier is trained on a small number of labeled fault samples based on a pre-trained feature extraction network. Assume there are K failure mode categories (normal, uniform wear, non-uniform wear, seal tooth breakage, impact wear, etc.), with only a small number of samples (1 to 5) for each category.

[0073] The implementation method of the prototype network is as follows: For each category, calculate the class prototype, and average the fused feature vectors of all samples in that class to obtain the representative feature vector of that class; For the sample to be classified, calculate the Euclidean distance between its fused features and each class prototype, and calculate the squared Euclidean distance between the feature vector of the sample to be classified and the vector of each class prototype; Use the softmax function to calculate the probability of belonging to each class, and use the negative distance value as input to calculate the probability distribution through the exponential function and normalization; The classification loss adopts the cross-entropy loss, and the one-hot encoding of the true label and the predicted probability are cross-entropy calculated.

[0074] Data Preprocessing Instructions: Failure mode category labels require one-hot encoding. Specifically, the original classification labels (such as string labels like "normal," "uniform wear," "non-uniform wear," "seal tooth breakage," and "impact wear") are mapped to numerical indices (0, 1, 2, 3, 4, etc.), and then converted into one-hot vectors. For example, for a 5-category problem, "uniform wear" (index 1) is converted to the vector [0, 1, 0, 0, 0], and "impact wear" (index 4) is converted to the vector [0, 0, 0, 0, 1]. This encoding method ensures that there is no numerical relationship between different categories, preventing the model from incorrectly learning the order relationship between categories.

[0075] A transfer learning strategy is employed, using labeled data from similar equipment (such as other compressor models or simulation data) to pre-train the model, and then fine-tuning it on a small amount of data from the target equipment. To reduce the distributional differences between the source and target domains, the maximum mean difference (MMD) is used as the domain adaptation loss. The domain adaptation loss is calculated as follows: the average value of the source domain sample features and the average value of the target domain sample features are calculated separately, the difference vector between the two average values ​​is calculated, and the squared L2 norm of this difference vector is used as the domain adaptation loss.

[0076] Furthermore, the Model-Agnostic Meta-Learning (MAML) algorithm, based on meta-learning, can be used to replace the prototype network for few-sample classification, thereby further improving the model's ability to adapt quickly in cases with very few samples. The MAML algorithm is implemented as follows: Inner loop update: For each task, gradient descent is used to update the model parameters based on the support set, calculating task-specific parameter updates, where the inner loop learning rate is typically 0.01; Outer loop evaluation: Performance is evaluated on the query set, calculating the meta-loss, which is the sum of the losses for all tasks; Meta-parameter update: The initial parameters of the model are updated by minimizing the meta-loss for all tasks, where the outer loop learning rate is typically 0.001.

[0077] In the application of flow-through seal diagnosis, the diagnosis of different devices or different fault degrees is treated as different tasks, and MAML is used to learn a general feature representation across tasks. When faced with new devices or new fault samples, only a few samples are needed for a few gradient updates to achieve rapid adaptation and accurate diagnosis. In specific implementation, a classification head is added to the feature extraction network in step four, and meta-training is performed using the MAML training framework. The inner loop performs gradient updates on the support set of each task, while the outer loop calculates the meta-loss and updates the initial parameters on the query set. After training, for real-time diagnostic tasks, the trained initial parameters are used as a starting point, and fine-tuning is performed using a small number of labeled samples from the target device to quickly obtain an adapted diagnostic model.

[0078] The total loss function is calculated as follows: calculate the loss of each subtask, including classification loss, contrastive learning loss, physical consistency loss, and domain adaptation loss; set the weight coefficients for each task, and determine the weight coefficients using cross-validation. Typically, the weight for classification loss is 1.0, the weight for contrastive learning loss is 0.5, the weight for physical consistency loss is 0.3, and the weight for domain adaptation loss is 0.2; calculate the weighted total loss by linearly combining the losses of each subtask according to the weight coefficients to obtain the total loss function.

[0079] The Adam optimizer was used for training, with a learning rate of 0.001, a batch size of 32, and 100 training epochs. An early stopping strategy was employed during training: training was stopped when the validation set loss did not decrease for 10 consecutive epochs to prevent overfitting. After training, a physically guided small-sample diagnostic model was obtained, which can achieve accurate failure mode identification even when fault samples are scarce.

[0080] Step 5: Input the optimized feature vector into the diagnostic model for anomaly detection, failure mode identification, and severity assessment; calculate the health score and confidence score to obtain the stratified diagnostic results. Specifically, it includes the following: Based on the feature extraction network trained in step 4, the real-time collected monitoring data is preprocessed and feature extracted in steps 2 and 3 to obtain fused features.

[0081] The implementation method of One-Class SVM anomaly detection is as follows: During the training phase, only normal samples are used to construct an optimal hypersphere or hyperplane in the feature space of normal samples to enclose them. An implicit mapping is performed using the radial basis function (RBF) kernel to map the samples in the original feature space to a high-dimensional space, making the linearly inseparable data in the original space linearly separable in the high-dimensional space. During the detection phase, a decision function value is calculated for real-time features, i.e., the distance between the real-time features after kernel mapping and the trained hyperplane is calculated. Anomaly judgment is performed when the decision function value is less than 0, and when it is greater than or equal to 0, it is judged as normal. The anomaly score is defined as the negative value of the decision function value; the larger the anomaly score, the further it deviates from the normal state.

[0082] When the first layer determines an anomaly, the second layer of failure mode recognition is activated. The prototype network classifier trained in step 4 is used to calculate the distance between the real-time features and each failure mode prototype, and the probability of belonging to each category is calculated through softmax.

[0083] The implementation method of failure mode recognition is as follows: calculate the Euclidean distance between the real-time features and the prototypes of each failure mode, use the softmax function to convert the distance into a probability distribution, select the category with the highest probability as the recognition result, and output the failure mode category and the corresponding probability distribution.

[0084] For the identified failure modes, assess their severity and define a health index with a value range of 0 to 1, where 1 represents complete health and 0 represents complete failure.

[0085] Health is calculated based on the degree of feature deviation and the change in key physical parameters. The health calculation process includes calculating the Euclidean distance between the real-time feature and the normal state prototype; normalizing the distance and dividing it by the modulus of the normal state prototype; and calculating the health using an exponential function, where the decay coefficient is set to 2.

[0086] The severity is divided into five levels based on health status: Healthy (health status greater than 0.9), Good (health status between 0.7 and 0.9), Caution (health status between 0.5 and 0.7), Warning (health status between 0.3 and 0.5), and Dangerous (health status less than or equal to 0.3).

[0087] The confidence level of the diagnostic results is assessed to quantify the reliability of the diagnosis. The confidence level assessment includes three aspects: Anomaly detection confidence: calculated based on the distance between the anomaly score and a threshold; the greater the distance, the higher the confidence. The calculation steps are: taking the absolute value of the anomaly score, calculating the confidence level using an exponential function, where the confidence level growth rate is 0.5. Failure mode recognition confidence: calculated based on the entropy of the probability distribution; the smaller the entropy, the more certain the classification, and the higher the confidence. The calculation steps are: the probability distribution entropy calculation process includes taking the logarithm of each category probability and multiplying it by the probability, summing the results, and taking the negative value; the normalized confidence level calculation process includes subtracting the ratio of the entropy value to the maximum possible entropy value from 1. Severity confidence level: Based on the consistency calculation of assessment results of multiple physical parameters, the health status is assessed using multiple independent indicators (such as sealing gap, aerodynamic stiffness, vibration amplitude, etc.), and their standard deviations are calculated. The smaller the standard deviation, the more consistent the assessment and the higher the confidence level. The calculation steps are as follows: perform standardized preprocessing on physical parameters of different dimensions using the Z-score standardization method; calculate the mean and standard deviation of the standardized health status; calculate the confidence level using an exponential function, with the ratio of the standard deviation to the mean as the negative value of the exponent.

[0088] Multi-index consistency calculation requires standardization preprocessing of physical parameters with different dimensions. Specifically, this involves converting each physical parameter (sealing gap, aerodynamic stiffness, vibration amplitude, etc.) into corresponding health indicators, using the Z-score standardization method, where the original health value is subtracted from the historical mean and then divided by the historical standard deviation; the average and standard deviation of the standardized health indicators are then calculated for consistency assessment. The standardized health indicators have the same dimensions and distribution characteristics, allowing for meaningful statistical analysis; the overall confidence level is calculated as a weighted average of three confidence levels, with weighting coefficients of 0.3, 0.4, and 0.3, respectively.

[0089] The weighting coefficients need to be normalized to ensure that their sum is 1. This process involves determining initial weight values ​​based on experience or cross-validation, calculating the sum of the weights, and then performing normalization. This normalization ensures the probabilistic nature of the weighting coefficients, giving the calculation of the overall confidence level a clear mathematical meaning.

[0090] Based on the diagnostic results and physical model, key physical parameters are inverted. According to the predicted sealing gap output from the regression branch of the network in step four, combined with the physical model and measured pressure and temperature data, the gap estimate is updated using Bayesian inference. The Bayesian inference method is as follows: a normal distribution is constructed as the prior distribution based on the sealing gap predicted by the network; a likelihood function is constructed based on the physical model and measurement noise; the prior distribution is multiplied by the likelihood function to obtain the posterior distribution; and the Markov Chain Monte Carlo (MCMC) method is used to sample the posterior distribution to obtain the optimal estimate of the gap and the uncertainty interval. Similarly, other key parameters such as wear depth and effective number of sealing teeth are inverted, outputting complete physical parameter evaluation results. The output is a hierarchical diagnostic result, including abnormal state, abnormal score, failure mode, failure probability, severity, health, overall confidence, and physical parameter estimates.

[0091] Step 6: Based on the health sequence and operating parameters in the stratified diagnostic results, establish a sealing performance degradation model, use the Bayesian method to estimate the degradation model parameters, and use Monte Carlo simulation to predict the remaining service life to obtain the remaining service life prediction results. Specifically, it includes the following: Based on historical diagnostic data, a sealing performance degradation model considering the influence of operating conditions is established. A proportional risk model is used to describe the degradation pattern of health over time.

[0092] The proportional hazards model is established as follows: Define the health degradation differential equation: the derivative of health with respect to time is equal to the negative baseline degradation rate multiplied by the acceleration factor and then multiplied by the power of health; establish the acceleration factor: the acceleration factor is a function of speed, pressure, and temperature, including the power terms of speed ratio, pressure ratio, and temperature influence terms based on the Arrhenius equation; introduce random disturbances: introduce random disturbance terms into the differential equation, and use stochastic differential equations to describe the randomness of the degradation process, where the diffusion coefficient is proportional to health.

[0093] Furthermore, a degradation model based on the Wiener process can be used instead of the proportional hazards model to better describe the stochastic fluctuations in sealing performance degradation. The Wiener process degradation model is established as follows: Defining the degradation process: The degradation process of health is modeled as the initial health minus a drift term and a diffusion term, where the drift term reflects the average degradation rate and the diffusion term reflects the degree of random fluctuation; Considering the influence of operating conditions: The degradation drift rate is modeled as a function of operating condition parameters, with the operating condition acceleration factor having the same form as the proportional hazards model; Parameter estimation: Based on historical health data, the parameters are estimated using the maximum likelihood estimation method. The health increment follows a normal distribution, and the parameters are estimated by maximizing the log-likelihood. The remaining useful life (RUL) prediction uses the first-arrival time distribution; the time when the health first drops to the failure threshold is the RUL. The first-arrival time distribution of the Wiener process is an inverse Gaussian distribution, allowing for analytical calculation of the probability density function and expected value of RUL. This method is suitable for cases where the degradation process exhibits significant random fluctuations, model parameter estimation is simpler, RUL prediction has an analytical solution, and computational efficiency is higher.

[0094] Based on historical health status sequences and corresponding operating condition sequences, Bayesian methods are used to estimate the parameters of the degradation model. The implementation steps of Bayesian parameter estimation are as follows: Establishing the prior distribution of parameters: For cases lacking historical information, weak information priors (such as normal or uniform distributions) are used. If data from similar equipment or expert knowledge is available, an informational prior can be constructed. Defining the likelihood function: Based on the solution of the stochastic differential equation, the conditional distribution of health status approximates a normal distribution, with its mean being the current health status minus the degradation term, and its variance proportional to the time interval. Calculating the posterior distribution: According to Bayes' theorem, the posterior distribution of parameters is proportional to the product of the likelihood function and the prior distribution. Parameter sampling: The Markov chain Monte Carlo method is used, specifically the Metropolis-Hastings algorithm, to perform posterior sampling and obtain the posterior distribution samples of the parameters. Point estimation of parameters uses the posterior mean.

[0095] Based on the current health status and estimated degradation model parameters, the future health status evolution and remaining useful life (RUL) are predicted. The implementation method for RUL prediction is as follows: a failure threshold is set, and when the health status drops below 0.3, it is determined to be a sealing failure; Monte Carlo simulation is performed, and for each simulation, parameters are sampled from the posterior distribution of parameters, and the working condition sequence is sampled from the statistical distribution of future working conditions. The stochastic differential equation is solved, and the calculation is iterated until the health status first crosses the failure threshold; the remaining useful life is recorded, and the time of the first crossing of the threshold is recorded, and the remaining useful life is calculated; statistical analysis is performed, and the RUL sample set is obtained by repeating the simulation 10,000 times. The expected value of RUL, i.e., the point estimate, and the confidence interval are calculated, i.e., the 5% and 95th quantiles are taken as the 90% confidence interval.

[0096] Based on the Monte Carlo simulation results, a health degradation trend curve is plotted. The implementation method for visualizing the degradation trend is as follows: calculate the expected trajectory, and average the health at each time point for all simulated trajectories; calculate the confidence interval, and take the 5th and 95th percentiles of the health at each time point; plot the trend graph, simultaneously displaying the expected trajectory and confidence intervals, intuitively showing the health degradation trend and uncertainty.

[0097] Based on the Remaining Life (RUL) prediction results and maintenance strategy, maintenance decision recommendations are provided. The decision criteria are as follows: when the expected RUL is greater than the safety threshold (e.g., 100 hours), continued operation and routine monitoring are recommended; when the expected RUL is between the warning threshold (e.g., 50 hours) and the safety threshold, increased monitoring frequency and preparation of maintenance resources are recommended; when the expected RUL is less than or equal to the warning threshold, timely maintenance shutdown is recommended to avoid sudden failures. Considering maintenance costs and downtime losses, an optimal maintenance timing decision model based on cost-benefit analysis is used to output recommended maintenance times. The output includes the expected RUL, confidence interval, degradation trend curve, and maintenance decision recommendations.

[0098] Step 7: Integrate the diagnostic model and the remaining life prediction model into an intelligent diagnostic system, and use online real-time diagnosis to obtain a diagnostic report; Specifically, it includes the following: Based on the edge computing architecture, edge computing nodes are deployed close to the data acquisition end to perform data acquisition, preprocessing and feature extraction in steps 1 to 3 in real time.

[0099] Edge nodes utilize industrial-grade embedded computers equipped with multi-core processors and GPU accelerator cards, enabling real-time processing of multi-channel, high-sampling-rate data. Immediately after data acquisition, filtering, denoising, order analysis, and feature extraction are performed to reduce the dimensionality of the high-dimensional raw data into low-dimensional feature vectors, minimizing data transmission volume. These feature vectors are then transmitted to a cloud server via industrial Ethernet or a wireless network, using MQTT or OPC UA protocols to ensure data real-time performance and reliability.

[0100] The cloud server receives feature vectors transmitted from edge nodes and executes diagnostic model inference and hierarchical diagnostic decision-making in steps 4 and 5. A pre-trained deep learning model is deployed in the cloud to infer real-time features and output diagnostic results. The inference process is accelerated using GPUs, with a single inference time of less than 100 milliseconds, meeting real-time requirements.

[0101] The cloud regularly collects new monitoring data and diagnostic results to incrementally update the model. Once a certain amount of new data has accumulated, model retraining is triggered, updating model parameters and improving the model's adaptability and accuracy. Using online learning or federated learning methods, the model can be updated without interrupting diagnostic services.

[0102] Based on cloud-based diagnostic results, the system provides a visual representation via a web interface or mobile app. The visualization includes: a real-time monitoring dashboard displaying key parameters such as current speed, pressure, and temperature, as well as diagnostic results such as health indicators and anomaly scores; historical trend curves showing the changing trends of health and key characteristic parameters over time, facilitating observation of the evolution of sealing performance; failure mode identification results displaying the probability distribution of each failure mode in pie chart or bar chart format, along with the currently identified failure mode type and severity; and remaining lifetime prediction displaying the expected value and confidence interval of RUL, as well as a degradation trend prediction curve.

[0103] When diagnostic results indicate that the seal is in a warning or dangerous state, the system automatically triggers an alarm and notifies maintenance personnel via SMS, email, or app push notification. Alarm information includes alarm time, failure mode, severity, and recommended actions.

[0104] Based on long-term accumulated diagnostic data, the system automatically generates diagnostic reports periodically. The reports include: Statistical analysis of monitoring data: summarizing the distribution of operating conditions and statistical characteristics of key parameters over a period (e.g., one month); Fault event records: recording abnormal events, failure mode identification results, handling measures, and their effects; Health evolution analysis: analyzing the trend of health changes and assessing the rate of seal performance degradation; Remaining Life (RUL) prediction updates: updating RUL predictions based on the latest data, comparing them with previous predictions, and evaluating prediction accuracy; Maintenance recommendations: providing maintenance recommendations for the next stage based on the current health status and RUL predictions, including suggested maintenance time, maintenance content, and spare parts preparation. The diagnostic reports are output in PDF format and can be automatically sent to a specified email address or viewed and downloaded within the system.

[0105] Through the above seven steps, this invention realizes multi-source information collaborative monitoring of flow guide seals, condition-adaptive feature extraction, physically guided intelligent diagnosis, hierarchical multi-granularity decision-making, and remaining life prediction. It constructs a complete intelligent diagnosis method and system for flow guide seals, effectively solves many technical problems in flow guide seal fault diagnosis, and improves the accuracy, real-time performance, and reliability of diagnosis.

[0106] A flow-guiding seal intelligent diagnostic system is used to perform the above-mentioned flow-guiding seal intelligent diagnostic method, such as... Figure 2 As shown, it includes: The information acquisition module is used to collect raw monitoring datasets, which include vibration signals, pressure signals, temperature signals, and rotational speed signals. The preprocessing module is used to perform condition-adaptive preprocessing on the original monitoring dataset to obtain a preprocessed dataset. The feature extraction module extracts multi-scale time-frequency-space collaborative features based on the preprocessed dataset, and uses feature selection methods to reduce dimensionality to obtain optimized feature vectors. The diagnostic model building module, based on optimized feature vectors, adopts a physical-guided deep learning diagnostic algorithm, and is trained using physical consistency constraints and few-shot learning methods to obtain a physical-guided few-shot diagnostic model. The diagnostic analysis module is used to input optimized feature vectors into the diagnostic model for anomaly detection, failure mode identification, and severity assessment, calculate health status and confidence level, and obtain stratified diagnostic results. The life prediction module establishes a sealing performance degradation model based on the health sequence and operating parameters in the hierarchical diagnosis results, estimates the degradation model parameters using the Bayesian method, and uses Monte Carlo simulation to predict the remaining service life, thus obtaining the remaining service life prediction results. The results output module is used to integrate the diagnostic model and the remaining life prediction model into an intelligent diagnostic system, which uses online real-time diagnosis to obtain a diagnostic report.

[0107] In one embodiment of the present invention, a specific example is provided: To verify the effectiveness of the method of this invention, a 180-day field test was conducted at a test base. During the test, a multi-channel data acquisition system was deployed, including 16 vibration sensors, 8 pressure sensors, 4 temperature sensors, and 2 speed sensors. The test area covered approximately 500 square meters of the high-pressure compressor test bench. The compressor rotor speed range of the test bench was 12,000-18,000 rpm. The flow guide seal was a 4-stage comb-tooth seal structure, with 36 sealing teeth per stage, a sealing diameter of 450 mm, and a designed clearance of 0.5 mm.

[0108] Following step one, install the sensors by placing a triaxial accelerometer at both the front and rear of the bearing housing, and pressure and temperature sensors on the inlet and outlet pipelines. The sampling frequency is 6000Hz, and data is collected every 10 seconds.

[0109] Examples of vibration signal data acquisition are shown in Table 1: Table 1: Example of vibration signal data acquisition; Examples of data acquisition under different operating conditions are shown in Table 2: Table 2: Examples of data acquisition under different operating conditions; As can be seen from the above examples, the intelligent diagnostic method for flow-guiding seals provided by the present invention can effectively identify the failure mode of the seal, accurately assess its health status, and reliably predict its remaining life. It has good performance and practical value in practical engineering applications.

[0110] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A method for intelligent diagnosis of flow-guiding seals, characterized in that, Includes the following steps: The original monitoring dataset is collected, which includes measured data and parameters extrapolated by virtual sensing. The measured data includes vibration signals, pressure signals, temperature signals, and rotational speed signals. The parameters extrapolated by virtual sensing include the current sealing gap, aerodynamic stiffness coefficient, and aerodynamic damping coefficient. Specifically, based on the labyrinth seal leakage calculation model, the leakage mass flow rate is calculated according to the inlet and outlet pressures, temperatures, and rotational speeds. The current sealing gap is inverted using the measured pressure difference and temperature. Based on rotor dynamics theory, the aerodynamic stiffness coefficient and aerodynamic damping coefficient of the seal are inverted using parameter identification methods based on the measured rotor vibration response and rotational speed. The original monitoring dataset is subjected to adaptive preprocessing under working conditions to obtain a preprocessed dataset. The adaptive preprocessing under working conditions includes: calculating the instantaneous rotational speed and interpolating the discrete rotational speed signal into a continuous instantaneous rotational speed curve; performing angular domain resampling on the time-domain vibration signal and converting equal time interval sampling into equal angular interval sampling; performing order analysis and extracting the amplitude of key orders as feature parameters to achieve working condition normalization. The dimensionless pressure difference coefficient is calculated based on the measured values ​​of inlet and outlet pressures, the dimensionless temperature rise coefficient is calculated based on outlet and inlet temperatures, the linear velocity at the seal is obtained by multiplying the rotor speed by the seal diameter as the characteristic velocity, the radial clearance of the seal is used as the characteristic length, and the product of the inlet gas density, characteristic velocity and characteristic length is divided by the gas dynamic viscosity to obtain the corrected Reynolds number. Multi-scale time-frequency-space-coordinated features are extracted from the preprocessed dataset, and dimensionality reduction is performed using a feature selection method to obtain an optimized feature vector. The multi-scale time-frequency-space-coordinated features include: calculating time-domain statistical features for each channel's order-domain signal, where each channel's order-domain signal includes the order-domain signals of each vibration signal channel and the pressure pulsation signal channel after angular domain resampling; performing time-frequency analysis on the vibration signal to extract time-frequency domain features; calculating spatial correlation features based on vibration signals from different measuring points; standardizing and concatenating the time-domain statistical features, order-domain features, time-frequency domain features, and spatial correlation features; and then using a feature selection method to reduce dimensionality to obtain an optimized feature vector. Based on optimized feature vectors, a physical-guided deep learning diagnostic algorithm is employed, trained using physical consistency constraints and few-shot learning methods to obtain a physical-guided few-shot diagnostic model. The physical-guided deep learning diagnostic algorithm includes: constructing a dual-channel feature extraction network, comprising a physical channel and a data channel; the physical channel inputs physical parameters inferred from virtual sensing, dimensionless pressure difference coefficients, dimensionless temperature rise coefficients, and corrected Reynolds numbers, which are then mapped through a fully connected neural network to obtain physical features; the data channel uses a multilayer perceptron structure to perform nonlinear transformation on the optimized feature vectors to obtain data features; the implementation method of physical consistency constraints involves adding a regression branch to predict the sealing gap using fused features; the physical features and data features are weighted and fused using an attention mechanism to obtain fused features; feature representation pre-training is performed on normal operating data using a contrastive learning method; a physical consistency loss is added to the total training loss, which is the square of the difference between the gap predicted by the calculation network and the gap calculated by the physical model; and a few-shot classifier is used for training to obtain the physical-guided few-shot diagnostic model. The optimized feature vectors are input into the diagnostic model for anomaly detection, failure mode identification, and severity assessment. Health score and confidence score are calculated to obtain hierarchical diagnostic results. This includes: using a One-Class SVM for anomaly detection, classifying anomalies as those with a decision function value less than 0; using a prototype network for failure mode identification, calculating the Euclidean distance between real-time fused features and various prototypes, and outputting the probability of belonging to each failure mode using a softmax function, selecting the category with the highest probability as the identification result; the prototype class is the average of the fused feature vectors of all labeled samples in that class; the health score calculation process includes calculating the Euclidean distance between real-time fused features and normal state prototypes; normalizing the distance, dividing it by the modulus of the normal state prototype, and calculating the health score using an exponential function; classifying severity into different levels based on health score; and calculating a comprehensive confidence score based on indicator consistency, where the comprehensive confidence score is a weighted average of anomaly detection confidence, failure mode identification confidence, and severity confidence. A sealing performance degradation model is established based on the health status sequence and operating condition parameters from the hierarchical diagnostic results. Bayesian methods are used to estimate the degradation model parameters, and Monte Carlo simulation is used to predict the remaining service life, yielding the remaining service life prediction results. The operating condition parameters include rotational speed, pressure, and temperature. The sealing performance degradation model uses a proportional hazards model to describe the degradation law of health over time. A prior distribution of the degradation model parameters is established using Bayesian methods, and posterior sampling is performed using Markov chain Monte Carlo methods to obtain posterior distribution samples of the parameters. Based on the current health status, the posterior distribution of the degradation model parameters, and the statistical distribution of future operating conditions, the evolution of health status is predicted through Monte Carlo simulation. When the health status first crosses the failure threshold, it is determined as sealing failure, and the time of the first crossing is recorded as the remaining service life. The expected value and confidence interval of the remaining service life are statistically obtained. The diagnostic model and the remaining life prediction model are integrated into an intelligent diagnostic system, which uses online real-time diagnosis to obtain a diagnostic report.

2. The intelligent diagnostic method for flow guiding and sealing according to claim 1, characterized in that, The use of online real-time diagnosis includes: Based on the edge computing architecture, edge computing nodes perform data acquisition, preprocessing, and feature extraction in real time close to the data acquisition end; The cloud server receives feature vectors transmitted from edge nodes and performs diagnostic model inference and hierarchical diagnostic decisions. The cloud regularly collects new monitoring data and diagnostic results to incrementally update the model, and uses online learning methods to update the model parameters.

3. The intelligent diagnostic method for flow guiding seals according to claim 1, characterized in that, The anomaly detection, failure mode identification, and severity assessment also include: The health status calculation is based on the degree of feature deviation and the amount of change in key physical parameters, and calculates the distance between the real-time features and the normal state prototype; the key physical parameters include sealing gap, wear depth and effective number of sealing teeth; Based on the diagnostic results and key physical parameters inverted from the physical model, the gap estimate is updated using Bayesian inference. The physical model includes a labyrinth seal leakage calculation model. The gap estimate is updated based on the seal gap predicted by the regression branch. A prior distribution is constructed using this predicted seal gap. A likelihood function of the observed data is constructed based on the physical model and measurement noise. The prior distribution is multiplied by the likelihood function to obtain the posterior distribution. The updated seal gap estimate and uncertainty interval are obtained by sampling the posterior distribution through Markov chain Monte Carlo.

4. A flow-guiding and sealing intelligent diagnostic system, characterized in that, include: The information acquisition module is used to collect raw monitoring datasets, which include measured datasets and parameters extrapolated by virtual sensing. The measured datasets consist of vibration signals, pressure signals, temperature signals, and rotational speed signals. The parameters extrapolated by virtual sensing include the current sealing gap, aerodynamic stiffness coefficient, and aerodynamic damping coefficient. Specifically, based on the labyrinth seal leakage calculation model, the leakage mass flow rate is calculated according to the inlet and outlet pressures, temperatures, and rotational speeds. The current sealing gap is inverted using measured pressure difference and temperature. Based on rotor dynamics theory, the aerodynamic stiffness coefficient and aerodynamic damping coefficient of the seal are inverted using parameter identification methods based on the measured rotor vibration response and rotational speed. The preprocessing module is used to perform adaptive preprocessing on the original monitoring dataset to obtain a preprocessed dataset. The adaptive preprocessing includes: calculating the instantaneous rotational speed and interpolating the discrete rotational speed signal into a continuous instantaneous rotational speed curve; performing angular domain resampling on the time-domain vibration signal and converting equal time interval sampling into equal angular interval sampling; performing order analysis and extracting the amplitude of key orders as feature parameters to achieve operating condition normalization; calculating the dimensionless pressure difference coefficient based on the measured values ​​of inlet and outlet pressures, calculating the dimensionless temperature rise coefficient based on the outlet and inlet temperatures, multiplying the rotor speed by the sealing diameter to obtain the linear velocity at the seal as the feature velocity, using the sealing radial clearance as the feature length, and dividing the product of the inlet gas density, feature velocity, and feature length by the gas dynamic viscosity to obtain the corrected Reynolds number. The feature extraction module is used to extract multi-scale time-frequency-space-coordinated features based on the preprocessed dataset, and to perform dimensionality reduction using a feature selection method to obtain an optimized feature vector. The multi-scale time-frequency-space-coordinated features include: calculating time-domain statistical features for each channel's order-domain signal, where each channel's order-domain signal includes the order-domain signals of each vibration signal channel and the pressure pulsation signal channel after angular domain resampling; performing time-frequency analysis on the vibration signal to extract time-frequency domain features; calculating spatial correlation features based on vibration signals from different measuring points; and standardizing and concatenating the time-domain statistical features, order-domain features, time-frequency domain features, and spatial correlation features, and then using a feature selection method to reduce dimensionality to obtain an optimized feature vector. A diagnostic model construction module is used to obtain a physical-guided small-sample diagnostic model based on optimized feature vectors and employing a physical-guided deep learning diagnostic algorithm. This algorithm utilizes physical consistency constraints and few-shot learning methods for training. The physical-guided deep learning diagnostic algorithm includes: constructing a dual-channel feature extraction network, comprising a physical channel and a data channel; the physical channel inputs physical parameters inferred from virtual sensing, dimensionless pressure difference coefficients, dimensionless temperature rise coefficients, and corrected Reynolds numbers, which are then mapped through a fully connected neural network to obtain physical features; the data channel uses a multilayer perceptron structure to perform nonlinear transformation on the optimized feature vectors to obtain data features; the physical consistency constraint is implemented by adding a regression branch to predict the sealing gap using fused features; the physical features and data features are weighted and fused using an attention mechanism to obtain fused features; feature representation pre-training is performed on normal operating data using a contrastive learning method; a physical consistency loss is added to the total training loss, which is the square of the difference between the gap predicted by the calculation network and the gap calculated by the physical model; and a few-shot classifier is used for training to obtain the physical-guided small-sample diagnostic model. The diagnostic analysis module is used to input optimized feature vectors into the diagnostic model for anomaly detection, failure mode identification, and severity assessment, calculating health and confidence levels to obtain hierarchical diagnostic results. This includes: using a One-Class SVM for anomaly detection, classifying anomalies as those with a decision function value less than 0; using a prototype network for failure mode identification, calculating the Euclidean distance between real-time fused features and various prototypes, and outputting the probability of belonging to each failure mode using a softmax function, selecting the category with the highest probability as the identification result; the prototype class is the average of the fused feature vectors of all labeled samples in that class; the health calculation process includes calculating the Euclidean distance between real-time fused features and normal state prototypes; normalizing the distance, dividing it by the modulus of the normal state prototype, and calculating the health using an exponential function; classifying severity into different levels based on health; and calculating a comprehensive confidence level based on indicator consistency, where the comprehensive confidence level is a weighted average of anomaly detection confidence, failure mode identification confidence, and severity confidence. The life prediction module is used to establish a sealing performance degradation model based on the health status sequence and operating condition parameters in the hierarchical diagnostic results. It uses a Bayesian method to estimate the degradation model parameters and employs Monte Carlo simulation to predict the remaining service life, obtaining the remaining service life prediction result. The operating condition parameters include rotational speed, pressure, and temperature. The sealing performance degradation model uses a proportional hazards model to describe the degradation law of health status over time. A prior distribution of the degradation model parameters is established using a Bayesian method, and a posterior sampling method using a Markov chain Monte Carlo method is used to obtain the posterior distribution sample of the parameters. Based on the current health status, the posterior distribution of the degradation model parameters, and the statistical distribution of future operating conditions, the evolution of health status is predicted through Monte Carlo simulation. When the health status first crosses the failure threshold, it is determined as sealing failure, and the time of the first crossing is recorded as the remaining service life. The expected value and confidence interval of the remaining service life are statistically obtained. The results output module is used to integrate the diagnostic model and the remaining life prediction model into an intelligent diagnostic system, which uses online real-time diagnosis to obtain a diagnostic report.

Images