Equipment supervision data real-time analysis method and system

Abstract

The invention provides an equipment supervision data real-time analysis method and system, and the method comprises the steps: calculating an energy value of a frequency band corresponding to a sub-band signal, determining the energy characteristics of the surface roughness and dust deposition distribution uniformity of a blade in each sub-band in a square sum form, and generating energy distribution vectors which are arranged according to a frequency band sequence; performing time-frequency analysis on the vibration signal by using the acquired real-time rotating speed information of the gas turbine, identifying a frequency component corresponding to the eccentric distance value of the rotor and the vibration frequency offset, and determining the amplitude of the frequency component; according to the classification result and the real-time rotating speed information, the comprehensive influence index of the material density change rate and the ash deposition layer thickness difference is calculated, the fault early warning level is determined through a weighted average method, and an early warning signal is generated; and according to the change trend of the early warning signal and the blade edge wear degree, a random forest algorithm is adopted to predict a development curve of the vibration frequency offset in the future target time, and real-time equipment supervision analysis data are obtained.

Description

Technical Field

[0001] The present invention relates to the field of information technology, and particularly to a method and system for real-time analysis of equipment supervision data. Background Art

[0002] When researching a real-time fault warning method based on vibration data of a large gas turbine, a specific technical problem needs to be deeply explored: how to accurately capture the frequency component synchronized with the rotational speed in the vibration signal and the characteristics of its amplitude changing with the unbalance degree in the scenario where the rotor has uneven mass distribution due to blade fouling, and extract the energy distribution characteristics of different frequency bands through wavelet packet decomposition technology. Finally, an early warning model is constructed by combining the rotational speed information. The energy characteristics of different frequency bands reflect the response state of the rotor system under different excitation frequencies. Among them, the fundamental frequency energy is mainly related to the unbalance state of the rotor, the harmonic energy may reflect the nonlinear characteristics of the rotor system or faults such as friction and looseness between components, and the high-frequency energy may be related to local turbulence, impact and other phenomena caused by blade fouling. The complexity of the problem lies in that during the operation of the gas turbine, the impeller may increase uneven additional weight due to corrosion or fouling, and the unevenness of the blade fouling degree will lead to the dynamic evolution of rotor unbalance. The vibration signal not only contains the frequency component synchronized with the rotational speed, but may also be superimposed with sub-harmonics or high-frequency noise caused by the asymmetry of the fouling distribution, resulting in the difficulty of traditional warning methods based on single frequency characteristics to accurately identify the early signs of faults, and being easily affected by noise interference, resulting in false alarms or missed alarms. When existing signal processing methods decompose frequency bands, it is difficult to adaptively match the real-time change of the rotor speed, resulting in insufficient resolution of energy feature extraction. Especially when the unbalance is weak in the initial stage of fouling, the characteristic signal is easily masked by background noise. In addition, although wavelet packet decomposition can refine the frequency band, how to determine the specific frequency band directly related to the rotor unbalance fault mode and maintain the stability of feature extraction under rotational speed fluctuations is still a difficult point. When correlating the energy characteristics of these frequency bands with the rotational speed information, the problem of time sequence alignment of data fusion needs to be solved, because there may be a slight delay between the rotational speed acquisition and the vibration signal sampling, which will interfere with the model's judgment of the fault trend. Finally, how to design an early warning threshold that is sensitive to fouling-induced unbalance faults and robust to other interferences in the real-time monitoring of the operating state has become a technical contradiction to be solved urgently. Summary of the Invention

[0003] To solve the above technical problems, the present invention provides a method for real-time analysis of equipment supervision data, which mainly includes:

[0004] Collect continuous vibration time series data of the impeller during the operation of the gas turbine through sensors, and perform multi-scale decomposition on the corrosion pit depth distribution and the difference in fouling layer thickness using wavelet packet decomposition to obtain sub-band signals of different frequency bands;

[0005] Calculate the energy value of the corresponding frequency band of the sub-band signal, and use the form of the sum of squares to determine the energy characteristics of the blade surface roughness and the ash deposition uniformity in each sub-band, and generate an energy distribution vector arranged in the order of frequency bands;

[0006] Perform time-frequency analysis on the vibration signal using the obtained real-time rotational speed information of the gas turbine, identify the frequency components corresponding to the rotor eccentricity distance value and the vibration frequency offset amount, and determine the amplitude of the frequency components;

[0007] Construct a feature matrix based on the energy distribution vector and the amplitude of the frequency components, and analyze the change trend of the blade edge wear degree and the ash deposition time difference in combination with the impeller center of gravity offset amount and the ash particle density value obtained by real-time monitoring to obtain a preliminary anomaly score;

[0008] If the preliminary anomaly score exceeds the preset threshold, then compare the feature matrix with the historical operation data, calculate the correlation coefficients of the current matrix with the corrosion area ratio and the ash amount on the windward side of the blade, and determine the rotor imbalance probability value;

[0009] Based on the rotor imbalance probability value, combined with the local stress concentration coefficient obtained by finite element analysis simulation of the blade, use the support vector machine algorithm to determine whether the current operating state is unbalanced due to the corrosion deposit thickness or the blade surface adhesion force, and obtain the classification result;

[0010] Through the classification result and the real-time rotational speed information, calculate the comprehensive influence index of the material density change rate and the ash layer thickness difference, use the weighted average method to determine the fault warning level, and generate a warning signal;

[0011] According to the warning signal and the change trend of the blade edge wear degree, use the random forest algorithm to predict the development curve of the vibration frequency offset amount within the future target time to obtain real-time equipment supervision analysis data.

[0012] Further, continuous vibration time series data of the impeller during the operation of the gas turbine is collected by sensors. Wavelet packet decomposition is used to perform multi-scale decomposition on the corrosion pit depth distribution and the difference in the ash deposit layer thickness, and sub-band signals in different frequency bands are obtained, including: obtaining the vibration time series data measured at a fixed sampling interval from the gas turbine impeller sensor, compensating the original time series data according to data integrity and sampling accuracy, obtaining the waveform integrity score by calculating the root mean square value, peak factor, and waveform index of the sequence data. When the waveform integrity score is less than the preset threshold, this section of data is marked as invalid data. A wavelet packet decomposition tree is constructed for the compensated effective time series data, the decomposition level is set to four layers, and the db4 wavelet basis function is selected to perform multi-level decomposition on the signal to obtain sub-band signals in sixteen frequency bands. Calculate the energy value and energy proportion of each sub-band signal, sort the sub-band signals according to the energy proportion to obtain the importance ranking, and select the sub-band signals with an energy proportion exceeding the preset threshold as the feature sub-bands. Perform wavelet threshold denoising on the feature sub-band signals, determine the optimal threshold using the Bayesian threshold criterion, and obtain the denoised feature sub-band signals. For the denoised feature sub-band signals, extract statistical features such as mean, standard deviation, skewness, and kurtosis, and construct a corrosion depth feature matrix and an ash deposit layer thickness feature matrix. Use the principal component analysis method to reduce the dimension of the feature matrix, and retain the principal components with a cumulative contribution rate reaching the preset threshold as the feature vectors after dimension reduction. Based on the feature vectors after dimension reduction, train a support vector regression model, and establish a corrosion pit depth prediction model and an ash deposit layer thickness prediction model respectively. Input the feature vectors into the trained prediction models to obtain the corrosion pit depth distribution value and the difference value of the ash deposit layer thickness.

[0013] Further, calculate the energy value of the corresponding frequency band of the sub-band signal, determine the energy characteristics of the blade surface roughness and the ash deposition distribution uniformity in each sub-band in the form of the sum of squares, and generate an energy distribution vector arranged in the order of frequency bands, including: extracting sub-band data from the gas turbine blade vibration signal, performing four-layer decomposition on the vibration signal according to wavelet packet transform, calculating the spectral energy mean value for the sub-band data of sixteen frequency bands to obtain the first frequency band energy matrix. Perform smoothing filtering on the first frequency band energy matrix through a sliding window with a length of 128 points to obtain the second frequency band energy matrix, perform envelope extraction on the second frequency band energy matrix, and obtain the local maximum points of the envelope curve. Construct a frequency band sequence according to the amplitude and position information of the local maximum points, perform segmentation processing on the frequency band sequence, calculate the mean value, variance and skewness of each segment of data to obtain the initial feature sequence. Calculate the variance cumulant for the initial feature sequence, calculate the local entropy value of the sub-band data using the information entropy formula, and construct the blade surface roughness feature sequence and the ash deposition distribution uniformity feature sequence according to the variance cumulant and the entropy value. Use the linear weighted method to fuse the two feature sequences, and the weight coefficient is obtained by optimizing through the least squares method to obtain the frequency band energy distribution weight sequence. Perform zero-mean normalization on the frequency band energy distribution weight sequence, extract the main feature components through the principal component analysis method, and obtain the energy feature vector in the order of frequency bands according to the descending order of the eigenvalues. The gas turbine blade vibration signal contains rich state information. Through wavelet packet decomposition, the signal is divided into sub-bands of sixteen different frequency bands, and each sub-band reflects the vibration characteristics in different frequency ranges.

[0014] Furthermore, time-frequency analysis is performed on the vibration signal using the acquired real-time rotational speed information of the gas turbine to identify the frequency components corresponding to the rotor eccentricity distance value and the vibration frequency offset amount, and the amplitude of the frequency components is determined, including: collecting the real-time rotational speed data of the rotor from the gas turbine speed sensor, calculating the sampling frequency adjustment coefficient according to the rotational speed data, resampling the rotational speed data for the adjusted sampling frequency to obtain the first rotational speed sequence. Performing Hilbert transform on the first rotational speed sequence, obtaining the second rotational speed sequence through instantaneous phase calculation, determining the fundamental frequency value according to the time-domain characteristics of the second rotational speed sequence to obtain the rotational speed fundamental frequency data. Segmenting the vibration signal using a Hanning window with a length of 256 points, performing short-time Fourier transform on each segment of the signal according to the rotational speed fundamental frequency data, obtaining the frequency peak points through triple root mean square threshold detection to obtain the first frequency amplitude data. Calculating the multiple frequency ratio according to the first frequency amplitude data, constructing the multiple frequency spectrum using the ratio sequence, determining the segmentation threshold of the multiple frequency spectrum through the autocorrelation coefficient to obtain the second frequency amplitude data. Performing phase demodulation on the second frequency amplitude data, calculating the initial eccentricity distance value according to the phase difference, linearly correcting the initial eccentricity distance value using the least squares method to obtain the corrected eccentricity distance data. Constructing a frequency offset amount feature vector for the corrected eccentricity distance data, extracting the frequency offset feature through wavelet transform, and dynamically estimating the feature data using Kalman filter to obtain the amplitude result of the frequency component.

[0015] Further, a feature matrix is constructed based on the energy distribution vector and the amplitude of the frequency component. Combining the impeller center of gravity offset and the ash particle density value obtained from real-time monitoring, the change trend of the blade edge wear degree and the ash deposition time difference is analyzed to obtain a preliminary anomaly score, including: constructing a state feature matrix based on the energy distribution vector and the amplitude of the frequency component, using the principal component dimensionality reduction method to retain the feature components whose cumulative contribution rate reaches a preset threshold, and obtaining a compressed state matrix. Extract the time-domain features of the compressed state matrix, calculate statistical parameters such as mean, variance, and kurtosis using a sliding window, and extract the time-domain features through a multi-layer perceptron to obtain the impeller state parameters. Perform Fourier transform on the displacement data collected by the impeller center of gravity offset sensor, use a band-pass filter with a center frequency near the fundamental frequency to denoise the spectral data, calculate the offset amplitude and phase difference according to the denoised spectral curve, and obtain the center of gravity offset parameters. Collect density data through the ash particle density sensor, perform four-layer wavelet packet decomposition on the density data, calculate the ratio of the signal variance to the mean of each sub-band, and obtain the density distribution parameters. Calculate the cross-correlation coefficient according to the density distribution parameters and the deposition time series, and use the exponential weighting method to smooth the cross-correlation sequence to obtain the ash deposition parameters. Normalize the maximum and minimum values of the impeller state parameters, center of gravity offset parameters, and ash deposition parameters, and use the weight calculation method based on information gain to construct a fusion weight vector. Weight and combine the normalized parameters according to the fusion weight vector to obtain a characterization vector of the blade edge wear degree and the ash deposition time difference. Perform probability density estimation on the characterization vector using the generalized extreme value distribution, determine the location parameter and scale parameter by the maximum likelihood method, and calculate the cumulative distribution function value as the preliminary anomaly score.

[0016] Further, if the preliminary anomaly score exceeds the preset threshold, the correlation coefficients of the current matrix with the corrosion area ratio and the ash content on the windward side of the blade are calculated by comparing the feature matrix with the historical operation data, and the rotor imbalance probability value is determined, including: if the anomaly score value in the current state matrix exceeds the preset threshold, the historical operation data is segmented by a sliding window with a length of 512 points, the Euclidean distance between each segment of data and the current state matrix is calculated, and the first similarity matrix is obtained. The first similarity matrix is normalized, and the optimal path distance is calculated using the dynamic programming method to obtain the second similarity matrix. The autocorrelation coefficient is calculated according to the second similarity matrix, a weight vector is constructed for the correlation coefficient, and the corrosion area data in the historical data is weighted and averaged to obtain the first probability sequence. The exponential weighting method is used to smooth the blade ash content data in the historical data, and the distribution characteristics of the smoothed data are calculated by Gaussian kernel density estimation to obtain the second probability sequence. The first probability sequence and the second probability sequence are standardized, and a probability mapping function is constructed using polynomial regression to obtain the joint probability value of the corrosion area ratio and the ash content distribution. A log-likelihood function is constructed according to the joint probability value, and the posterior probability distribution of the rotor imbalance state is calculated by maximum likelihood estimation to obtain the imbalance probability value.

[0017] Further, according to the rotor imbalance probability value, combined with the local stress concentration coefficient obtained by finite element analysis and simulation of the blade, the support vector machine algorithm is used to judge whether the current operating state is unbalanced due to the corrosion deposit thickness or the blade surface adhesion, and the classification result is obtained, including: a blade stress state matrix is constructed according to the rotor imbalance probability value, the blade surface is meshed by the tetrahedral mesh generation method, the node stress value is calculated for the mesh element, and the first stress distribution matrix is obtained. The gradient of the first stress distribution matrix is calculated, the stress gradient mutation points are marked by the region growing method, and the stress concentration region is extracted according to the mutation point distribution to obtain the second stress distribution matrix. The principal stress direction and value are calculated for the second stress distribution matrix, and the local stress concentration coefficient is extracted by the stress tensor decomposition method to obtain the stress concentration characteristic data. A spectrum analysis window is constructed according to the stress concentration characteristic data, the instantaneous frequency characteristic is calculated by Hilbert transform, and the frequency band range is determined by the adaptive segmentation method to obtain the stress spectrum characteristic data. The discrete wavelet transform is used to perform multi-scale analysis on the stress spectrum characteristic data, the characteristic coefficients are screened by the variance threshold method to obtain the multi-scale characteristic vector. The characteristic weight is calculated according to the multi-scale characteristic vector, the kernel principal component method is used to reduce the dimension of the characteristic data, and a sample matrix is constructed by combining the sediment thickness data and the surface adhesion data. The support vector machine with a Gaussian kernel function is used for feature learning on the sample matrix, the penalty factor and the kernel parameter are determined by cross-validation to obtain the classification decision boundary. The current state data is classified according to the classification decision boundary to obtain the imbalance classification label caused by corrosion deposition.

[0018] Furthermore, collect the blade ash deposition distribution data to obtain the initial distribution characteristics, analyze the initial distribution characteristics to determine the sub-harmonic and high-frequency noise components. If the sub-harmonic and high-frequency noise exceed the preset threshold, decompose the signal through wavelet transform to obtain the denoised characteristics, calculate the rotor imbalance degree and judge its dynamic change trend, compare with the historical trend to determine the abnormal distribution pattern, classify the abnormal distribution pattern, and obtain the early signs of faults, including: construct the first feature matrix according to the distribution data collected by the blade surface ash sensor, perform spectral decomposition on the matrix data through Fourier transform to obtain the fundamental frequency component, sub-harmonic component and high-frequency noise component. Calculate the amplitude ratio for the fundamental frequency component and sub-harmonic component, calculate the root mean square of the high-frequency noise component, mark the over-standard area according to the preset fundamental frequency ratio threshold and noise root mean square threshold, and obtain the second feature matrix. Perform four-layer decomposition on the over-standard area in the second feature matrix by using multi-scale wavelet transform, select the optimal decomposition scale according to the energy density, perform noise reduction on the wavelet coefficients by using the soft threshold method, and obtain the third feature matrix. Calculate the imbalance index for the third feature matrix, fit the imbalance sequence by using the least squares method, calculate the imbalance change rate through numerical differentiation, and obtain the first trend sequence. Construct a time window according to the first trend sequence, segment the historical data by using the sliding window method, calculate the similarity between the current trend and the historical trend through the Euclidean distance, and obtain the second trend sequence. Perform normalization processing on the second trend sequence, classify the trend data by using the K-means clustering algorithm, determine the abnormal distribution pattern according to the inter-class distance, and obtain the third trend sequence. Adopt the time series correlation analysis method to extract features from the third trend sequence, construct a fault mode library according to the feature combination, and obtain the early signs of fault labels.

[0019] Furthermore, based on the classification results and real-time rotational speed information, calculate the comprehensive influence index of the material density change rate and the ash deposit layer thickness difference, and use the weighted average method to determine the fault warning level and generate a warning signal, including: constructing a state matrix based on the classification result label and real-time rotational speed data, calculating the spectral characteristics of the material density change using Fourier transform, calculating the change rate of the ash deposit layer thickness through a difference operator, and obtaining a state feature vector. Perform maximum-minimum normalization on the state feature vector, set a sliding window according to the sampling period of the rotational speed signal, and smooth the normalized data using the window mean to obtain a smoothed feature vector. Perform three-layer wavelet packet decomposition on the smoothed feature vector, calculate the contribution degree of each sub-band signal using information entropy, sort the sub-band signals according to the contribution degree, and obtain a multi-scale feature vector. Calculate the covariance matrix for the multi-scale feature vector, extract the main feature components using eigenvalue decomposition, determine the feature dimension through the cumulative variance contribution rate, and obtain a dimensionality-reduced feature vector. Perform hierarchical clustering on the dimensionality-reduced feature vector, determine the fault warning threshold according to the inter-class distance, calculate the warning index score through threshold comparison, and obtain a warning feature vector. Construct a fuzzy membership function for the warning feature vector, calculate the membership degree values of each index using triangular membership, determine the fault warning level according to the membership degree values, and obtain a warning level vector. Perform exponential weighting on the warning level vector, calculate the weighted warning score using the geometric mean method, and generate a warning signal value according to the warning score.

[0020] Furthermore, based on the change trends of the warning signal and the blade edge wear degree, use the random forest algorithm to predict the development curve of the vibration frequency offset in the future target time to obtain real-time equipment supervision analysis data, including: constructing a time state matrix based on the warning signal value, segmenting the blade edge wear degree data using a sliding window with a length of 256 points, calculating the wear change rate through third-order polynomial fitting, and obtaining a wear change matrix. Smooth the data for the wear change matrix, use the exponential weighted moving average method to eliminate short-term fluctuations, calculate the change trend slope based on the smoothed data, and obtain a slope feature vector. Perform wavelet decomposition on the slope feature vector, use the soft threshold method to denoise the wavelet coefficients, and reconstruct the signal through the optimal scale to obtain a trend feature vector. Calculate the temporal correlation for the trend feature vector, construct a feature correlation matrix using the Pearson correlation coefficient, select significant features according to the correlation matrix, and obtain a prediction feature vector. Standardize the data for the prediction feature vector, construct a predictor using the random forest algorithm, determine the optimal number of trees and feature sampling ratio through cross-validation, and obtain prediction model parameters. Calculate the future change value of the frequency offset according to the prediction model parameters, smooth the prediction result using cubic spline interpolation, determine the prediction interval through confidence testing, and obtain a prediction curve vector. Segment the prediction curve vector, calculate the slope of each segment using piecewise linear regression, identify the development inflection point according to the slope change, and obtain a monitoring data matrix.

[0021] The present invention provides a real-time analysis system for equipment supervision data, mainly including:

[0022] A wavelet packet decomposition module, which is used to collect continuous vibration time series data of the impeller during the operation of the gas turbine through a sensor, and perform multi-scale decomposition on the depth distribution of corrosion pits and the difference in ash deposit layer thickness by using wavelet packet decomposition to obtain sub-band signals of different frequency bands;

[0023] An energy feature extraction module, which is used to calculate the energy values of the corresponding frequency bands of the sub-band signals, determine the energy features of the blade surface roughness and the ash deposit distribution uniformity in each sub-band in the form of the sum of squares, and generate an energy distribution vector arranged in the order of frequency bands;

[0024] A time-frequency analysis module, which is used to perform time-frequency analysis on the vibration signal by using the obtained real-time rotational speed information of the gas turbine, identify the frequency components corresponding to the rotor eccentricity distance value and the vibration frequency offset amount, and determine the amplitude of the frequency components;

[0025] A feature matrix construction module, which is used to construct a feature matrix according to the energy distribution vector and the amplitude of the frequency components, analyze the change trend of the blade edge wear degree and the ash deposit time difference in combination with the impeller center of gravity offset amount and the ash particle density value obtained by real-time monitoring, and obtain a preliminary anomaly score;

[0026] A preliminary anomaly scoring module, which is used to, if the preliminary anomaly score exceeds a preset threshold, calculate the correlation coefficients of the current matrix with the corrosion area ratio and the ash amount on the windward surface of the blade by comparing the feature matrix with historical operation data, and determine the rotor imbalance probability value;

[0027] A rotor imbalance probability calculation module, which is used to, according to the rotor imbalance probability value, combine the local stress concentration coefficient obtained by finite element analysis simulation of the blade, and use the support vector machine algorithm to judge whether the current operating state is unbalanced due to the corrosion deposit thickness or the blade surface adhesion force, and obtain a classification result;

[0028] A fault warning level determination module, which is used to calculate the comprehensive influence index of the material density change rate and the difference in ash deposit layer thickness through the classification result and the real-time rotational speed information, and use the weighted average method to determine the fault warning level and generate a warning signal;

[0029] A vibration frequency offset prediction module, which is used to predict the development curve of the vibration frequency offset in the future target time by using the random forest algorithm according to the warning signal and the change trend of the blade edge wear degree, and obtain real-time equipment supervision analysis data.

[0030] The technical solution provided by the embodiment of the present invention may include the following beneficial effects:

[0031] The present invention discloses a method for real-time analysis of equipment supervision data. This method collects the vibration time series data of the impeller, combines wavelet packet decomposition and time-frequency analysis to construct a feature matrix. According to the change trends of the blade edge wear degree and the ash deposition time difference, a preliminary anomaly score is calculated. When the score exceeds the threshold, the probability of rotor imbalance is further analyzed, and combined with the finite element analysis results, it is judged whether the imbalance is caused by corrosion deposits or blade surface adhesion. By analyzing the ash deposition distribution characteristics and denoising, the early signs of faults are identified. Finally, considering the material density change rate and the ash layer thickness difference, the fault warning level is determined, and the development trend of the future vibration frequency offset is predicted. The present invention can timely identify the corrosion and ash fouling faults of gas turbine blades, improving the equipment operation reliability and maintenance efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 It is a flowchart of a method for real-time analysis of equipment supervision data of the present invention.

[0033] Figure 2 It is a schematic diagram of a method for real-time analysis of equipment supervision data of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0034] The technical solutions of the present invention will be clearly and completely described below in conjunction with the embodiments. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0035] As Figure 1-2 , a method and system for real-time analysis of equipment supervision data in this embodiment may specifically include:

[0036] S101. During the operation of the gas turbine, collect the vibration time series data of the impeller through sensors and perform preprocessing. Use wavelet packet decomposition technology to perform multi-scale decomposition on the effective data to extract frequency band characteristics, construct a feature matrix reflecting the corrosion depth and ash thickness, and generate a prediction value.

[0037] S1011. During the operation of the gas turbine, collect the vibration time series data of the impeller in real time through high-precision vibration sensors. Set the sampling interval to 0.1 millisecond to ensure the continuity of signal capture. For the data missing caused by interference or equipment jitter during the collection process, use linear interpolation method for compensation processing. Calculate the root mean square value of the time series to reflect the signal energy level, the peak factor to characterize the impact characteristics, and the waveform index to describe the fluctuation degree, and comprehensively generate a waveform integrity score. If the score is lower than 0.85, it is determined that this section of data is invalid and excluded, thereby ensuring the data quality for subsequent analysis.

[0038] S1012. Construct a wavelet packet decomposition tree for the preprocessed effective vibration time series data. Set the decomposition level to four layers and select the db4 wavelet basis function for multi-level decomposition to obtain sub-band signals in sixteen frequency bands. Calculate the energy values of each sub-band signal and their proportions in the total energy. Screen out the characteristic sub-band signals according to the criterion that the proportion exceeds 5%. Perform noise reduction processing on these characteristic sub-band signals using the Bayesian threshold criterion to obtain an optimal threshold range between 0.5 and 2.0. Extract statistical features such as the mean, standard deviation, skewness, and kurtosis of the noise-reduced signal, and construct a feature matrix for characterizing the corrosion pit depth and the ash deposit layer thickness.

[0039] S1013. Perform dimensionality reduction processing on the constructed feature matrix using the principal component analysis method. Retain the principal components with a cumulative contribution rate reaching 90% as the dimensionality-reduced feature vectors. Based on these feature vectors, train a support vector regression model. Select the radial basis function as the kernel function and set the penalty factor to 10 and the epsilon value to 0.1. Optimize the model parameters through cross-validation to control the root mean square error of the validation set within 0.2. Input the dimensionality-reduced feature vectors into the trained model to predict the corrosion pit depth distribution value between 0.5 mm and 2.0 mm and the ash deposit layer thickness difference value between 0.3 mm and 1.5 mm respectively, and verify that the correlation coefficient between the prediction results and the actual measured values reaches 0.92.

[0040] In the embodiment of the present invention, the acquisition and preprocessing of vibration data ensure the effectiveness of the signal, and the application of the wavelet packet decomposition technology realizes the fine extraction of frequency band characteristics. Through the noise reduction and statistical feature analysis of the characteristic sub-band signals, combined with the principal component analysis and the support vector regression model, the corrosion and ash deposit states of the impeller can be accurately predicted. The prediction results show that when running at high speed, the main vibration frequency is concentrated between 500 Hz and 800 Hz, and when running at low speed, it is distributed between 100 Hz and 300 Hz. The corrosion depth is positively correlated with the energy distribution, and the ash deposit thickness significantly affects the characteristics of the high-frequency band. It provides a reliable data basis for the judgment of unbalance faults.

[0041] It can be understood that the embodiment of the present invention does not overly limit the specific sensor type or interpolation algorithm details for data acquisition, which can be adjusted by technicians according to the actual scenario. The noise reduction threshold and the feature extraction method can also be flexibly set according to the operating conditions of the equipment to adapt to the monitoring requirements of different gas turbines. Further analyze the rotor state based on this feature matrix.

[0042] S102. For the sub-band data extracted from the vibration signals of the gas turbine blades through wavelet packet decomposition, calculate the energy values of each frequency band and generate a feature vector reflecting the surface roughness and ash deposit distribution uniformity of the blades. Use smoothing filtering and envelope analysis to optimize the energy matrix, and fuse the statistical features and entropy values to construct a frequency band energy distribution weight sequence.

[0043] S1021. Extract sub-band data from the vibration signals collected from the gas turbine impeller. Use wavelet packet decomposition technology to decompose the signals into four layers to generate sub-band signals in sixteen frequency bands. Calculate the mean value of the spectral energy for each sub-band to form the first frequency band energy matrix. By setting the sampling frequency to 2048 Hz and collecting 60 seconds of vibration data, ensure that the signals contain complete blade operating state information. The calculated mean energy reflects the vibration intensity of each frequency band. Among them, the low-frequency band is associated with the fundamental frequency vibration, the middle frequency band corresponds to the surface state, and the high-frequency band reveals the ash fouling characteristics.

[0044] S1022. Apply a sliding window with a length of 128 points to the first frequency band energy matrix for smoothing filtering. Set the window overlap rate to 50% to reduce random noise interference and improve signal smoothness. After generating the second frequency band energy matrix, perform envelope extraction on it. Obtain the time position and amplitude information of the envelope curve by identifying the local maximum points with an amplitude greater than 0.3. These maximum points characterize the dynamic changes in vibration intensity. For example, it is detected that the energy value in the 500 - 700 Hz frequency band rises from 0.15 to 0.45, indicating a significant enhancement in the vibration characteristics of this frequency band.

[0045] S1023. Construct a frequency band sequence for the local maximum points in chronological order and segment them according to a length of 256 points. Calculate the mean value of each segment to reflect the average level of the signal, the variance to reflect the fluctuation amplitude, and the skewness to describe the distribution asymmetry to generate an initial feature sequence. It is found that the variance in the 500 - 700 Hz frequency band reaches 0.25, indicating severe vibration, while the variance in the 200 - 400 Hz frequency band is only 0.08, showing stability. Through segmental analysis, further reveal the vibration characteristic differences in different frequency bands to provide support for the refined assessment of the blade state. In the embodiment of the present invention, further calculate the variance cumulative amount for the initial feature sequence to quantify the central tendency of the energy distribution. When the cumulative amount exceeds 85%, it indicates that the main vibration characteristics have been captured. At the same time, use the information entropy formula to calculate the local entropy value to measure the signal complexity. For example, a high entropy value indicates a diverse vibration pattern in the 700 - 900 Hz frequency band. Based on this, construct the blade surface roughness feature sequence and the ash fouling distribution uniformity feature sequence respectively. Among them, the roughness feature appears when the energy in the middle frequency band of 300 - 600 Hz increases, and the ash fouling uniformity is prominent when the 700 - 900 Hz frequency band fluctuates in the high-frequency range.

[0046] S1024. The variance cumulant and local entropy value are fused using a linear weighting method to generate a frequency band energy distribution weight sequence. The weight coefficients are optimized to 0.6 and 0.4 by the least squares method to balance the contributions of the two types of features. The fused sequence is processed by zero-mean normalization to eliminate the influence of dimensions. The first three principal components with a cumulative contribution rate of 90% are extracted by principal component analysis to generate an energy feature vector arranged in the order of frequency bands. Among them, the eigenvalue near 450 Hz is the largest and the contribution rate reaches 28%, accurately characterizing the comprehensive state of blade vibration, surface roughness, and ash deposition distribution.

[0047] In the embodiment of the present invention, wavelet packet decomposition refines the vibration signal into sixteen frequency bands. The energy distribution of each frequency band reflects the characteristics of different blade states. For example, the increase in energy in the middle frequency band is related to the increase in roughness, and the fluctuations in the high frequency band correspond to uneven ash deposition. The energy matrix optimized by smoothing filtering and envelope analysis improves the robustness of feature extraction, and the weight sequence generated after fusing statistical features and entropy values comprehensively characterizes the operating characteristics of the blade. In practical applications, when it is detected that the energy in the 450 Hz frequency band rises abnormally, combined with the roughness eigenvalue of 0.85 and the ash deposition uniformity of 0.35, the phenomenon of increased local surface roughness and unbalanced ash deposition distribution of the blade is identified in a timely manner. The advantage of this method is that it enhances the sensitivity to early signs of faults through multi-level analysis, providing a reliable basis for subsequent imbalance assessment. It can be understood that the embodiment of the present invention does not overly limit the specific implementation details of the sliding window length or weight optimization algorithm, which can be adjusted by technicians according to the actual equipment requirements. For example, the window length can be flexibly set according to the signal noise level, and the optimization method can also be replaced by the gradient descent method to improve the calculation efficiency. Based on this weight sequence, the rotor state is further analyzed, and the specific implementation method will be described in detail in other embodiments.

[0048] S103. Real-time speed data is collected through a gas turbine speed sensor and combined with the vibration signal for time-frequency analysis to extract the frequency components and their amplitudes related to the rotor eccentricity distance and frequency offset.

[0049] In the embodiment of the present invention, rotor speed data is obtained during the operation of the gas turbine, and the high-precision speed sensor continuously collects speed values at intervals of 0.1 milliseconds. For example, when the speed increases from 3000 revolutions per minute to 4500 revolutions per minute, the calculated sampling frequency adjustment coefficient is dynamically adjusted from 1.0 to 1.5 to ensure that the sampling frequency is synchronized with the speed change. The original 50 Hz sampling rate is increased to 75 Hz through resampling technology to generate the first speed sequence, so as to more accurately capture the speed fluctuation characteristics.

[0050] S1031. Apply the Hilbert transform to the first rotational speed sequence to calculate the instantaneous phase to generate a second rotational speed sequence. Determine the rotational speed fundamental frequency data based on the phase difference. For example, when the rotational speed is stable at 3600 revolutions per minute, extract a fundamental frequency of 60 Hz. Segment the vibration signal using a 256-point Hanning window and set a 50% window overlap rate to ensure the continuity of time-frequency analysis. Use the short-time Fourier transform to generate a time-frequency spectrum and screen the significant frequency peaks with a three-times root mean square threshold of 0.5 mm / s to obtain the first frequency amplitude data. Among them, the fundamental frequency amplitude is approximately 2.5 mm / s, and the second harmonic amplitude is approximately 1.8 mm / s, reflecting the periodic characteristics of rotor vibration.

[0051] S1032. Perform phase demodulation on the first frequency amplitude data to extract the phase difference information and calculate the initial eccentricity distance value. For example, when the maximum phase difference reaches 45 degrees, the initial value is 0.3 mm. Use the least squares method to linearly correct the initial value to generate corrected eccentricity distance data of 0.28 mm. At the same time, calculate the harmonic ratio, such as the ratio of the second harmonic to the fundamental frequency is 0.72, and the ratio of the third harmonic to the fundamental frequency is 0.36. Determine the harmonic spectrum segmentation threshold of 0.8 through an autocorrelation coefficient greater than 0.85, and generate the second frequency amplitude data to further refine the frequency characteristics.

[0052] S1033. Construct a frequency offset feature vector for the corrected eccentricity distance data and the second frequency amplitude data. Apply wavelet transform for multi-scale decomposition and extract significant offset features from the third-layer coefficients. Use the Kalman filter to dynamically estimate the feature data with a state noise covariance of 0.01 and an observation noise covariance of 0.1 to obtain the frequency component amplitude result. For example, the fundamental frequency amplitude fluctuates between 2.3 and 2.7 mm / s, accurately characterizing the dynamic changes of rotor operation.

[0053] In the embodiments of the present invention, the resampling of the rotational speed data and the Hilbert transform ensure the accuracy of the fundamental frequency extraction, while the short-time Fourier transform and the phase demodulation refine the correlation between the frequency components and the eccentricity distance. For example, when the rotational speed increases from 3000 revolutions per minute to 4200 revolutions per minute, the eccentricity distance increases from 0.25 mm to 0.32 mm, the frequency offset increases from 0 Hz to 2.5 Hz, and the amplitude increases from 2.2 mm / s to 3.1 mm / s. This changing trend clearly reflects the coupling relationship between the rotor vibration and the operating parameters. The combination of the wavelet transform and the Kalman filter further improves the stability and accuracy of the characteristic data, enabling the analysis results to effectively reveal the dynamic characteristics of the rotor under different operating conditions. It can be understood that the embodiments of the present invention do not overly limit the specific settings of the Hanning window length or the filtering parameters, which can be flexibly adjusted by those skilled in the art according to the actual rotational speed range and the signal characteristics. For example, the window length can vary according to the frequency resolution requirements, and the Kalman filter parameters can also be optimized through experiments to adapt to different noise environments. Based on these characteristic data, the rotor imbalance state is further evaluated, and the specific steps will be elaborated in other embodiments.

[0054] S104. Construct a state characteristic matrix by fusing the energy distribution vector and the frequency component amplitude, and combine the real-time monitored impeller center of gravity offset and ash particle density data to analyze the changing trend of the blade edge wear degree and the ash deposition time difference and generate a preliminary anomaly score.

[0055] S1041. According to the energy distribution vector obtained by wavelet packet decomposition and the frequency component amplitude extracted by time-frequency analysis, construct a state characteristic matrix reflecting the vibration characteristics of the gas turbine blade. Use the principal component analysis method to perform dimensionality reduction on the matrix. Compress the original 16-dimensional data to 6 dimensions by retaining the characteristic components with a cumulative contribution rate of 90%, ensuring that the main vibration information is retained while reducing the computational complexity. For example, in the range of 0 - 1000 Hz, the energy ratio of the 200 - 400 Hz frequency band is 35% and that of the 400 - 600 Hz frequency band is 28%. These mid-frequency band characteristics dominate the blade state characterization.

[0056] S1042. Calculate the time-domain statistical parameters for the compressed state matrix using a 256-point sliding window, including the mean reflecting the signal central tendency, the variance measuring the fluctuation amplitude, and the kurtosis characterizing the sharpness, to generate the impeller state parameters. For example, the mean is 0.45, the variance is 0.08, and the kurtosis is 3.2. Use a multi-layer perceptron to perform non-linear feature extraction on these parameters to further explore the deep patterns of the signal and improve the sensitivity of the state parameters to blade anomalies.

[0057] In an embodiment of the present invention, a high-precision displacement sensor is used to collect the impeller center-of-gravity offset data at a sampling frequency of 1000 Hz. The time-domain signal is converted into the frequency domain through Fourier transform. A band-pass filter with a center frequency of 50 Hz and a passband range of 45 - 55 Hz is set to perform noise reduction processing on the spectrum. The calculated offset amplitude is 0.15 mm and the phase difference is 28 degrees. These parameters reveal the vibration offset characteristics caused by blade imbalance. The filtered signal more clearly reflects the correlation between the center-of-gravity offset and the fundamental frequency.

[0058] S1043. The density data collected by the ash particle density sensor is decomposed into 16 sub-bands by using four-layer wavelet packet decomposition. The variance-to-mean ratio of each sub-band signal is calculated to characterize the density distribution parameter. For example, when the ratio of the 8th sub-band reaches 1.3, it indicates uneven ash deposition in the corresponding frequency band. The cross-correlation coefficient calculated with the deposition time series reaches 0.82, showing a strong correlation between the two on a 4-hour scale. The ash deposition parameter is generated by exponentially weighted smoothing of the cross-correlation sequence.

[0059] S1044. The impeller state parameters, center-of-gravity offset parameters, and ash deposition parameters are normalized by the maximum and minimum values and mapped to the 0 - 1 interval. The information gain method is used to calculate the fusion weights. For example, the center-of-gravity offset weight is 0.4, the impeller state weight is 0.35, and the ash deposition weight is 0.25. A vector characterizing the edge wear degree and the ash deposition time difference is generated through weighted combination. For example, the wear degree index is 0.72 and the time difference is 2.5 hours. The generalized extreme value distribution is used for probability density estimation, and the location parameter 0.65 and the scale parameter 0.15 are determined by the maximum likelihood method. The cumulative distribution function value 0.78 is calculated as the preliminary anomaly score.

[0060] In an embodiment of the present invention, the multi-parameter fusion method comprehensively characterizes the blade state through energy distribution, center-of-gravity offset, and ash characteristics. Practice shows that when the score exceeds 0.85, the edge wear depth of the blade often exceeds 0.5 mm and the ash thickness difference is greater than 1.2 mm, which is consistent with the 0.3 mm wear depth and 0.8 mm thickness difference found in actual maintenance. This method not only improves the accuracy of anomaly detection but also enhances the calculation efficiency and robustness through dimensionality reduction and probability estimation, providing reliable support for real-time supervision. It can be understood that the embodiment of the present invention does not strictly limit the window length or the filtering range, and those skilled in the art can adjust according to the equipment working conditions. For example, increasing the window can improve the resolution, or optimizing the weight calculation to adapt to different operating scenarios.

[0061] S105. If the preliminary anomaly score exceeds the preset threshold, then through the comparison and analysis of the current feature matrix and historical operation data, the correlation between the corrosion area ratio and the ash amount on the windward side of the blade is calculated to determine the rotor imbalance probability value.

[0062] In an embodiment of the present invention, when the abnormal score of the operating state of the gas turbine exceeds 0.85, an in-depth analysis process is triggered. The historical operating data is segmented by a sliding window with a length of 512 points, and a 50% window overlap rate is set to ensure data continuity. The Euclidean distance between each segment of historical data and the current state matrix is calculated to generate a first similarity matrix, and the distance value ranges from 0.2 to 1.8. For example, when the distance is less than 0.3, it indicates that the historical data is highly similar to the current state. Subsequently, the matrix is normalized to map the values to the 0-1 interval.

[0063] S1051. The dynamic programming method is used to calculate the optimal matching path for the first similarity matrix to obtain a second similarity matrix. For example, when the path cost value is 0.25, it indicates that the temporal pattern of the current state is highly consistent with the historical data. The autocorrelation coefficient is calculated based on the second similarity matrix. For example, the correlation coefficient with the data 18 months ago reaches 0.92. Based on this, a weight vector is constructed, where the weight of the most similar segment is 0.4 and the weight of the second most similar segment is 0.25. The corrosion area in the historical data is weighted and averaged to generate a first probability sequence, which reflects the probability distribution of the corrosion area evolving over time.

[0064] S1052. For the ash quantity data on the windward side of the blade in the historical data, the exponential weighting method is used for smoothing to weaken the noise impact. The distribution characteristics of the smoothed data are analyzed by Gaussian kernel density estimation to generate a second probability sequence. For example, the ash quantity distribution shows a right-skewed characteristic, and the peak value is 1.0 kg per square meter. After standardizing the first probability sequence and the second probability sequence, a probability mapping function is constructed by applying third-order polynomial regression, and the goodness of fit reaches 0.95. The joint probability value of the corrosion area ratio and the ash quantity distribution is calculated.

[0065] S1053. A log-likelihood function is constructed based on the joint probability value, and the maximum likelihood estimation method is used to calculate the posterior probability distribution of the rotor imbalance state. For example, when the corrosion area ratio is in the range of 15% to 18% and the ash quantity is between 0.9 and 1.1 kg per square meter, the imbalance probability increases significantly. The final posterior probability value is 0.82, and this result is consistent with the case where the rotor vibration amplitude increases by 50% in the historical cases, indicating that corrosion and ash accumulation are the key influencing factors.

[0066] Through multi-dimensional probability analysis, not only can the degree of abnormality of the current state be quantified, but also the fault development trend can be revealed. For example, historical data shows that when the corrosion area ratio increases from 12% to 18% and the ash content increases from 0.8 kg per square meter to 1.2 kg per square meter, the probability of rotor imbalance increases significantly. Gaussian kernel density estimation shows the bimodal characteristics of the corrosion distribution, with the main peak at 15% and the secondary peak at 20%, reflecting the phased development of corrosion. Practice shows that when the imbalance probability exceeds 0.8, the vibration of the unit often intensifies within the next 200 to 300 hours, providing a warning basis for maintenance decisions. It can be understood that the embodiments of the present invention do not fixedly limit the sliding window length or the order of the regression model, and can be adjusted according to the data characteristics. For example, increasing the window can improve the resolution or optimizing the kernel density parameters to adapt to different working conditions, and further judging the cause of the fault based on this probability value.

[0067] S106. According to the rotor imbalance probability value and the blade stress concentration characteristics obtained by finite element analysis, use the support vector machine algorithm to judge whether the imbalance is caused by the thickness of corrosion deposits or surface adhesion force. At the same time, through the spectrum analysis and trend comparison of the ash deposition distribution data, extract the early signs of the fault and classify the abnormal patterns.

[0068] S1061. Construct a blade stress state matrix according to the rotor imbalance probability value, use the tetrahedral mesh generation method to finely divide the blade surface, set the mesh element size to 0.5 mm, and the total number of meshes reaches 520,000. Generate the first stress distribution matrix by calculating the node stress values, and identify that the maximum stress in the root transition region reaches 180 MPa when the probability value is 0.85. Then calculate the stress gradient value and mark the mutation points by the region growing method. For example, the gradient exceeds 12 MPa per mm at 30 mm from the leading edge to the root of the blade, and extract the stress concentration region to generate stress concentration characteristic data, reflecting the local stress abnormality caused by corrosion and ash deposition.

[0069] In the embodiments of the present invention, Hilbert transform is used for spectrum analysis of the stress concentration characteristic data, the instantaneous frequency is calculated using a 200 ms window, and it is found that 80% of the vibration energy is concentrated in the frequency band of 40 - 70 Hz, and the main frequency of 55 Hz coincides with the rotational speed. After determining the frequency band range by the adaptive segmentation method, generate stress spectrum characteristic data, and perform five-layer multi-scale analysis in combination with discrete wavelet decomposition. The variance contribution rates of the third and fourth layers reach 35% and 28% respectively. After screening the characteristic coefficients, construct a multi-scale feature vector, and reduce the dimension to 8 dimensions by kernel principal component analysis. Integrate the data of the sediment thickness from 0.5 to 1.2 mm and the adhesion force from 0.8 to 2.5 MPa to generate a support vector machine training sample matrix.

[0070] S1062. Use a support vector machine with a Gaussian kernel function for feature learning on the sample matrix. Set the kernel parameter to 0.1 and the penalty factor to 10. Optimize the classification boundary through five-fold cross-validation, with an accuracy rate of 93%. When making a determination on the current state data, for example, when the ash accumulation thickness is 1.1 mm and the adhesion force is 1.2 MPa, the classification result points to the imbalance caused by corrosion and deposition, which is consistent with the subsequent maintenance. At the same time, based on the 16×32 distribution data collected by the blade surface ash sensor, decompose the fundamental frequency amplitude of 0.85, the sub-harmonic amplitude of 0.32, and the noise root mean square of 0.15 through Fourier transform, and mark the exceeding standard areas where the sub-harmonic ratio exceeds 0.4 and the noise exceeds 0.12.

[0071] S1063. Use the db4 wavelet basis function to decompose the exceeding standard area into four layers. The energy density of the third layer accounts for 42%. Denoise with a threshold of 0.08 through the soft threshold method, and the noise amplitude is reduced by 85%. The reconstructed signal shows that the periodic change of the leading edge ash accumulation thickness reaches 1.2 mm. Calculate the imbalance index to be 0.72, fit the sequence with the least squares method, obtain the change rate of 0.015 per hour through numerical differentiation, construct a 20-minute time window for comparison with the historical trend, with a similarity of 0.88, and apply K-means clustering to divide it into three categories. The current state is classified as an abnormal mode.

[0072] In the embodiment of the present invention, trend features are extracted through time series correlation analysis and a fault mode library is constructed. For example, when the imbalance increases from 0.65 to 0.78 in only 4 hours and the standard deviation of the leading edge ash accumulation thickness increases from 0.3 mm to 0.8 mm, the similarity with the "ash accumulation accelerated deposition type" mode reaches 0.92. Practical verification shows that this method can issue an early warning at the initial stage of the exacerbation of uneven ash accumulation. For example, when the change rate exceeds 0.012 per hour and the sub-harmonic ratio exceeds 0.42, it gives a 24-hour early warning, which saves time for maintenance and reflects the prediction ability of multi-source data fusion and trend analysis.

[0073] It can be understood that the embodiment of the present invention does not strictly limit the grid size or the number of decomposition layers. Those skilled in the art can adjust according to the blade geometry and working conditions. For example, increasing the grid density can improve the accuracy or optimizing the threshold can adapt to different ash accumulation characteristics, and further optimize the early warning mechanism based on the classification result.

[0074] S107. Construct a state matrix through the classification result and the real-time rotational speed information, calculate the comprehensive influence index of the change rate of the material density and the difference in the ash accumulation layer thickness, and use the multi-scale decomposition and fuzzy logic method to determine the fault early warning level and generate an early warning signal.

[0075] S1071. Construct a state matrix based on the support vector machine classification result label and the real-time speed data collected by the speed sensor. For example, when the speed is 3600 rpm, the spectrum characteristics of material density change are analyzed by Fourier transform. The main frequency is 45 Hz and the amplitude is 0.15. At the same time, the difference operator is used to calculate the thickness change rate of the dust layer. The leading edge area reaches 0.02 mm per hour, which is much higher than the 0.005 mm per hour in other areas. A state feature vector containing density and thickness changes is generated.

[0076] In an embodiment of the present invention, the state eigenvector is normalized to its maximum and minimum values, the density change index is mapped to 0.2 to 0.8, the thickness change index is mapped to 0.3 to 0.9, a 256-point sliding window is set and a 50% overlap rate is maintained, the data is smoothed by the window mean, the standard deviation is reduced from 0.15 to 0.08, and then the smoothed eigenvector is decomposed into three layers of wavelet packets to generate 8 sub-bands, of which the information entropy of the third frequency band reaches 0.35, and the cumulative contribution rate of the first three sub-bands after sorting is 85%, and multi-scale eigenvectors are extracted to reflect the low-medium frequency change characteristics.

[0077] S1072. Calculate the covariance matrix for the multi-scale eigenvectors, extract the principal components through eigenvalue decomposition, the first principal component contribution rate is 45%, the second principal component is 28%, and the third principal component is 15%. Select three dimensions with cumulative contribution rates exceeding 85% to reduce them to 3-dimensional eigenvectors, and divide the warning thresholds through hierarchical clustering. The Euclidean distance less than 0.3 is normal, 0.3 to 0.6 is slightly abnormal, and greater than 0.6 is seriously abnormal. For example, a detection distance of 0.52 is determined to be a slightly abnormal state.

[0078] S1073. Construct a fuzzy membership function for the reduced-dimensional feature vector, use the triangular membership function to set the core range from 0.4 to 0.6, calculate the material density change membership of 0.75 and the dust thickness difference membership of 0.82, assign a density change weight of 0.6 and a thickness difference weight of 0.4 through exponential weighting, calculate the weighted warning score of 0.78 using the geometric mean method, and finally determine the second-level warning and generate a warning signal value of 0.85, reflecting the potential risk of the equipment.

[0079] In an embodiment of the present invention, the method captures the periodic characteristics of material density changes through spectral analysis, and extracts the multi-scale characteristics of thickness differences through wavelet packet decomposition. Fuzzy logic quantifies the parameter coupling effect. For example, when the density change rate is 0.012 per hour and the thickness difference is 0.8 mm during a certain operation, the early warning signal promptly prompts the risk, and the maintenance verifies the compound fault of uneven density and abnormal dust distribution, thereby improving the accuracy of fault identification and maintenance efficiency.

[0080] It is understandable that the embodiments of the present invention do not strictly limit the window length or the form of the membership function, which can be adjusted according to the speed range or fault sensitivity. For example, the window can be extended to improve the smoothing effect or the weights can be optimized to adapt to different working conditions, and the maintenance strategy can be further optimized based on the warning signal.

[0081] S108. According to the change trend of the warning signal and the blade edge wear degree, use the random forest algorithm to predict the development curve of the vibration frequency offset in the future target time, and generate real-time equipment supervision analysis data.

[0082] S1081. Construct a time-state matrix according to the warning signal value, use a 256-point sliding window to segment the blade edge wear degree data at an overlapping rate of 50%, calculate the wear rate change rate through third-order polynomial fitting. For example, when the warning signal is 0.85, the wear rate increases from 0.02 mm per hour to 0.05 mm per hour, generate a wear change matrix, and then apply the exponential weighted moving average method to smooth the data with a weight of 0.3, and the slope increases from 0.015 to 0.038, obtaining a slope feature vector to reflect the trend change.

[0083] In the embodiments of the present invention, for the slope feature vector, perform four-layer wavelet decomposition using the db4 wavelet basis function, select the third layer as the optimal reconstruction scale, denoise by the soft threshold method, retain the trend features and reduce the root mean square error by 85%. The correlation coefficient between the signal and the original data reaches 0.92. After reconstruction, generate a trend feature vector, and calculate its Pearson correlation coefficient with the warning signal of 0.85. Screen the significant features with a correlation coefficient exceeding 0.6, including the wear rate and frequency offset, and construct a prediction feature vector.

[0084] S1082. Perform standardization processing on the prediction feature vector, construct a predictor through the random forest algorithm, optimize the parameters using ten-fold cross-validation, determine the number of trees as 100 and the feature sampling ratio as 0.8, the prediction accuracy reaches 91%, the root mean square error is lower than 0.15. According to the model parameters, predict that the frequency offset in the next 24 hours will increase from 2.5 Hz to 3.8 Hz, use cubic spline interpolation to smooth the curve at 1-hour nodes, and the interval difference remains 0.5 Hz under a 95% confidence level, generating a prediction curve vector.

[0085] S1083. Perform piecewise linear regression on the prediction curve vector, identify the inflection points where the slope suddenly changes from 0.08 Hz per hour to 0.10 Hz per hour at 8 hours and increases to 0.15 Hz per hour at 16 hours, obtain a monitoring data matrix, reflecting the accelerated stage of state deterioration. For example, when the wear degree of a certain unit is 0.45 mm and the offset is 2.2 Hz, predict that it will reach 3.5 Hz after 24 hours, and the actual measurement is 3.6 Hz, with an error of 2.8%, verifying the reliability of the prediction.

[0086] In an embodiment of the present invention, the method improves data quality through smoothing and noise reduction. Random forests combined with multi-parameter features capture trend changes. Practice shows that when the warning signal exceeds 0.8 and the wear rate exceeds 0.04 millimeters per hour, the frequency offset often increases significantly within 24 hours. After inspection and confirmation, the wear depth reaches 0.85 millimeters.

[0087] This is consistent with the prediction and provides an accurate basis for maintenance decisions. It can be understood that the embodiments of the present invention do not fixedly limit the window length or the number of trees, which can be adjusted according to the data characteristics. For example, increasing the window can improve the resolution or optimizing the sampling ratio to adapt to different prediction periods, further improving the supervision strategy.

[0088] The present invention provides a real-time analysis system for equipment supervision data, mainly including:

[0089] A wavelet packet decomposition module, which is used to collect continuous vibration time series data of the impeller during the operation of the gas turbine through a sensor, and perform multi-scale decomposition on the corrosion pit depth distribution and the difference in the ash deposit layer thickness by using wavelet packet decomposition to obtain sub-band signals in different frequency bands;

[0090] An energy feature extraction module, which is used to calculate the energy values of the corresponding frequency bands of the sub-band signals, determine the energy features of the blade surface roughness and the ash deposit distribution uniformity in each sub-band in the form of the sum of squares, and generate an energy distribution vector arranged in the order of frequency bands;

[0091] A time-frequency analysis module, which is used to perform time-frequency analysis on the vibration signal by using the obtained real-time rotational speed information of the gas turbine, identify the frequency components corresponding to the rotor eccentricity distance value and the vibration frequency offset amount, and determine the amplitude of the frequency components;

[0092] A feature matrix construction module, which is used to construct a feature matrix based on the energy distribution vector and the amplitude of the frequency components, analyze the change trends of the blade edge wear degree and the difference in the ash deposit time according to the impeller center of gravity offset amount and the ash particle density value obtained by real-time monitoring, and obtain a preliminary anomaly score;

[0093] A preliminary anomaly scoring module, which is used to, if the preliminary anomaly score exceeds a preset threshold, calculate the correlation coefficients of the current matrix with the corrosion area ratio and the ash amount on the windward side of the blade by comparing the feature matrix with historical operation data, and determine the rotor imbalance probability value;

[0094] A rotor imbalance probability calculation module, which is used to, according to the rotor imbalance probability value, combine the local stress concentration coefficient obtained by finite element analysis simulation of the blade, and use the support vector machine algorithm to determine whether the current operating state is unbalanced due to the corrosion deposit thickness or the blade surface adhesion force, and obtain a classification result;

[0095] A fault warning level determination module, which is used to calculate the comprehensive influence index of the material density change rate and the difference in the ash accumulation layer thickness through the classification result and the real-time rotational speed information, determine the fault warning level by using the weighted average method, and generate a warning signal;

[0096] A vibration frequency offset prediction module, which is used to predict the development curve of the vibration frequency offset within the future target time by using the random forest algorithm according to the warning signal and the change trend of the blade edge wear degree, and obtain the real-time equipment supervision analysis data.

[0097] As mentioned above, the above embodiments are only used to illustrate the technical solutions of the present application, rather than limiting them; although the present application has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A real-time analysis method for equipment supervision data, characterized in that: The method comprises: The continuous vibration time series data of the impeller during the operation of the gas turbine is collected by sensors, and wavelet packet decomposition is used to perform multi-scale decomposition on the corrosion pit depth distribution and the thickness difference of the dust layer to obtain sub-band signals in different frequency bands. Calculate the energy value of the frequency band corresponding to the sub-band signal, determine the energy characteristics of the blade surface roughness and dust accumulation distribution uniformity in each sub-band in the form of square sum, and generate an energy distribution vector arranged in the order of frequency bands; The obtained real-time speed information of the gas turbine is used to perform time-frequency analysis on the vibration signal, identify the frequency components corresponding to the rotor eccentricity distance value and the vibration frequency offset, and determine the amplitude of the frequency components; The characteristic matrix is ​​constructed based on the energy distribution vector and the frequency component amplitude. The change trend of blade edge wear and dust deposition time difference is analyzed by combining the impeller gravity center offset and dust particle density value obtained by real-time monitoring, and the preliminary abnormality score is obtained. If the preliminary abnormality score exceeds the preset threshold, the correlation coefficient between the current matrix and the area ratio of the corrosion area and the amount of ash on the windward side of the blade is calculated by comparing the characteristic matrix with the historical operation data to determine the rotor imbalance probability value; According to the rotor imbalance probability value and the local stress concentration coefficient obtained by finite element analysis simulation of the blade, the support vector machine algorithm is used to determine whether the current operating state is unbalanced due to the thickness of corrosion deposits or the adhesion of the blade surface, and the classification result is obtained; The comprehensive impact index of the material density change rate and the dust layer thickness difference is calculated through the classification results and real-time speed information, and the fault warning level is determined by the weighted average method to generate a warning signal; According to the changing trend of the early warning signal and the blade edge wear, the random forest algorithm is used to predict the development curve of the vibration frequency offset within the future target time to obtain real-time equipment supervision and analysis data.

2. The method according to claim 1, characterized in that The continuous vibration time series data of the impeller during the operation of the gas turbine is collected by the sensor, and the multi-scale decomposition is performed on the corrosion pit depth distribution and the thickness difference of the dust layer by wavelet packet decomposition to obtain sub-band signals of different frequency bands, including: Acquire vibration time series data collected by a gas turbine impeller sensor, and judge the time series data as invalid data if the waveform integrity score is less than a preset threshold; Construct a wavelet packet decomposition tree for the compensated valid time series data, perform multi-level decomposition according to the set number of decomposition layers, and obtain several sub-band signals; The energy value and energy proportion of each sub-band signal are calculated, the sub-band signals are sorted according to their importance according to their energy proportion, and the sub-band signals whose energy proportion exceeds the preset threshold are selected for wavelet threshold denoising to obtain the denoised sub-band signals.

3. The method according to claim 1, characterized in that The energy value of the frequency band corresponding to the sub-band signal is calculated, and the energy characteristics of the blade surface roughness and dust accumulation distribution uniformity in each sub-band are determined in the form of square sum, and the energy distribution vector arranged in the frequency band order is generated, including: Decomposing the gas turbine blade vibration signal in four layers according to wavelet packet transform, calculating the spectrum energy mean of the vibration signal, and obtaining a first frequency band energy matrix; Performing smoothing filtering on the first frequency band energy matrix through a sliding window to obtain a second frequency band energy matrix, performing envelope extraction on the second frequency band energy matrix to obtain a local maximum point of the envelope curve; Constructing a frequency band sequence according to the local maximum point, calculating the mean, variance and skewness of the frequency band sequence, and obtaining an initial feature sequence; The variance accumulation amount and the local entropy value are calculated for the initial feature sequence, and the variance accumulation amount and the local entropy value are fused by using a linear weighting method to obtain a frequency band energy distribution weight sequence.

4. The method according to claim 1, characterized in that The method uses the acquired real-time speed information of the gas turbine to perform time-frequency analysis on the vibration signal, identifies the frequency component corresponding to the rotor eccentricity distance value and the vibration frequency offset, and determines the amplitude of the frequency component, including: Acquire real-time rotor speed data from a gas turbine speed sensor, calculate a sampling frequency adjustment coefficient according to the speed data, and resample the speed data using the sampling frequency adjustment coefficient to obtain a first speed sequence; Performing a Hilbert transform on the first speed sequence, calculating an instantaneous phase through the Hilbert transform, and obtaining a second speed sequence according to the instantaneous phase; The second speed sequence is segmented by using a Hanning window of a preset length, a short-time Fourier transform is performed according to the segmented sequence, and first frequency amplitude data is obtained by triple root mean square threshold detection; Performing phase demodulation on the first frequency amplitude data, calculating an initial value of the eccentric distance, and performing linear correction on the initial value of the eccentric distance using a least square method to obtain corrected eccentric distance data; The frequency offset feature vector is constructed for the corrected eccentricity distance data, the frequency offset features are extracted by wavelet transform, and the Kalman filter is used to dynamically estimate the feature data to obtain the frequency component amplitude.

5. The method according to claim 1, characterized in that: The characteristic matrix is ​​constructed according to the energy distribution vector and the frequency component amplitude, and the change trend of blade edge wear and dust deposition time difference is analyzed by combining the impeller gravity center offset and dust particle density value obtained by real-time monitoring, and a preliminary abnormality score is obtained, including: The state characteristic matrix is ​​constructed according to the energy distribution vector and the frequency component amplitude, and the state characteristic matrix is ​​subjected to dimension reduction processing by the principal component analysis method, and then feature extraction is performed by a multi-layer perceptron to obtain the impeller state parameters; Band-pass filtering is performed on the impeller state parameters and the center of gravity offset data, and the sub-band variance mean ratio is calculated for the filtered data using wavelet packet decomposition to obtain density distribution parameters; A fusion weight vector is constructed according to the density distribution parameters, and the probability density of the fusion weight vector is estimated using generalized extreme value distribution. The location parameter and scale parameter are determined by maximum likelihood method, and the cumulative distribution function value is calculated as the preliminary anomaly score.

6. The method according to claim 1, characterized in that If the preliminary abnormality score exceeds the preset threshold, the correlation coefficient between the current matrix and the corrosion area ratio and the amount of ash on the windward side of the blade is calculated by comparing the characteristic matrix with the historical operation data to determine the rotor imbalance probability value, including: The historical operation data is segmented by a sliding window method, and the Euclidean distance is calculated based on the segmented data and the feature matrix to obtain a first similarity matrix; The first similarity matrix is ​​normalized, and the optimal path distance is calculated by using a dynamic programming method to obtain a second similarity matrix; Calculating an autocorrelation coefficient according to the second similarity matrix, constructing a weight vector for the autocorrelation coefficient, performing weighted average processing on the corrosion area in the historical operation data, and obtaining a first probability sequence; Gaussian kernel density estimation is used to perform joint probability calculation on the first probability sequence and the gray distribution data, and the rotor imbalance probability value is obtained through the log-likelihood function and maximum likelihood estimation.

7. The method according to claim 1, characterized in that According to the rotor imbalance probability value, combined with the local stress concentration coefficient obtained by finite element analysis simulation of the blade, the support vector machine algorithm is used to determine whether the current operating state is unbalanced due to the thickness of corrosion deposits or the adhesion of the blade surface, and the classification results are obtained, including: The blade stress state matrix is ​​constructed according to the rotor imbalance probability value, and the blade surface is meshed by a tetrahedral meshing method to obtain a first stress distribution matrix; Calculating stress gradient values ​​for the first stress distribution matrix, marking stress gradient mutation points by using a regional growing method, extracting stress concentration areas according to the distribution of the mutation points, and obtaining stress concentration feature data; Using Hilbert transform to perform spectrum analysis on the stress concentration characteristic data, determining the frequency band range through an adaptive segmentation method, and obtaining stress spectrum characteristic data; A support vector machine with a Gaussian kernel function is used to perform feature learning on the stress spectrum feature data, and a category determination is performed on the current state data according to the classification decision boundary to obtain a corrosion deposition imbalance classification label; It also includes: collecting blade dust distribution data to obtain initial distribution characteristics, analyzing the initial distribution characteristics to determine subharmonic and high-frequency noise components, if the subharmonic and high-frequency noise exceed a preset threshold, decomposing the signal through wavelet transform to obtain denoised characteristics, calculating the degree of rotor imbalance and judging its dynamic change trend, comparing with historical trends to determine abnormal distribution patterns, classifying abnormal distribution patterns, and obtaining early signs of failure.

8. The method according to claim 1, characterized in that: The method calculates the comprehensive impact index of the material density change rate and the dust layer thickness difference through the classification results and the real-time rotation speed information, determines the fault warning level by using the weighted average method, and generates a warning signal, including: A state matrix is ​​constructed according to the classification result label and the real-time rotation speed data, and a frequency spectrum feature of the material density change in the state matrix is ​​calculated by Fourier transform to obtain a state feature vector; Performing maximum and minimum value normalization processing on the state feature vector, and smoothing the normalized state feature vector using a window mean to obtain a smoothed feature vector; Performing three-layer wavelet packet decomposition on the smoothed feature vector, using information entropy to calculate the contribution value of each sub-band signal after the wavelet packet decomposition, sorting the sub-band signals according to the contribution value, and obtaining a multi-scale feature vector; A fuzzy membership function is constructed for the multi-scale feature vector, the membership value of each warning indicator is calculated using the triangle membership, the fault warning level is determined according to the membership value, and a warning signal is generated.

9. The method according to claim 1, characterized in that: According to the change trend of the early warning signal and the blade edge wear, the random forest algorithm is used to predict the development curve of the vibration frequency offset within the future target time to obtain real-time equipment supervision and analysis data, including: According to the blade edge wear data, a sliding window is used to perform segmented processing, and the wear change rate is calculated by third-order polynomial fitting to obtain a wear change matrix; The exponentially weighted moving average method is used to perform data smoothing on the wear change matrix, and the slope of the change trend is calculated based on the smoothed data to obtain a slope characteristic vector; Performing wavelet decomposition on the slope feature vector, performing noise reduction processing on the wavelet coefficients by using a soft threshold method, and obtaining a trend feature vector by reconstructing the signal through an optimal scale; A random forest predictor is constructed according to the trend feature vector, and the predictor parameter value is determined by cross validation to obtain real-time equipment supervision and analysis data.

10. A real-time analysis system for equipment supervision data, characterized in that: The system comprises: The wavelet packet decomposition module is used to collect the continuous vibration time series data of the impeller during the operation of the gas turbine through sensors, and uses wavelet packet decomposition to perform multi-scale decomposition on the corrosion pit depth distribution and the thickness difference of the dust layer to obtain sub-band signals of different frequency bands; The energy feature extraction module is used to calculate the energy value of the frequency band corresponding to the sub-band signal, determine the energy characteristics of the blade surface roughness and dust accumulation distribution uniformity in each sub-band in the form of square sum, and generate an energy distribution vector arranged in the order of frequency bands; The time-frequency analysis module is used to perform time-frequency analysis on the vibration signal using the acquired real-time speed information of the gas turbine, identify the frequency component corresponding to the rotor eccentricity distance value and the vibration frequency offset, and determine the amplitude of the frequency component; The feature matrix construction module is used to construct a feature matrix based on the energy distribution vector and the frequency component amplitude, and analyze the change trend of blade edge wear and dust deposition time difference by combining the impeller gravity center offset and dust particle density value obtained by real-time monitoring, and obtain a preliminary abnormality score; The preliminary anomaly scoring module is used to compare the feature matrix with the historical operation data, calculate the correlation coefficient between the current matrix and the corrosion area ratio and the ash amount on the windward side of the blade, and determine the rotor imbalance probability value if the preliminary anomaly score exceeds the preset threshold; The rotor imbalance probability calculation module is used to determine whether the current operating state is unbalanced due to the thickness of corrosion deposits or the adhesion of the blade surface based on the rotor imbalance probability value and the local stress concentration factor obtained by finite element analysis simulation of the blade, and obtain the classification result by using the support vector machine algorithm; The fault warning level determination module is used to calculate the comprehensive impact index of the material density change rate and the dust accumulation layer thickness difference through the classification results and real-time rotation speed information, determine the fault warning level by weighted average method, and generate a warning signal; The vibration frequency offset prediction module is used to predict the development curve of the vibration frequency offset within the future target time based on the changing trend of the early warning signal and the blade edge wear, and obtain real-time equipment supervision and analysis data using the random forest algorithm.

Images