Real-time state data processing method for alternating current variable pitch system

Abstract

This invention discloses a real-time status data processing method for an AC pitch system, specifically relating to the field of data processing. The method includes: S1, parallel acquisition of multi-dimensional data from the pitch system, dynamically matching interpolation orders of different precision based on the comprehensive fluctuation level of real-time operating conditions, and performing time-base resampling alignment on heterogeneous data with different sampling rates; S2, converting speed-related time-domain signals into equally spaced angle-domain signals based on rotor position information, parallel solving of deep fault feature parameters across multiple physical dimensions, and obtaining feature data representing the operating status of multiple components. This invention employs a multi-order interpolation synchronization alignment method with adaptive operating condition change rate, solving the problems of large synchronization errors and redundant computing power in steady-state operation caused by existing fixed-order interpolation methods during sudden changes in operating conditions. It reduces overall computing power consumption while improving synchronization accuracy under rapidly changing operating conditions, providing a reliable time base for multi-source feature fusion.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and more specifically, to a method for processing real-time status data of an AC pitch system. Background Technology

[0002] In industrial data acquisition and processing, big data processing technology for the operational status of new energy equipment is an important application branch of industrial-specific data processing technology in wind power operation and maintenance scenarios. For monitoring the operational status of wind power AC pitch systems, existing solutions mostly employ general digital signal processing frameworks to complete the acquisition and preprocessing of multi-source heterogeneous data. Through basic data processing methods such as time-domain feature statistics, frequency-domain Fourier transform, envelope demodulation, symmetric component decomposition, and threshold discrimination, basic operational status assessment and fault detection functions are achieved. Overall, this represents a standardized application of conventional industrial data processing methods in equipment operation and maintenance scenarios.

[0003] Existing data processing methods possess advantages such as mature algorithms and ease of deployment, enabling the identification of relatively obvious faults under steady-state conditions and meeting the basic condition monitoring requirements of pitch systems. However, they still have shortcomings in practical applications: First, multi-source data synchronization and alignment often employs general interpolation algorithms of fixed order, failing to fully consider the impact of pitch system operating condition fluctuations on interpolation accuracy. Synchronization errors increase significantly during sudden changes in operating conditions, making it difficult to provide a reliable time reference for subsequent fusion analysis. Simultaneously, feature extraction often uses general signal processing algorithms in the time or frequency domains, failing to fully consider the characteristics of pitch systems operating at varying speeds. Speed ​​fluctuations can lead to spectral broadening and feature amplitude distortion, easily masking early, weak fault characteristics. Second, feature normalization often employs fixed thresholds or simple linear lookup table methods, lacking consideration for the multi-tasking characteristics of pitch systems. The nonlinear characteristics of the fault range are partitioned for adaptation, resulting in insufficient normalization accuracy across the entire operating range. Furthermore, multi-feature fusion often adopts general fusion models based on direct weighting or root mean square, lacking computational logic that incorporates the physical propagation path of the fault, leading to low matching degree between the fusion results and the actual evolution law of the fault. Thirdly, cross-domain cross-verification is based solely on numerical synthesis of feature amplitudes, failing to fully consider the propagation time sequence characteristics of the fault in different physical domains, resulting in limited misjudgment suppression. In addition, fault trend prediction often uses linear fitting methods with fixed windows, making it difficult to distinguish between the stable evolution and abrupt evolution stages of the fault, resulting in insufficient accuracy in evolution rate calculation. At the same time, data storage scheduling often adopts passive threshold-triggered hierarchical strategies, lacking forward-looking scheduling that incorporates the fault development trend, leading to low utilization efficiency of storage and bandwidth resources. Summary of the Invention

[0004] To overcome the aforementioned deficiencies in the prior art, this invention proposes a real-time status data processing method for AC pitch systems. Targeted optimizations and improvements are made from multiple dimensions, including synchronization alignment, feature calculation, operating condition normalization, fault fusion, cross-verification, trend calculation, and data scheduling. Each improvement is adapted to the operating characteristics of the pitch system, and the overall collaborative approach achieves high-precision, highly adaptable, and highly reliable pitch status data processing.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for processing real-time status data of an AC pitch system, comprising: S1. Parallel acquisition of multi-dimensional data from the pitch system, dynamic matching of interpolation orders with different precision based on the comprehensive fluctuation of real-time operating conditions, and time base resampling and alignment of heterogeneous data with different sampling rates to generate a unified time series synchronous dataset. S2. Based on rotor position information, convert the rotational speed-related time-domain signal into an angle-domain equally spaced signal, and solve the deep fault feature parameters of multiple physical dimensions in parallel to obtain feature data that characterizes the operating status of multiple components. S3. Call the corresponding nonlinear baseline model according to the working condition interval division rules, correct the real-time baseline value in combination with the dynamic change rate of the working condition, perform relative deviation normalization analysis on each characteristic parameter, and obtain normalized deviation data to eliminate working condition interference. S4. Divide the fault physical propagation path into multi-level propagation nodes, obtain the single-level feature coupling degree step by step according to the node level and perform cascade synthesis to obtain cascade coupling degree data that characterizes the severity of the fault. S5. For dual-domain associated faults, analyze the fault coupling degree of the electrical domain and the mechanical domain respectively, and combine the temporal synchronization degree of abnormal features to correct the amplitude cross-verification results to obtain confidence data characterizing the reliability of the fault. S6. Using dual sliding windows of different lengths to fit the long-term evolution trend and short-term abrupt change characteristics of the fault respectively, and combining change point detection to dynamically adjust the short window length, the comprehensive fault evolution rate is synthesized and the fault development trend is predicted. S7. Based on fault confidence and comprehensive evolution rate, obtain storage priority, and combine fault trend prediction results to perform forward-looking sampling, storage and upload bandwidth pre-scheduling to achieve differentiated hierarchical data management.

[0006] Preferably, the multidimensional data includes five types of raw state data and two types of operating condition data. The five types of raw state data are: three-phase instantaneous current of the pitch motor, DC bus voltage of the pitch driver, vibration acceleration of the pitch reducer input housing, acoustic emission signal of the outer ring of the pitch bearing, and rotor position angle of the pitch motor. The two types of operating condition data are: pitch angle at the current moment. and the current speed of the pitch motor The degree of real-time operating condition fluctuation is quantitatively characterized by the overall operating condition change rate: ,in, This represents the overall rate of change in operating conditions. This is the rated pitch angle of the pitch control system; This is the first derivative of the pitch angle with respect to time, i.e., the rate of change of the pitch angle; This refers to the rated speed of the pitch motor; The first derivative of motor speed with respect to time is the rate of change of speed. The interpolation order is divided into three accuracy levels: first-order linear interpolation is used for stable operating conditions, second-order parabolic interpolation is used for gradually changing operating conditions, and third-order Lagrange interpolation is used for rapidly changing operating conditions. Hysteresis margin is set for the switching threshold of both order levels. When the comprehensive rate of change of operating conditions rises to the gradually changing threshold, it switches to second-order interpolation, and when it rises to the rapidly changing threshold, it switches to third-order interpolation. When the comprehensive rate of change of operating conditions decreases, it must fall below the corresponding threshold minus the hysteresis margin before switching back to the lower order interpolation. The time reference resampling alignment uses the time series of the current acquisition channel as the reference time axis, and uses the matched interpolation order to resample the data of other channels with different sampling rates.

[0007] Preferably, the speed-correlated time-domain signals include three-phase current signals and vibration envelope signals. The characteristic amplitudes of both types of signals are affected by speed fluctuations, resulting in spectral broadening and distortion. Angle-domain conversion is needed to eliminate speed interference. The angle-domain equally spaced signals are generated through equal-angle resampling. Based on the rotor position angle, the time-domain uniformly sampled signals are converted into signals with uniform angle intervals, eliminating the spectral broadening effect of speed fluctuations and restoring the true amplitude of the fault characteristics. The multi-physical-dimensional deep fault characteristic parameters are solved in parallel across four dimensions. The four solutions are executed independently and synchronously to improve real-time processing: the first dimension is the stator current frequency domain and symmetrical component feature solution. Based on the synchronized instantaneous values ​​of the three-phase current, the signals are first converted into angle-domain signals, and then a fast Fourier transform is performed to extract the harmonic amplitudes at the fundamental frequency order ± gear meshing order, obtaining the fundamental frequency sideband harmonic amplitude of the pitch motor U-phase current. Simultaneously, symmetrical component decomposition is performed to obtain the amplitude of the negative sequence component of the three-phase composite current. ,in m is the integer index of the sampling point in the angle domain; is the instantaneous value of the U-phase current at point m after resampling in the angle domain; M is the number of sampling points in a single electrical cycle within the angle domain; The order of the sideband harmonics is determined by the number of meshing teeth in the reducer gears and the number of pole pairs in the motor; j is the imaginary unit; the second dimension is the DC bus ripple characteristic calculation, which obtains the peak-to-peak value of the 300Hz ripple of the pitch driver DC bus based on the synchronized DC bus voltage data. The third dimension involves the calculation of vibration envelope and acoustic emission characteristics. Based on the synchronized vibration acceleration data, a 5kHz~15kHz bandpass filter is first performed, and then Hilbert demodulation is performed on the filtered signal to obtain the vibration envelope time-domain signal. Combined with the rotor position angle, the vibration envelope time-domain signal is resampled into an angle-domain equally spaced signal. The peak amplitude of the envelope signal in the angle domain is extracted to obtain the demodulated amplitude of the vibration envelope of the pitch reducer input housing. Acoustic emission signal energy rate at the fixed end of the outer ring of the pitch bearing The fourth dimension involves solving the rotor flux cross-axis component characteristics, and obtaining the estimated value of the rotor flux cross-axis component of the pitch motor in real time. The specific feature data representing the operating status of the various components are as follows: These correspond to the gear meshing state of the reducer, the symmetrical state of the stator winding, the aging state of the bus capacitor, the impact state of the reducer bearing, the crack propagation state of the bearing raceway, and the orientation of the rotor magnetic field and the state of the permanent magnet.

[0008] Preferably, the operating condition interval division rule divides the entire operating condition range of the pitch system into three operating condition intervals: low-speed small pitch, medium-speed medium pitch, and high-speed large pitch. If the operating condition point falls on the boundary between two intervals, baseline calculation is performed according to the higher-level interval. The nonlinear baseline model is a second-order polynomial fitting model corresponding to each interval. The model coefficients are obtained by fitting the historical normal operating data within each operating condition interval, respectively adapting to the characteristic change patterns of different intervals. The real-time baseline value is obtained by superimposing a dynamic correction term on the static baseline fitting value. The static baseline fitting value is calculated by the nonlinear baseline model of the corresponding operating condition interval, and the dynamic correction term is the product of the dynamic operating condition correction coefficient and the comprehensive change rate of the operating condition. The relative deviation normalization analysis uses the real-time baseline value as a benchmark, converting the measured values ​​of each characteristic parameter into relative deviation values ​​relative to the real-time baseline, eliminating the dual interference of steady-state differences and dynamic fluctuations in the operating condition. ,in, This represents the normalized deviation value of the k-th feature parameter; is the measured solution value of the k-th feature parameter; z is the index of the interval to which the current working condition belongs; This represents the static baseline fitting value of the kth feature parameter within the zth operating condition interval; This is the dynamic operating condition correction coefficient for the k-th characteristic parameter.

[0009] Preferably, the physical propagation path of the fault is derived based on the actual evolution mechanism of five types of faults: gear meshing fault, bearing fault, stator winding fault, bus capacitor fault, and permanent magnet demagnetization fault. Gear meshing faults are divided into three levels: gear damage node, load pulsation node, and current modulation node; bearing faults are divided into two levels: raceway damage node and vibration impact node; stator winding faults are divided into two levels: winding impedance change node and magnetic field distortion node; bus capacitor faults are divided into one level: capacitance attenuation node and ripple increase node; and permanent magnet demagnetization faults are divided into two levels: flux attenuation node and flux linkage offset node. The multi-level propagation nodes are the characteristic association nodes of each fault type along the physical propagation path, with each level node corresponding to a set of associated characteristic parameters. The single-level feature coupling degree is calculated using the root mean square algorithm, fusing the normalized deviation values ​​of all associated features within a single node. The cascade synthesis uses a geometric average method, multiplying the single-level coupling degree level by level along the propagation path and then raising the order root to obtain the final cascade coupling degree. ,in, Let f be the cascade coupling degree of the type f fault; is the total number of levels of the fault propagation path corresponding to the type f fault; s is the level number of the fault propagation path; The number of feature parameters corresponding to the s-th level propagation node of the f-th type of fault; This is the normalized deviation value of the kth associated feature parameter within the corresponding node; the cascade coupling degree data covers all five types of faults and is output to the subsequent confidence analysis and trend prediction stages.

[0010] Preferably, the dual-domain associated faults only include gear meshing faults, which are associated with both mechanical and electrical domain features. The other four types of faults are single-domain faults, with bearing faults being purely mechanical domain faults, and stator winding faults, bus capacitor faults, and permanent magnet demagnetization faults being purely electrical domain faults. Single-domain faults are assigned a base confidence level of 0.6, representing evidence supported only by a single dimension, thus having limited credibility. The fault coupling degree between the electrical and mechanical domains is obtained based on cascaded coupling degree data. The electrical domain coupling degree extracts the cascaded calculation results of the fault's electrical propagation path, and the mechanical domain coupling degree extracts the cascaded calculation results of the fault's mechanical propagation path. The temporal synchronization degree of the abnormal features is quantified using the Pearson correlation coefficient to obtain the temporal correlation coefficient of the fault coupling degree sequence between the electrical and mechanical domains. If the calculated Pearson correlation coefficient is negative, it is determined that the two-domain anomaly has no synchronous evolutionary characteristics, and the value is taken as... The final value range is [0,1]; the amplitude cross-verification result is quantified by the amplitude consistency coefficient, calculated based on the matching degree of the two-domain coupling degree. The closer the amplitudes of the two-domain coupling degree are, the higher the consistency; the final cross-verification confidence is obtained after correction by the degree of time synchronization: ,in, The cross-confirmation confidence level for type f faults; The electrical-mechanical characteristic time-series correlation coefficient corresponding to the f-th type of fault; The electrical domain characteristic coupling degree corresponding to the f-th type of fault; The mechanical domain feature coupling degree corresponds to the type f fault. When the timing is completely synchronized and the amplitudes of the two domains are completely equal, the confidence level takes the maximum value of 1, which is significantly higher than the basic confidence level of the single-domain fault. The confidence level data that characterizes the credibility of the fault covers all five types of faults. After sorting based on the confidence level values, the fault type identification result and the corresponding confidence level are output and transmitted to the subsequent data scheduling stage.

[0011] Preferably, the two sliding windows of different lengths are a long trend window and a short acceleration window, respectively. The two windows slide and update synchronously, continuously storing the coupling degree time series data of each type of fault within the window. The long-term evolution trend and short-term mutation characteristics of the fault are obtained by least-squares linear fitting within the long and short windows, respectively, and the long-term evolution rate of the fault corresponding to the long trend window is output. Short-term fault evolution rate corresponding to short acceleration window The change point detection employs a sliding t-test algorithm, using time-series data of fault coupling degree within a short window as input to test the significance of the difference in means between two equal-length sub-windows. When the test statistic exceeds a preset significance threshold, a fault mutation node is identified, and the short window length is shortened to half its original length to enhance mutation capture capability. If no mutation node is detected within three consecutive short window periods, the short window length is restored to its original set length. The comprehensive fault evolution rate is calculated using a vector synthesis method, taking into account both long-term trends and short-term acceleration characteristics. ,in, The comprehensive evolution rate of the coupling degree of the f-th type of fault is given. The predicted fault development trend is calculated based on the comprehensive evolution rate and the current fault coupling degree. The remaining time for the fault characteristics to reach the warning threshold is calculated, and the fault development trend prediction result and the remaining time data are finally output.

[0012] Preferably, the storage priority is calculated based on the combined evolution rate of the current highest fault confidence and the corresponding fault, and normalized using a baseline evolution rate as a reference: ,in, This is the storage priority value for the current state data; Let f be the cross-confirmation confidence level for the type f fault, and take the maximum value among all fault categories when substituting into the formula; This represents the overall evolution rate of the corresponding fault; The system uses a preset baseline evolution rate reference value, which is the background fluctuation rate of fault coupling under normal operating conditions, obtained through historical normal data statistics. The fault trend prediction result uses the remaining time for the fault to reach the warning threshold output by S6 as the triggering basis for pre-scheduling. The forward-looking sampling, storage, and upload bandwidth pre-scheduling execution rules are as follows: when the remaining time of the fault is less than the preset pre-scheduling threshold, the sampling density and storage priority of the corresponding fault-related data channel are increased in advance, and cloud upload bandwidth is pre-allocated; when the remaining time is greater than the pre-scheduling threshold, the normal sampling and storage levels are maintained. The differentiated hierarchical data management uses the baseline priority value under normal fault-free operating conditions as a reference, and is divided into three levels according to the storage priority value: more than twice the baseline priority is high priority, corresponding to a state with high fault confidence and fast evolution rate, and full original data storage and real-time upload are executed; 1 to 2 times the baseline priority is medium priority, corresponding to a state with medium fault confidence and slow evolution rate, and only feature values ​​and fault identification results are stored and uploaded periodically; less than 1 times the baseline priority is low priority, corresponding to a state with no obvious fault characteristics, and only operating condition and health status labels are stored and uploaded periodically.

[0013] The technical effects and advantages of this invention are as follows: 1. This invention adopts a multi-order interpolation synchronization alignment method with adaptive operating condition change rate. It dynamically matches the interpolation order according to the degree of fluctuation of the pitch system operating condition and avoids frequent order oscillation by combining a two-level hysteresis switching mechanism. It solves the problems of large synchronization error and redundant computing power in steady state when the existing fixed order interpolation changes suddenly. While reducing the overall computing power consumption, it improves the synchronization accuracy under drastic operating conditions and provides a reliable time reference for multi-source feature fusion. 2. This invention adopts a rotational speed tracking angle domain feature calculation method. It performs angle domain transformation on rotational speed-related feature quantities such as current and vibration envelope before performing frequency domain analysis. This solves the problems of spectrum broadening and feature amplitude distortion in existing general time domain algorithms under variable speed scenarios, effectively eliminates the interference of rotational speed fluctuations on feature extraction, and improves the sensitivity of early weak fault feature identification. 3. This invention adopts an adaptive normalization method for nonlinear fitting of working condition zones, combined with the static baseline of the zones and the dynamic rate of change correction, which solves the problem that the existing linear lookup table normalization is difficult to adapt to the nonlinear characteristics of the whole working condition, realizes the accurate quantification of characteristic deviations in the whole working condition range, and greatly improves the working condition adaptability of fault diagnosis. 4. This invention adopts a fault propagation path cascade coupling degree calculation model, which synthesizes the coupling degree step by step according to the fault physical propagation level. This solves the problem of low matching degree between the existing general fusion model and the fault evolution law, making the feature fusion result more consistent with the physical mechanism of fault occurrence and development, and improving the accuracy of fault severity measurement. 5. This invention adopts a cross-domain cross-verification method based on temporal correlation matching, which calculates the fault confidence from two dimensions: amplitude consistency and temporal synchronization. This solves the problems of high misjudgment rate and inconsistent confidence scale of existing cross-verification methods that rely solely on amplitude synthesis. Only when dual-domain anomalies match simultaneously in amplitude and temporal sequence will a high confidence level be output, effectively reducing the probability of fault misjudgment caused by single-domain interference. 6. This invention adopts a dual-window linkage fault evolution rate calculation method, which takes into account both long-term evolution trends and short-term abrupt changes. It combines change point detection to dynamically adjust the short window length and sets up a recovery mechanism, which solves the problem that existing single-window linear fitting cannot distinguish between stationary evolution and abrupt evolution. This improves the calculation accuracy of fault development rate and provides a more reliable prediction basis for preventive maintenance. 7. This invention adopts a prediction-driven state data pre-scheduling and storage method, which combines fault trend prediction to adjust sampling and storage strategies in advance. This solves the problems of slow response and low resource utilization in existing passive hierarchical scheduling, and significantly improves the utilization efficiency of storage and bandwidth resources while ensuring the integrity of fault data. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of the method steps of the present invention. Detailed Implementation

[0015] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0016] As attached Figure 1 The method for processing real-time status data of the AC pitch system shown includes the following steps: S1. Parallel acquisition of multi-dimensional data from the pitch system, dynamic matching of interpolation orders with different precision based on the comprehensive fluctuation of real-time operating conditions, and time base resampling and alignment of heterogeneous data with different sampling rates to generate a unified time series synchronous dataset. It should be noted that the multidimensional data includes five types of raw state data and two types of operating condition data. The five types of raw state data are: three-phase instantaneous current of the pitch motor, DC bus voltage of the pitch driver, vibration acceleration of the pitch reducer input housing, acoustic emission signal of the outer ring of the pitch bearing, and rotor position angle of the pitch motor. The two types of operating condition data are: the pitch angle at the current moment. and the current speed of the pitch motor All data are acquired in parallel through the corresponding acquisition modules; the overall fluctuation of the real-time operating conditions is quantitatively characterized by the overall rate of change of the operating conditions. ,in, The comprehensive rate of change of operating conditions, in units of It is used to characterize the severity of fluctuations in the operating conditions of the pitch system; The current propeller pitch angle; This is the rated pitch angle of the pitch control system; This is the first derivative of the pitch angle with respect to time, i.e., the rate of change of the pitch angle; This represents the current motor speed. This refers to the rated speed of the pitch motor; The first derivative of motor speed with respect to time is the rate of change of speed. The interpolation order is divided into three accuracy levels: first-order linear interpolation is used for stable operating conditions, second-order parabolic interpolation is used for gradually changing operating conditions, and third-order Lagrange interpolation is used for rapidly changing operating conditions. Hysteresis margin is set for the switching threshold of both order levels, and the hysteresis margin is 5%~10% of the corresponding threshold. When the comprehensive rate of change of operating conditions rises to the gradually changing threshold, it switches to second-order interpolation, and when it rises to the rapidly changing threshold, it switches to third-order interpolation. When the comprehensive rate of change of operating conditions decreases, it must fall below the corresponding threshold minus the hysteresis margin before switching back to the lower order interpolation to avoid frequent order switching near the threshold. The time reference resampling alignment uses the time series of the current acquisition channel as the reference time axis, adds a high-precision hardware timestamp to the original data frames of each acquisition channel, and uses the matched interpolation order to resample the data of other channels with different sampling rates. The synchronized dataset of the unified time series serves as the unified input for all subsequent feature solving stages, ensuring that the time reference of multi-source features is completely consistent.

[0017] In detail: This step is executed by the data preprocessing module built into the edge data processing unit. It is carried out in real time, sampling cycle by cycle, throughout the entire operation of the pitch system, and applies to the scenario of unifying the time base of multi-source heterogeneous data. This module first acquires five types of raw operating status data and two types of operating condition data in parallel through the corresponding acquisition module at the front end. The five types of raw status data are: pitch motor three-phase instantaneous current, pitch driver DC bus voltage, pitch reducer input housing vibration acceleration, pitch bearing outer ring acoustic emission signal, and pitch motor rotor position angle. The two types of operating condition data are: current pitch angle and pitch motor speed. Then, based on the current pitch angle and speed values, the rate of change of both with respect to time is calculated. The overall operating condition rate of change is then calculated using a vector synthesis method with the square root of the sum of squares, used to characterize the severity of operating condition fluctuations. Subsequently, corresponding interpolation is matched according to the numerical range of the overall operating condition rate of change. The interpolation order is determined as follows: first-order linear interpolation is used for stable operating conditions, second-order parabolic interpolation is used for gradually changing operating conditions, and third-order Lagrange interpolation is used for rapidly changing operating conditions. A hysteresis margin of 5% to 10% is set when switching between orders. When the rate of change of the operating condition rises to the threshold, it switches to a higher-order interpolation. When it falls, it must be below the threshold minus the hysteresis margin before switching back to a lower-order interpolation to avoid frequent switching and oscillations near the threshold. Finally, using the time series of the current acquisition channel as the reference time axis, hardware timestamps are added to the raw data of each channel. The data of other channels with different sampling rates are resampled using the matched interpolation order, and a synchronized dataset with a completely unified time reference is output as the input for all subsequent feature calculation stages.

[0018] S2. Based on rotor position information, convert the rotational speed-related time-domain signal into an angle-domain equally spaced signal, and solve the deep fault feature parameters of multiple physical dimensions in parallel to obtain feature data that characterizes the operating status of multiple components. It should be noted that: the rotor position information is derived from the rotor position angle data of the pitch motor in the synchronization dataset, serving as the position reference for the conversion of the time-domain signal to the angle domain, ensuring the position accuracy of the angle domain resampling; the speed-related time-domain signals include three-phase current signals and vibration envelope signals, both of which are affected by speed fluctuations, resulting in spectral broadening and distortion, which needs to be eliminated through angle domain conversion; the angle-domain equally spaced signals are generated through an equal-angle resampling method, converting the uniformly sampled time-domain signals into signals with uniform angle intervals based on the rotor position angle, eliminating the spectral broadening effect of speed fluctuations, and restoring the true amplitude of the fault characteristics; the deep fault feature parameters of the multi-physical dimensions are solved in parallel across four dimensions, with the four solutions executed independently and synchronously to improve real-time processing: the first dimension is the stator current frequency domain and symmetrical component feature solution, based on the instantaneous values ​​of the synchronized three-phase currents, first converted into angle-domain signals, and then a fast Fourier transform is performed to extract the harmonic amplitudes at the fundamental frequency order ± gear meshing order, obtaining the fundamental frequency sideband harmonic amplitude of the pitch motor U-phase current. Simultaneously, symmetrical component decomposition is performed to obtain the amplitude of the negative sequence component of the three-phase composite current. The amplitude of the fundamental frequency sideband harmonics of the U-phase current of the pitch motor , m is the integer index of the sampling point in the angle domain, and its value range is . ; is the instantaneous value of the U-phase current at point m after resampling in the angle domain; M is the number of sampling points in a single electrical cycle within the angle domain; The order of the sideband harmonics is determined by the number of meshing teeth in the reducer gears and the number of pole pairs in the motor; j is the imaginary unit; the second dimension is the DC bus ripple characteristic calculation, based on the synchronized DC bus voltage data, the ripple component in the 280Hz~320Hz frequency band is extracted by digital bandpass filtering, and the difference between the maximum and minimum values ​​of the ripple waveform within a unit time window is statistically analyzed to obtain the peak-to-peak value of the 300Hz ripple of the pitch drive DC bus. The third dimension involves the calculation of vibration envelope and acoustic emission characteristics. Based on the synchronized vibration acceleration data, a 5kHz~15kHz bandpass filter is first performed, and then Hilbert demodulation is performed on the filtered signal to obtain the vibration envelope time-domain signal. Combined with the rotor position angle, the vibration envelope time-domain signal is resampled into an angle-domain equally spaced signal. The peak amplitude of the envelope signal in the angle domain is extracted to obtain the demodulated amplitude of the vibration envelope of the pitch reducer input housing. Simultaneously, based on the synchronized acoustic emission signal data, the ratio of the cumulative energy value of all acoustic emission events within a unit time window to the window duration is calculated to obtain the acoustic emission signal energy rate at the fixed end of the outer ring of the pitch bearing. The fourth dimension involves the calculation of the rotor flux cross-axis component characteristics. Based on the synchronized stator three-phase current, DC bus voltage, and rotor position angle data, the estimated value of the rotor flux cross-axis component of the pitch motor is obtained in real time through the permanent magnet synchronous motor voltage model and the full-order observer algorithm. The specific feature data representing the operating status of the various components are as follows: These correspond to the gear meshing state of the reducer, the symmetrical state of the stator winding, the aging state of the bus capacitor, the impact state of the reducer bearing, the crack propagation state of the bearing raceway, and the orientation of the rotor magnetic field and the state of the permanent magnet. All of these are output to the subsequent normalization process.

[0019] In detail: This step is executed by the parallel feature calculation module built into the edge data processing unit. Parallel computation in four dimensions is started immediately after the synchronous dataset is generated, which is applied to the deep fault feature extraction scenario in multiple physical domains. The first dimension is stator current characteristic calculation. The module first resamples the time-domain current signal into an angle-domain equally spaced signal based on the synchronized three-phase current data and rotor position angle data, eliminating spectral broadening distortion caused by speed fluctuations. Then, it performs a fast Fourier transform on the angle-domain signal to extract the harmonic amplitudes at the fundamental frequency order superimposed on the gear meshing order, obtaining the fundamental frequency sideband harmonic amplitude of the pitch motor's U-phase current. Simultaneously, it calculates the amplitude of the negative-sequence component of the three-phase composite current using a symmetrical component decomposition method, used to characterize the degree of three-phase asymmetry in the stator windings. The second dimension is DC bus ripple characteristic calculation. The module performs digital bandpass filtering on the synchronized DC bus voltage data, extracting the ripple component in the 280Hz to 320Hz frequency band. It statistically analyzes the difference between the maximum and minimum values ​​of the ripple waveform within a unit time window to obtain the peak-to-peak value of the 300Hz ripple on the pitch drive's DC bus. The third dimension is vibration and acoustic emission characteristic calculation. The module first performs vibration and acoustic emission characteristic calculation on the synchronized... The speed data undergoes bandpass filtering from 5 kHz to 15 kHz, and then Hilbert demodulation is performed on the filtered signal to obtain the vibration envelope time-domain signal. Next, the vibration envelope time-domain signal is resampled into an angle-domain equally spaced signal by combining the rotor position angle. The peak amplitude of the envelope signal in the angle domain is extracted to obtain the vibration envelope demodulated amplitude of the pitch reducer input housing. At the same time, for the synchronized acoustic emission signal data, the ratio of the cumulative energy of all acoustic emission events within a unit time window to the window duration is calculated to obtain the acoustic emission signal energy rate of the fixed end of the outer ring of the pitch bearing. The fourth dimension is the rotor flux linkage feature calculation. Based on the synchronized stator three-phase current, DC bus voltage and rotor position angle data, the module estimates the cross-axis component of the pitch motor rotor flux linkage in real time through the permanent magnet synchronous motor voltage model and the full-order observer algorithm. After the synchronous calculation of the four dimensions is completed, a total of six types of feature data characterizing the operating status of the components are output, all of which are sent to the subsequent operating condition normalization stage.

[0020] S3. Call the corresponding nonlinear baseline model according to the working condition interval division rules, correct the real-time baseline value in combination with the dynamic change rate of the working condition, perform relative deviation normalization analysis on each characteristic parameter, and obtain normalized deviation data to eliminate working condition interference. It should be noted that: the operating condition interval division rule divides the entire operating condition range of the pitch system into three operating condition intervals: low speed with small pitch, medium speed with medium pitch, and high speed with large pitch. If the operating point falls on the boundary between two intervals, baseline calculation is performed according to the higher-level interval to avoid baseline jumps at the boundary; the nonlinear baseline model is a second-order polynomial fitting model corresponding to each interval, and the model coefficients are obtained by fitting the historical normal operating data within each operating condition interval to adapt to the characteristic change patterns of different intervals; the operating condition dynamic change rate adopts the comprehensive operating condition change rate output by S1. This is used to characterize the impact of dynamic fluctuations in operating conditions on the offset of the characteristic baseline. The real-time baseline value is obtained by superimposing a dynamic correction term on the static baseline fitting value. The static baseline fitting value is calculated through a nonlinear baseline model of the corresponding operating condition interval, and the dynamic correction term is the product of the dynamic operating condition correction coefficient and the comprehensive rate of change of the operating condition. The relative deviation normalization analysis uses the real-time baseline value as a benchmark to convert the measured values ​​of each characteristic parameter into relative deviation values ​​relative to the real-time baseline, eliminating the dual interference of steady-state differences and dynamic fluctuations in operating conditions. ,in, is the normalized deviation value of the kth feature parameter, which is a dimensionless quantity; is the measured solution value of the k-th feature parameter; z is the index of the interval to which the current working condition belongs; This represents the static baseline fitting value of the kth feature parameter within the zth operating condition interval; , is the dynamic operating condition correction coefficient of the kth characteristic parameter. Its dimension is consistent with the dimension of the corresponding characteristic quantity multiplied by time. Its physical meaning is the characteristic baseline offset caused by the unit operating condition change rate. The data was obtained through fitting and calibration using historical data of varying operating conditions under normal operating conditions. During fitting, linear regression was performed with the rate of change of operating conditions as the independent variable and the characteristic baseline offset as the dependent variable. The regression coefficients are... ; The comprehensive rate of change of operating conditions is consistent with the definition in S1; the normalized deviation data for eliminating operating condition interference is the data corresponding to each characteristic parameter. All values ​​are output to the subsequent fault coupling degree calculation stage.

[0021] In detail: This step is executed by the adaptive processing module built into the edge data processing unit. It is carried out immediately after the generation of the six types of feature data and is used to eliminate the interference of operating condition fluctuations on the feature baseline. The module first divides the entire operating condition range of the pitch system into three operating condition intervals: low speed with small pitch, medium speed with medium pitch, and high speed with large pitch. If the operating point falls on the boundary between two intervals, the baseline calculation is performed according to the higher-level interval to avoid baseline jumps at the boundary. Each operating condition interval has a pre-established second-order polynomial nonlinear baseline fitting model corresponding to each feature parameter. The model coefficients are obtained offline by fitting the normal operating history data within each operating condition interval. Then, the module obtains the pitch angle and speed values ​​at the current moment, determines the operating condition interval, calls the nonlinear baseline model of the corresponding interval, calculates the static baseline values ​​of each feature parameter under the current operating condition, and then introduces the comprehensive operating condition change rate as a dynamic correction term. The baseline offset is calculated by multiplying the dynamic operating condition correction coefficient and the comprehensive operating condition change rate, and then superimposed on the static baseline value to obtain the real-time baseline value. The dynamic operating condition correction coefficient is calibrated by fitting historical data of variable operating conditions under normal operating conditions. During fitting, a linear regression is performed with the operating condition change rate as the independent variable and the characteristic baseline offset as the dependent variable. The regression coefficient is the dynamic correction coefficient of the corresponding feature. Finally, based on the real-time baseline value, the measured value of each feature parameter is subtracted from the real-time baseline value and then divided by the real-time baseline value to obtain the relative deviation value of each feature relative to the current operating condition baseline, i.e., the normalized deviation data. This eliminates the dual interference of steady-state differences and dynamic fluctuations in the operating conditions and is output to the subsequent fault coupling degree calculation stage.

[0022] Table 1 shows the relative fitting errors of different baseline models under typical operating conditions. The data is taken from the measured data of a 2MW doubly-fed induction generator pitch system operating continuously for 30 days. 1000 sets of U-phase current fundamental frequency sideband harmonic feature samples were extracted for each operating point, and the relative fitting errors of the three baseline models for this feature were statistically analyzed. The data in the table show that the partitioned second-order polynomial baseline superposition operating condition dynamic change rate correction scheme adopted in this paper has significantly lower fitting errors than the commonly used static linear baseline in the entire operating condition range. Specifically, the fitting error under high-speed, large-pitch operating conditions decreased from 7.8% to 1.9%, a reduction of 75.6%. This effectively eliminates the interference of steady-state differences and dynamic fluctuations on the characteristic baseline offset, avoiding misjudgments of faults caused by inaccurate baseline calculations.

[0023] Table 1: Relative Error of Feature Fitting for Different Baseline Models under Typical Operating Conditions

[0024] S4. Divide the fault physical propagation path into multi-level propagation nodes, obtain the single-level feature coupling degree step by step according to the node level and perform cascade synthesis to obtain cascade coupling degree data that characterizes the severity of the fault. It should be noted that the physical propagation path of the faults described is based on the actual evolution mechanism of five types of faults: gear meshing faults, bearing faults, stator winding faults, bus capacitor faults, and permanent magnet demagnetization faults. Gear meshing faults are further divided into three levels: gear damage nodes, load pulsation nodes, and current modulation nodes; bearing faults are divided into two levels: raceway damage nodes and vibration impact nodes; stator winding faults are divided into two levels: winding impedance change nodes and magnetic field distortion nodes; bus capacitor faults are divided into one level: capacitance attenuation nodes and ripple increase nodes; and permanent magnet demagnetization faults are divided into magnetic flux attenuation nodes. The system comprises two levels of flux linkage offset nodes; the multi-level propagation nodes are feature-related nodes of each type of fault, decomposed along the physical propagation path, with each level node corresponding to a set of associated feature parameters; the single-level feature coupling degree is calculated using the root mean square algorithm, fusing the normalized deviation values ​​of all associated features within a single node to quantify the anomalous significance of the single-level node; the cascade synthesis uses a geometric mean method, multiplying the single-level coupling degree along the propagation path level by level and then raising the order root. Only when all levels of nodes exhibit anomalies will the overall coupling degree increase significantly, conforming to the physical law of fault cascade propagation. The final cascade coupling degree calculation formula is as follows: ,in, Let be the cascade coupling degree of the f-th type of fault, and be a dimensionless quantity that characterizes the severity of the fault. is the total number of levels of the fault propagation path corresponding to the type f fault; s is the level number of the fault propagation path; The number of feature parameters corresponding to the s-th level propagation node of the f-th type of fault; This is the normalized deviation value of the kth associated feature parameter within the corresponding node; the cascade coupling degree data covers all five types of faults and is output to the subsequent confidence analysis and trend prediction stages.

[0025] In detail: This step is executed by the fault fusion calculation module built into the edge data processing unit. It is carried out after the normalized deviation data is generated and is applied to the mechanistic quantification scenario of multi-feature fault severity. The module first sorts out the physical propagation path of the fault based on the actual evolution mechanism of five types of faults: gear meshing fault, bearing fault, stator winding fault, bus capacitor fault, and permanent magnet demagnetization fault. Among them, gear meshing fault is divided into three levels of nodes along the propagation path: gear damage, load pulsation, and current modulation; bearing fault is divided into two levels of nodes: raceway damage and vibration impact; stator winding fault is divided into two levels of nodes: winding impedance change and magnetic field distortion; bus capacitor fault is divided into one level of nodes: capacitance attenuation and ripple increase; and permanent magnet demagnetization fault is divided into two levels of nodes: magnetic flux attenuation and flux linkage shift. Each level of node corresponds to a set of associated feature parameters. Then, for each level of propagation node of each type of fault, The normalized deviation values ​​of all associated features of the node are extracted, and the root mean square algorithm is used to calculate the single-level node coupling degree to quantify the anomaly significance of the single-level node. Subsequently, the single-level coupling degrees of each level of nodes are cascaded along the fault propagation path. The geometric mean method is used to multiply the single-level coupling degrees of each level and then take the total series power. Only when all levels of propagation nodes show anomalies will the overall fault coupling degree increase significantly, which perfectly matches the physical law of fault cascade propagation. Finally, the cascade coupling degree corresponding to each type of fault is obtained to characterize the severity of the corresponding fault. The cascade coupling degree data of all faults can be directly used for fault level determination, and can also be output to the subsequent cross-confirmation confidence calculation stage.

[0026] S5. For dual-domain associated faults, analyze the fault coupling degree of the electrical domain and the mechanical domain respectively, and combine the temporal synchronization degree of abnormal features to correct the amplitude cross-verification results to obtain confidence data characterizing the reliability of the fault. It should be noted that: the dual-domain associated faults only include gear meshing faults, which are associated with both mechanical and electrical domain features; the other four types of faults are single-domain faults, with bearing faults being purely mechanical domain faults, and stator winding faults, bus capacitor faults, and permanent magnet demagnetization faults being purely electrical domain faults; single-domain faults are directly assigned a base confidence level of 0.6, representing evidence supported only by a single dimension, with limited credibility; the fault coupling degree between the electrical and mechanical domains is obtained based on cascaded coupling degree data, with the electrical domain coupling degree extracting the cascaded calculation results of the fault's electrical propagation path, and the mechanical domain coupling degree extracting the cascaded calculation results of the fault's mechanical propagation path; the temporal synchronization degree of the abnormal features is quantified using the Pearson correlation coefficient, taking the time series data of the 30 seconds before the current moment as the calculation window to obtain the temporal correlation coefficient of the fault coupling degree sequence between the electrical and mechanical domains. If the calculated Pearson correlation coefficient is negative, it is determined that the two-domain anomaly has no synchronous evolutionary characteristics, and the value is taken as... The final value range is [0,1]; the amplitude cross-verification result is quantified by the amplitude consistency coefficient, calculated based on the matching degree of the two-domain coupling degree. The closer the amplitudes of the two-domain coupling degree are, the higher the consistency; the final cross-verification confidence is obtained after correction by the degree of time synchronization: ,in, Let f be the cross-confirmation confidence level for the type f fault, which is a dimensionless quantity with a value range of [0,1]. The electrical-mechanical characteristic time-series correlation coefficient corresponding to the f-th type of fault; The electrical domain characteristic coupling degree corresponding to the f-th type of fault; The mechanical domain feature coupling degree corresponds to the type f fault. When the timing is completely synchronized and the amplitudes of the two domains are completely equal, the confidence level takes the maximum value of 1, which is significantly higher than the basic confidence level of the single-domain fault. The confidence level data that characterizes the credibility of the fault covers all five types of faults. After sorting based on the confidence level values, the fault type identification result and the corresponding confidence level are output and transmitted to the subsequent data scheduling stage.

[0027] In detail: This step is executed by the fault confidence analysis module built into the edge data processing unit. It is carried out after the cascaded coupling degree data is generated and is applied to scenarios that reduce the fault misjudgment rate through multi-dimensional cross-verification. The module first divides the fault into domains according to the physical attributes of the features. Only gear meshing faults are associated with both mechanical and electrical domain features and are classified as dual-domain faults. The other four types of faults are single-domain faults: bearing faults are purely mechanical domain faults, while stator winding faults, bus capacitor faults, and permanent magnet demagnetization faults are purely electrical domain faults. The basic confidence value for single-domain faults is directly set to 0.6, representing that there is only single-dimensional evidence to support the fault and the credibility is limited. For dual-domain gear meshing faults, the module first extracts the corresponding electrical and mechanical coupling degrees from the cascaded coupling degree data, ensuring complete time alignment between the two sets of data. Then, using the Pearson correlation coefficient calculation method, the module takes the time-series data from the previous 30 seconds as the calculation window to calculate the time-series correlation coefficient between the electrical and mechanical coupling degree sequences. This coefficient characterizes the degree of matching between the abnormal trends of the two domains. If the calculated correlation coefficient is negative, it is determined that the dual-domain anomaly lacks synchronous evolution characteristics, and the correlation coefficient is directly set to zero. Ultimately, the correlation coefficient value is fixed between zero and one. The amplitude consistency coefficient is then calculated based on the current dual-domain coupling degree. This coefficient is calculated by dividing the geometric mean of the dual-domain coupling degree by the arithmetic sum of the two domain coupling degrees. The closer the amplitudes of the dual-domain coupling degrees are, the higher the consistency coefficient, which also ranges from zero to one. Finally, the temporal correlation coefficient is multiplied by the amplitude consistency coefficient to obtain the cross-verification confidence score for dual-domain faults. Only when dual-domain anomalies simultaneously meet the criteria for amplitude matching and temporal synchronization will a high confidence score be output. When the temporal synchronization is complete and the dual-domain amplitudes are completely equal, the confidence score reaches its maximum value of one, significantly higher than the baseline confidence score for single-domain faults. After all fault confidence scores are calculated, they are sorted from highest to lowest, and the fault type identification result and corresponding confidence score for the current moment are output. Simultaneously, the confidence score data is transmitted to the subsequent data scheduling stage.

[0028] Table 2 shows a comparison of fault identification performance for different confidence level calculation schemes. The data comes from the test set of the pitch gearbox fault injection semi-physical test bench, which includes 1200 sets of normal operating condition samples and 600 sets of gear meshing fault samples, covering five typical field scenarios of wind farms: light / severe faults, operating condition fluctuations, electromagnetic interference, and pitch control. The data in the table shows that the proposed dual-dimensional cross-verification formula combining amplitude consistency and temporal synchronization significantly reduces the false positive and false negative rates in all test scenarios compared to the traditional single-amplitude cross-verification scheme. Specifically, the false positive rate in strong electromagnetic noise scenarios decreased from 16.3% to 4.7%, a reduction of 71.2%, and the false positive rate in scenarios with frequent operating condition fluctuations decreased by 74.5%. This effectively suppresses false fault signals caused by random interference and operating condition disturbances, significantly reducing the probability of fault misjudgment.

[0029] Table 2: Comparison of Fault Identification Performance of Different Confidence Calculation Schemes

[0030] S6. Using dual sliding windows of different lengths to fit the long-term evolution trend and short-term abrupt change characteristics of the fault respectively, and combining change point detection to dynamically adjust the short window length, the comprehensive fault evolution rate is synthesized and the fault development trend is predicted. It should be noted that the two sliding windows of different lengths are a long-trend window and a short-acceleration window, respectively. The long-trend window is typically set to 30 days to capture the long-term evolution pattern of the fault; the short-acceleration window is typically set to 7 days to capture the recent abrupt change trend of the fault. The two windows slide and update synchronously, continuously storing the coupling degree time-series data of each type of fault within the window. The long-term evolution trend and short-term abrupt change characteristics of the fault are obtained by least-squares linear fitting within the long and short windows, respectively, and the corresponding output is the long-term evolution rate. With short-term evolution rate The change point detection employs a sliding t-test algorithm, using time-series data of fault coupling degree within a short window as input to test the significance of the difference in means between two equal-length sub-windows. When the test statistic exceeds a preset significance threshold, a fault mutation node is identified, and the short window length is shortened to half its original length to enhance mutation capture capability. If no mutation node is detected within three consecutive short window periods, the short window length is restored to its original set length. The comprehensive fault evolution rate is calculated using a vector synthesis method, taking into account both long-term trends and short-term acceleration characteristics. ,in, The overall evolution rate of the coupling degree of the f-th type of fault is given by , in units of ; This represents the long-term evolution rate of the fault corresponding to the long-term trend window. The short-term evolution rate of the fault corresponds to the short acceleration window; the predicted fault development trend is calculated based on the comprehensive evolution rate and the current fault coupling degree, and the remaining time for the fault characteristics to reach the warning threshold is calculated; the warning threshold is calibrated by statistical analysis of historical fault data, and is taken as 3 to 5 times the maximum coupling degree under normal operating conditions, and finally the fault development trend prediction result and the remaining time data are output.

[0031] In detail: This step is executed by the trend prediction module built into the edge data processing unit. It updates periodically with a fixed sliding step size during continuous system operation, and is applied to the quantification and prediction of fault evolution trends. The module first sets two sliding time windows of different lengths: a longer window to capture the long-term evolution patterns of faults, and a shorter window to capture recent abrupt changes in fault trends. The two windows slide and update synchronously, continuously storing the coupling degree time-series data for each type of fault within the window. Then, least-squares linear fitting is performed on the fault coupling degree time-series data within the two windows, yielding the corresponding long-term and short-term fault evolution rates. Simultaneously, the module embeds a sliding t-test change point detection logic. Using the fault coupling degree time-series data within the short window as input, it tests the significance of the difference in the means between two equal-length sub-windows. When the test statistic exceeds a preset significance threshold, a fault abrupt change node is identified. At this point, the short window length is shortened to half its original length to enhance the ability to capture abrupt changes. When no abrupt change node is detected within three consecutive short window periods, the short window length is restored to its original setting to prevent the window from continuously shortening. Subsequently, a vector synthesis method using the square root of the sum of squares is employed to fuse the long-term and short-term evolution rates, yielding a comprehensive evolution rate for each type of fault coupling, while simultaneously considering both the long-term development trend and short-term acceleration characteristics of the fault. Finally, based on the comprehensive evolution rate and the current fault coupling value, the remaining time before the fault characteristics reach the warning threshold is calculated. The warning threshold is calibrated through statistical analysis of historical fault data, taking three to five times the maximum coupling degree under normal operating conditions. The final output is the fault development trend prediction result and the remaining time data.

[0032] S7. Based on fault confidence and comprehensive evolution rate, obtain storage priority, and combine fault trend prediction results to perform forward-looking sampling, storage and upload bandwidth pre-scheduling to achieve differentiated hierarchical data management.

[0033] It should be noted that the storage priority is calculated based on the combined evolution rate of the current highest fault confidence and the corresponding fault, and normalized using the baseline evolution rate as a reference. ,in, The storage priority value for the current state data is a dimensionless quantity. Let f be the cross-confirmation confidence level for the type f fault, and take the maximum value among all fault categories when substituting into the formula; This represents the overall evolution rate of the corresponding fault; The baseline evolution rate reference value preset by the system is taken as the background fluctuation rate of fault coupling under normal operating conditions, obtained through historical normal data statistics; the fault trend prediction result uses the remaining time of the fault reaching the warning threshold output by S6 as the triggering basis for pre-scheduling; the forward-looking sampling, storage and upload bandwidth pre-scheduling execution rules are as follows: when the remaining time of the fault is less than the preset pre-scheduling threshold, the sampling density and storage priority of the corresponding fault-related data channel are increased in advance, and cloud upload bandwidth is pre-allocated; when the remaining time is greater than the pre-scheduling threshold, the normal sampling and storage level is maintained; the pre-scheduling threshold is set according to the operation and maintenance response cycle, and the engineering conventional value is 2. The time frame is 4-72 hours. The differentiated hierarchical data management uses the baseline priority value under normal fault-free operating conditions as a reference, and divides it into three levels according to the storage priority value: more than twice the baseline priority is high priority, corresponding to a state with high fault confidence and fast evolution rate, and performs full raw data storage and real-time upload; 1-2 times the baseline priority is medium priority, corresponding to a state with medium fault confidence and slow evolution rate, and only stores feature values ​​and fault identification results and uploads them periodically; less than 1 times the baseline priority is low priority, corresponding to a state with no obvious fault characteristics, and only stores operating condition and health status labels and uploads them periodically; ultimately, it realizes collaborative data processing between the edge and the cloud.

[0034] In detail: This step is executed by the data scheduling and management module built into the edge data processing unit. Priority is calculated synchronously after the fault confidence and overall evolution rate are updated. The pre-scheduling logic is executed according to a fixed inspection cycle, and it applies to the optimization configuration of edge storage and bandwidth resources. The module first calculates the storage priority based on the highest current fault confidence and the corresponding fault's overall evolution rate. During the calculation, the evolution rate is normalized by dividing it by a system-preset baseline evolution rate reference value, and then combined with the confidence level using the square root of the sum of squares to synthesize the storage priority value. The baseline evolution rate is taken as the background fluctuation rate of fault coupling under normal operating conditions, obtained through historical normal data statistics. Next, the module performs forward-looking pre-scheduling based on the remaining time predicted by the fault trend. When the remaining time before a fault reaches the warning threshold is less than the preset pre-scheduling threshold, the sampling density and storage priority of the corresponding fault-related data channel are increased in advance, and cloud upload bandwidth is pre-allocated. When the remaining time is greater than the pre-scheduling threshold, the normal sampling and storage levels are maintained. The pre-scheduling threshold is set according to the operation and maintenance response cycle, with a typical engineering value of 24 to 72 hours. The module then uses a baseline priority value under normal, fault-free operating conditions as a reference to divide the status data into three storage levels: High priority (greater than twice the baseline priority), corresponding to states with high fault confidence and rapid evolution, performs full raw data storage and real-time cloud upload; Medium priority (between one and two times the baseline priority), corresponding to states with moderate fault confidence and slow evolution, stores only feature values ​​and fault identification results and uploads them periodically; Low priority (less than one times the baseline priority), corresponding to states without obvious fault characteristics, stores only operating condition information and health status tags and uploads them periodically. Ultimately, through differentiated hierarchical storage and upload scheduling, collaborative data processing between the edge and cloud is achieved, improving the utilization efficiency of storage and bandwidth resources while ensuring the integrity of fault data.

[0035] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for processing real-time status data of an AC pitch system, characterized in that, include: S1. Parallel acquisition of multi-dimensional data from the pitch system, dynamic matching of interpolation orders with different precision based on the comprehensive fluctuation of real-time operating conditions, and time base resampling and alignment of heterogeneous data with different sampling rates to generate a unified time series synchronous dataset. S2. Based on rotor position information, convert the rotational speed-related time-domain signal into an angle-domain equally spaced signal, and solve the deep fault feature parameters of multiple physical dimensions in parallel to obtain feature data that characterizes the operating status of multiple components. S3. Call the corresponding nonlinear baseline model according to the working condition interval division rules, correct the real-time baseline value in combination with the dynamic change rate of the working condition, perform relative deviation normalization analysis on each characteristic parameter, and obtain normalized deviation data to eliminate working condition interference. S4. Divide the fault physical propagation path into multi-level propagation nodes, obtain the single-level feature coupling degree step by step according to the node level and perform cascade synthesis to obtain cascade coupling degree data that characterizes the severity of the fault. S5. For dual-domain associated faults, analyze the fault coupling degree of the electrical domain and the mechanical domain respectively, and combine the temporal synchronization degree of abnormal features to correct the amplitude cross-verification results to obtain confidence data characterizing the reliability of the fault. S6. Using dual sliding windows of different lengths to fit the long-term evolution trend and short-term abrupt change characteristics of the fault respectively, and combining change point detection to dynamically adjust the short window length, the comprehensive fault evolution rate is synthesized and the fault development trend is predicted. S7. Based on fault confidence and comprehensive evolution rate, obtain storage priority, and combine fault trend prediction results to perform forward-looking sampling, storage and upload bandwidth pre-scheduling to achieve differentiated hierarchical data management.

2. The method for processing real-time status data of an AC pitch system according to claim 1, characterized in that: The multidimensional data includes five types of raw state data and two types of operating condition data. The five types of raw state data are: three-phase instantaneous current of the pitch motor, DC bus voltage of the pitch driver, vibration acceleration of the pitch reducer input housing, acoustic emission signal of the outer ring of the pitch bearing, and rotor position angle of the pitch motor. The two types of operating condition data are: the pitch angle at the current moment. and the current speed of the pitch motor ; The degree of real-time overall fluctuation in operating conditions is quantitatively characterized by the overall rate of change in operating conditions: ,in, This represents the overall rate of change in operating conditions. This refers to the rated pitch angle of the pitch control system. This is the first derivative of the pitch angle with respect to time, i.e., the rate of change of the pitch angle; This refers to the rated speed of the pitch motor; The first derivative of motor speed with respect to time is the rate of change of speed. The interpolation order is divided into three accuracy levels: first-order linear interpolation is used for stable operating conditions, second-order parabolic interpolation is used for gradually changing operating conditions, and third-order Lagrange interpolation is used for rapidly changing operating conditions. Hysteresis margin is set for the switching threshold of both order levels. When the comprehensive rate of change of operating conditions rises to the gradually changing threshold, it switches to second-order interpolation, and when it rises to the rapidly changing threshold, it switches to third-order interpolation. When the comprehensive rate of change of operating conditions decreases, it must fall below the corresponding threshold minus the hysteresis margin before switching back to the lower order interpolation. The time reference resampling alignment uses the time series of the current acquisition channel as the reference time axis, and uses the matched interpolation order to resample the data of other channels with different sampling rates.

3. The method for processing real-time status data of an AC pitch system according to claim 1, characterized in that: The speed-correlated time-domain signals include three-phase current signals and vibration envelope signals. The characteristic amplitudes of both types of signals are affected by speed fluctuations, resulting in spectral broadening and distortion. Angle domain conversion is needed to eliminate speed interference. The angle-domain equally spaced signals are generated by equal-angle resampling. Based on the rotor position angle, the time-domain uniformly sampled signals are converted into signals with uniform angle intervals, eliminating the spectral broadening effect of speed fluctuations and restoring the true amplitude of fault characteristics. The deep fault characteristic parameters of the multi-physical dimensions are solved in parallel across four dimensions. The four calculation processes are executed independently and synchronously to improve real-time processing: The first dimension is the stator current frequency domain and symmetrical component characteristic calculation. Based on the instantaneous values ​​of the three-phase current after synchronization, they are first converted into angle domain signals, and then a fast Fourier transform is performed to extract the harmonic amplitude values ​​at the fundamental frequency order ± gear meshing order, thus obtaining the fundamental frequency sideband harmonic amplitude values ​​of the pitch motor U-phase current. Simultaneously, symmetrical component decomposition is performed to obtain the amplitude of the negative sequence component of the three-phase composite current. ,in m is the integer index of the sampling point in the angle domain; is the instantaneous value of the U-phase current at point m after resampling in the angle domain; M is the number of sampling points in a single electrical cycle within the angle domain; The order of the sideband harmonics is determined by the number of meshing teeth in the reducer gears and the number of pole pairs in the motor; j is the imaginary unit; the second dimension is the DC bus ripple characteristic calculation, which obtains the peak-to-peak value of the 300Hz ripple of the pitch driver DC bus based on the synchronized DC bus voltage data. The third dimension involves the calculation of vibration envelope and acoustic emission characteristics. Based on the synchronized vibration acceleration data, a 5kHz~15kHz bandpass filter is first performed, and then Hilbert demodulation is performed on the filtered signal to obtain the vibration envelope time-domain signal. Combined with the rotor position angle, the vibration envelope time-domain signal is resampled into an angle-domain equally spaced signal. The peak amplitude of the envelope signal in the angle domain is extracted to obtain the demodulated amplitude of the vibration envelope of the pitch reducer input housing. Acoustic emission signal energy rate at the fixed end of the outer ring of the pitch bearing ; The fourth dimension involves solving the cross-axis component characteristics of the rotor flux linkage, and obtaining the estimated value of the cross-axis component of the rotor flux linkage of the pitch motor in real time. The specific feature data representing the operating status of the various components are as follows: These correspond to the gear meshing state of the reducer, the symmetrical state of the stator winding, the aging state of the bus capacitor, the impact state of the reducer bearing, the crack propagation state of the bearing raceway, and the orientation of the rotor magnetic field and the state of the permanent magnet.

4. The method for processing real-time status data of an AC pitch system according to claim 2, characterized in that: The operating condition interval division rule divides the entire operating condition range of the pitch system into three intervals: low-speed small pitch, medium-speed medium pitch, and high-speed large pitch. If the operating point falls on the boundary between two intervals, baseline calculation is performed according to the higher-level interval. The nonlinear baseline model is a second-order polynomial fitting model corresponding to each interval. The model coefficients are obtained by fitting the historical normal operating data within each interval to adapt to the characteristic change patterns of different intervals. The real-time baseline value is obtained by superimposing a dynamic correction term on the static baseline fitting value. The static baseline fitting value is calculated by the nonlinear baseline model of the corresponding operating condition interval, and the dynamic correction term is the product of the dynamic operating condition correction coefficient and the comprehensive operating condition change rate. The relative deviation normalization analysis uses the real-time baseline value as a benchmark to convert the measured values ​​of each characteristic parameter into relative deviation values ​​relative to the real-time baseline, eliminating the dual interference of steady-state differences and dynamic fluctuations in the operating condition. ,in, This represents the normalized deviation value of the k-th feature parameter; is the measured solution value of the k-th feature parameter; z is the sequence number of the interval to which the current working condition belongs; This represents the static baseline fitting value of the kth feature parameter within the zth operating condition interval; This is the dynamic operating condition correction coefficient for the k-th characteristic parameter.

5. The method for processing real-time status data of an AC pitch system according to claim 4, characterized in that: The physical propagation path of the fault is derived from the actual evolution mechanism of five types of faults, namely gear meshing fault, bearing fault, stator winding fault, bus capacitor fault, and permanent magnet demagnetization fault. The gear meshing faults are divided into three levels: gear damage nodes, load pulsation nodes, and current modulation nodes; bearing faults are divided into two levels: raceway damage nodes and vibration impact nodes; stator winding faults are divided into two levels: winding impedance change nodes and magnetic field distortion nodes; bus capacitor faults are divided into one level: capacitance attenuation nodes and ripple increase nodes; and permanent magnet demagnetization faults are divided into two levels: magnetic flux attenuation nodes and magnetic flux offset nodes. The multi-level propagation nodes are the feature association nodes of each type of fault along the physical propagation path, and each level node corresponds to a set of associated feature parameters. The single-level feature coupling degree is calculated using the root mean square algorithm, and the normalized deviation values ​​of all associated features within a single node are fused.

6. The method for processing real-time status data of an AC pitch system according to claim 5, characterized in that: The cascade coupling is synthesized using a geometric average method, by multiplying the single-stage coupling degree along the propagation path and then raising the result to the power of the series number to obtain the final cascade coupling degree. ,in, Let f be the cascade coupling degree of the type f fault; is the total number of levels of the fault propagation path corresponding to the type f fault; s is the level number of the fault propagation path; The number of feature parameters corresponding to the s-th level propagation node of the f-th type of fault; This is the normalized deviation value of the kth associated feature parameter within the corresponding node; the cascade coupling degree data covers all five types of faults and is output to the subsequent confidence analysis and trend prediction stages.

7. The method for processing real-time status data of an AC pitch system according to claim 6, characterized in that: The dual-domain associated faults only include gear meshing faults, which are associated with both mechanical and electrical domain features. The other four types of faults are single-domain faults, with bearing faults being purely mechanical domain faults, and stator winding faults, bus capacitor faults, and permanent magnet demagnetization faults being purely electrical domain faults. Single-domain faults are directly assigned a base confidence level of 0.6, representing evidence supported only by a single dimension, and thus have limited credibility. The fault coupling degree between the electrical and mechanical domains is obtained based on cascaded coupling degree data. The electrical domain coupling degree extracts the cascaded calculation results of the electrical propagation path of the fault, and the mechanical domain coupling degree extracts the cascaded calculation results of the mechanical propagation path of the fault. The temporal synchronization degree of the aforementioned abnormal features is quantified using the Pearson correlation coefficient to obtain the temporal correlation coefficient of the fault coupling sequence between the electrical and mechanical domains. If the calculated Pearson correlation coefficient is negative, it is determined that the two-domain anomaly has no synchronous evolutionary characteristics, and the value is taken as... The final value range is [0,1].

8. The method for processing real-time status data of an AC pitch system according to claim 7, characterized in that: The amplitude cross-verification results are quantified by the amplitude consistency coefficient, which is calculated based on the matching degree of the two-domain coupling degree. The closer the amplitudes of the two-domain coupling degree are, the higher the consistency. The final cross-confirmation confidence level is obtained after adjusting for timing synchronization: ,in, The cross-confirmation confidence level for type f faults; The electrical-mechanical characteristic time-series correlation coefficient corresponding to the f-th type of fault; The electrical domain characteristic coupling degree corresponding to the f-th type of fault; The mechanical domain feature coupling degree corresponding to the f-th type of fault; When the timing is completely synchronized and the amplitudes of the two domains are completely equal, the confidence level takes the maximum value of 1, which is significantly higher than the basic confidence level of a single-domain fault. The confidence level data that characterizes the credibility of the fault covers all five types of faults. After sorting based on the confidence level values, the fault type identification result and the corresponding confidence level are output and transmitted to the subsequent data scheduling stage.

9. The method for processing real-time status data of an AC pitch system according to claim 1, characterized in that: The two sliding windows of different lengths are a long trend window and a short acceleration window, respectively. The two windows slide and update synchronously, continuously storing the coupling degree time series data of each type of fault within the window. The long-term evolution trend and short-term mutation characteristics of the fault are obtained by least-squares linear fitting within the long and short windows, respectively, and the long-term evolution rate of the fault corresponding to the long trend window is output. Short-term fault evolution rate corresponding to short acceleration window ; The change point detection uses a sliding t-test algorithm, taking the fault coupling time series data within a short window as input, to test the significance of the difference in the mean between two equal-length sub-windows. When the test statistic exceeds a preset significance threshold, a fault mutation node is identified. At this point, the short window length is shortened to half its original length to enhance mutation detection capability. If no mutation node is detected within three consecutive short window periods, the short window length is restored to its original setting. The comprehensive fault evolution rate is calculated using a vector synthesis method, taking into account both long-term trends and short-term acceleration characteristics. ,in, The comprehensive evolution rate of the coupling degree of the f-th type of fault is given. The predicted fault development trend is calculated based on the comprehensive evolution rate and the current fault coupling degree. The remaining time for the fault characteristics to reach the warning threshold is calculated, and the fault development trend prediction result and the remaining time data are finally output.

10. The method for processing real-time status data of an AC pitch system according to claim 9, characterized in that: The storage priority is calculated based on the combined evolution rate of the current highest fault confidence and the corresponding fault, and normalized with a baseline evolution rate as a reference: ,in, This is the storage priority value for the current state data; Let f be the cross-confirmation confidence level for the type f fault, and take the maximum value among all fault categories when substituting into the formula; This represents the overall evolution rate of the corresponding fault; The system uses a preset baseline evolution rate reference value, which is the background fluctuation rate of fault coupling under normal operating conditions, obtained through historical normal data statistics. The fault trend prediction result uses the remaining time for the fault to reach the warning threshold output by S6 as the triggering basis for pre-scheduling. The forward-looking sampling, storage, and upload bandwidth pre-scheduling execution rules are as follows: when the remaining time of the fault is less than the preset pre-scheduling threshold, the sampling density and storage priority of the corresponding fault-related data channel are increased in advance, and cloud upload bandwidth is pre-allocated; when the remaining time is greater than the pre-scheduling threshold, the normal sampling and storage levels are maintained. The differentiated hierarchical data management uses the baseline priority value under normal fault-free operating conditions as a reference, and is divided into three levels according to the storage priority value: more than twice the baseline priority is high priority, corresponding to a state with high fault confidence and fast evolution rate, and full original data storage and real-time upload are executed; 1 to 2 times the baseline priority is medium priority, corresponding to a state with medium fault confidence and slow evolution rate, and only feature values ​​and fault identification results are stored and uploaded periodically; less than 1 times the baseline priority is low priority, corresponding to a state with no obvious fault characteristics, and only operating condition and health status labels are stored and uploaded periodically.

Images