Wind generating set bearing monitoring method based on vibration data analysis

Abstract

The invention relates to the technical field of bearing monitoring, in particular to a wind generating set bearing monitoring method based on vibration data analysis, and the method comprises the steps: dividing a mixed signal into narrow-band components with independent energy through variational mode decomposition, separating dynamic responses of different frequency bands, weakening the interference effect of random noise, and enhancing the recognizability of a weak characteristic signal; the difference between the inner and outer ring load effects is displayed in a time sequence matrix in an offset trend, quantitative characterization of the energy and offset relation is enhanced, time alignment is carried out on the raceway load and the frequency amplitude by executing dynamic time warping, time sequence offset caused by rotating speed fluctuation is corrected, and the synchronization characteristics of energy and phase in the structure are displayed. Response abnormal characteristics under working condition changes are extracted, and signal de-aliasing, time sequence reconstruction and structure response accurate matching are realized, so that vibration signal identification sensitivity is improved, energy and phase change correlation reliability is enhanced, and abnormal part identification certainty is improved.

Description

Technical Field

[0001] This invention relates to the field of bearing monitoring technology, and in particular to a method for monitoring wind turbine bearings based on vibration data analysis. Background Technology

[0002] The field of bearing monitoring technology mainly studies the technical system for real-time assessment and fault identification of bearing operating status during the operation of rotating machinery using measurement, analysis, and modeling methods. The core content includes vibration signal acquisition, signal feature extraction, modeling of the correspondence between features and fault modes, health status assessment, and early warning strategy design. In wind turbine generators, bearings, as key rotating components, bear the load of the main shaft and the torque of the transmission. Their operating status directly affects the reliability and power generation efficiency of the entire machine. The research focus of this field is on the comprehensive application of highly sensitive signal acquisition, multi-scale feature fusion, dynamic threshold modeling, and intelligent diagnostic algorithms to achieve highly reliable monitoring and life management of equipment operation.

[0003] The basic principle of the wind turbine bearing monitoring method based on vibration data analysis is to deploy triaxial acceleration sensors in the main bearing housing and gearbox input shaft to collect multidimensional vibration data containing time-domain, frequency-domain, and order information. Signal decomposition and pattern recognition technologies are then used to analyze the bearing's operating status in real time. The aim is to accurately identify early, minor bearing faults, including raceway pitting, roller cracks, and lubricating oil film degradation, under varying operating conditions and high noise environments. This enables early fault warning, fault type localization, and health trend assessment. By establishing an adaptive diagnostic model based on a health baseline, the deviation between the current state and the normal operating state is calculated in real time, achieving intelligent diagnosis from experience-based judgment to data-driven approaches, thereby improving the safety and economy of wind turbine operation and maintenance.

[0004] Existing technologies rely heavily on static configurations and empirical thresholds for mapping acquired features to fault modes in actual operation. When faced with speed fluctuations and load changes, the characteristic peak migration caused by frequency drift and amplitude disturbances affects the stability of the criteria, leading to false alarms and missed alarms. Multidimensional vibration data exhibits temporal inconsistencies, lacks unified index constraints between measurement points, causes mismatches in cross-channel comparisons, results in a wide fault location range, and under strong noise backgrounds, sudden energy interference can cover weak fault signs, leading to decreased alarm stability and increased early identification lag. The mapping link from features to parts lacks path-based expression, making it difficult to trace the direction and order of load action, resulting in unfocused maintenance goals and excessive maintenance time. Furthermore, the lack of a unified labeling mechanism for stable intervals under cross-operating conditions causes baseline drift with operating conditions, leading to periodic repetitions in long-term trend assessments and affecting the accuracy of life management and spare parts planning. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies and propose a method for monitoring wind turbine generator bearings based on vibration data analysis.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a method for monitoring wind turbine generator bearings based on vibration data analysis, comprising the following steps: S1: Record the radial and axial vibration amplitudes of the rolling element through the vibration signal sequence output by the triaxial accelerometer, perform amplitude difference through time index alignment and calculate frequency distribution, identify the stable interval index after determining the direction of speed change, and generate the dynamic characteristic sequence of the working condition. S2: Based on the dynamic feature sequence of the working condition, after extracting the stable mode components by variational mode decomposition, the rolling body energy is divided into partitions and the offset rate is calculated. The offset rate and amplitude change rate of the inner and outer rings are calculated and the high offset interval is screened to construct the displacement trend structure matrix. S3: Based on the displacement trend structure matrix, the roller load and frequency amplitude are compared after time alignment by dynamic time warping. The main and slave responses are divided by direction matching and the index is recorded. The energy ratio and phase difference are calculated. The indexes are integrated to form a propagation path table, and the structural response distribution table is obtained. S4: Based on the displacement trend structure matrix and the structure response distribution table, the rolling element energy ratio and rotational speed gradient are jointly compared. High gradient segments are identified through difference calculation, and the frequency band energy difference and offset mean are extracted. The index of the out-of-limit interval and the measurement point number are recorded to form a set of active signal intervals. S5: Based on the set of active signal intervals, determine the intersection of the energy ratio and phase difference of the rolling element measuring points, confirm the double high amplitude interval through trend matching, screen continuous abnormal measuring points, record the measuring point number and offset amplitude to form an index table, and output the bearing abnormal part index list.

[0007] As a further embodiment of the present invention, the dynamic characteristic sequence of the working condition includes a time index sequence, a vibration amplitude sequence, a frequency distribution sequence, and a rotational speed direction identifier; the displacement trend structure matrix includes a time index dimension, an energy partition dimension, and an offset rate recording unit; the structural response distribution table includes a master-slave response index, an energy ratio record, and a phase difference table; the signal active interval set includes a high gradient segment index table, a frequency band energy difference table, and an offset mean record table; and the bearing abnormal part index list includes a dual high amplitude measurement point index table, a continuous abnormal measurement point set, and an offset amplitude mapping table.

[0008] As a further aspect of the present invention, the specific steps for generating the dynamic feature sequence of the working condition are as follows: By recording the radial and axial vibration amplitudes of the rolling body through the vibration signal sequence output by the triaxial accelerometer, the signals are extracted in segments and aligned with the time index. The amplitudes are differentially analyzed and the ratios are calculated. The index positions are marked in chronological order to generate a multi-directional vibration amplitude sequence set. Based on the multi-directional vibration amplitude sequence set, the frequency distribution is calculated and the speed change is judged. The frequency range is determined by period measurement and the direction of amplitude change is identified. After determining the speed rise and fall trend, the index position of the stable range is recorded, and the dynamic feature sequence of the working condition is generated.

[0009] As a further aspect of the present invention, the specific steps for constructing the displacement trend structure matrix are as follows: Based on the dynamic characteristic sequence of the working condition, variational mode decomposition is used to decompose the original vibration signal into multiple stable mode components, the rolling body energy is partitioned, the energy is accumulated and the density is statistically calculated according to the time index fixed window, and the segment difference and ratio are calculated and the change index is recorded to generate a rolling body energy distribution table. Based on the rolling element energy distribution table, load difference and rotational speed offset are matched, two values ​​are paired by the same time index, the difference is calculated and the average offset is calculated by aggregating by segment, the offset direction is determined and the index position is marked to generate an offset rate matching sequence. Based on the offset rate matching sequence, the offset rate of the inner and outer circles and the amplitude change rate are calculated. The offset and vibration amplitude are extracted in each time segment and proportional calculation is performed. High offset segments are selected, and the corresponding indices are written into the matrix row and column structure to generate a displacement trend structure matrix.

[0010] As a further aspect of the present invention, the specific execution steps of the variational mode decomposition are as follows: first, the original vibration signal corresponding to the dynamic characteristic sequence of the working condition is read, the sampling time range is determined, and the amplitude of the signal is normalized. Then, the signal is decomposed into independent narrowband mode components according to the frequency distribution characteristics. During the decomposition process, the decomposition step size is adjusted according to the energy distribution density to keep the frequency bands of each mode independent. After the decomposition is completed, the instantaneous amplitude and frequency information of each mode are extracted one by one, and the difference between the local energy and the average energy of each mode is calculated to form a multimode energy matrix.

[0011] As a further aspect of the present invention, the matching of load difference and speed offset is specifically implemented in the following steps: First, load data and speed data for the same time period in the rolling element energy distribution table are read, and a correspondence between the two is established based on the time index. Then, the average load value for each time period is calculated and the load difference between adjacent time periods is obtained. At the same time, the speed change for the corresponding time period is calculated as the speed offset value. The load difference and speed offset are paired according to the same time index to generate a load-speed corresponding sequence. In the sequence, the numerical pairs of adjacent index segments are aggregated and the average offset is calculated. The matching results and index positions of each segment are recorded.

[0012] As a further aspect of the present invention, the specific steps for obtaining the structural response distribution table are as follows: Based on the displacement trend structure matrix, dynamic time warping is used to align the raceway load and frequency amplitude synchronization sequence in time. The system is partitioned by time index, the amplitude direction difference is calculated and a direction sequence is generated. The direction change and the corresponding index position are marked. The main direction and the secondary direction are extracted and the matching index is recorded to generate a master-slave dynamic response index set. Based on the master-slave response index set, the load energy is accumulated according to the index interval, the frequency amplitude peak is extracted, the energy amplitude ratio is calculated to form a ratio sequence, the phase change angle is generated to form an angle sequence, the direction change index is marked, the difference matrix is ​​sorted and the energy mapping is established to generate an energy phase difference ratio table. Based on the energy phase difference ratio table, energy directions are classified according to index intervals and a continuous sequence is established. After arranging the energy change paths, turning points are marked. Adjacent intervals are integrated to establish a hierarchical mapping. The index order is verified and the numbers are filled in. The path data is summarized and the structural response distribution table is output.

[0013] As a further aspect of the present invention, the specific execution process of the dynamic time warping is as follows: First, read the synchronization sequence of the rolling load and frequency amplitude in the displacement trend structure matrix, determine the time index range of the two sequences and establish a corresponding mapping table, then calculate the amplitude difference and frequency difference between adjacent time nodes, accumulate the differences in sequence and generate a time distance matrix, compare the distance between adjacent nodes row by row in the matrix and record the matching path index, and align and resample the time axes of the two sequences according to the accumulated difference of the path valley values ​​to generate a synchronized aligned sequence after time warping.

[0014] As a further aspect of the present invention, the specific steps for constructing the set of active signal intervals are as follows: Based on the displacement trend structure matrix and the structure response distribution table, the rolling element energy ratio and speed gradient synchronization sequence are extracted, the time index partitions are aligned, the energy difference and speed increment of adjacent points are calculated, the abrupt segment index and measurement point number are merged, and a high gradient segment index table is generated. Based on the high gradient segment index table, extract the corresponding segment frequency band energy sequence and perform interval integration, calculate the frequency band energy difference to generate a difference sequence, extract the offset number to calculate the average offset, mark the energy difference and offset mean value exceeding the limit interval index, organize the measurement point numbers and output the signal active interval set.

[0015] As a further aspect of the present invention, the specific steps for outputting the bearing abnormality index list are as follows: Based on the set of active signal intervals, extract the energy ratio and phase difference sequence of the rolling body measuring points, align the time index and divide the segments, calculate the energy amplitude difference and phase difference change, perform cross-comparison and lock the double high amplitude segments, record the index and measuring point number, and generate a double high amplitude measuring point index table. Based on the dual high amplitude measurement point index table, the energy amplitude and offset amplitude sequence of continuous measurement points are extracted, the amplitude is differentially analyzed and the offset increment is statistically calculated, abrupt measurement points are screened and adjacent abnormal sections are integrated, a mapping between measurement point number and offset amplitude is established, the index order is sorted and the bearing abnormal part index list is output.

[0016] Compared with the prior art, the advantages and positive effects of the present invention are as follows: 1. In this invention, variational mode decomposition divides the mixed signal into energy-independent narrowband components, separates the dynamic response of different frequency bands, weakens the interference effect of random noise, enhances the discernibility of weak feature signals, and makes the difference between inner and outer ring loads appear in the time series matrix as an offset trend, thus strengthening the quantitative characterization of the relationship between energy and offset. 2. In this invention, the raceway load and frequency amplitude are time-aligned by performing dynamic time warping, which corrects the timing offset caused by speed fluctuations, so that the active and passive responses can be compared on a unified scale, showing the synchronous characteristics of energy and phase within the structure, and refining the abnormal response characteristics under changing operating conditions. 3. In this invention, by achieving signal dealiasing, timing reconstruction, and precise matching of structural response, the sensitivity of vibration signal identification is improved, the fault separation capability under noisy conditions is enhanced, the alignment accuracy is improved when the rotational speed is not constant, the reliability of the correlation between energy and phase changes is enhanced, and the certainty of abnormal part identification is improved. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the main steps of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0019] Example 1 Please see Figure 1 This invention provides a technical solution: a method for monitoring wind turbine generator bearings based on vibration data analysis, comprising the following steps: S1: Record the radial and axial vibration amplitudes of the rolling element through the vibration signal sequence output by the triaxial accelerometer, perform amplitude difference through time index alignment and calculate frequency distribution, identify the stable interval index after determining the direction of speed change, and generate the dynamic characteristic sequence of the working condition. S2: Based on the dynamic feature sequence of the working condition, after extracting the stable mode components by variational mode decomposition, the rolling body energy is divided into partitions and the offset rate is calculated. The offset rate and amplitude change rate of the inner and outer rings are calculated and the high offset interval is screened to construct the displacement trend structure matrix. S3: Based on the displacement trend structure matrix, the roller load and frequency amplitude are compared after time alignment by dynamic time warping. The main and slave responses are divided by direction matching and the index is recorded. The energy ratio and phase difference are calculated. The indexes are integrated to form a propagation path table and obtain the structural response distribution table. S4: Based on the displacement trend structure matrix and structural response distribution table, the rolling element energy ratio and rotational speed gradient are jointly compared. High gradient segments are identified through difference calculation, and the frequency band energy difference and offset mean are extracted. The index of the out-of-limit interval and the measurement point number are recorded to form a set of active signal intervals. S5: Based on the active signal interval set, determine the intersection of the energy ratio and phase difference of the rolling element measuring point, confirm the double high amplitude interval through trend matching, screen continuous abnormal measuring points, record the measuring point number and offset amplitude to form an index table, and output the bearing abnormal part index list.

[0020] The dynamic characteristic sequence of the working condition includes a time index sequence, vibration amplitude sequence, frequency distribution sequence and rotational speed direction identifier; the displacement trend structure matrix includes a time index dimension, energy partition dimension and offset rate recording unit; the structural response distribution table includes a master-slave response index, energy ratio record and phase difference table; the signal active interval set includes a high gradient segment index table, frequency band energy difference table and offset mean record table; the bearing abnormal part index list includes a dual high amplitude measurement point index table, continuous abnormal measurement point set and offset amplitude mapping table.

[0021] The specific steps for generating the dynamic feature sequence of operating conditions are as follows: By recording the radial and axial vibration amplitudes of the rolling body through the vibration signal sequence output by the triaxial accelerometer, the signals are extracted in segments and aligned with the time index. The amplitudes are differentially analyzed and the ratios are calculated. The index positions are marked in chronological order to generate a multi-directional vibration amplitude sequence set. Based on the multi-directional vibration amplitude sequence set, the frequency distribution is calculated and the speed change is judged. The frequency range is determined by period measurement and the direction of amplitude change is identified. After determining the speed rise and fall trend, the stable interval index position is recorded to generate the dynamic characteristic sequence of the working condition. Based on the vibration signal sequence output by the triaxial accelerometer, the sampling frequency was set to 5120 Hz, the recording duration was 60 seconds, the segment length was 2000 points, the overlap between segments was 200 points, the time index started at 0 seconds and ended at 60 seconds, the time step was 0.001 seconds, the triaxial time index was uniformly mapped, the interpolation boundary adopted the endpoint preservation strategy, the amplitude change of adjacent sampling points was calculated, the amplitude ratio was taken as the ratio of the differential amplitude to the original amplitude, the fixed threshold was 0.05, the index number started from 1 and incremented, the axial signal and radial signal were processed independently and then merged on the time axis, the amplitude unit was g, the time unit was second, the outlier shielding threshold was set to empty samples exceeding ±10g, the samples within the segment were arranged in time order, the data between segments were connected in the order of the starting time, the order of the triaxial channels was kept fixed, and the final output was a multi-directional vibration amplitude sequence set; Based on a multi-directional vibration amplitude sequence set, frequency distribution and rotational speed changes are calculated. The window type is a Hanning window with a length of 1024 points and an overlap of 512 points. The frequency resolution is 5 Hz, and power is measured in power density. The peak retrieval threshold is set to 0.002, the minimum inter-peak interval is 10 Hz, and main peak tracking uses a sliding window width of 3 frequency grids. The frequency gradient threshold is ±20 Hz per second, and the rise and fall trends are indicated by positive and negative signs. Period calculation uses a preset lookup table with frequencies set to 100 and 200 Hz. 300, 500, and 1000 Hz correspond to periods of 0.010, 0.005, 0.003, 0.002, and 0.001 seconds, respectively. Frequency intervals are divided using main peak clustering with a cluster radius of 15 Hz and a minimum number of samples per cluster of 5. The stability threshold is defined as frequency fluctuations not exceeding 2% and a duration of at least 3 seconds. Stable intervals are arranged in numerical order, rotational speed is marked with an upward or downward trend, frequency intervals are divided by the main peak, and time indexes are aligned to milliseconds to generate a dynamic feature sequence of operating conditions.

[0022] The specific steps for constructing the displacement trend structure matrix are as follows: Based on the dynamic characteristic sequence of the working condition, variational mode decomposition is used to decompose the original vibration signal into multiple stable mode components. The rolling body energy is partitioned, and the energy is accumulated and the density is statistically calculated according to the time index fixed window. The segment difference and ratio are calculated and the change index is recorded to generate the rolling body energy distribution table. Based on the rolling element energy distribution table, load difference and rotational speed offset are matched. The two values ​​are paired by the same time index, the difference is calculated and the average offset is calculated by aggregating by segment. After determining the offset direction, the index position is marked to generate an offset rate matching sequence. Based on the offset rate matching sequence, the offset rate of the inner and outer circles and the amplitude change rate are calculated. The offset and vibration amplitude are extracted in each time segment and proportional calculation is performed. High offset segments are screened, and the corresponding indices are written into the matrix row and column structure to generate a displacement trend structure matrix. Based on the dynamic feature sequence of the working condition, the variational mode decomposition method is used to decompose the original vibration signal. The sampling frequency is set to 5120 Hz, the input signal length is set to 307200 points, the signal boundary is supplemented with the last 20 samples using a mirror extension method, the number of modes is set to 8, the penalty parameter is set to 2000, and the convergence threshold is set to... The maximum number of iterations is set to 500. The initial center frequency is distributed at equal intervals from 0 to half the sampling frequency. An exponential step size of 0.25 is used to distribute the initial value. The weight coefficient is set to 0.8. The update method adopts an alternating direction multiplier iteration mode. In each iteration, the modal component bandwidth is calculated and the center frequency is updated until the modal change is less than the threshold and the iteration stops. The instantaneous energy of each modal component obtained by decomposition is calculated. The sampling window width is set to 0.5 seconds and the window interval is set to 0.25 seconds. The energy accumulation adopts the sliding window summation method. Each time index corresponds to an energy accumulation value. The energy partition threshold is divided into five segments according to the total energy range with a step size of 20%. The energy density distribution in each segment is statistically analyzed. The energy difference and ratio of adjacent segments are calculated. The difference threshold is set to 0.1 and the ratio threshold is set to 1.2. The change segment number is recorded in the order of time index to generate the rolling body energy distribution table. Based on the rolling element energy distribution table, load difference and speed offset are matched. The load data sampling frequency is set to 100 Hz, and the speed sampling frequency is set to 50 Hz. The time alignment method is linear interpolation with an interpolation interval of 0.02 seconds. The synchronization range is consistent with the energy table index. After alignment, two sets of values ​​are paired according to the same time index. The average load difference and speed offset are calculated. The difference calculation accuracy is retained to three decimal places. The time segment width is set to 1 second. Mean aggregation is performed within each segment with an aggregation threshold of 0.05. Segments with an average offset greater than the threshold are recorded as offset active segments. The offset direction is determined as positive or negative based on the consistency of the signs of the load difference and speed offset. The offset direction label is recorded at the corresponding index position to generate an offset rate matching sequence. Based on the offset rate matching sequence, the offset rate of the inner and outer circles and the amplitude change rate are calculated. The offset and corresponding vibration amplitude samples are extracted in each time segment. The amplitude range is set to ±10g, and the offset range is set to ±0.05. The ratio is calculated using the ratio of the offset to the amplitude difference, and the rounding precision is set to 0.001. The segment with an offset rate greater than 0.03 is selected as the high offset segment. The index step size is set to 1. The matrix rows represent the time index, and the columns represent the offset category, including the inner circle and the outer circle. The matrix cell value is the offset rate. The matrix dimension is adaptively generated according to the number of segments. The rows and columns are arranged in ascending order of time. The corresponding index number is written into the matrix structure. The matrix is ​​filled with zero values ​​for the unvalued parts, generating a displacement trend structure matrix.

[0023] The specific execution steps of variational mode decomposition are as follows: First, read the original vibration signal corresponding to the dynamic characteristic sequence of the working condition, determine the sampling time range and normalize the amplitude of the signal. Then, according to the frequency distribution characteristics, decompose the signal into independent narrowband modal components. During the decomposition process, adjust the decomposition step size according to the energy distribution density to keep the frequency bands of each mode independent. After the decomposition is completed, extract the instantaneous amplitude and frequency information of each mode one by one, and calculate the difference between the local energy and the average energy of each mode to form a multimode energy matrix. Variational mode decomposition, according to the formula:

[0024] in: For indexing within the time window Time Energy value of each mode Modal subscripts are used to identify the modal subscript. One modal component, The time window index is used to identify the first... A time-sliding window, The time sampling point index is used to identify the first sampling point. Each sampling point location, Here, represents the total number of sampling points, and represents the total length of signal sampling. For sampling points at time The corresponding number The penalty coefficient for each modality, Here, is the penalty base value, and is a constant parameter controlling the modal bandwidth. This is a time-differential operator used to calculate the rate of change of a signal over time. For sampling points at time First One modal component signal, The base of the natural constant, The imaginary unit, For the first The central angular frequency of each mode The weights are used to reconstruct the constraint weights and balance the weights of the bandwidth term and the reconstruction error term. For sampling points at time The input vibration signal at the location, The total number of modes represents the number of modes decomposed. Execution process: Input vibration signal Sampling was performed at a frequency of 5120 Hz, with a total sampling time of 60 seconds. The signal amplitude is normalized to the range of 0 to 1, and the endpoints are mirrored and extended by 20 points to fill the boundary. The total number of modes is set. Center angular frequency The modes are incremented by 2π×80 radians per second, ranging from 2π×20 to 2π×580 radians per second. Establish corresponding penalty coefficients At each time sampling point The penalty base value is 2000, the time weighting parameter is 0.15, and the frequency correction parameter is applied. Calculate values ​​and initialize each modal component. The result is the bandpass filtered result of the original signal, followed by an alternating direction multiplier update loop to calculate the th... The time derivative of each mode , and complex exponential terms Multiply by the product, take the square of the modulus, and then multiply by the current sampling time. Corresponding penalty coefficient Accumulated bandwidth-constrained energy terms are formed, and signal reconstruction errors are calculated. Multiply by weight And add the energy summation formula in the time dimension. The above summation is performed to obtain the current time window. Energy results for each mode .

[0025] In the specific execution process of matching load difference and speed offset, firstly, load data and speed data for the same time period in the rolling element energy distribution table are read, and the correspondence between the two is established based on the time index. Then, the average load value for each time period is calculated and the load difference between adjacent time periods is obtained. At the same time, the speed change for the corresponding time period is calculated as the speed offset value. The load difference and speed offset are paired according to the same time index to generate a load-speed corresponding sequence. In the sequence, the numerical pairs of adjacent index segments are aggregated and the average offset is calculated. The matching results and index positions of each segment are recorded.

[0026] The specific steps to obtain the structural response distribution table are as follows: Based on the displacement trend structure matrix, dynamic time warping is used to align the raceway load and frequency amplitude synchronization sequence in time. The system is partitioned by time index, the amplitude direction difference is calculated and a direction sequence is generated. The direction change and the corresponding index position are marked. The main direction and the secondary direction are extracted and the matching index is recorded to generate a master-slave dynamic response index set. Based on the master-slave response index set, the load energy is accumulated according to the index interval, the frequency amplitude peak is extracted, the energy amplitude ratio is calculated to form a ratio sequence, the phase change angle is generated to form an angle sequence, the direction change index is marked, the difference matrix is ​​sorted and the energy mapping is established to generate an energy phase difference ratio table. Based on the energy phase difference ratio table, energy directions are classified according to index intervals and a continuous sequence is established. After arranging the energy change path, the turning point is marked. Adjacent intervals are integrated to establish a hierarchical mapping. The index order is checked and the numbers are filled in. The path data is summarized and the structural response distribution table is output. Based on the displacement trend structure matrix, a dynamic time warping algorithm is used to align the raceway load and frequency amplitude synchronization sequence in time. The input sequence length is set to 10,000 points, the sampling interval is set to 0.001 seconds, the time index range is set to 0 to 10 seconds, the distance metric is set to Euclidean distance, the step size limit is set to ±1, the cumulative distance matrix dimension is set to 10,000 × 10,000, the diagonal distance is initialized to zero, the boundary expansion uses zero-padding to fill 50 points at each end, the path constraint window width is set to 100 points, the matching path is the path with the minimum cumulative cost, the dynamic programming search order adopts a recursive approach from top left to bottom right, the recursion step size is 1 step, and the time series is on the path After alignment, samples are resampled to a unified index interval of 0.001 seconds. The amplitude difference is calculated by direct difference between adjacent alignment points. The calculation direction is determined by the sign of the difference: positive values ​​are marked as the upward direction, negative values ​​as the downward direction, and equal values ​​as the steady-state direction. The length of the direction sequence is consistent with the alignment sequence. The threshold for direction change points is set to more than 3 consecutive changes in direction. The index number is incremented in chronological order. The portion with no less than 5 consecutive points in the direction change segment is extracted. The main direction and the secondary direction are extracted based on the consistency of the amplitude change sign. The main direction is identified as the average positive amplitude segment, and the secondary direction is identified as the average negative amplitude segment. The matching number is recorded according to the corresponding time index to generate the master-slave dynamic response index set. Based on the master-slave response index set, the load energy is accumulated according to the index interval. The input energy sampling frequency is set to 5120 Hz, the interval width is set to 0.2 seconds, the energy calculation formula uses the amplitude square integral form, the absolute value of the amplitude is taken and then squared and accumulated, the integration precision is set to four decimal places, the peak value is extracted by locating the maximum value within the interval, the number of peak values ​​in the frequency amplitude sequence is limited to the first 3 peak values ​​of each segment, the energy amplitude ratio is calculated by the ratio of the average energy of the master and slave segments, the ratio precision is set to three decimal places, the phase angle is calculated by using the arctangent function to obtain the phase delay between the master and slave, the angle output range is set to -180 to 180 degrees, the phase resolution is set to 0.1 degrees, the direction change index is marked by the angle change exceeding 30 degrees, the difference matrix is ​​constructed by combining the energy ratio row and the phase angle column, the matrix element value is the product of the ratio and the angle, the matrix size is automatically adjusted according to the number of indexes, the energy mapping adopts the interval normalization method, the mapping step size is set to 0.05, and an energy phase difference ratio table is generated. Based on the energy phase difference ratio table, energy directions are categorized according to index intervals. The energy direction threshold is set to 0 as the boundary, with positive values ​​classified as ascending directions and negative values ​​as descending directions. The continuous sequence identification threshold is set to a continuous index difference of no more than 2. Energy paths are arranged in ascending order of time. The turning point labeling threshold is set to an energy direction reversal lasting at least 0.1 seconds. Turning point numbers are incremented in chronological order. When integrating adjacent intervals, the interval overlap ratio is set to 50%. The hierarchical mapping is established with the first-level path as the base node, the second-level path as the energy reversal interval, and the third-level path as the direction repetition segment. The index order is checked using chronological logic. Missing numbers are filled by adding 1 to the adjacent preceding number. The path data is summarized in chronological order and organized into a hierarchical structure for output, generating a structural response distribution table.

[0027] The specific execution process of dynamic time warping is as follows: First, read the synchronization sequence of the roller load and frequency amplitude in the displacement trend structure matrix, determine the time index range of the two sequences and establish a corresponding mapping table, then calculate the amplitude difference and frequency difference between adjacent time nodes, accumulate the differences in turn and generate a time distance matrix, compare the distance between adjacent nodes row by row in the matrix and record the matching path index, and align and resample the time axis of the two sequences according to the accumulated difference of the path valley value to generate a synchronized aligned sequence after time warping. Dynamic time warping, according to the formula:

[0028] in: To minimize the cumulative path distance, The path step index indicates the current matching path step number. Each calculation location, This represents the total number of steps on the path. The amplitude difference weighting coefficient, The difference weighting coefficient between adjacent points. This is the index offset weight coefficient. The path smoothing penalty coefficient is... For the synchronization sequence of the roller load in the path number Step corresponding index The amplitude at that point, For frequency amplitude synchronization sequence in path number 1 Step corresponding index The amplitude at that point, For the synchronization sequence of the roller load in the path number Step corresponding index The amplitude at that point, For frequency amplitude synchronization sequence in path number 1 Step corresponding index The amplitude at that point, For the raceway load sequence index number, in the number of... Step corresponds to the sample position, For the frequency amplitude sequence index number, in the number of... Step corresponds to the sample position, For path direction change terms, it represents the first term. The absolute value of the angle difference between the step and the previous direction; Execution process: Extract the synchronous sequence of raceway load and frequency amplitude from the displacement trend structure matrix, denoted as follows: and The sampling length was set to 10,000 points, and linear interpolation was performed to ensure uniform time index. The time index range of each sequence was set to 0 to 10 seconds, the sampling interval was 0.001 seconds, and the weight parameters were initialized. , , , The weights are determined by the criterion of minimizing the matching error of multiple sets of sample data, and are assigned to each step. Calculate the four measures, including the absolute difference in magnitude. Difference between adjacent points Index offset item and direction change terms Multiply each item by its corresponding weight. And gradually accumulate to the total path distance. In the middle. The entire accumulation process starts from the path origin. To the finish line The process is performed sequentially, with the path direction at each step automatically updated based on the index difference between adjacent points to ensure consistent time order. The result is output after calculation. .

[0029] The specific steps for constructing a set of active signal intervals are as follows: Based on the displacement trend structure matrix and structural response distribution table, the rolling element energy ratio and speed gradient synchronization sequence are extracted, the time index partitions are aligned, the energy difference and speed increment of adjacent points are calculated, the abrupt segment index and measurement point number are merged, and a high gradient segment index table is generated. Based on the high gradient segment index table, extract the corresponding segment frequency band energy sequence and perform interval integration, calculate the frequency band energy difference to generate a difference sequence, extract the offset number to calculate the average offset, mark the energy difference and offset mean value exceeding the limit interval index, organize the measurement point number and output the signal active interval set; Based on the displacement trend structure matrix and structural response distribution table, the synchronous sequence of rolling element energy ratio and rotational speed gradient is extracted. The input sampling duration is set to 60 seconds, the sampling frequency is set to 5120 Hz, the start of the time index interval is set to 0 seconds, the end of the time index interval is set to 60 seconds, and the time step is set to 0.001 seconds. The energy data adopts a synchronous time alignment method with an alignment accuracy of 0.001 seconds and the alignment range is limited to the entire time domain. When calculating the energy difference between adjacent sampling points, a fixed interval between samples is used with an interval of 5 points. The rotational speed increment is calculated by the difference between the preceding and following points. The gradient value is the ratio of the difference result to the sampling time interval, and the gradient unit is Hz per second. The abrupt change threshold is set to 20 Hz per second. When marking the abrupt change interval, 5 consecutive sampling points exceeding the threshold are used as the minimum length of the abrupt change segment. The measurement point number range is set to 1 to 64, and the numbers are arranged according to the equipment layout order. The abrupt change segment index number is incremented in chronological order. The spacing between adjacent abrupt change segments is merged with a merging interval threshold of 0.2 seconds. The time index, abrupt change segment number, and measurement point number are summarized to generate a high gradient segment index table. Based on the high gradient segment index table, the corresponding segment frequency band energy sequence is extracted and interval integration is performed. The frequency band division range is set to 10 to 1000 Hz, the division step size is set to 10 Hz, the integration window length is set to 0.5 seconds, and the window interval is set to 0.25 seconds. The cumulative sum of frequency band energy is calculated within each window. The energy difference is calculated as the difference of cumulative values ​​between adjacent windows. The difference precision is set to three decimal places. The difference sequence length is consistent with the number of windows. The offset number is extracted by counting the number of energy difference sign changes. The offset mean is calculated by averaging the absolute offsets. The calculation precision is set to two decimal places. The energy difference threshold is set to 0.05, and the offset mean threshold is set to 0.02. The logic for judging the over-limit interval is that the interval where the energy difference is greater than the threshold and the offset mean is greater than the threshold is marked as the over-limit interval. The index number is incremented in chronological order, and the measurement point number follows the input table number. The energy difference, offset mean, over-limit index, and measurement point number are integrated into the output record sequence to generate a set of active signal intervals.

[0030] The specific steps for generating a list of abnormal bearing locations are as follows: Based on the active signal interval set, the energy ratio and phase difference sequence of the rolling body measurement points are extracted, the time index is aligned and the segments are divided, the change in energy amplitude difference and phase difference is calculated, cross-comparison is performed and the double high amplitude segment is locked, the index and measurement point number are recorded, and a double high amplitude measurement point index table is generated. Based on the dual high amplitude measurement point index table, the energy amplitude and offset amplitude sequence of continuous measurement points are extracted, the amplitude is differentially analyzed and the offset increment is statistically calculated, abrupt measurement points are screened and adjacent abnormal sections are integrated, a mapping between measurement point number and offset amplitude is established, the index order is sorted and the bearing abnormal part index list is output. Based on the active signal interval set, the energy ratio and phase difference sequences of the rolling body measurement points are extracted. The sampling time is set to 60 seconds, the sampling frequency is set to 5120 Hz, the start of the time index interval is set to 0 seconds, the end of the time index interval is set to 60 seconds, and the time step is set to 0.001 seconds. The energy ratio sequence is generated by normalizing the energy amplitude mean, and the normalization interval is set to 0 to 1. The phase difference sequence is obtained by sine and cosine inversion. The angle output range is set to -180 to 180 degrees, the angle resolution is set to 0.1 degrees, the time alignment is performed using linear interpolation, the interpolation interval is set to 0.001 seconds, the segment length is set to 0.5 seconds, and the overlap rate of adjacent segments is set to 50%. The energy amplitude difference is calculated using the absolute difference method. The amplitude is taken as the difference between the average values ​​of adjacent segments of the sample point, and the precision is set to three decimal places. The phase difference change is calculated based on the average phase difference between adjacent time periods, in degrees, and the change threshold is set to 15 degrees. The energy amplitude difference and phase difference change use cross-comparison logic. When both exceed their respective thresholds, the segment is locked as a double high amplitude segment. When the locking condition is met continuously for no less than 3 segments, it is marked as a stable double high segment. The time index number and the measurement point number are recorded. The measurement point number range is set to 1 to 64, and the numbering order is arranged according to the physical deployment order of the sensor. Finally, the recording sequence is output and a double high amplitude measurement point index table is generated. Based on the dual high-amplitude measurement point index table, the energy amplitude and offset amplitude sequences of continuous measurement points are extracted. The input data length is set to 64 measurement points, the number of measurement point samples is set to 60,000 points, the time index range is set to 0 to 60 seconds, the amplitude unit is set to g, and the offset unit is set to millimeters. When performing differential operations on the amplitude sequence, the direct difference method of adjacent sampling points is used, with a difference interval of 1 point. The absolute value of the difference result is retained to three decimal places. The offset increment is calculated by the difference of the average offset of adjacent time windows. The time window length is set to 0.5 seconds, the sliding interval is set to 0.25 seconds, and the increment threshold is set to 0.02 millimeters. When screening abrupt change measurement points... The mutation condition is defined as the measurement point number where the increment of three consecutive windows exceeds the threshold. For measurement points that meet the condition, adjacent segments are integrated with an integration interval of 0.2 seconds. Adjacent abnormal segments are merged into a single segment. The starting index is the earliest abnormal time point, and the ending index is the last abnormal time point. When establishing the mapping relationship between measurement point number and offset amplitude, the rows of the mapping matrix represent measurement point numbers, and the columns represent offset amplitude intervals. The interval division step size is set to 0.01 mm, and unobserved values ​​are filled with zeros. The index order is arranged in ascending order of time. The corresponding record sequence of time index, measurement point number, and offset amplitude is output to generate an index list of abnormal bearing parts.

[0031] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for monitoring wind turbine generator bearings based on vibration data analysis, characterized in that, Includes the following steps: S1: Record the radial and axial vibration amplitudes of the rolling element through the vibration signal sequence output by the triaxial accelerometer, perform amplitude difference through time index alignment and calculate frequency distribution, identify the stable interval index after determining the direction of speed change, and generate the dynamic characteristic sequence of the working condition. S2: Based on the dynamic feature sequence of the working condition, after extracting the stable mode components by variational mode decomposition, the rolling body energy is divided into partitions and the offset rate is calculated. The offset rate and amplitude change rate of the inner and outer rings are calculated and the high offset interval is screened to construct the displacement trend structure matrix. S3: Based on the displacement trend structure matrix, the roller load and frequency amplitude are compared after time alignment by dynamic time warping. The main and slave responses are divided by direction matching and the index is recorded. The energy ratio and phase difference are calculated. The indexes are integrated to form a propagation path table, and the structural response distribution table is obtained. S4: Based on the displacement trend structure matrix and the structure response distribution table, the rolling element energy ratio and rotational speed gradient are jointly compared. High gradient segments are identified through difference calculation, and the frequency band energy difference and offset mean are extracted. The index of the out-of-limit interval and the measurement point number are recorded to form a set of active signal intervals. S5: Based on the set of active signal intervals, determine the intersection of the energy ratio and phase difference of the rolling element measuring points, confirm the double high amplitude interval through trend matching, screen continuous abnormal measuring points, record the measuring point number and offset amplitude to form an index table, and output the bearing abnormal part index list.

2. The method for monitoring wind turbine generator bearings based on vibration data analysis according to claim 1, characterized in that, The dynamic characteristic sequence of the working condition includes a time index sequence, a vibration amplitude sequence, a frequency distribution sequence, and a rotational speed direction identifier. The displacement trend structure matrix includes a time index dimension, an energy partition dimension, and an offset rate recording unit. The structural response distribution table includes a master-slave response index, an energy ratio record, and a phase difference table. The signal active interval set includes a high gradient segment index table, a frequency band energy difference table, and an offset mean record table. The bearing abnormal part index list includes a dual high amplitude measurement point index table, a continuous abnormal measurement point set, and an offset amplitude mapping table.

3. The method for monitoring wind turbine generator bearings based on vibration data analysis according to claim 1, characterized in that, The specific steps for generating the dynamic feature sequence of the operating condition are as follows: By recording the radial and axial vibration amplitudes of the rolling body through the vibration signal sequence output by the triaxial accelerometer, the signals are extracted in segments and aligned with the time index. The amplitudes are differentially analyzed and the ratios are calculated. The index positions are marked in chronological order to generate a multi-directional vibration amplitude sequence set. Based on the multi-directional vibration amplitude sequence set, the frequency distribution is calculated and the speed change is judged. The frequency range is determined by period measurement and the direction of amplitude change is identified. After determining the speed rise and fall trend, the index position of the stable range is recorded, and the dynamic feature sequence of the working condition is generated.

4. The method for monitoring wind turbine generator bearings based on vibration data analysis according to claim 1, characterized in that, The specific steps for constructing the displacement trend structure matrix are as follows: Based on the dynamic characteristic sequence of the working condition, variational mode decomposition is used to decompose the original vibration signal into multiple stable mode components, the rolling body energy is partitioned, the energy is accumulated and the density is statistically calculated according to the time index fixed window, and the segment difference and ratio are calculated and the change index is recorded to generate a rolling body energy distribution table. Based on the rolling element energy distribution table, load difference and rotational speed offset are matched, two values ​​are paired by the same time index, the difference is calculated and the average offset is calculated by aggregating by segment, the offset direction is determined and the index position is marked to generate an offset rate matching sequence. Based on the offset rate matching sequence, the offset rate of the inner and outer circles and the amplitude change rate are calculated. The offset and vibration amplitude are extracted in each time segment and proportional calculation is performed. High offset segments are selected, and the corresponding indices are written into the matrix row and column structure to generate a displacement trend structure matrix.

5. The method for monitoring wind turbine generator bearings based on vibration data analysis according to claim 1, characterized in that, The specific execution steps of the variational mode decomposition are as follows: First, the original vibration signal corresponding to the dynamic feature sequence of the working condition is read, the sampling time range is determined, and the amplitude of the signal is normalized. Then, the signal is decomposed into independent narrowband mode components according to the frequency distribution characteristics. During the decomposition process, the decomposition step size is adjusted according to the energy distribution density to keep the frequency bands of each mode independent. After the decomposition is completed, the instantaneous amplitude and frequency information of each mode are extracted one by one, and the difference between the local energy and the average energy of each mode is calculated to form a multimode energy matrix.

6. The method for monitoring wind turbine generator bearings based on vibration data analysis according to claim 4, characterized in that, Specifically, the matching of load difference and speed offset involves first reading the load data and speed data for the same time period from the rolling element energy distribution table, establishing a correspondence between the two based on the time index, then calculating the average load for each time period and obtaining the load difference between adjacent time periods, while simultaneously calculating the speed change for the corresponding time period as the speed offset value. The load difference and speed offset are then paired according to the same time index to generate a load-speed correspondence sequence. In the sequence, the numerical pairs of adjacent index segments are aggregated and the average offset is calculated. The matching results and index positions of each segment are recorded.

7. The method for monitoring wind turbine generator bearings based on vibration data analysis according to claim 1, characterized in that, The specific steps to obtain the structural response distribution table are as follows: Based on the displacement trend structure matrix, dynamic time warping is used to align the raceway load and frequency amplitude synchronization sequence in time. The system is partitioned by time index, the amplitude direction difference is calculated and a direction sequence is generated. The direction change and the corresponding index position are marked. The main direction and the secondary direction are extracted and the matching index is recorded to generate a master-slave dynamic response index set. Based on the master-slave response index set, the load energy is accumulated according to the index interval, the frequency amplitude peak is extracted, the energy amplitude ratio is calculated to form a ratio sequence, the phase change angle is generated to form an angle sequence, the direction change index is marked, the difference matrix is ​​sorted and the energy mapping is established to generate an energy phase difference ratio table. Based on the energy phase difference ratio table, energy directions are classified according to index intervals and a continuous sequence is established. After arranging the energy change paths, turning points are marked. Adjacent intervals are integrated to establish a hierarchical mapping. The index order is verified and the numbers are filled in. The path data is summarized and the structural response distribution table is output.

8. The method for monitoring wind turbine generator bearings based on vibration data analysis according to claim 1, characterized in that, The specific execution process of the dynamic time warping is as follows: First, read the synchronization sequence of the rolling load and frequency amplitude in the displacement trend structure matrix, determine the time index range of the two sequences and establish a corresponding mapping table, then calculate the amplitude difference and frequency difference between adjacent time nodes, accumulate the differences in turn and generate a time distance matrix, compare the distance between adjacent nodes row by row in the matrix and record the matching path index, and align and resample the time axis of the two sequences according to the accumulated difference of the path valley value to generate a synchronized aligned sequence after time warping.

9. The method for monitoring wind turbine generator bearings based on vibration data analysis according to claim 1, characterized in that, The specific steps for constructing the set of active signal intervals are as follows: Based on the displacement trend structure matrix and the structure response distribution table, the rolling element energy ratio and speed gradient synchronization sequence are extracted, the time index partitions are aligned, the energy difference and speed increment of adjacent points are calculated, the abrupt segment index and measurement point number are merged, and a high gradient segment index table is generated. Based on the high gradient segment index table, extract the corresponding segment frequency band energy sequence and perform interval integration, calculate the frequency band energy difference to generate a difference sequence, extract the offset number to calculate the average offset, mark the energy difference and offset mean value exceeding the limit interval index, organize the measurement point numbers and output the signal active interval set.

10. The method for monitoring wind turbine generator bearings based on vibration data analysis according to claim 1, characterized in that, The specific steps for outputting the index list of abnormal bearing parts are as follows: Based on the set of active signal intervals, extract the energy ratio and phase difference sequence of the rolling body measuring points, align the time index and divide the segments, calculate the energy amplitude difference and phase difference change, perform cross-comparison and lock the double high amplitude segments, record the index and measuring point number, and generate a double high amplitude measuring point index table. Based on the dual high amplitude measurement point index table, the energy amplitude and offset amplitude sequence of continuous measurement points are extracted, the amplitude is differentially analyzed and the offset increment is statistically calculated, abrupt measurement points are screened and adjacent abnormal sections are integrated, a mapping between measurement point number and offset amplitude is established, the index order is sorted and the bearing abnormal part index list is output.

Images