Online Early Warning Method for Wind Turbine Blade Faults Based on Multi-Physical Quantity Sensing

By constructing a rotating phase label and a digital twin benchmark model, and combining evidence theory with multi-source fault evidence, online early warning of wind turbine blade faults was achieved. This solved the accuracy problem of early warning based on a single physical quantity in existing technologies, and improved the accuracy and reliability of fault diagnosis.

CN123109971APending Publication Date: 2026-11-17DATANG FUYU NEW ENERGY POWER GENERATION CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611509613.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-28
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing wind turbine blade fault early warning methods rely on a single physical quantity, making it difficult to achieve accurate early warning under complex operating conditions. Furthermore, the lack of a unified rotating phase reference and operating condition normalization mechanism leads to frequent missed and false alarms.

Method used

By constructing a rotating phase label as a unified time reference, collecting multi-physical quantity sensing data, preprocessing and synchronizing it, generating a working condition state space grid, constructing a blade response digital twin benchmark model, and using evidence theory to fuse multi-source fault evidence for graded early warning.

Benefits of technology

It achieves accurate alignment and fault diagnosis of multiple physical quantities under complex working conditions, significantly improving the accuracy and reliability of fault diagnosis and avoiding false alarms and missed alarms.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123109971A_ABST
    Figure CN123109971A_ABST
Patent Text Reader

Abstract

This invention discloses an online early warning method for wind turbine blade faults based on multi-physical quantity sensing, belonging to the field of wind power equipment monitoring technology. This invention acquires multi-physical quantity sensing data of wind turbine blades in real time, divides sensing channels, and constructs a rotating phase label as a unified time reference to obtain rotational cycle data frames. The multi-physical quantity sensing data within each rotational cycle data frame is preprocessed sequentially to obtain a rotational cycle synchronous data volume. The rotational cycle synchronous data volume undergoes operating condition state space reconstruction to generate an operating condition state space grid. Based on the operating condition state space grid, a blade response digital twin benchmark model is constructed to generate fault features, which are then converted into multi-source fault evidence. Based on evidence theory, the multi-source fault evidence is fused to obtain a comprehensive fault confidence level, and corresponding graded early warnings are performed. This invention effectively eliminates the interference of operating condition fluctuations on fault identification, improving the accuracy and reliability of blade fault diagnosis.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of wind power equipment monitoring technology, specifically to an online early warning method for wind turbine blade faults based on multi-physical quantity sensing. Background Technology

[0002] Wind turbine blades are subjected to complex environmental loads such as wind load, lightning strike, icing, salt spray and wake for a long time, which makes them prone to fatigue cracks, leading edge erosion, icing imbalance and loose connection. If the early warning is delayed, it may reduce power generation efficiency or even cause major safety accidents such as blade breakage and tower collapse. Existing blade fault early warning methods mostly rely on threshold judgment of a single physical quantity or offline analysis of a single sensor signal. Under complex operating conditions such as sudden wind speed changes, pitch regulation, and wake interference, sensor signals are severely affected by operating condition fluctuations and environmental noise interference, which can easily lead to missed or false alarms. The lack of a unified rotating phase reference and operating condition normalization mechanism among different physical quantities makes it difficult to form a chain of evidence that corroborates each other and conforms to the law of fault evolution, and also makes it impossible to achieve an online closed loop from anomaly detection and trend tracking to maintenance decision-making. How to achieve online early warning of wind turbine blade faults based on the coordination of multiple physical quantities under complex and variable operating conditions is a problem we need to solve. To this end, we now provide an online early warning method for wind turbine blade faults based on the perception of multiple physical quantities. Summary of the Invention

[0003] The purpose of this invention is to provide an online early warning method for wind turbine blade faults based on multi-physical quantity sensing.

[0004] The objective of this invention can be achieved through the following technical solution: an online early warning method for wind turbine blade faults based on multi-physical quantity sensing, comprising the following steps: Step S1: Collect multi-physical quantity sensing data of wind turbine blades in real time, divide the sensing channels, and then construct a rotation phase label as a unified time reference to obtain rotation cycle data frames, and map each sensing channel to the rotation phase label. Step S2: Preprocess the multi-physical quantity sensing data corresponding to each sensing channel in each rotation cycle data frame sequentially to obtain the rotation cycle synchronization data volume; Step S3: Reconstruct the operating condition state space of the rotation period synchronization data volume, generate the operating condition state space grid, and construct the blade response digital twin benchmark model based on the operating condition state space grid. Step S4: Generate fault features based on the blade response digital twin baseline model and convert the fault features into multi-source fault evidence; Step S5: Based on evidence theory, integrate multi-source fault evidence to obtain comprehensive fault confidence, and then conduct graded early warning based on comprehensive fault confidence.

[0005] Furthermore, the process of real-time acquisition of multi-physical quantity sensing data from wind turbine blades and the division of sensing channels includes: The system collects strain data, vibration data, acoustic emission data, rotating phase pulse data, wind speed, and wind direction from strain sensors, vibration acceleration sensors, acoustic emission sensors, rotary encoders, and anemometers, and reads rotor speed, generator power, pitch angle, and pitch control commands from the unit controller in real time. After adding timestamps and blade numbers, it forms multi-physical quantity sensing data. Based on the sampling method, the sensing channel is divided into a continuous waveform channel and an event channel. Strain data and vibration data are classified into the continuous waveform channel, while acoustic emission events and pitch commands are classified into the event channel.

[0006] Furthermore, the process of constructing a rotation phase label as a unified time reference to obtain rotation period data frames includes: The phase reference blade number and impeller azimuth angle are set to obtain the relative phase of the blades; Rotational phase labels are constructed using blade numbers and relative blade phases. The rotation cycle is defined as the time period corresponding to one revolution of the impeller, and the multi-physical quantity sensing data of each sensing channel is divided into rotation cycle data frames.

[0007] Furthermore, the process of mapping each sensing channel to a rotating phase tag includes: Angular domain resampling is performed on the continuous waveform channel with the relative phase of the blades as the independent variable. Before angular domain resampling, anti-aliasing filtering is performed on the continuous waveform channel. For the event channel, the trigger time of each acoustic emission event is converted into the relative phase of the blade, and the number of acoustic emission events of each blade in each rotation cycle is counted according to the phase window.

[0008] Furthermore, the process of preprocessing the multi-physical quantity sensing data corresponding to each sensing channel within each rotation cycle data frame to obtain the rotation cycle synchronization data volume includes: For the multi-physical quantity sensing data corresponding to each sensing channel in each rotation cycle data frame, outlier removal, detrending, and bandpass filtering preprocessing are performed sequentially. After preprocessing, the strain data, vibration data, acoustic emission data, as well as impeller speed, generator power, pitch angle and pitch command within the same rotational cycle data frame are reordered according to the rotational phase label to form a rotational cycle synchronization data body for each blade to complete one rotation.

[0009] Furthermore, the process of reconstructing the operating condition state space of the rotational periodic synchronization data volume to generate the operating condition state space mesh includes: Using the rotation cycle synchronization data volume as the basic sample, extract the working condition feature vector corresponding to each rotation cycle synchronization data volume; A self-organizing map neural network is used to cluster the operating condition feature vectors of all historical rotation cycle synchronization data volumes. The historical samples are then grouped into operating condition units according to the winning neurons activated in the self-organizing map neural network to generate an operating condition state space grid.

[0010] Furthermore, the process of constructing a digital twin baseline model of the blade response based on the operating condition state space grid includes: For each working condition cell in the working condition state space grid, extract the historical rotation cycle synchronization data volume within that working condition cell that is determined to be fault-free as a modeling sample. The modeling samples are divided into training samples and validation samples according to a preset ratio; The theoretical response of each measuring point at the relative phase of the blade is derived based on the aerodynamic load formation mechanism of the blade. By statistically correcting the theoretical response using training samples, the mean and standard deviation of the normal response at the relative phase of the blade at the measurement point under the working condition unit are obtained. The normal response range is verified using validation samples. When the validation coverage meets the preset requirements, the construction of the blade response digital twin benchmark model is completed. Otherwise, historical normal samples are added and statistical correction is performed again. After traversing all operating condition units, a blade response digital twin benchmark model covering each operating condition unit is obtained.

[0011] Furthermore, the process of generating fault features based on the blade response digital twin baseline model and converting these fault features into multi-source fault evidence includes: The measured physical quantities at the measuring point and the relative phase of the blade in the current rotation cycle synchronization data volume are compared with the normal response benchmark of the corresponding working condition unit to obtain the corresponding physical consistency residual. Based on the physical consistency residual, the fault feature vector corresponding to the current rotation cycle synchronization data volume is extracted. The fault feature vector is input into the evidence generator constructed by the Gaussian membership function to generate multi-source fault evidence. At the same time, sensor confidence is assigned to each sensing channel. The multi-source fault evidence includes acoustic emission evidence, strain evidence, vibration evidence and modal frequency evidence.

[0012] Furthermore, the process of obtaining comprehensive fault confidence based on evidence theory by fusing multi-source fault evidence includes: Establish an identification framework for evidence theory and construct a basic probability assignment function; The multi-source fault evidence is organized into a cross-physical quantity causal evidence chain according to the physical evolution sequence of blade damage. The basic probability allocation function and the corresponding conflict coefficient of the cross-physical quantity causal evidence chain are obtained according to the basic probability allocation function. When the conflict coefficient is lower than the preset conflict threshold, the components are merged step by step according to the physical evolution order. When the conflict coefficient is not lower than the preset conflict threshold, check whether the first phase of the multi-source fault evidence appears in the order of physical evolution. If the order is inconsistent, reduce the confidence of the sensor corresponding to the evidence source that caused the conflict, and reconstruct the basic probability allocation function according to the updated sensor confidence before fusing again. The proposition corresponding to the maximum value of the basic probability allocation function in the fusion result is used as the judgment result of the current rotation cycle synchronization data body, and the basic probability allocation function value corresponding to the proposition is used as the comprehensive fault confidence. The determination result of the current rotation cycle synchronization data volume and the corresponding comprehensive fault confidence are obtained based on the maximum value of the basic probability allocation function; A sliding window analysis is performed on the comprehensive fault confidence sequence of continuous rotation cycle synchronization data volumes. The sliding window is formed by taking the comprehensive fault confidence of several consecutive rotation cycle synchronization data volumes as the endpoint. The regression slope of the window built-in confidence over time is estimated using the least squares method to obtain the fault evolution trend index.

[0013] Furthermore, the process of tiered early warning based on comprehensive fault confidence includes: When the overall fault confidence level is not less than the third confidence threshold, the alarm status is directly output; When the overall fault confidence level is not less than the third confidence threshold and the fault evolution trend index continues to rise, an alarm status is output. When the overall fault confidence level is not less than the second confidence threshold and the fault evolution trend index is not less than the second trend threshold, an early warning status is output and an early warning information is pushed. When the overall fault confidence level is not less than the first confidence threshold or the fault evolution trend index is not less than the first trend threshold, output the attention status; Otherwise, output the normal status; Among them, the first confidence threshold, the second confidence threshold, and the third confidence threshold increase sequentially, and the first trend threshold is less than the second trend threshold.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: The present invention constructs a rotating phase label as a unified time reference, and maps the multi-physical quantity sensing data to the blade relative phase in a unified manner, thereby achieving accurate alignment of different sensing channels within the rotation cycle. The continuous waveform channel performs angular domain resampling with the blade relative phase as the independent variable, avoiding the data inaccuracy problem caused by impeller speed fluctuations in traditional time domain analysis methods. The unified phase mapping of the two channels provides a high-precision aligned data foundation for subsequent digital twin modeling and fault diagnosis based on the rotation cycle synchronous data volume. A self-organizing map network is used to cluster the operating condition feature vectors of all historical rotation cycle synchronization data volumes to generate an operating condition state space grid. Each operating condition unit corresponds to a typical operating condition. A blade response digital twin benchmark model is trained for each operating condition unit. It can calculate the normal response value and normal response standard deviation of each measuring point of the blade based on operating condition parameters such as wind speed, turbulence intensity, impeller speed, and pitch angle. By pre-reproducing the normal response fluctuations caused by changes in operating conditions, the subsequent fault feature extraction only retains the real abnormal components caused by blade faults. This effectively avoids misjudging normal operating condition fluctuations such as pitch regulation and sudden wind speed changes as blade faults, and significantly improves the accuracy of fault diagnosis. Based on evidence theory, a multi-source fault evidence fusion framework is constructed. Acoustic emission evidence, strain evidence, vibration evidence, and modal frequency evidence are organized into a cross-physical quantity causal evidence chain according to the physical evolution order of blade damage. During fusion, the evidence is fused step by step according to this physical evolution order. A conflict coefficient is introduced to verify the consistency of the evidence. This effectively avoids decision-making errors caused by false alarms from a single sensor or contradictions between multiple sources of evidence, and significantly improves the reliability of fault determination. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0016] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0017] like Figure 1 As shown, the online early warning method for wind turbine blade faults based on multi-physical quantity sensing includes the following steps: Step S1: Collect multi-physical quantity sensing data of wind turbine blades in real time, divide the sensing channels, and then construct a rotation phase label as a unified time reference to obtain rotation cycle data frames, and map each sensing channel to the rotation phase label. Step S2: Preprocess the multi-physical quantity sensing data corresponding to each sensing channel in each rotation cycle data frame sequentially to obtain the rotation cycle synchronization data volume; Step S3: Reconstruct the operating condition state space of the rotation period synchronization data volume, generate the operating condition state space grid, and construct the blade response digital twin benchmark model based on the operating condition state space grid. Step S4: Generate fault features based on the blade response digital twin baseline model and convert the fault features into multi-source fault evidence; Step S5: Based on evidence theory, integrate multi-source fault evidence to obtain comprehensive fault confidence, and then conduct graded early warning based on comprehensive fault confidence.

[0018] It should be further explained that, in the specific implementation process, the real-time acquisition of multi-physical quantity sensing data from wind turbine blades and the division of sensing channels include: Strain sensors, vibration acceleration sensors, and acoustic emission sensors are deployed in the root, middle, and tip sections of the wind turbine blades, respectively. A rotary encoder is deployed at the hub, and an anemometer is deployed on the top of the nacelle. The system collects strain data, vibration data, acoustic emission data, rotating phase pulse data, wind speed, and wind direction from strain sensors, vibration acceleration sensors, acoustic emission sensors, rotary encoders, and anemometers, and reads rotor speed, generator power, pitch angle, and pitch control commands from the unit controller in real time. After adding timestamps and blade numbers, it forms multi-physical quantity sensing data. Among them, the multi-physical quantity sensing data is a set of data obtained by uniformly adding timestamps and blade numbers to the original data of each sensing channel, the impeller speed is the number of revolutions per minute of the impeller, the generator power is the active power currently output by the wind turbine, the pitch angle is the angle between the blade chord and the impeller rotation plane, and the pitch command is the target pitch angle output by the unit's pitch control system. Based on the sampling method, the sensing channel is divided into a continuous waveform channel and an event channel. The sensing channel refers to the data acquisition channel corresponding to the sensing data of multiple physical quantities. Strain data and vibration data are signals that are continuously sampled at a fixed sampling rate and are classified as continuous waveform channels. Acoustic emission events and pitch control commands are recorded according to the time of occurrence and categorized into event channels; Among them, an acoustic emission event is a single transient acoustic emission record obtained after triggering and identifying acoustic emission data. Each acoustic emission event records at least the trigger time, peak amplitude, event energy, duration, and the blade number to which it belongs.

[0019] It should be further explained that, in the specific implementation process, the process of constructing a rotating phase label as a unified time reference to obtain rotating period data frames includes: Let the phase reference blade be numbered as The impeller azimuth angle is The blade number is If the rotational phase interval between adjacent blades is 120°, then the first... The relative phase of the leaf blades is denoted as ,Right now: ; The impeller azimuth angle is the impeller rotation angle accumulated by the rotary encoder with zero position as reference. The rotation phase is the angular position of the blade during impeller rotation. The relative phase of the blade is the [missing information]. The phase value obtained after converting the branch blade relative to the phase reference blade; The phase reference blade is a blade pre-specified during impeller assembly to establish a rotating phase reference, and its radial axis is aligned with the vertical upward direction and corresponds to the zero position of the rotary encoder; Construct a rotational phase label, which includes the blade number and the relative phase of the blade. The relative phase of the blade is expressed in degrees or radians, with a value range of 0° to 360°. Divide the 360° phase interval of one revolution of the impeller into three 120° sub-intervals according to the blade number. Each sub-interval corresponds to the complete rotational phase range of one blade. The rotating phase label is a unified identifier that identifies the blade number and the relative phase of the blade corresponding to the sampling time, and is used as a unified reference for aligning various physical quantities in time and space. Using the time period corresponding to one revolution of the impeller as the rotation cycle, the multi-physical quantity sensing data of each sensing channel is divided into rotation cycle data frames. Each rotation cycle data frame retains the original sampling time of each sensing channel and the corresponding impeller azimuth angle. The rotation cycle data frame is a data container that takes one revolution of the impeller as the rotation cycle, contains the original sampling time and the impeller azimuth angle, and has not yet completed angular domain resampling.

[0020] It should be further explained that, in the specific implementation process, the process of mapping each sensing channel to the rotating phase label includes: The continuous waveform channel performs angular domain resampling with the relative phase of the blades as the independent variable. The purpose of this angular domain resampling is to convert the non-uniform time-domain sampled data into data with uniform intervals in the rotating phase domain, for the target phase. Its linear interpolation formula is: ; in, Indicates the target phase The signal value obtained by resampling, Indicates the target phase to be interpolated. and These represent the two original sampled phases adjacent to the target phase, and These represent the original sampling phases. and The corresponding signal value at that location; When the target phase is less than the first original sampling phase in the rotation period or greater than the last original sampling phase in the rotation period, the original sampling point at the other end of the rotation period is used for interpolation according to the periodicity of the phase. When the target phase crosses the 0° / 360° phase seam, the end sample of this rotation cycle and the beginning sample of the next rotation cycle are regarded as adjacent samples for interpolation according to the phase periodicity. Before corner domain resampling, anti-aliasing filtering is performed on the continuous waveform channel. The function of the anti-aliasing filter is to filter out high-frequency components that may cause frequency aliasing after corner domain resampling. Specifically: Let the original sampling frequency of the continuous waveform channel be... The time it takes for the impeller to rotate once is The equivalent time sampling frequency after corner domain resampling is Anti-aliasing filtering uses a low-pass filter with a cutoff frequency of satisfy: ; in, The preset maximum impeller rotation period, This indicates the preset number of equally spaced resampling points within a single rotation cycle; the filter uses a zero-phase low-pass filter. The equivalent time sampling frequency is the equivalent time domain sampling frequency calculated by the impeller rotation period after angular domain resampling, and the same number of equally spaced resampling points are set within a single rotation period. The acoustic emission event channel does not perform continuous waveform interpolation. Instead, it converts the trigger time of each acoustic emission event into the relative phase of the blade and counts them according to the phase window, thereby avoiding physical meaningless interpolation of discrete events. The phase window refers to the window obtained by dividing the relative phase interval of the blades from 0° to 360° within a single impeller rotation cycle according to a preset phase width. When converting the trigger time of each acoustic emission event into the relative phase of the blade, first convert the impeller azimuth angle corresponding to the trigger time of the acoustic emission event into the relative phase of the blade, and then classify it into the corresponding rotation period data frame according to the blade to which the acoustic emission event belongs. When the same acoustic emission event spans multiple phase windows or crosses the 0° / 360° phase seam, it is assigned once according to the phase window where the event is triggered or the peak phase is located, and is not counted repeatedly. The number of acoustic emission events per blade and per rotation cycle is counted according to the assigned results.

[0021] It should be further explained that, in the specific implementation process, the process of preprocessing the multi-physical quantity sensing data corresponding to each sensing channel within each rotation cycle data frame to obtain the rotation cycle synchronization data volume includes: The multi-physical quantity sensing data corresponding to each sensing channel in each rotation cycle data frame are preprocessed sequentially to obtain a preprocessed signal. The preprocessing includes outlier removal processing, detrending processing, and bandpass filtering preprocessing. For continuous waveform data of each sensing channel, first calculate the median of the waveform data of that channel, then calculate the median of the absolute deviation of each data point from the median, and take the interval of the absolute median difference plus or minus a preset multiple as the effective data range. Data points that fall outside this range are judged as outliers and removed. After removal, linear interpolation is used to fill the missing positions. For the waveform data after outlier removal, a low-order polynomial fitting method is used to extract the trend term of the signal. The trend term reflects the low-frequency component in the waveform data that changes slowly with the operating parameters. The least squares fitting is performed with the sampling time series as the independent variable and the waveform amplitude as the dependent variable to obtain the trend term estimate. The original waveform data is subtracted from the trend term estimate to obtain the waveform data after removing the trend term. For the detrended waveform data, the passband range of the bandpass filter is set, and a zero-phase digital filter is used for filtering; After preprocessing, the strain data, vibration data, acoustic emission data, as well as impeller speed, generator power, pitch angle and pitch command within the same rotation cycle data frame are reordered according to the rotation phase label to form a rotation cycle synchronization data body for each blade to complete one rotation. The rotation period synchronization data volume is a data set organized by blade number for three blades within the same rotation period data frame under the same rotation phase label.

[0022] It should be further explained that, in the specific implementation process, the process of reconstructing the operating condition state space of the rotation period synchronization data volume and generating the operating condition state space mesh includes: Using the rotational cycle synchronization data volume as the basic sample, the operating condition feature vector corresponding to each rotational cycle synchronization data volume is extracted. This operating condition feature vector consists of wind speed, turbulence intensity, impeller speed, generator power, pitch angle, and pitch command, denoted as […]. ,Right now: ; in, Indicates wind speed. Indicates turbulence intensity. Indicates the impeller speed. Indicates generator power. Indicates the pitch angle. Indicates pitch control command; The wind speed is the incoming wind speed measured by the aforementioned anemometer, and the turbulence intensity is the ratio of the standard deviation of the wind speed to the average wind speed. A self-organizing mapping neural network is used to cluster the operating condition feature vectors of all historical rotation cycle synchronization data volumes to generate an operating condition state space grid. The operating condition state space grid is a set of operating condition units obtained by merging historical samples according to the similarity of operating conditions. Each operating condition unit corresponds to a set of similar operating conditions. Specifically: Self-organizing map neural networks are a type of unsupervised competitive learning neural network that can map high-dimensional operating condition feature vectors to neurons on a two-dimensional topological structure. Adjacent neurons correspond to similar operating conditions. The number of nodes in the input layer of the self-organizing map neural network is equal to the dimension of the operating condition feature vector. The output layer adopts an m×m grid structure, where m is the number of rows and columns of neurons in the output layer, and m≥5. After training, for any input sample, the Euclidean distance between it and the weight vectors of each neuron in the output layer is calculated. The neuron with the smallest distance is defined as the winning neuron of the sample. Historical samples are grouped into operating condition units according to the activated winning neurons. The input sample is the operating condition feature vector of the historical rotation cycle synchronization data volume; The historical rotation cycle synchronization data volume is merged according to the neurons activated in the self-organizing map neural network based on its operating condition feature vector to obtain each operating condition unit. To avoid forcibly assigning new samples that are far from existing working cells to cells that are too far away, the activation distance is calculated and denoted as . ,Right now: ; in, The input sample is a feature vector representing the operating conditions. The weight vector of the winning neuron; The activation distance is used to measure the degree of deviation between the input sample and the cluster center represented by the winning neuron. The smaller the activation distance, the higher the degree of matching between the input sample and the working unit. When the activation distance is less than the preset activation distance threshold, the input sample will be assigned to the corresponding working condition unit. When the activation distance is not less than the preset activation distance threshold, the input sample is marked as a sample outside the working condition coverage and the extended acquisition mode is started. Wherein, the samples outside the working condition coverage are input samples that are not classified into any working condition unit but whose activation distance is not less than a preset activation distance threshold; the extended acquisition mode is to increase the sampling frequency and extend the continuous acquisition time for the working condition interval with insufficient working condition coverage. Operating condition coverage is obtained by measuring the proportion of historical samples within a single operating condition unit to the total number of historical samples, denoted as . ,Right now: ; in, Indicates operating condition unit Number of historical samples This indicates the total number of historical samples, where the historical samples are historical rotation cycle synchronization data volumes. The operating condition coverage is used to characterize the sample proportion of each operating condition unit in all historical data. A low coverage of an operating condition unit indicates that the operating condition has a low frequency of occurrence in historical operation or that the sample acquisition is insufficient. When the operating condition coverage of the operating condition unit to which the real-time sample belongs is lower than the preset coverage threshold, the sampling frequency of that operating condition interval is increased and the continuous collection time is extended.

[0023] It should be further explained that, in the specific implementation process, the process of constructing a digital twin benchmark model of the blade response based on the operating condition state space grid includes: The blade response digital twin benchmark model is a set of blade normal response models covering each operating condition unit, with each operating condition unit corresponding to a set of normal response mean and normal response standard deviation. For each working condition cell in the working condition state space grid, extract the historical rotation cycle synchronization data volume within that working condition cell that is determined to be fault-free as a modeling sample. The modeling sample includes the operating condition feature vector corresponding to each rotation cycle under the operating condition unit and the multi-physical quantity sensing data of each measuring point on the blade aligned with the rotation phase label, wherein the measuring point is the monitoring position of the sensor channel deployed on the blade. The modeling samples for each working condition unit are divided into training samples and validation samples according to a preset ratio; Based on the aerodynamic load formation mechanism of the blade, the theoretical response at each measuring point at the relative phase of the blade is derived. During the rotation of the blade, it is subjected to gravity, centrifugal force, aerodynamic thrust, aerodynamic tangential force, and periodic loads caused by wind shear and tower shadow effect. The theoretical response is denoted as... ,Right now: ; in, Indicates the measurement point number. Indicates the blade number. Represents a constant term. This represents the amplitude coefficient of the cosine component of the h-th harmonic. This represents the amplitude coefficient of the sinusoidal component of the h-th harmonic. This indicates the relative phase of the j-th blade. Represents the harmonic order variable. This indicates the highest order of the harmonics involved in the fitting; Subsequently, statistical corrections were performed on the theoretical response using training samples to obtain the mean and standard deviation of the normal response at the relative phase of the blades under the operating condition unit, i.e.: ; ; in, This indicates that after statistical correction, the measurement points... relative phase of the blade Operating unit The mean of the normal response under the given conditions. Indicates the amplitude correction factor. This represents the bias correction coefficient. For the first Training samples at measurement points relative phase of the blades The actual response at the location, This represents the number of corresponding training samples, where n represents the training sample index. Verification is performed using a validation sample. If the actual response meets the formula for judging the normal response range, then the validation sample is determined to fall within the normal response range, i.e.: ; in, This indicates that the verification sample is at the measurement point. relative phase of the blades The actual response value at that location, Indicates operating condition unit Below, measuring point relative phase of the blade The standard deviation of the normal response at the given location, where λ is a preset multiple; The validation coverage rate is defined as the proportion of the number of validation samples that meet the above conditions to the total number of validation samples. When the verification coverage meets the preset requirements, the construction of the blade response digital twin benchmark model is completed; otherwise, historical normal samples are added and statistical correction is performed again. After traversing all operating condition units, a digital twin reference model of the blade response covering each operating condition unit is obtained. Real-time samples call the corresponding reference model according to their respective operating condition units.

[0024] It should be further explained that, in the specific implementation process, the process of generating fault features based on the blade response digital twin baseline model and converting the fault features into multi-source fault evidence includes: The measured physical quantities at the measuring point and blade relative phase of the current rotation cycle synchronization data volume are compared with the normal response benchmark of the corresponding working condition unit. The normalized physical consistency residual is calculated and denoted as... ,Right now: ; in, This indicates the current rotation cycle synchronization data volume at the measuring point. relative phase of the blades The measured physical quantity value at the location, and These are the mean and standard deviation of the normal response output by the digital twin benchmark model of the blade response for the respective operating condition unit; Physical consistency residuals are used to eliminate normal response fluctuations caused by changes in operating conditions and to highlight abnormal components caused only by blade failures. Based on the physical consistency residual, four types of fault features are extracted: blade root load imbalance deviation, vibration frequency band energy ratio deviation, acoustic emission event rate deviation, and modal frequency shift, which constitute a fault feature vector. The blade root load imbalance deviation is calculated by converting the strain data at the blade root of the three blades into the blade root bending moment amplitude after calibration, and then calculating the bending moment difference between each pair of the three blades. The deviation is determined based on the current measured bending moment difference and the normal blade root bending moment difference in the normal response reference of the corresponding working condition unit. Vibration data is extracted from the rotation period synchronization data volume. Two preset frequency bands are set respectively. The energy of the vibration data in the two preset frequency bands is calculated respectively. The ratio of the energy of the two frequency bands is taken as the vibration frequency band energy ratio. The deviation between the current measured energy ratio and the normal vibration frequency band energy ratio in the normal response benchmark of the unit is taken as the vibration frequency band energy ratio deviation. The acoustic emission event rate deviation is the ratio of the number of acoustic emission events in the current rotational cycle synchronization data body to the duration of the rotational cycle synchronization data body, and the deviation from the normal acoustic emission event rate in the normal response benchmark of the corresponding working condition unit. Vibration data is extracted from the rotation period synchronization data volume, and the first-order modal frequency of the blade is extracted using fast Fourier transform. The difference between the current measured modal frequency and the normal value of the modal frequency in the normal response reference of the unit under the corresponding working condition is taken as the modal frequency offset. The fault feature vector is input into the evidence generator constructed by the Gaussian membership function to generate acoustic emission evidence, strain evidence, vibration evidence and modal frequency evidence respectively; The acoustic emission evidence, strain evidence, vibration evidence, and modal frequency evidence are integrated to form multi-source fault evidence; The evidence generator operates independently for each type of fault feature. The input of each evidence generator is the fault feature value of that type, and the output is the evidence support degree corresponding to that fault feature value. The fault characteristic value refers to the specific numerical value of the above four types of fault characteristics at the current moment. It is a collective term for the four types of fault characteristics. For a certain fault characteristic value, the corresponding evidence support is denoted as... ,Right now: ; in, , This represents the average value of the corresponding fault characteristics under healthy blade conditions. This represents the standard deviation of the corresponding fault characteristics under healthy blade conditions; The further the fault characteristic value deviates from the healthy mean, the closer the evidence support is to 1; Simultaneously, a sensor confidence level is assigned to each sensing channel, with all preset initial values ​​set to [value missing]. .

[0025] It should be further explained that, in the specific implementation process, the process of obtaining the comprehensive fault confidence based on the fusion of multi-source fault evidence according to evidence theory includes: First, establish a framework for identifying evidence theories, denoted as: ,Right now: ; in, For the proposition of no-fault blades, to Propose a problem for a predetermined type of failure. The number of preset fault types; Construct the basic probability assignment function, denoted as ,Right now: ; Where j is the source of evidence number, This represents the corresponding candidate fault proposition. This indicates the degree of support of the corresponding evidence. This indicates the corresponding sensor reliability. The candidate fault proposition is the fault type proposition most likely supported by the j-th evidence source based on its own fault characteristics, and the evidence support is the evidence support obtained after the fault characteristics corresponding to the j-th evidence source are transformed by the evidence generator. The acoustic emission evidence, strain evidence, vibration evidence, and modal frequency evidence are organized into a chain of causal evidence across physical quantities according to the physical evolution of blade damage. Specifically: Acoustic emission evidence reflects the initiation of internal damage in the material; strain or blade root load evidence reflects the redistribution of structural loads caused by damage; vibration evidence reflects the changes in vibration characteristics caused by the decrease in structural stiffness; and modal frequency evidence reflects the physical evolution of the late-stage manifestations of the overall stiffness degradation of the blade. The fusion order constraint obtained from the multi-source evidence is arranged in the order of the physical evolution. During the fusion process, acoustic emission evidence and strain evidence are first fused together, the fusion result is then fused together with vibration evidence, and finally fused together with modal frequency evidence. Let the basic probability assignment functions of the two evidence sources to be fused be respectively and Then the fused basic probability allocation function is denoted as ,Right now: ; in, The goal proposition after integration, Jiao Yuan, as the primary source of evidence, Jiao Yuan, as the second source of evidence, The conflict coefficient; The focal element is the set of propositions supported by the basic probability assignment function; The conflict coefficient is used to measure the degree of conflict between two sources of evidence. satisfy: ; When the conflict coefficient is lower than the preset conflict threshold, the components are merged step by step in the above order. When the conflict coefficient is not lower than the preset conflict threshold, check whether the initial phase of the multi-source fault evidence (i.e., acoustic emission evidence, strain evidence, vibration evidence and modal frequency evidence) appears in the physical evolution order of acoustic emission evidence, strain evidence, vibration evidence and modal frequency evidence. The initial phase refers to the relative phase of the blade where the evidence of the multi-source fault first appears. If the order is consistent, the fusion result is retained. If the order is inconsistent, the confidence of the sensor corresponding to the source of conflicting evidence is reduced, and the basic probability allocation function is reconstructed according to the updated sensor confidence before fusion is performed again. The proposition corresponding to the maximum value of the basic probability allocation function in the fusion result is used as the judgment result of the current rotation cycle synchronization data body, and the basic probability allocation function value corresponding to the proposition is used as the comprehensive fault confidence, wherein the comprehensive fault confidence is the comprehensive credibility of the blade occurring the corresponding fault type after cross-physical quantity fusion. A sliding window analysis is performed on the comprehensive fault confidence sequence of the continuous rotation cycle synchronization data volume, taking the current rotation cycle synchronization data volume as the endpoint, and taking the continuous... The combined fault confidence of each rotational cycle synchronization data volume constitutes a sliding window; in, The length of the sliding window and ; The regression slope of the window's built-in confidence over time is estimated using the least squares method to obtain the fault evolution trend index, denoted as . ,Right now: ; in, Let i be the time sequence number of the synchronization data volume for the i-th rotation cycle within the sliding window. This corresponds to the overall fault confidence level; The fault evolution trend index is used to characterize the upward or downward trend of the overall fault confidence over time.

[0026] It should be further explained that, in the specific implementation process, the tiered early warning process based on the comprehensive fault confidence level includes: The tiered early warning system uses a priority order from high to low: The graded early warning signals are classified into normal state, attention state, early warning state and alarm state according to the severity of the fault, from low to high. When the overall fault confidence level is not less than the third confidence threshold, the alarm status is directly output; When the overall fault confidence is not less than the third confidence threshold and the fault evolution trend index continues to rise, an alarm status is output. The continuous rise means that the fault evolution trend index is greater than zero and increases sequentially within multiple consecutive sliding windows. When the overall fault confidence level is not less than the second confidence threshold and the fault evolution trend index is not less than the second trend threshold, an early warning status is output and an early warning information is pushed. When the overall fault confidence level is not less than the first confidence threshold or the fault evolution trend index is not less than the first trend threshold, output the attention status; Otherwise, output the normal status; Among them, the first confidence threshold, the second confidence threshold, and the third confidence threshold increase sequentially, and the first trend threshold is less than the second trend threshold.

[0027] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any modifications or equivalent substitutions 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 scope of the present invention.

Claims

1. A method for online early warning of wind turbine blade faults based on multi-physical quantity sensing, characterized in that, Includes the following steps: Step S1: Collect multi-physical quantity sensing data of wind turbine blades in real time, divide the sensing channels, and then construct a rotation phase label as a unified time reference to obtain rotation cycle data frames, and map each sensing channel to the rotation phase label. Step S2: Preprocess the multi-physical quantity sensing data corresponding to each sensing channel in each rotation cycle data frame sequentially to obtain the rotation cycle synchronization data volume; Step S3: Reconstruct the operating condition state space of the rotation period synchronization data volume, generate the operating condition state space grid, and construct the blade response digital twin benchmark model based on the operating condition state space grid. Step S4: Generate fault features based on the blade response digital twin baseline model and convert the fault features into multi-source fault evidence; Step S5: Based on evidence theory, integrate multi-source fault evidence to obtain comprehensive fault confidence, and then conduct graded early warning based on comprehensive fault confidence.

2. The online early warning method for wind turbine blade faults based on multi-physical quantity sensing according to claim 1, characterized in that, The process of real-time acquisition of multi-physical quantity sensing data from wind turbine blades and the division of sensing channels includes: The system collects strain data, vibration data, acoustic emission data, rotating phase pulse data, wind speed, and wind direction from strain sensors, vibration acceleration sensors, acoustic emission sensors, rotary encoders, and anemometers, and reads rotor speed, generator power, pitch angle, and pitch control commands from the unit controller in real time. After adding timestamps and blade numbers, it forms multi-physical quantity sensing data. Based on the sampling method, the sensing channel is divided into a continuous waveform channel and an event channel. Strain data and vibration data are classified into the continuous waveform channel, while acoustic emission events and pitch commands are classified into the event channel.

3. The online early warning method for wind turbine blade faults based on multi-physical quantity sensing according to claim 2, characterized in that, The process of constructing a rotation phase label as a unified time reference to obtain rotation period data frames includes: The phase reference blade number and impeller azimuth angle are set to obtain the relative phase of the blades; Rotational phase labels are constructed using blade numbers and relative blade phases. The rotation cycle is defined as the time period corresponding to one revolution of the impeller, and the multi-physical quantity sensing data of each sensing channel is divided into rotation cycle data frames.

4. The online early warning method for wind turbine blade faults based on multi-physical quantity sensing according to claim 3, characterized in that, The process of mapping each sensor channel to a rotating phase label includes: Angular domain resampling is performed on the continuous waveform channel with the relative phase of the blades as the independent variable. Before angular domain resampling, anti-aliasing filtering is performed on the continuous waveform channel. For the event channel, the trigger time of each acoustic emission event is converted into the relative phase of the blade, and the number of acoustic emission events of each blade in each rotation cycle is counted according to the phase window.

5. The online early warning method for wind turbine blade faults based on multi-physical quantity sensing according to claim 4, characterized in that, The process of preprocessing the multi-physical quantity sensing data corresponding to each sensing channel within each rotation cycle data frame to obtain the rotation cycle synchronization data volume includes: For the multi-physical quantity sensing data corresponding to each sensing channel in each rotation cycle data frame, outlier removal, detrending, and bandpass filtering preprocessing are performed sequentially. After preprocessing, the strain data, vibration data, acoustic emission data, as well as impeller speed, generator power, pitch angle and pitch command within the same rotational cycle data frame are reordered according to the rotational phase label to form a rotational cycle synchronization data body for each blade to complete one rotation.

6. The online early warning method for wind turbine blade faults based on multi-physical quantity sensing according to claim 5, characterized in that, The process of reconstructing the operating condition state space of the rotational periodic synchronization data volume and generating the operating condition state space mesh includes: Using the rotation cycle synchronization data volume as the basic sample, extract the working condition feature vector corresponding to each rotation cycle synchronization data volume; A self-organizing map neural network is used to cluster the operating condition feature vectors of all historical rotation cycle synchronization data volumes. The historical samples are then grouped into operating condition units according to the winning neurons activated in the self-organizing map neural network to generate an operating condition state space grid.

7. The online early warning method for wind turbine blade faults based on multi-physical quantity sensing according to claim 6, characterized in that, The process of constructing a digital twin baseline model of blade response based on the operating condition state space grid includes: For each working condition cell in the working condition state space grid, extract the historical rotation cycle synchronization data volume within that working condition cell that is determined to be fault-free as a modeling sample. The modeling samples are divided into training samples and validation samples according to a preset ratio; The theoretical response of each measuring point at the relative phase of the blade is derived based on the aerodynamic load formation mechanism of the blade. By statistically correcting the theoretical response using training samples, the mean and standard deviation of the normal response at the relative phase of the blade at the measurement point under the working condition unit are obtained. The normal response range is verified using validation samples. When the validation coverage meets the preset requirements, the construction of the blade response digital twin benchmark model is completed. Otherwise, historical normal samples are added and statistical correction is performed again. After traversing all operating condition units, a blade response digital twin benchmark model covering each operating condition unit is obtained.

8. The online early warning method for wind turbine blade faults based on multi-physical quantity sensing according to claim 7, characterized in that, The process of generating fault features based on the blade response digital twin baseline model and converting these fault features into multi-source fault evidence includes: The measured physical quantities at the measuring point and the relative phase of the blade in the current rotation cycle synchronization data volume are compared with the normal response benchmark of the corresponding working condition unit to obtain the corresponding physical consistency residual. Based on the physical consistency residual, the fault feature vector corresponding to the current rotation cycle synchronization data volume is extracted. The fault feature vector is input into the evidence generator constructed by the Gaussian membership function to generate multi-source fault evidence. At the same time, sensor confidence is assigned to each sensing channel. The multi-source fault evidence includes acoustic emission evidence, strain evidence, vibration evidence and modal frequency evidence.

9. The online early warning method for wind turbine blade faults based on multi-physical quantity sensing according to claim 8, characterized in that, The process of obtaining a comprehensive fault confidence score by fusing multi-source fault evidence based on evidence theory includes: Establish an identification framework for evidence theory and construct a basic probability assignment function; The multi-source fault evidence is organized into a cross-physical quantity causal evidence chain according to the physical evolution sequence of blade damage. The basic probability allocation function and the corresponding conflict coefficient of the cross-physical quantity causal evidence chain are obtained according to the basic probability allocation function. When the conflict coefficient is lower than the preset conflict threshold, the components are merged step by step according to the physical evolution order. When the conflict coefficient is not lower than the preset conflict threshold, check whether the first phase of the multi-source fault evidence appears in the order of physical evolution. If the order is inconsistent, reduce the confidence of the sensor corresponding to the evidence source that caused the conflict, and reconstruct the basic probability allocation function according to the updated sensor confidence before fusing again. The proposition corresponding to the maximum value of the basic probability allocation function in the fusion result is used as the judgment result of the current rotation cycle synchronization data body, and the basic probability allocation function value corresponding to the proposition is used as the comprehensive fault confidence. The determination result of the current rotation cycle synchronization data volume and the corresponding comprehensive fault confidence are obtained based on the maximum value of the basic probability allocation function; A sliding window analysis is performed on the comprehensive fault confidence sequence of continuous rotation cycle synchronization data volumes. The sliding window is formed by taking the comprehensive fault confidence of several consecutive rotation cycle synchronization data volumes as the endpoint. The regression slope of the window built-in confidence over time is estimated using the least squares method to obtain the fault evolution trend index.

10. The online early warning method for wind turbine blade faults based on multi-physical quantity sensing according to claim 9, characterized in that, The process of issuing graded early warnings based on comprehensive fault confidence includes: When the overall fault confidence level is not less than the third confidence threshold, the alarm status is directly output; When the overall fault confidence level is not less than the third confidence threshold and the fault evolution trend index continues to rise, an alarm status is output. When the overall fault confidence level is not less than the second confidence threshold and the fault evolution trend index is not less than the second trend threshold, an early warning status is output and an early warning information is pushed. When the overall fault confidence level is not less than the first confidence threshold or the fault evolution trend index is not less than the first trend threshold, output the attention status; Otherwise, output the normal status; Among them, the first confidence threshold, the second confidence threshold, and the third confidence threshold increase sequentially, and the first trend threshold is less than the second trend threshold.