Fan blade health state quantitative evaluation method, device and medium based on working condition adaptive multi-source fusion

Abstract

The application relates to the field of fan blade monitoring, and discloses a fan blade health state quantitative evaluation method, equipment and medium based on self-adaptive multi-source fusion of working conditions, which comprises the following steps: multi-source data synchronous acquisition; multi-index fusion intelligent working condition sensing; feature normalization; feature extraction and signal quality evaluation; self-adaptive dynamic weighted fusion based on an attention mechanism; health index calculation and intelligent state evaluation; condition-triggered intelligent baseline self-learning; and operation and maintenance decision output. The application significantly improves the classification accuracy under wind speed fluctuation and transition working conditions.

Description

Technical Field

[0001] This invention relates to the field of wind turbine blade monitoring, and in particular to a method, equipment, and medium for quantitative assessment of the health status of wind turbine blades based on condition-adaptive multi-source fusion. Background Technology

[0002] As a core component of wind turbine generators, the health of wind turbine blades directly affects the safety and power generation efficiency of the unit. However, traditional testing techniques have significant limitations: Static inspection is highly dependent on static inspection, while dynamic defects are difficult to identify. Existing image inspection methods (such as UAV inspection) require the blades to be stopped and fixed in position, which cannot reproduce the dynamic stress (such as wind load and centrifugal force) in actual operation. This results in a missed detection rate of more than 30% for dynamically expanding defects such as microcracks and delamination. Defects introduced in the blade layup process may be closed under static conditions and are difficult to identify by surface texture or tapping methods alone.

[0003] A single data source lacks information, leading to a high false alarm rate. Monitoring models based on SCADA data (such as power curve analysis) are not sensitive enough to early faults, while vibration analysis is easily affected by environmental noise. Although existing technologies integrate image and vibration data, they do not achieve dynamic correlation between acoustic signatures, SCADA data, and environmental parameters, resulting in a high false alarm rate.

[0004] The model has poor generalization ability and high maintenance costs. Traditional acoustic emission technology requires independent modeling for different blade models, relies on a massive number of fault samples, and cannot adapt to feature drift caused by equipment aging. Summary of the Invention

[0005] The purpose of this invention is to propose a method, equipment, and medium for quantitative assessment of the health status of wind turbine blades based on condition-adaptive multi-source fusion, which solves the technical problems of high false alarm rate, poor generalization ability, and difficulty in identifying dynamic defects in traditional wind turbine blade monitoring methods.

[0006] Specifically, this invention provides a method for quantitatively assessing the health status of wind turbine blades based on condition-adaptive multi-source fusion, comprising the following steps: S1. Synchronously collect multi-source sensor data during the operation of the wind turbine, including acoustic fingerprint signals, vibration signals and SCADA data, and achieve time synchronization through the IEEE 1588 protocol; S2. Construct a multi-dimensional working condition feature vector based on multi-source sensor data, classify the current working condition based on a lightweight neural network, calculate the working condition identification confidence level, and obtain the working condition category and working condition change indicator. S3. Extract feature values ​​from the acoustic signature signal, vibration signal and SCADA data respectively, and calculate the signal quality index of each signal source; S4. Normalize the extracted features according to the current working condition category so that they follow a standard normal distribution; S5. Input the operating condition category, signal quality index, and operating condition change flag into the dynamic weight generation network to generate adaptive weights for each feature source. S6. Calculate the Health Index (HI) and perform trend analysis and multi-level alarm judgment based on historical HI sequences; S7. Trigger the baseline self-learning mechanism when the preset conditions are met to update the statistical parameters required for feature normalization. S8. Output health status assessment results and operation and maintenance suggestions.

[0007] A storage device that stores instructions and data for implementing a method for quantitatively assessing the health status of wind turbine blades based on condition-adaptive multi-source fusion.

[0008] A device for quantitatively assessing the health status of wind turbine blades based on adaptive multi-source fusion under operating conditions includes: a processor and a storage device; the processor loads and executes instructions and data in the storage device to implement a method for quantitatively assessing the health status of wind turbine blades based on adaptive multi-source fusion under operating conditions.

[0009] The beneficial effects provided by this invention are: Intelligent operating condition perception technology that integrates multiple indicators: It abandons the traditional fixed threshold classification method that relies solely on wind speed and rotational speed, and innovatively introduces multi-dimensional operating parameters such as turbulence intensity, pitch angle change rate and wind direction change rate. It achieves accurate identification of operating conditions through a lightweight neural network, which significantly improves the classification accuracy under wind speed fluctuations and transitional operating conditions.

[0010] Dynamic weight generation technology based on attention mechanism: Completely abandoning the conventional approach of preset weight table, a weight generation network based on attention mechanism is proposed, which can dynamically generate fusion weights according to real-time operating conditions and the signal quality of each sensor, so that the system can automatically adjust the evaluation strategy under different operating conditions and signal interference conditions, and achieve true "operating condition adaptive" fusion.

[0011] Time-series health index analysis and multi-level early warning technology: Breaking through the limitations of single threshold assessment, it innovatively introduces a time-series health index analysis mechanism. By monitoring the evolution trend and rate of change of health status, it realizes a three-level alarm mechanism of early warning, alarm and emergency, which greatly improves the ability to detect early faults and reduces the false alarm rate.

[0012] Condition-triggered intelligent baseline self-learning technology: A strict multi-condition triggering mechanism is designed (30 consecutive days of health status, no alarm records, and small fluctuation range). Combined with the exponential moving average update algorithm and update amplitude limit, it ensures that the baseline update is based only on high-quality health data, effectively adapts to feature drift caused by long-term device aging, and avoids system instability caused by erroneous updates.

[0013] Closed-loop adaptive assessment system architecture: The above-mentioned innovations are organically integrated to build a complete closed-loop system of "condition perception - feature extraction - dynamic fusion - status assessment - baseline learning". It realizes full-process automation from data collection to operation and maintenance decision-making, and solves key problems in existing technologies such as low condition recognition accuracy, rigid fusion strategy, lack of time series analysis in assessment and imperfect baseline update mechanism. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of the process of this invention; Figure 2 This is a schematic diagram of the hardware device used in this application. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0016] Before formally describing the present invention, in order to better facilitate understanding of this application, this application will first provide a general description of conventional solutions for ease of understanding.

[0017] The details of this application are described below.

[0018] Example 1 Please refer to Figure 1 The invention provides a method for quantitatively assessing the health status of wind turbine blades based on adaptive multi-source fusion under operating conditions, comprising the following steps: S1. Synchronously collect multi-source sensor data during the operation of the wind turbine, including acoustic fingerprint signals, vibration signals and SCADA data, and achieve time synchronization through the IEEE 1588 protocol; Specifically, this invention first collects wind turbine operating data synchronously using multiple source sensors, providing a data foundation for subsequent analysis. This includes: Acoustic signal acquisition: A bone conduction sensor was used to acquire acoustic signals in the 20Hz-20kHz frequency band, with a sampling frequency of 40kHz, to capture the acoustic features caused by damage to the internal structure of the blade.

[0019] Vibration signal acquisition: A triaxial accelerometer is installed at the blade root to acquire vibration signals in the 0.5-5kHz frequency band, with a sampling frequency of 10kHz, for monitoring the vibration characteristics of the blade structure.

[0020] SCADA data acquisition: Obtain operating parameters such as wind speed, rotational speed, power, and blade pitch angle from the wind turbine monitoring system, with a sampling frequency of 1Hz.

[0021] Synchronization mechanism: The IEEE 1588 precision clock protocol is used to achieve time synchronization of multi-source data, ensuring that the time error is less than 1ms.

[0022] S2. Construct a multi-dimensional working condition feature vector based on multi-source sensor data, classify the current working condition based on a lightweight neural network, calculate the working condition identification confidence level, and obtain the working condition category and working condition change indicator. It should be noted that the multi-dimensional operating condition feature vector constructed in step S2 includes: real-time wind speed, real-time rotational speed, turbulence intensity, pitch angle change rate, and wind direction change rate.

[0023] Specifically, this invention abandons the traditional fixed threshold classification method based solely on wind speed and rotational speed, and proposes an adaptive working condition classification method based on multi-index fusion and transfer learning.

[0024] This invention defines a 5-dimensional working condition feature vector. ,in: Real-time wind speed, in m / s; Real-time rotational speed, in rpm; Turbulence intensity, calculated using the following formula: ;in, It is the standard deviation of wind speed within a 10-minute window. This is the average wind speed at that window; The rate of change of propeller pitch angle, in ° / s, is calculated using the following formula: ;in, The current propeller pitch angle. This is the time interval (usually 1 second). : Wind direction change rate, in ° / s, calculated in the same way as pitch angle change rate; In this invention, the working condition classification is achieved through a lightweight neural network. This invention designs a miniature neural network classifier with only 128 parameters, the structure of which is as follows:

[0025] in: It is a 5-dimensional feature vector of working conditions; This is the weight matrix from the input layer to the hidden layer; This is the hidden layer bias vector; This is the weight matrix from the hidden layer to the output layer; This is the output layer bias vector; For network parameter set; For activation functions; The activation function for the output layer; The network outputs a 4-dimensional probability vector. The predicted probabilities correspond to four operating conditions: shutdown, low wind speed, rated wind speed, and exceeding the rated wind speed.

[0026] This invention further introduces a working condition identification confidence level. Used to handle fuzzy conditions:

[0027] in: This is the probability distribution vector of the operating conditions output by the network; For the first Predicted probability of various working conditions ( =1 corresponds to shutdown. =2 corresponds to low wind speed. =3 corresponds to the rated wind speed. =4 corresponds to over-rated); For probability distribution Shannon entropy; when When the value is less than 0.7, the system enters the "fuzzy working condition" mode: it adopts the working condition label of the previous moment and triggers a manual review request; it also starts a short-term data accumulation mechanism (waiting 10 minutes before reclassification).

[0028] The condition change flag is set to 1 if the lightweight neural network classifies the current condition differently from the previous condition, and 0 otherwise.

[0029] S3. Extract feature values ​​from the acoustic signature signal, vibration signal and SCADA data respectively, and calculate the signal quality index of each signal source; Specifically, this invention calculates the kurtosis coefficient of the voiceprint from the voiceprint signal acquired by the bone conduction sensor. :

[0030] in: For voiceprint signal sampling points, =1,2,...,n; The average value of the voiceprint signal; The standard deviation of the voiceprint signal; This is the number of sampling points (usually 4096 points, corresponding to 0.1 seconds of voiceprint data). The vibration RMS value is calculated from the vibration signal collected by the blade root accelerometer:

[0031] in: For vibration acceleration signal sampling points, =1,2,..., ; This is the number of sampling points (usually 1024 points, corresponding to 0.1 seconds of vibration data).

[0032] Calculate the deviation between actual power and theoretical power from SCADA data:

[0033] in: This represents the actual measured power of the fan. This is the theoretical power curve; and The power curve coefficients were obtained by fitting historical health data using the least squares method.

[0034] in,

[0035] This refers to the number of historical data points.

[0036] Define a signal quality index Q for each feature source, as follows: Voiceprint signal quality : ,in: Signal-to-noise ratio, in dB; The background noise interference index has a value range of [0,1]. Vibration signal quality : ;in: For short-term RMS changes; The average RMS value of historical health data is used as a reference.

[0037] SCADA signal quality : ;in: The missing data rate is calculated as the ratio of the number of missing data points to the total number of data points.

[0038] S4. Normalize the extracted features according to the current working condition category so that they follow a standard positive distribution; It should be noted that, in order to eliminate the influence of different dimensions, each feature is normalized to the [0,1] interval: ;in: These are the original characteristic values ​​(acoustic kurtosis, vibration RMS, or power deviation). For this feature in the current operating condition The historical statistical average; For this feature in the current operating condition The historical standard deviation below; The normalized characteristic value represents the current operating condition category (shutdown, low wind speed, rated wind speed, exceeding rated wind speed). It follows a standard normal distribution, which facilitates subsequent fusion processing.

[0039] S5. Input the operating condition category, signal quality index, and operating condition change flag into the dynamic weight generation network to generate adaptive weights for each feature source. Specifically, this invention proposes a Dynamic Weight Generation Network (DWGN), whose mathematical expression is as follows:

[0040]

[0041]

[0042] in: For the current operating condition category, Its one-hot encoded vector (4 types of operating conditions, excluding shutdown state); For each feature source, the signal quality index is used. ; This is a sign indicating changes in operating conditions, with a value of 0 (stable) or 1 (changing). This represents a vector concatenation operation; The weight matrix is ​​input to the hidden layer; This is the hidden layer bias vector; This is the weight matrix from the hidden layer to the output layer; This is the output layer bias vector; Output for hidden layer; This is the importance score vector; This is the normalized weight vector; The dynamic weight generation network is trained by maximizing the health index discriminative power, and its loss function is:

[0043] in, HI discrimination for health / failure states; The mean HI values ​​for healthy and faulty states, respectively; The standard deviations of HI for healthy and faulty states, respectively; Weighting based on expert experience; For hyperparameters; These are the adaptive weights for each feature source after normalization.

[0044] To ensure that the generated weights conform to physical meaning, this invention designs a dual safeguard mechanism: 1. Weights and Constraints Naturally guaranteed by the Softmax layer .

[0045] 2. Operating condition sensitivity constraints For low wind speed conditions, add the following constraints: For overload conditions, add constraints: .

[0046] S6. Calculate the Health Index (HI) and perform trend analysis and multi-level alarm judgment based on historical HI sequences; The normalized feature vector and the dynamic weight vector are weighted and summed to obtain a health index of 0-100.

[0047] in: These are the normalized eigenvalues; The weights are dynamically generated. This represents the current health index, with a value range of [0, 100].

[0048] Furthermore, the system maintains a HI sequence for the past 7 days. .

[0049] Calculation of rate of change: Trend slope calculation: ; in, The HI value is the 7-day average.

[0050] The multi-level alarm logic is as follows: Early Warning: Alert: ;urgent: .

[0051] S7. Trigger the baseline self-learning mechanism when the preset conditions are met to update the statistical parameters required for feature normalization. The baseline self-learning mechanism of this invention is as follows: It receives health index (HI) data streams from the health index calculation and intelligent status assessment module, as well as raw feature value data from the feature normalization module. After processing, it outputs updated feature statistical parameters and feeds them back to the feature normalization module, forming a closed-loop system.

[0052] Input source: Health Index Sequence: A sequence of HI values ​​derived from the Health Index Calculation and Intelligent Status Assessment module; Raw feature data: raw feature values ​​such as kurtosis of the voiceprint, vibration RMS, and power deviation from the feature normalization module; Operating condition information: Operating condition category G from the intelligent operating condition perception module that integrates multiple indicators; Output destination: Updated feature statistics (mean μ and standard deviation σ): fed back to the feature normalization module for subsequent feature normalization calculations; Update status signal: Notifies the system that the baseline has been updated and triggers the switching of feature normalization parameters; A baseline update is triggered if and only if all of the following conditions are met: 1. Under the same operating condition G, the HI value for 30 consecutive days is greater than or equal to 90; 2. No alarms were recorded during the period; 3. The HI fluctuation range is less than 3% (i.e.) .

[0053] Simultaneously, the mean and standard deviation of the features are updated using the Exponential Moving Average (EMA) algorithm: The mean is updated as follows: Variance update: ;in: The mean value of this feature under operating condition G within the last 7 days after the triggering condition is met; The standard deviation of this feature under operating condition G within the last 7 days after the trigger condition is met; α is the learning rate, with a value of 0.1.

[0054] To prevent noise from being introduced during baseline updates, the present invention designs the following safeguard mechanism: Outlier filtering: in calculation and At that time, remove those exceeding the limit. Outliers within the range; Update range limitations: ; Specifically, it means: update the mean. Limit to 95% to 105% of the previous value; update the standard deviation. It is limited to between 95% and 105% of the previous value.

[0055] The clip function is a commonly used numerical restriction function, and its mathematical definition is:

[0056] in: For input values; The lower limit threshold; This is the upper limit threshold.

[0057] Updated baseline parameters and The data is sent back to the feature normalization module for subsequent feature normalization calculations.

[0058] This closed-loop mechanism enables the system to: Automatic adaptation to equipment aging: As the operating time of the wind turbine increases, the material properties of the blades change slowly, and the characteristic distribution drifts; Maintaining assessment accuracy: By regularly updating the baseline, ensure that the normalized eigenvalues ​​maintain a stable distribution; Avoid erroneous updates: Strict triggering conditions ensure that the baseline is updated only in a truly healthy state, preventing erroneous updates in a faulty state; Operating condition isolation: Baseline parameters are maintained independently for different operating conditions to avoid data contamination across operating conditions.

[0059] S8. Output health status assessment results and operation and maintenance suggestions.

[0060] Example 2 Please see Figure 2 Figure 2 The hardware device of this invention includes: a wind turbine blade health status quantitative assessment device 401 based on working condition adaptive multi-source fusion, a processor 402, and a storage device 403.

[0061] A wind turbine blade health status quantitative assessment device 401 based on operating condition adaptive multi-source fusion: The wind turbine blade health status quantitative assessment device 401 based on operating condition adaptive multi-source fusion implements the wind turbine blade health status quantitative assessment method based on operating condition adaptive multi-source fusion.

[0062] Processor 402: The processor 402 loads and executes the instructions and data in the storage device 403 to implement the method for quantitative evaluation of wind turbine blade health status based on condition-adaptive multi-source fusion.

[0063] Storage device 403: The storage device 403 stores instructions and data; the storage device 403 is used to implement the method for quantitative evaluation of the health status of wind turbine blades based on adaptive multi-source fusion under operating conditions.

[0064] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for quantitatively assessing the health status of wind turbine blades based on condition-adaptive multi-source fusion, characterized in that: The method includes the following steps: S1. Synchronously collect multi-source sensor data during the operation of the wind turbine, including acoustic fingerprint signals, vibration signals and SCADA data, and achieve time synchronization through the IEEE 1588 protocol; S2. Construct a multi-dimensional working condition feature vector based on multi-source sensor data, classify the current working condition based on a lightweight neural network, calculate the working condition identification confidence level, and obtain the working condition category and working condition change indicator. S3. Extract feature values ​​from the acoustic signature signal, vibration signal and SCADA data respectively, and calculate the signal quality index of each signal source; S4. Normalize the extracted features according to the current working condition category so that they follow a standard normal distribution; S5. Input the operating condition category, signal quality index, and operating condition change flag into the dynamic weight generation network to generate adaptive weights for each feature source. S6. Calculate the Health Index (HI) and perform trend analysis and multi-level alarm judgment based on historical HI sequences; S7. Trigger the baseline self-learning mechanism when the preset conditions are met to update the statistical parameters required for feature normalization. S8. Output health status assessment results and operation and maintenance suggestions; The dynamic weight generation network is trained by maximizing the health index discriminative power, and its loss function is: in, HI discrimination for health / failure states; The mean HI values ​​for healthy and faulty states, respectively; The standard deviations of HI for healthy and faulty states, respectively; Weighting based on expert experience; For hyperparameters; The adaptive weights for each feature source after normalization; Step S6 is as follows: The normalized feature vector and the dynamic weight vector are weighted and summed to obtain a health index of 0-100. in: These are the normalized eigenvalues; The weights are dynamically generated. This represents the current health index, with a value range of [0, 100].

2. The method for quantitatively assessing the health status of wind turbine blades based on adaptive multi-source fusion under operating conditions as described in claim 1, characterized in that: The multi-dimensional operating condition feature vector constructed in step S2 includes: real-time wind speed, real-time rotational speed, turbulence intensity, pitch angle change rate, and wind direction change rate.

3. The method for quantitatively assessing the health status of wind turbine blades based on adaptive multi-source fusion under operating conditions as described in claim 1, characterized in that: The lightweight neural network classifier structure described in step S2 is as follows: 5-dimensional input layer, 4-dimensional hidden layer, 4-dimensional output layer, and a total of 128 parameters.

4. The method for quantitatively assessing the health status of wind turbine blades based on adaptive multi-source fusion under operating conditions as described in claim 1, characterized in that: The inputs to the dynamic weight generation network in step S5 include: one-hot encoding of operating condition category, quality indicators of each signal, and indicator of operating condition change. The output is the normalized weights of each feature source.

5. The method for quantitative assessment of wind turbine blade health status based on adaptive multi-source fusion under operating conditions as described in claim 1, characterized in that: In step S7, after the baseline self-learning is triggered, the mean and standard deviation of the features are updated using the exponential moving average algorithm. At the same time, an update guarantee mechanism is introduced to filter outliers and limit the update magnitude.

6. A storage device, characterized in that: The storage device stores instructions and data to implement the method for quantitatively assessing the health status of wind turbine blades based on adaptive multi-source fusion of operating conditions, as described in any one of claims 1 to 5.

7. A device for quantitatively assessing the health status of wind turbine blades based on adaptive multi-source fusion under operating conditions, characterized in that: include: A processor and a storage device; the processor loads and executes instructions and data in the storage device to implement the method for quantitative evaluation of the health status of wind turbine blades based on adaptive multi-source fusion according to any one of claims 1 to 5.

Images