Wind turbine blade crack acoustic emission signal analysis method, device and storage medium

Abstract

The present invention relates to the field of voiceprint information processing, and discloses a method, device and storage medium for analyzing acoustic emission signals of cracks in fan blades. The method comprises the following steps: using a contact pickup to obtain acoustic emission signals generated at different positions of the fan blades; preprocessing the acoustic emission signals; segmenting the preprocessed signals; extracting dimensionless parameters; constructing feature vectors; fusing feature vectors of different contact pickups; constructing a machine learning model and pre-training; inputting the fused features into the trained model to obtain crack probability; determining the crack position according to the time difference of different contact pickups; and completing crack alarm according to the crack probability and crack position. The present invention has the beneficial effects of improving the accuracy and robustness of signal processing by extracting a series of dimensionless parameters, and realizing the precise positioning of cracks, thereby providing a reliable technical means for early warning of fan blade cracks.

Description

Technical Field

[0001] The present invention relates to the field of voiceprint information processing, and in particular to a method, device and storage medium for analyzing acoustic emission signals of cracks in fan blades. Background Art

[0002] Generally, cracks will appear on the surface of the blades of wind turbines after two to three years of operation. Each self-vibration or shutdown will further deepen and lengthen the cracks, and the dirt and sand in the air will also accelerate the expansion of the cracks, which may eventually cause the blades to break, seriously threatening the safety of the blades. Therefore, it is particularly important to monitor the cracks in the blades.

[0003] Current blade crack detection technologies, whether based on acoustic emission technology, vibration signal spectrum analysis or other methods, have certain limitations and are difficult to accurately measure and predict the development of cracks. Specifically:

[0004] 1) Based on acoustic emission technology: Although it can monitor the generation and development of cracks in real time, the signal is easily interfered by background noise and requires a special acoustic emission sensor and signal processing system.

[0005] 2) Spectral analysis based on vibration signals: Although it can provide rich fault information, when the crack is small or in the early stage, the vibration characteristics may not be obvious enough and difficult to detect.

[0006] 3) Crack detection based on image processing: It can automatically identify cracks, but it needs to be inspected in a stopped state, cannot achieve online monitoring, and is greatly affected by factors such as lighting conditions and lens viewing angle.

[0007] 4) Crack detection based on laser displacement sensor: Although it is a non-contact measurement, it may not be sensitive enough to detect very small cracks.

[0008] 5) Crack monitoring based on piezoelectric elements: Although the health status of the blade can be monitored in real time, its durability in extreme environments remains to be verified.

[0009] The limitations of these technologies have limited the accuracy and reliability of crack detection, which in turn affects the early detection of cracks and the formulation of preventive measures, and increases the cost and difficulty of equipment maintenance. Summary of the invention

[0010] The purpose of the present invention is to propose a method, device and storage medium for analyzing acoustic emission signals of cracks in fan blades, so as to solve the technical problem that the prior art is not accurate and reliable in detecting cracks in fan blades.

[0011] The present invention provides a method for analyzing acoustic emission signals of cracks in fan blades, comprising the following steps:

[0012] S1: Use contact pickup to obtain acoustic emission signals generated at different positions of the fan blades;

[0013] S2: preprocessing the acoustic emission signal to obtain a preprocessed signal;

[0014] S3: Segment the preprocessed signal to obtain multiple segmented signals;

[0015] S4: extracting dimensionless parameters of multiple segmented signals;

[0016] S5: construct feature vectors based on dimensionless parameters;

[0017] S6: Fusing feature vectors of different contact pickups to obtain fused features;

[0018] S7: Build a machine learning model and pre-train it to obtain a trained model;

[0019] S8: Input the fused features into the trained model to obtain the crack probability;

[0020] S9: Determine the crack location based on the time difference of different contact pickups;

[0021] S10: Complete crack alarm according to crack probability and crack position.

[0022] A storage medium stores instructions and data for implementing a method for analyzing acoustic emission signals of cracks in fan blades.

[0023] A wind turbine blade crack acoustic emission signal analysis device comprises: a processor and the storage medium; the processor loads and executes instructions and data in the storage medium to implement a wind turbine blade crack acoustic emission signal analysis method.

[0024] The beneficial effects provided by the present invention are: by extracting a series of dimensionless parameters, the accuracy and robustness of signal processing are improved, and the precise location of cracks is achieved. The method combines contact acoustic emission monitoring technology with advanced signal processing algorithms, providing a reliable technical means for early warning of fan blade cracks. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 It is a schematic flow chart of the method of the present invention;

[0026] Figure 2 It is a working schematic diagram of the hardware device of the present invention. DETAILED DESCRIPTION

[0027] To make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0028] Before formally describing the present invention, the scheme of the present invention is first generally described for easy understanding.

[0029] Please refer to Figure 1 , Figure 1 It is a schematic flow diagram of the method of the present invention;

[0030] The present invention provides a method for analyzing acoustic emission signals of cracks in fan blades, comprising the following steps:

[0031] S1: Use contact pickup to obtain acoustic emission signals generated at different positions of the fan blades;

[0032] As an example, a contact (bone conduction) acoustic emission sensor (pickup) is used to detect the acoustic emission signal generated by the fan blade. Bone conduction sensors are widely used in acoustic emission detection because they can effectively transmit high-frequency signals. This sensor directly contacts the fan blade and converts the acoustic emission signal emitted by the crack of the blade into an electrical signal.

[0033] S2: preprocessing the acoustic emission signal to obtain a preprocessed signal;

[0034] It should be noted that the preprocessing in step S2 specifically includes: discretization processing and bandpass filtering.

[0035] Acoustic emission signals are usually time series data, and the time series of these signals can be recorded using a digital signal processor (DSP) or other data acquisition system. Assume that the time series of the signal can be expressed as ,in is time. The signal is at the time point The sample value can be expressed as .

[0036] In the present invention, the time signal is discretized. A discrete signal can be represented as a vector, where each element represents the signal value at a specific time point. Sampling interval: Assume that the sampling interval is , then the continuous-time signal In time The sample value can be expressed as .

[0037] Sampling frequency: Sampling frequency is the number of samples per second, which can be expressed as .

[0038] In addition, the present invention also performs filtering processing on the collected acoustic emission signals to remove unnecessary noise. For the acoustic emission signals in the crack detection of fan blades, considering that the acoustic emission signals usually contain higher frequency components and the background noise may be distributed over a wider frequency band, a bandpass filter may be the most appropriate choice. The bandpass filter can effectively remove low-frequency and high-frequency noise while retaining the useful frequency components in the signal.

[0039] Bandpass filter design,Design a second-order bandpass filter, whose transfer function is:

[0040] (1)

[0041] in, is the quality factor, Is a complex frequency variable in the Laplace transform domain. Input signal: Assume that the original signal is , then the output signal after the bandpass filter is processed It can be expressed as:

[0042] (2)

[0043] in, It's a signal The Laplace transform of represents the inverse Laplace transform.

[0044] Digital filter design: To convert the above analog filter into a digital filter, the bilinear transformation method can be used. Assume that the sampling frequency is , then the bilinear transformation will Replace with The relationship is:

[0045] (3)

[0046] in, is the sampling interval.

[0047] The transfer function of the digital bandpass filter is:

[0048] The transfer function of the above analog bandpass filter is converted into the transfer function of the digital filter through bilinear transformation:

[0049] (4)

[0050] in, and They are the numerator and denominator coefficients of the filter, respectively, and can be calculated using digital filter design tools or formulas.

[0051] The parameter selections are as follows:

[0052] Center frequency : Select the appropriate center frequency according to the spectral characteristics of the acoustic emission signal.

[0053] Quality Factor : Determines the bandwidth of the filter, the larger the A higher value means a narrower bandwidth, i.e. tighter frequency selectivity.

[0054] For example: Assuming that the main frequency components of the acoustic emission signal are concentrated between 10 kHz and 50 kHz, we can design a bandpass filter with a center frequency of 30 kHz and a quality factor of 10. In this case, the center frequency .

[0055] In the crack detection of wind turbine blades, the use of bandpass filters can effectively remove noise and retain the useful frequency components in the signal. By reasonably selecting the center frequency and quality factor, a bandpass filter suitable for a specific application can be designed. This filter can improve the accuracy of subsequent dimensionless parameter extraction and signal analysis.

[0056] S3: Segment the preprocessed signal to obtain multiple segmented signals;

[0057] It should be noted that the segmentation process in step S3 includes: sliding window segmentation and fixed window segmentation.

[0058] As an embodiment, the present invention divides the continuous signal into multiple time windows, and the signal in each window is used for subsequent analysis. The signal can be divided by a sliding window or a fixed window method.

[0059] Among them, the specific process of using the sliding window is as follows:

[0060] Assume that the original signal is , the sampling interval is The signal in a window can be expressed as:

[0061] (5)

[0062] Among them, the window size , step length , It is The starting time point of a window can be expressed as:

[0063] (6)

[0064] The specific process of using a fixed window is as follows:

[0065] If a fixed window is used, the window size Keep the step size constant Also remains unchanged, the overlap between windows depends on the step size and window size .

[0066] Example: Assuming the original signal , the sampling rate is , that is, the sampling interval ,Window size sampling points, step length sampling points, the time length of each window is .

[0067] S4: extracting dimensionless parameters of multiple segmented signals;

[0068] It should be noted that, assuming the input signal is , the sampling time is , then the discrete signal can be expressed as . There are the following statistical calculation parameters:

[0069] ① Root mean square value (RMS):

[0070] (7)

[0071] ② Mean Absolute Value (MAV):

[0072] (8)

[0073] ③ Peak Value:

[0074] (9)

[0075] ④ Mean (\mu):

[0076] (10)

[0077] ⑤ Standard Deviation (\sigma):

[0078] (11)

[0079] According to the above parameters, dimensionless parameters are obtained. The dimensionless parameters in step S4 include: waveform index, pulse index, peak index, kurtosis, and normalized energy entropy.

[0080] As an example, the shape factor (SF) specifically refers to: a measure of the shape of a signal waveform, which is usually related to the root mean square value (RMS) and absolute mean of the signal. The formula is as follows:

[0081] (12)

[0082] The waveform indicator reflects the sharpness of the signal waveform and is more sensitive to impact signals.

[0083] Impulse Factor (IF) specifically refers to: describing the intensity of the pulse component in the signal, which is related to the peak value and RMS value of the signal. The formula is as follows:

[0084] (13)

[0085] The larger the pulse index is, the stronger the pulse component contained in the signal is, and it is more sensitive to transient events such as crack extension.

[0086] Crest Factor (CF) specifically refers to the ratio of the signal peak value to the RMS value, which is used to measure the peak characteristics of the signal. The formula is as follows:

[0087] (14)

[0088] Among them, a higher peak index indicates that the maximum amplitude of the signal is larger relative to the average energy level, which is usually related to the transient characteristics of the signal.

[0089] Kurtosis (K) specifically refers to: a statistic that describes the shape of the signal distribution, especially the degree to which the tail of the signal deviates from the normal distribution. The formula is as follows:

[0090] (15)

[0091] in, is the signal value, is the mean, is the standard deviation, Express expectations.

[0092] Among them, kurtosis can sensitively reflect the impact component in the signal and has a good ability to identify transient events such as crack extension.

[0093] Normalized Energy Entropy (NEE) specifically refers to: an indicator that measures the uniformity of signal energy distribution, which can help identify changes in signal complexity. The formula is as follows:

[0094] (16)

[0095] in It is The probability of an energy segment, is the number of energy segments.

[0096] The significance of normalized energy entropy is that by calculating the energy distribution of the signal and dividing it into multiple energy segments, the characteristics of the crack signal can be extracted, which can then be used for crack identification and early warning.

[0097] As an embodiment, the calculation process of normalized energy entropy is as follows:

[0098] ①Signal energy distribution: Divide the energy distribution of the signal into several energy segments.

[0099] ②Energy segment probability: Calculate the probability of each energy segment .

[0100] (17)

[0101] in, It is The signal value of the sampling point, It is The signal value of the sampling point, is the total number of sampling points. Example: If the energy distribution of the signal is divided into 5 energy segments, then (n = 5).

[0102] ③Energy entropy: calculate the entropy of signal energy distribution.

[0103] (18)

[0104] ④ Normalization: Normalize the entropy to make it within a reasonable numerical range.

[0105] (19)

[0106] Usually, multiple acoustic emission sensors are used to monitor a blade. When using multiple sensors for acoustic emission signal analysis and early warning of wind turbine blade cracks, the signal correlation and consistency between different sensors need to be considered.

[0107] S5: construct feature vectors based on dimensionless parameters;

[0108] As an example, the feature vector is as follows:

[0109]

[0110] in, , respectively The sensors are waveform index, pulse index, peak index, kurtosis, and normalized energy entropy.

[0111] S6: Fusing feature vectors of different contact pickups to obtain fused features;

[0112] It should be noted that the fusion strategies in step S6 include: average fusion, maximum fusion, voting fusion, direct fusion and weighted fusion.

[0113] As an embodiment, the present invention can significantly improve the ability of wind turbine blade crack monitoring and reduce the false alarm rate by using multiple acoustic emission sensors and different data fusion strategies. These methods can improve the accuracy of crack location and enhance the robustness and accuracy of the system by comprehensively analyzing the data of all sensors.

[0114] Specifically, a variety of strategies can be used to fuse data from different sensors to improve detection capabilities and reduce false alarm rates.

[0115] The average fusion is as follows:

[0116] Mean eigenvector: (twenty one)

[0117] in, is the mean eigenvector; It is The feature vector of each sensor; is the number of sensors.

[0118] The maximum value is fused as follows:

[0119] Maximum eigenvector: (twenty two)

[0120] in, is the maximum eigenvector; They are the maximum waveform index, pulse index, peak index, kurtosis and normalized energy entropy extracted from all sensors.

[0121] Voting merge is as follows:

[0122] Voting feature vector: (twenty three)

[0123] in, They are the representative values ​​of waveform index, pulse index, peak index, kurtosis and normalized energy entropy selected through the voting mechanism.

[0124] Direct fusion is as follows:

[0125] Eigenvector matrix: (twenty four)

[0126] in, is the matrix of all sensor feature vectors; It is The feature vector of each sensor.

[0127] The weighted fusion is as follows:

[0128] Weighted eigenvectors: (25)

[0129] in, is the weighted eigenvector; It is The weight of each sensor; It is The feature vector of each sensor.

[0130] Choose the appropriate fusion strategy: Choose the most suitable fusion strategy based on the actual situation and data characteristics. For example, if the signal quality of some sensors is high, you can use weighted fusion; if you want to reduce false alarms, you can use voting fusion.

[0131] In addition, for the convenience of description, the feature vectors after data fusion are uniformly named .

[0132] S7: Build a machine learning model and pre-train it to obtain a trained model;

[0133] Using machine learning or deep learning techniques, such as support vector machine (SVM), convolutional neural network (CNN), etc., the feature vector is trained to identify whether there is a crack signal.

[0134] Take the support vector machine as an example: it is used to solve the optimal classification hyperplane. That is, the following optimization problem:

[0135] (26)

[0136] The constraints are:

[0137] (27)

[0138] in, is the weight vector, which determines the direction of the classification hyperplane; is the bias term, which determines the distance between the classification hyperplane and the origin; is a slack variable used to allow some data points to be within the classification boundary or misclassified to increase the robustness of the model; C is a penalty parameter used to control the trade-off between misclassification and model complexity; is the number of training samples; It is The class label of each sample (for binary classification problems, it is usually -1 or 1); After fusion The feature vector of samples; represents the Euclidean norm of the vector, that is ; Represents the dot product of vectors.

[0139] The goal of this optimization problem is to find a set of parameters , so that the objective function is minimized while satisfying all constraints. This objective function consists of two parts:

[0140] Part of it is , which represents the complexity of the model; the other part is , Represents the penalty for misclassification. By adjusting the parameter C, a balance can be achieved between the complexity of the model and misclassification.

[0141] Cross-validation techniques were used to evaluate the performance of the model and ensure that the model had good generalization ability. Fold cross validation: Divide the dataset into Subsets, each time using The subset is used as the training set, and the remaining subset is used as the test set.

[0142] Repeat this process times, each time using a different subset as the test set.

[0143] The performance indicators of cross validation are as follows:

[0144] Accuracy: (28)

[0145] in, : True Positives, : True Negatives, : False Positives, : False Negatives.

[0146] Recall: (29)

[0147] Precision: (30)

[0148] F1 score: (31)

[0149] By combining data fusion and support vector machine (SVM), the ability to monitor cracks in wind turbine blades can be significantly improved and the false alarm rate can be reduced. By comprehensively analyzing the data from all sensors and evaluating the model performance through cross-validation, it is possible to ensure that the model has good generalization ability and achieve accurate crack location and predictive maintenance.

[0150] S8: Input the fused features into the trained model to obtain the crack probability;

[0151] S9: Determine the crack location based on the time difference of different contact pickups;

[0152] As an example, after the blade crack fault is confirmed, the crack position can be determined using the time difference of multiple sensors. The sensor positions are arranged reasonably to cover different areas of the entire blade to ensure comprehensive signal coverage.

[0153] The basic steps of position estimation calculation are:

[0154] ① Determine the sensor location: Assume that the sensor and The locations are and .

[0155] ② Measuring time difference: Measuring sensor and The time difference between receiving the same signal .

[0156] ③ Signal propagation speed: Assume that the signal propagation speed in the medium is .

[0157] ④ Calculate distance difference: Calculate the distance difference between the two sensors and the signal source based on the time difference and signal propagation speed.

[0158] (34)

[0159] in, It is a sensor and sensors The distance difference to the signal source.

[0160] ⑤ Position estimation: Use the time difference and position information of at least three sensors to solve the position of the signal source through mathematical methods.

[0161] For example: Suppose there are two sensors and , whose locations are and The signal source position is unknown, denoted as .

[0162] Measuring time differences: Measuring sensors and The time difference of receiving the same signal .

[0163] Calculate the distance difference: (35)

[0164] Position estimation:

[0165] Using the time difference and position information of at least three sensors, the position of the signal source is estimated by solving an overdetermined system of equations. Overdetermined system of equations:

[0166] (36)

[0167] in, is the location of the third sensor; and Sensors and With sensor and The distance difference to the signal source. The above overdetermined equations can be solved using the least squares method to obtain the signal source position. best estimate of .

[0168] S10: Complete crack alarm according to crack probability and crack position.

[0169] As an embodiment, the crack alarm is specifically as follows:

[0170] (37)

[0171] in: is the average value vector of the feature vector after data fusion; is the standard deviation vector of the feature vector after data fusion; is the threshold factor, typically taken as 2 or 3, corresponding to approximately 95% or 99.7% confidence levels.

[0172] when Call the police.

[0173] See also Figure 2 , Figure 2 It is a schematic diagram of the working of the hardware device of an embodiment of the present invention, and the hardware device specifically includes: a wind turbine blade crack acoustic emission signal analysis device 401, a processor 402 and a storage medium 403.

[0174] A wind blade crack acoustic emission signal analysis device 401: The wind blade crack acoustic emission signal analysis device 401 implements the wind blade crack acoustic emission signal analysis method.

[0175] Processor 402: The processor 402 loads and executes the instructions and data in the storage medium 403 to implement the wind turbine blade crack acoustic emission signal analysis method.

[0176] Storage medium 403: the storage medium 403 stores instructions and data; the storage medium 403 is used to implement the wind turbine blade crack acoustic emission signal analysis method.

[0177] The beneficial effects of the present invention are: by extracting a series of dimensionless parameters, the accuracy and robustness of signal processing are improved, and the precise location of cracks is achieved. The method combines contact acoustic emission monitoring technology with advanced signal processing algorithms, providing a reliable technical means for early warning of fan blade cracks.

[0178] 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 principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for analyzing acoustic emission signals of cracks in fan blades, characterized in that: The method comprises the following steps: S1: Use contact pickup to obtain acoustic emission signals generated at different positions of the fan blades; S2: preprocessing the acoustic emission signal to obtain a preprocessed signal; S3: Segment the preprocessed signal to obtain multiple segmented signals; S4: extracting dimensionless parameters of multiple segmented signals; The dimensionless parameters in step S4 include: waveform index, pulse index, peak index, kurtosis, and normalized energy entropy; The calculation process of normalized energy entropy is as follows: ①Signal energy distribution: divide the energy distribution of the signal into several energy segments; ②Energy segment probability: Calculate the probability of each energy segment ; （17） in, It is The signal value of the sampling point, It is The signal value of the sampling point, is the total number of sampling points; ③Energy entropy: calculate the entropy of signal energy distribution; （18） ④Normalization: Normalize the entropy to make it within a reasonable numerical range; （19） in, is the total number of energy segments; S5: construct feature vectors based on dimensionless parameters; S6: Fusing feature vectors of different contact pickups to obtain fused features; The fusion strategies in step S6 include: average fusion, maximum fusion, voting fusion, direct fusion and weighted fusion; S7: Build a machine learning model and pre-train it to obtain a trained model; S8: Input the fused features into the trained model to obtain the crack probability; S9: Determine the crack location based on the time difference of different contact pickups; Step S9 is specifically as follows: Obtain the time difference of the received signals of at least three contact pickups A, B, and C, and use the least squares method to solve the crack position according to the following formula r : in , , is the distance difference of the sensor, which is equal to the signal propagation speed multiplied by the time difference between the pickups receiving the signal; These are the locations of the three contact pickups; S10: Complete crack alarm according to crack probability and crack position.

2. A method for analyzing acoustic emission signals of cracks in fan blades according to claim 1, characterized in that: The preprocessing in step S2 specifically includes: discretization processing and bandpass filtering.

3. A method for analyzing acoustic emission signals of cracks in fan blades according to claim 1, characterized in that: The segmentation process in step S3 includes: sliding window segmentation and fixed window segmentation.

4. A method for analyzing acoustic emission signals of cracks in fan blades according to claim 1, characterized in that: When the model is trained in step S7, cross validation is used to evaluate its performance.

5. A storage medium, characterized in that: The storage medium stores instructions and data for being executed by a processor to implement a method for analyzing acoustic emission signals of cracks in wind turbine blades as described in any one of claims 1 to 4.

6. A wind turbine blade crack acoustic emission signal analysis device, characterized in that: include: Processor and storage medium; the processor loads and executes instructions and data in the storage medium to implement a method for analyzing acoustic emission signals of cracks in fan blades as described in any one of claims 1 to 4.

Images