A method for detecting crack faults in compressor blades
Through mixed feature selection and long-term memory neural network, combined with time domain, frequency domain and spectral centroid energy migration characteristics, the problem of low feature extraction efficiency and accuracy in compressor blade crack fault detection is solved, achieving high accuracy and high efficiency detection effect.
Patent Information
- Application Number
- CN202210001059.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-01-04
- Publication Date
- 2024-11-26
- Estimated Expiration
- 2042-01-04
AI Technical Summary
The prior art is difficult to accurately detect compressor blade crack failures, especially under strong background noise interference. Traditional feature extraction efficiency is low and the accuracy is not high, and the features are prone to conflicts, resulting in inaccurate detection results.
The hybrid feature selection method combines the time domain, frequency domain, acoustic emission and spectral centroid energy migration characteristics to realize the crack fault detection of compressor blades through long and short-term memory neural networks, and use Laplace scores and random forest scores to select the optimal feature sub Set, optimize feature extraction and detection processes.
It improves the accuracy of crack fault detection of compressor blades, reduces the number of feature extraction, and achieves high accuracy detection at different speeds, ensuring the reliability and efficiency of detection.
Smart Images

Figure CN114528868B_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the technical field of rotating machinery acoustic emission signal analysis, in particular to a compressor blade crack fault detection method. Background Art
[0002] Compressors are widely used in fields such as petrochemicals and electric power. Blades, as core components, are susceptible to cracking due to centrifugal forces, friction, and unstable airflow loads, which can affect the normal operation of the entire compressor. Therefore, timely detection of blade crack failures is crucial for ensuring safe and stable compressor operation. Furthermore, in actual engineering applications, the presence of strong background noise in the acoustic emission signals can obscure the crack failure information, making compressor blade crack detection extremely complex and difficult. Therefore, it is worthy of further research to effectively extract features that reflect blade crack failures and accurately detect compressor blade crack failures using optimal features.
[0003] Traditional acoustic emission features (including energy and amplitude) are very sensitive to cracks, but may not be able to accurately detect blade cracks due to the presence of strong noise. Although common time domain, frequency domain and acoustic emission features can reflect blade crack fault information to a certain extent, they are very prone to errors under the interference of noise. Extracting too many features also causes a long time problem, which directly affects the detection efficiency. In addition, once there is a conflict or oversaturation between the characteristics, the fault detection accuracy will be affected. Therefore, there is an urgent need for a feature that can accurately reflect the blade crack fault, and adaptively select features according to the signal characteristics, and realize a reliable and accurate compressor blade crack fault detection method. Summary of the Invention
[0004] In response to the above-mentioned shortcomings, the present invention proposes a compressor blade crack fault detection method to overcome the current situation where existing features are difficult to accurately reflect the characteristics of compressor blade crack faults and conflicts between features make it impossible to accurately detect blade crack faults, and effectively realize feature selection and compressor blade crack fault detection.
[0005] In order to solve the above technical problems, the present invention adopts the following technical solutions:
[0006] A method for detecting a crack fault in a compressor blade comprises the following steps:
[0007] Step 1: Collect two-channel acoustic emission signals at the compressor outlet and divide them into two-channel training samples and two-channel test samples;
[0008] Step 2: Extract features of the two-channel training samples, including acoustic emission features, time domain features, frequency domain features, and spectral centroid energy migration features;
[0009] Step 3: Use a hybrid feature selection method to select the features and establish an optimal feature subset;
[0010] Step 4: extracting features of the two-channel training samples and the two-channel test samples according to the optimal feature subset, and establishing a two-channel training sample feature subset and a two-channel test sample feature subset;
[0011] Step 5: Merge the two-channel training sample feature subsets and input them into the long short-term memory neural network for training;
[0012] Step 6: Merge the two-channel test sample feature subsets and input them into the trained long short-term memory neural network for crack fault classification and detection, thereby realizing crack fault detection of compressor blades.
[0013] In step 2, the acoustic emission characteristics, time domain characteristics, and frequency domain characteristics include:
[0014] Acoustic emission rise time: AE ris =t max -t th1 , acoustic emission duration: AE dur =t th2 -t th1 , maximum amplitude of acoustic emission: AE amp =|max x(t)|, acoustic emission energy: Acoustic Emission Energy Count AE evc , acoustic emission ringing count AE ric , RMS value: Zero crossing rate ZCR, time domain energy: Time domain energy entropy: Frequency domain energy entropy: Spectral centroid: Spectral flux: Spectral attenuation:
[0015] Among them, x i (t) represents the amplitude at the i-th time, N represents the signal length, f i represents the i-th frequency value, X i (f) represents the frequency f i The amplitude at Indicates the frequency f at the previous moment i The amplitude at t max Indicates the moment when the acoustic emission amplitude is the highest, t th1 Indicates the moment when the acoustic emission process signal rises above the threshold, t th2 Indicates the moment when the acoustic emission process signal drops below the threshold.
[0016] Preferably, the expression of the spectral centroid energy transfer characteristic in step 2 is:
[0017]
[0018]
[0019] Among them, ER i (f) represents the frequency f i Energy ringing at E i (f) represents the frequency f i energy, T represents the energy threshold, and SCES represents the spectrum centroid energy transfer feature.
[0020] Preferably, the specific steps of selecting the features using the hybrid feature selection method in step 3 are:
[0021] Step 3.1, calculate the Laplace score of all features. The Laplace score expression is:
[0022]
[0023] Among them, L t represents the Laplace score of feature t, s t represents the variance of feature t in the training sample, t i represents the value of feature t in the i-th training sample, S ij Represents the weight coefficients of the i-th and j-th training samples.
[0024] Step 3.2, calculate the random forest score of all features. The random forest score expression is:
[0025]
[0026] Among them, R t represents the random forest score of feature t, er1(i) represents the out-of-bag estimation error of the i-th random tree, and er 2t (i) represents the second-type out-of-bag estimation error of the i-th random tree for feature t.
[0027] Step 3.3, calculate the mixture fraction of all features and select features with a mixture fraction greater than 0.5 to form the potential feature set. The mixture fraction expression is
[0028]
[0029] Among them, TS t represents the mixture fraction of feature t, and Represents the Laplace and random forest scores of feature t after normalization from 0 to 1.
[0030] In step 3.4, the detection accuracy and its rate of change under different number of feature combinations are calculated, and the number of features at the inflection point is selected as the number of features in the optimal feature subset.
[0031] Step 3.5: Create several potential feature subsets from the potential feature set according to the number of features in the optimal feature subset, calculate the detection accuracy of each potential feature subset based on the random forest algorithm, and the potential feature subset with the highest detection accuracy is the optimal feature subset.
[0032] Preferably, the two-channel training sample feature subsets in step 5 are merged into direct splicing.
[0033] Preferably, the dimension of the long short-term memory neural network input feature in step 5 is twice the number of features in the optimal feature subset, the number of hidden gates of the long short-term memory neural network is 200, the classification label of the long short-term memory neural network is 3 categories, and the long short-term memory neural network training cycle is 100.
[0034] The beneficial effects of the present invention are as follows:
[0035] 1) This paper provides a compressor blade crack fault detection method based on spectral centroid energy migration and hybrid feature selection. This method extracts blade crack fault information from time domain, frequency domain, acoustic emission, and spectral centroid energy migration features, optimizes these features using a hybrid feature selection method, and finally uses a long-short-term memory neural network to detect compressor blade crack faults. This method overcomes the limitation of traditional features in accurately reflecting blade crack faults and improves the accuracy of compressor blade crack fault detection.
[0036] 2) By analyzing the changes in spectral centroid and energy when cracks appear on the blade, the present invention provides a spectral centroid energy migration feature that can accurately reflect the characteristics of blade crack failure and, in combination with other common features, realize accurate detection of compressor blade crack failure.
[0037] 3) The hybrid feature selection method provided by the present invention combines the Laplace score and the random forest score to select the features that best reflect the blade crack fault, and uses the random forest algorithm to determine the optimal feature subset, thereby reducing the number of feature extractions and achieving highly accurate detection of blade crack faults with the optimal features. BRIEF DESCRIPTION OF THE DRAWINGS
[0038] Figure 1 is a flowchart of an embodiment of the present invention;
[0039] Figure 2 2 is a graph showing the feature score results in an embodiment of the present invention.
[0040] Figure 3 1 is a graph showing the relationship between the number of features, detection accuracy, and change rate in an embodiment of the present invention. DETAILED DESCRIPTION
[0041] The following description is merely exemplary in nature and is not intended to limit the present disclosure, application, or uses. It should be understood that throughout the drawings, corresponding reference numerals indicate identical or corresponding parts and features.
[0042] like Figure 1 As shown, the compressor blade crack fault detection method according to an embodiment of the present invention includes the following steps:
[0043] Step 1: Collect two-channel acoustic emission signals at the compressor outlet and divide them into two-channel training samples and two-channel test samples;
[0044] Step 2: Extract features of the two-channel training samples, including acoustic emission features, time domain features, frequency domain features, and spectral centroid energy transfer features;
[0045] Step 3: Use hybrid feature selection method to select features and establish the optimal feature subset;
[0046] Step 4: Extract the features of the two-channel training samples and the two-channel test samples according to the optimal feature subset, and establish a two-channel training sample feature subset and a two-channel test sample feature subset.
[0047] Step 5: Combine the feature subsets of the two-channel training samples and input them into the long short-term memory neural network for training.
[0048] Step 6: Combine the feature subsets of the two-channel test samples and input them into the trained long short-term memory neural network for crack fault classification and detection to achieve crack fault detection of compressor blades.
[0049] In the above embodiment, two acoustic emission sensors collect two-channel acoustic emission signals at the compressor outlet, divide them into training samples and two-channel test samples, and then extract the time domain, frequency domain, acoustic emission characteristics and spectral centroid energy migration characteristics of the training samples. The optimal feature subset is established by mixing the features, and the optimal feature subsets of the training samples and test samples are extracted. Finally, a long short-term memory neural network is used for training and testing to realize compressor blade crack fault detection.
[0050] It should be noted that the two acoustic emission sensors are installed in a relative manner at the compressor outlet. The time domain signal of the acoustic emission signal is converted into a frequency domain signal through Fourier transform. There may be certain differences between the two-channel training samples and the two-channel test samples due to the randomness of the division, but it does not affect the subsequent processing.
[0051] In step 2, the acoustic emission characteristics, time domain characteristics, frequency domain characteristics and spectral centroid energy migration characteristics include:
[0052] Acoustic emission rise time: AE ris =t max -t th1 , acoustic emission duration: AE dur =t th2 -t th1 , maximum amplitude of acoustic emission: AE amp =|max x(t)|, acoustic emission energy: Acoustic Emission Energy Count AE evc , acoustic emission ringing count AE ric , RMS value: Zero crossing rate ZCR, time domain energy: Time domain energy entropy: Frequency domain energy entropy: Spectral centroid: Spectral flux: Spectral attenuation:
[0053]
[0054] Among them, x i (t) represents the amplitude at the i-th time, N represents the signal length, f i represents the i-th frequency value, X i (f) represents the frequency f i The amplitude at Indicates the frequency f at the previous moment i The amplitude at t max Indicates the moment when the acoustic emission amplitude is the highest, t th1 Indicates the moment when the acoustic emission process signal rises above the threshold, t th2 Indicates the moment when the acoustic emission process signal drops below the threshold.
[0055] It should be noted that the acoustic emission characteristics take the maximum value of the characteristics in all acoustic emission processes. The acoustic emission characteristics include the maximum acoustic emission amplitude, acoustic emission rise time, acoustic emission duration and acoustic emission energy.
[0056] In step 2, the expression of the spectral centroid energy migration characteristic is:
[0057]
[0058]
[0059] Among them, ER i (f) represents the frequency f i Energy ringing at E i (f) represents the frequency fi energy, T represents the energy threshold, and SCES represents the spectrum centroid energy transfer feature.
[0060] It should be noted that the energy threshold is set to 10% of the maximum amplitude.
[0061] Step 3: The specific steps of selecting the features using the hybrid feature selection method are as follows:
[0062] Step 3.1, calculate the Laplace score of all features. The Laplace score expression is:
[0063]
[0064] Among them, L t represents the Laplace score of feature t, s t represents the variance of feature t in the training sample, t i represents the value of feature t in the i-th training sample, S ij Represents the weight coefficients of the i-th and j-th training samples.
[0065] Step 3.2, calculate the random forest score of all features. The random forest score expression is:
[0066]
[0067] Among them, R t represents the random forest score of feature t, er1(i) represents the out-of-bag estimation error of the i-th random tree, and er 2t (i) represents the second-type out-of-bag estimation error of the i-th random tree for feature t.
[0068] Step 3.3, calculate the mixture fraction of all features and select features with a mixture fraction greater than 0.5 to form the potential feature set. The mixture fraction expression is
[0069]
[0070] Among them, TS t represents the mixture fraction of feature t, and Represents the Laplace and random forest scores of feature t after normalization from 0 to 1.
[0071] In step 3.4, the detection accuracy and its rate of change under different number of feature combinations are calculated, and the number of features at the inflection point is selected as the number of features in the optimal feature subset.
[0072] In step 3.5, several potential feature subsets are created from the potential feature set according to the number of features in the optimal feature subset. The detection accuracy of each potential feature subset is calculated based on the random forest algorithm. The potential feature subset with the highest detection accuracy is the optimal feature subset.
[0073] It should be noted that the weight coefficient S ij The expression is:
[0074]
[0075]
[0076] Among them, D ij It represents the distance between the i-th and j-th training samples, and d represents the proportional coefficient (usually 1).
[0077] In step 5, the feature subsets of the two-channel training samples are merged into direct concatenation.
[0078] In step 5, the dimension of the long short-term memory network input feature is twice the number of features in the optimal feature subset, the number of hidden gates of the long short-term memory network is 200, the classification label of the long short-term memory network is 3 categories, and the long short-term memory network training cycle is 100.
[0079] In order to further verify the effectiveness of the invention method, the scheme of this application is further explained using a compressor blade crack test bench:
[0080] The compressor blade crack test bench has three types of blade crack lengths: 0, 10mm and 20mm, and operates at four speeds of 1200rpm, 1500rpm, 1800rpm and 2100rpm. The two acoustic emission sensors are installed at the compressor outlet and divided into two-channel training samples and two-channel test samples. A total of 12 features are extracted, including the time domain, frequency domain, acoustic emission features and spectral centroid energy migration features, and the Laplace, random forest and mixed scores of all features are calculated. The results after normalization of the scores of all features at 1200rpm are as follows Figure 2 As shown in Table 1. Analyze the relationship between the number of features and fault detection accuracy, and determine the number of features in the optimal feature subset, as shown in Figure 3As shown. Then, the optimal feature subset is determined, and a training sample feature subset and a two-channel test sample feature subset are established. The two-channel training sample feature subsets are merged and input into the long short-term memory neural network for training. Finally, the merged two-channel test sample feature subsets are input into the trained long short-term memory network for crack fault classification and detection. The results of compressor blade crack fault detection at four speeds are shown in Table 2. At the four speeds, the accuracy of blade crack fault detection exceeds 96%, and the accuracy reaches 100% at 1200rpm and 2100rpm, verifying the accuracy and reliability of the method of this application.
[0081] Table 1 Feature selection results
[0082] Serial number feature Laplace fraction Random Forest Score Mixed fractions 1 Acoustic emission rise time 0.3503 0.9819 -0.0723 2 Acoustic emission duration 0.3070 0.9761 0.1870 3 Maximum acoustic emission amplitude 0.2248 0.9709 0.1800 4 Acoustic emission energy 0.7450 0.9889 1.4688 5 Acoustic emission energy counting 0.1111 0.9599 0.4877 6 Acoustic emission ringing count 0.5431 0.9741 1.6198 7 RMS value 0.8565 0.9920 1.8092 8 Zero-crossing rate 0.2868 0.9745 0.2113 9 Time domain energy 0.5679 0.9909 0.3504 10 Time domain energy entropy 0.0365 0.9622 -0.1107 11 Frequency domain energy entropy 0.6314 0.9781 1.7646 12 Spectral centroid 0.7543 0.9806 2.2125 13 Spectral flux 0.0769 0.9635 0 14 Spectral attenuation 0.3879 0.9776 0.4975 15 Spectral center of mass energy transfer 0.9150 0.9866 2.5821
[0083] Table 2 Compressor blade crack fault detection results
[0084] Speed 1200rpm 1500rpm 1800rpm 2100rpm Accuracy 100% 99.3% 96.4% 100%
[0085] Compared with existing technologies, the present invention's compressor blade crack fault detection method based on spectral centroid energy migration and hybrid feature selection reflects blade crack fault information by extracting features from the time domain, frequency domain, acoustic emission, and spectral centroid energy migration. This method, combined with a hybrid feature selection method, optimizes features and finally uses a long-short-term memory neural network to detect compressor blade crack faults. This method overcomes the limitation of traditional features in accurately reflecting blade crack faults and improves the accuracy of compressor blade crack fault detection.
[0086] The specific embodiments described in this invention are merely preferred embodiments of the invention and are not intended to limit the scope of the invention. That is, any equivalent changes and modifications made within the scope of the patent application of this invention should be considered as the technical scope of this invention.
Claims
1. A method for detecting crack faults in compressor blades, It is characterized in that The following steps are involved: Step 1: Collect at least two channels of acoustic emission signals at the compressor outlet and divide them into two-channel training samples and two-channel test samples; Step 2: extracting features from the two-channel training samples, including acoustic emission features, time domain features, frequency domain features, and spectral centroid energy migration features; Step 3: Use the hybrid feature selection method to select the extracted features and establish the optimal feature subset; Step 4: extracting the features of the two-channel training samples and the two-channel test samples according to the optimal feature subset, and establishing a two-channel training sample feature subset and a two-channel test sample feature subset; Step 5: Combining the two-channel training sample feature subsets and inputting them into the neural network for training; Step 6: Combining the two-channel test sample feature subsets and inputting them into the trained neural network for crack fault classification and detection, thereby realizing crack fault detection of compressor blades; In step 3, the specific steps of selecting the features using the hybrid feature selection method are: Step 3.1, calculate the Laplacian scores of all features: Among them, L t represents the Laplace score of feature t, s t represents the variance of feature t in the training sample, t q represents the value of feature t in the qth training sample, S qr Represents the weight coefficient of the qth and rth training samples; Step 3.2, calculate the random forest score for all features: Among them, R t represents the random forest score of feature t, er 1 (v) represents the out-of-bag estimation error of the vth random tree, er 2t (v) represents the second-type out-of-bag estimation error of the v-th random tree for feature t; Step 3.3, calculate the mixed fraction of all features, and select features with mixed fractions greater than a preset value to form a potential feature set, where the mixed fraction is: Among them, TS t represents the mixture fraction of feature t, and represents the Laplace and random forest scores of feature t after normalization between 0 and 1, and a represents the weight of the random forest score after standardization; Step 3.4, calculate the detection accuracy and its change rate under different number of feature combinations, and select the number of features at the inflection point as the number of features in the optimal feature subset; Step 3.5, creating several potential feature subsets from the potential feature set according to the number of features in the optimal feature subset, and calculating the detection accuracy of the potential feature subsets based on the random forest algorithm, and the potential feature subset with the highest detection accuracy is the optimal feature subset.
2. The compressor blade crack fault detection method according to claim 1, It is characterized in that In step 3.3, the default value is set to 0.
5.
3. The compressor blade crack fault detection method according to claim 1, It is characterized in that In step 2: The acoustic emission characteristics include: acoustic emission rise time: AE ris =t max -t th1 , acoustic emission duration: AE dur =t th2 -t th1 , maximum amplitude of acoustic emission: AE amp =|max x(t)|, acoustic emission energy: Acoustic Emission Energy Count AE evc , acoustic emission ringing count AE ric , RMS value: Zero crossing rate ZCR; The time domain features include: time domain energy: Energy entropy in time domain: The frequency domain features include: frequency domain energy entropy: The spectrum centroid energy migration characteristics include: spectrum centroid: Spectral throughput: Spectral attenuation: Among them, x i (t) represents the amplitude at the i-th time, N represents the signal length, and f i represents the i-th frequency value, X i (f) represents frequency f i The amplitude at Indicates the frequency f at the previous moment k The amplitude at t max Indicates the moment when the acoustic emission amplitude is the highest, t th1 Indicates the time when the acoustic emission process signal rises above the threshold, t th2 Indicates the moment when the acoustic emission process signal drops below the threshold.
4. The compressor blade crack fault detection method according to claim 1, It is characterized in that In step 2, the expression of the spectrum center of mass energy migration characteristic is: Among them, SCES represents the energy transfer characteristics of the spectrum center of mass, SC represents the above spectrum center of mass, and ER i (f) represents frequency f i Energy ringing at Among them, E i (f) represents frequency f i energy, T represents the energy threshold.
5. The compressor blade crack fault detection method according to claim 1, It is characterized in that In step 5, the two-channel training sample feature subsets are merged into direct concatenation.
6. The compressor blade crack fault detection method according to claim 1, It is characterized in that In step 5, the neural network is a long short-term memory neural network.
7. The compressor blade crack fault detection method according to claim 6, It is characterized in that The dimension of the LSTM neural network input feature is twice the number of features in the optimal feature subset, the number of hidden gates of the LSTM neural network is 200, the classification label of the LSTM neural network is 3 categories, and the training cycle of the LSTM neural network is 100.
Citation Information
Patent Citations
Method for detecting girder cracks based on image processing
CN105787486A
Blade tip timing and random forest-based blade crack online measurement method
CN112364886A
System and method for real-time detection of gas turbine or aircraft engine blade problems
US20090228230A1
Method for Fault Diagnosis of an Aero-engine Rolling Bearing Based on Random Forest of Power Spectrum Entropy
US20200200648A1
