Furnace tube leakage detection method based on KL distance weight optimization multi-channel acoustic characteristics

Abstract

The invention provides a furnace tube leakage detection method based on KL distance weight optimization multipath acoustic characteristics. The method comprises the following steps: S1, collecting thermal state background noise under a leakage-free condition and a tiny leakage sound signal under the thermal state background noise; s2, selecting acoustic signal characteristics used in the field of furnace tube leakage detection; s3, in the presence or absence of leakage, for each sound signal feature selected in the step S2, respectively calculating a KL distance reflecting the value distribution difference of the sound signal features; s4, performing feature sorting on the multiple paths of sound signals based on the KL distance weight; and S5, determining the number of optimized features.

Description

Technical Field

[0001] The present invention belongs to the technical field of boiler pipeline safety monitoring, and relates to a method for detecting furnace tube leakage based on the optimization of multiple acoustic features with Kullback-Leible (KL) distance weights. This method is mainly used to detect minute leakage of furnace tubes in the background noise environment of a hot boiler, and is particularly suitable for accurately identifying early leakage during the operation of power plant boilers. Background Art

[0002] A boiler is the main equipment in the thermal power production process, responsible for generating steam to drive the generator set, and its safe and stable operation is crucial for the power plant. Due to the high-temperature and high-pressure environment inside the boiler, furnace tube leakage accidents occur frequently. If the furnace tube leakage cannot be detected in time, it will cause chain damage to the boiler, bringing huge economic losses and safety hazards to the power plant production. Therefore, it is of great significance to detect furnace tube leakage promptly and accurately in the initial stage of leakage.

[0003] Due to the high sensitivity of acoustic sensors, most power plant boilers are equipped with acoustic detection equipment for furnace tube leakage [1]. The existing technologies in power plants usually set the signal energy threshold as the standard for leakage judgment based on historical acoustic signal data. However, there is strong background noise in the furnace, and the minute leakage acoustic signals of furnace tubes will be submerged by the noise, and this technology is prone to false negative phenomena.

[0004] In order to improve the accuracy of detecting minute leakage of furnace tubes, more sensitive acoustic signal features need to be sought. Some studies have proposed methods for judging leakage by using time-domain, frequency-domain statistical features and wavelet packet energy ratio features of acoustic signals, combined with machine learning algorithms [2]-[3]. Although these methods have obtained high detection accuracy, they do not fully consider the information of multiple acoustic sensors, resulting in insufficient generalization ability of the methods. Moreover, these methods are only applicable to laboratory conditions without thermal background noise and are difficult to be applied to the actual boiler site.

[0005] In order to make full use of the information of multiple acoustic sensors to improve the detection accuracy of minute leakage, the present invention designs a weight calculation formula based on KL distance to optimize multiple types of time-domain and frequency-domain features and wavelet packet energy ratio features of multiple acoustic signals, and then inputs the optimized features into a Support Vector Machine (SVM) for leakage detection.

[0006] References

[0007] [1]S Panday R, Indrawan N, Shadle L J, et al. Leak detection in a subcritical boiler[J]. Applied Thermal Engineering, 2021, 185: 116371.

[0008] [2]ohaib M, Kim J M. Data driven leakage detection and classification of a boiler tube[J]. Applied Sciences, 2019, 9(12): 2450.

[0009] [3]Kim Y H, Kim J, Kim J M. Leakage detection of a boiler tube using a genetic algorithm-like method and support vector machines[C]. Proceedings of the Tenth International Conference on Soft Computing and Pattern Recognition(SoCPaR 2018)10. Springer International Publishing, 2020: 86-93. Summary of the Invention

[0010] The present invention proposes a detection method for preferentially selecting multi-channel acoustic features based on KL distance weights, aiming to solve the problem of low accuracy in detecting minute leaks in a hot background noise environment. The present invention first selects the acoustic signal features commonly used in furnace tube leak detection, then preferentially selects various features of multi-channel acoustic signals according to the designed KL distance weight calculation formula, and finally combines SVM for leak detection. The technical solution of the present invention is as follows:

[0011] A method for detecting furnace tube leaks by preferentially selecting multi-channel acoustic features based on KL distance weights, comprising the following steps:

[0012] S1, Collect the hot background noise without leakage and the minute leak acoustic signals under the hot background noise;

[0013] S2, Select the acoustic signal features used in the field of furnace tube leak detection;

[0014] S3. For each of the acoustic signal features selected in step S2, calculate the KL distance, denoted as d, which reflects the difference in the numerical distribution of the acoustic signal features, in both the cases of leakage and no leakage. KL value;

[0015] S4. Feature ranking of multi-channel acoustic signals based on KL distance weights, with the method as follows:

[0016] S41. Construct a training data set, and perform a descending order arrangement based on the KL distance for each of the M types of features of the N-channel acoustic signals in each case of the leakage source position.

[0017] S42. For the M types of features of the N-channel acoustic signals, calculate and take the maximum value of the d KL value of each type of acoustic signal feature among the N-channel acoustic signals to obtain a descending order arrangement of the M types of features in each case of the leakage source position; comprehensively consider the feature ranking results in all cases of the leakage source position and take into account the KL distance weights to obtain the final feature ranking.

[0018] S5. Determination of the optimal number of features.

[0019] Furthermore, in step S2, for each of the collected acoustic signals, construct a sample x i ∈R N×M where N is the number of acoustic channels and M is the number of selected feature types, for subsequent feature optimization.

[0020] Furthermore, in step S2, the selected acoustic signal features for the field of furnace tube leakage detection include multiple types of time-domain and frequency-domain features and wavelet packet energy ratio features.

[0021] Furthermore, in step S3, the formula for calculating the d KL value is:

[0022]

[0023] where w1 and w2 represent the two cases of leakage and no leakage; v represents a certain feature component of the sample; F is the probability distribution function of the corresponding feature; the greater the difference in the numerical distribution of the feature in the cases of leakage and no leakage, the larger the d KL value, indicating that the feature has a higher discrimination degree for the two cases.

[0024] Furthermore, in step S42, according to the designed KL distance weight calculation formulas (1) and (2), obtain the final feature ranking:

[0025]

[0026] where represents the d of the i-th feature vi in the j-th case of the leakage source positionKL Numerical value; θ i,j Denote d of the i-th feature vi under the j-th leakage source position KL Weight; θ i Denote the average d of the feature vi considering all situations KL Weight; T represents the number of all leakage source position situations in the training dataset.

[0027] Furthermore, the method in step S5 is as follows: According to the final sorting result in step S5, add features to the SVM classifier one by one to train the detection model, and use the leakage detection accuracy rate of the training dataset as the measurement standard of the model performance. Comprehensively weigh the number of features and the detection performance, and finally determine the number of preferred features.

[0028] Furthermore, in step S5, the calculation formula of the detection accuracy rate is:

[0029]

[0030] Among them, TP and TN respectively represent the number of data points correctly judged as having leakage and no leakage, and FP and FN respectively represent the number of data points misjudged as having leakage and no leakage. Description of the Drawings

[0031] Figure 1 : Calculation results of the feature d under 3 leakage source positions KL Calculation results

[0032] Figure 2 : Feature sorting results of multi-channel acoustic signals based on KL distance weights

[0033] Figure 3 : Detection accuracy rates under different numbers of features Detailed Embodiment

[0034] The present invention will be further described below in conjunction with the drawings and embodiments.

[0035] In this example, 4 acoustic sensors and 4 leakage sources S1 - S4 at different positions are set up on the experimental platform. The leakage nozzle diameter is 0.6 mm, and the leakage pressures include three situations of 0.2 MPa, 0.3 MPa, and 0.4 MPa, combined into 12 kinds of micro-leakage working conditions. The sampling frequency is set to 40 kHz.

[0036] 1) Fuse the boiler hot-state background noise signal and the micro-leakage acoustic signal collected in the laboratory to obtain a noisy leakage acoustic signal. Among them, in the case of leakage sources S1 - S3, 80% of the data is used for model training, and the remaining 20% of the data is used for model testing; the data in the case of leakage source S4 is used to verify the generalization ability of the model.

[0037] 2) Twenty-nine types of acoustic signal features were selected, including 9 time-domain statistical features, 4 frequency-domain statistical features, and 16 energy ratio features in different frequency bands after four-layer wavelet packet decomposition, as shown in Table 1. The 29 types of features of the four acoustic signals were combined as one sample x i ∈R 116 .

[0038] Table 1 Acoustic signal features

[0039]

[0040] 4) Calculate d of 116 acoustic signal features under three cases of leakage source positions KL , as Figure 1 shown.

[0041] 5) Considering the three cases of leakage source positions comprehensively, according to the designed KL distance weight calculation formula, the ranking of the 29 types of features was obtained, and the ranking results are as Figure 2 shown.

[0042] 6) According to the ranking results, the features were added to the SVM classifier one by one, and the radial basis kernel function was selected to obtain the leakage detection accuracy when different numbers of features were input, as Figure 3 shown. When the number of input features is greater than or equal to 16, the increase in accuracy gradually levels off, and the classification performance of the SVM model basically reaches the optimal. Therefore, the first 16 features of the four signals were selected for leakage detection.

[0043] 7) The detection results of the test dataset are shown in Table 2.

[0044] Table 2 Detection accuracy of the test dataset

[0045]

[0046] 8) The verification results of the generalization ability of the method are shown in Table 3.

[0047] Table 3 Detection accuracy under leakage source S4

[0048]

[0049] The present invention proposes a detection method for preferentially selecting multi-channel acoustic features based on KL distance weight, which makes full use of multi-channel acoustic sensor information. Under the hot background noise, the detection accuracy of small leaks in furnace tubes is improved, and the generalization ability of the method is enhanced. The present invention reduces the risk of large-scale safety accidents and improves the emergency rescue efficiency of boiler pipeline accidents.

Claims

1. A method for detecting furnace tube leakage by preferentially selecting multi-channel acoustic features based on KL distance weights, comprising the following steps: S1, Collect the hot background noise and the tiny leakage sound signal under the hot background noise without leakage. S2, Select the acoustic signal features for the field of furnace tube leakage detection. S3. For each acoustic signal feature selected in step S2, calculate the KL distance, denoted as d, which reflects the difference in the numerical distribution of the acoustic signal feature, in both the case of leakage and non-leakage. KL value; S4, Feature ranking of multi-channel acoustic signals based on KL distance weights. The method is as follows: S41, Construct a training data set, and perform descending order arrangement based on KL distance for the M types of features of the N-channel acoustic signals in each case of leakage source position. S42. For the M types of features of N-channel acoustic signals, calculate and take the maximum value of the d values of each type of acoustic signal feature among the N-channel acoustic signals, and obtain the descending order of the M types of features in each case of the leakage source position; comprehensively consider the feature sorting results in all cases of the leakage source position, and consider the KL distance weight to obtain the final feature sorting. KL ​ S5, Determination of the number of preferentially selected features.

2. The method for detecting furnace tube leakage according to claim 1, characterized in that, In step S2, for each of the collected acoustic signals, a sample x of the selected types of acoustic signal features is constructed i ∈R N×M , where N is the number of acoustic channels and M is the number of selected feature types, for subsequent feature optimization.

3. The method for detecting furnace tube leakage according to claim 1, characterized in that, In step S2, the selected acoustic signal features for the field of furnace tube leakage detection include multiple types of time-domain and frequency-domain features and wavelet packet energy ratio features.

4. The furnace tube leakage detection method according to claim 1, characterized in that In step S3, calculate d KL The formula for the numerical value is: Among them, w1 and w2 represent the two situations of leakage and no leakage; v represents a certain characteristic component of the sample; F is the probability distribution function of the corresponding characteristic. The greater the difference in the numerical distribution of the characteristic under the situations of leakage and no leakage, the KL larger the value of d is, indicating that the characteristic has a higher discrimination degree for the two situations.

5. The method for detecting furnace tube leakage according to claim 1, wherein, In step S42, according to the designed KL distance weight calculation formulas (1) and (2), obtain the final feature ranking: Among them, represents the d value of the i-th feature vi in the case of the j-th leakage source location KL value; θ i,j represents the d value of the i-th feature vi in the case of the j-th leakage source location KL weight; θ i represents the average d value of the feature vi considering all cases KL weight; T represents the number of all leakage source location cases in the training dataset.

6. The method for detecting furnace tube leakage according to claim 1, wherein, The method of step S5 is as follows: According to the final ranking result in step S5, add features to the SVM classifier one by one to train the detection model, and use the leakage detection accuracy of the training data set as the measure of the model performance. Comprehensively weigh the number of features and the detection performance, and finally determine the number of preferentially selected features.

7. The method for detecting furnace tube leakage according to claim 6, characterized in that, In step S5, the calculation formula for the detection accuracy is: Among them, TP and TN respectively represent the number of data points correctly judged as having leakage and no leakage, and FP and FN respectively represent the number of data points misjudged as having leakage and no leakage.

Images