Thermal power plant fan equipment fault diagnosis method based on fusion fault diagnosis model

By constructing SVM, DF, and LGBM models and using improved D-S evidence theory and multi-core maximum average distance method, the problems of low accuracy of single classifiers and insufficient conflict and fusion between evidence in the fault diagnosis of fan equipment in thermal power plant are solved, achieving high accuracy Rate and robust fault diagnosis results.

CN119066614AActive Publication Date: 2024-12-03SHANDONG UNIV OF SCI & TECH

Patent Information

Application Number
CN202411196055.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-29
Publication Date
2024-12-03
Estimated Expiration
2044-08-29

AI Technical Summary

Technical Problem

In the diagnosis of fan equipment in thermal power plants, the accuracy of a single classifier is low, and the traditional D-S evidence theory has insufficient conflicts and fusion between evidence, resulting in inaccurate diagnosis results.

Method used

Three algorithms: SVM, DF, and LGBM are used to build a fusion fault diagnosis model. The weight of the evidence body is calculated through the improved D-S evidence theory and the multi-core maximum average distance method, and combined with the Dempster decision fusion, the weighted fusion between the evidence is achieved.

Benefits of technology

It improves the accuracy and robustness of fault diagnosis, reduces the occurrence of conflict events, improves the effectiveness of detection of abnormal operating conditions of fans in thermal power plants, and enhances the rationality and effectiveness of diagnostic results.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119066614A_ABST
    Figure CN119066614A_ABST
Patent Text Reader

Abstract

The invention relates to the technical field of fault diagnosis and detection, in particular to a thermal power plant fan equipment fault diagnosis method based on a fusion fault diagnosis model. According to the method, multiple time domain feature vectors are collected from two sensor signals, and then a fusion fault diagnosis model is constructed by adopting three algorithms of a support vector machine, a deep forest and a lightweight gradient boosting tree; and finally, calculating Euclidean distance and MK-MMD by using an improved D-S evidence theory method, and performing decision fusion on fault information of the three models. According to the method, when classifiers are selected for combination, three algorithms with good classification performance in the field of fault diagnosis are selected as evidence bodies, and the principles of the algorithms aim at different emphasis points during data classification, so that data can be processed and analyzed from multiple angles; and when the weight coefficient is obtained, the internal and external relations between the evidences are comprehensively considered, so that the occurrence of conflict events is reduced, misdiagnosis and missed diagnosis phenomena are avoided, and the diagnosis accuracy is improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of fault diagnosis and detection, and in particular to a method for diagnosing fan equipment faults in a thermal power plant based on a fusion fault diagnosis model. Background Art

[0002] Thermal power generation occupies an important position in my country's power energy structure. Large-capacity, low-energy thermal power units are key equipment in the field of power generation. Efficient operation is of great significance to promoting the transformation of energy structure. Among them, wind turbine equipment is an indispensable auxiliary equipment in large thermal power plants. Its structure is complex and there are many monitoring parameters. It runs under various extreme working conditions for a long time, resulting in frequent abnormal working conditions. The occurrence of faults will not only reduce the operating efficiency of auxiliary equipment, but also cause serious safety accidents. Therefore, the research on state monitoring and fault warning of wind turbines has important theoretical and practical significance for improving the power generation efficiency, overall operation and maintenance, and safe production of thermal power plants. Traditional wind turbine equipment fault diagnosis mainly relies on the method of mechanism model, which is to establish a precise input and output mathematical model of the system, and the model mechanism is clear and interpretable. However, for non-Gaussian, nonlinear, and complex equipment such as thermal power plant wind turbines, modeling is difficult. With the rapid development of industrial intelligence, it is convenient to obtain actual equipment operation data, and data-driven fault diagnosis methods have become a research hotspot. This method does not require a clear physical model and prior knowledge of the system. It uses sensors to collect historical data for analysis and processing to detect abnormal conditions, which can overcome the uncertainty and time-varying problems of the system. Related research is mainly aimed at the optimization and improvement of a single algorithm. However, the use of a single model for abnormal monitoring or fault diagnosis has limitations and one-sidedness, and cannot identify abnormal conditions of the system from multiple angles and multiple theories. Therefore, how to reasonably and effectively integrate multiple algorithms to improve the accuracy of diagnosis and achieve comprehensive risk assessment of the entire system is the focus and difficulty of current research. Information fusion theory effectively solves the problem of collaborative processing of multiple information sources, making the same thing more comprehensive and objective. Since different algorithm principles may lead to different preferences for data mining and analysis, it is critical to choose a suitable fusion method for different research problems. Dempster-Shafer (DS) evidence theory is also a widely used information fusion method. Compared with traditional probability theory, it is easier to obtain prior data, has weaker restrictions and more standard theoretical derivations. It can directly express "uncertainty" and "unknown", and can integrate information from multiple evidence sources to quantify the uncertainty of events. On the one hand, the DS method can integrate multiple algorithm models. On the other hand, multi-sensor information fusion technology is also developing rapidly, which can effectively solve the problem of multi-source information processing of large and complex equipment. In addition, the DS evidence theory also has shortcomings. When there is a high degree of conflict between evidences, there is a high probability of wrong judgment results. For this reason, scholars have proposed many solutions. One is to improve the rules for merging evidence, such as Semts, Yager, Voorbraak, etc. Another is to pre-process the evidence before fusion based on the Dempster rule, and assign weights to the evidence according to cosine similarity and gray correlation analysis. In order to solve the above technical problems, technicians in this field have made the following efforts.For example, the Chinese patent publication number is CN112036301A, which discloses a method for constructing a drive motor fault diagnosis model based on intra-class feature transfer learning and multi-source information fusion. In view of the problem that there are certain uncertainties in the signal acquisition of a single sensor, the DS evidence theory is used to fuse the decision-making layer of the drive motor multi-source information, and the diagnosis results of the vibration and current signals on the model are subjected to secondary DS evidence fusion. However, the above scheme still has the following problems that have not been solved: as an important auxiliary equipment in a thermal power plant, the fan often fails during its operation. When a single classifier is used for fault diagnosis, it has low accuracy and randomness, and the traditional DS evidence theory has deficiencies in conflict fusion between evidence. Summary of the invention

[0003] The technical problem to be solved by the present invention is to overcome the deficiencies of the prior art and provide a method for diagnosing fan equipment faults in a thermal power plant based on a fusion fault diagnosis model.

[0004] The technical solution adopted by the present invention is as follows: A method for diagnosing fan equipment faults in a thermal power plant based on a fusion fault diagnosis model comprises the following steps: S1, collecting vibration signals and torque signals of the fan sensor of the power plant, then extracting multiple time domain feature vectors from the vibration signals and torque signals simultaneously to obtain a time domain feature data set and perform standardization processing, and randomly dividing the standardized time domain feature data set into a training data set and a test data set; S2, input the training data set in S1 into the SVM model to train the model to obtain the SVM diagnostic model; input the training data set in S1 into the DF model to train the model to obtain the DF diagnostic model; input the training data set in S1 into the LGBM model to train the model to obtain the LGBM diagnostic model; S3: Input the test data set divided in S1 into the trained SVM diagnosis model for preliminary fault diagnosis, use the final output result of the SVM diagnosis model as the basic probability distribution function value, and use the SVM diagnosis model as the evidence body. m 2. Input the test data set divided in S1 into the trained DF diagnosis model for preliminary fault diagnosis, use the final output result of the DF diagnosis model as the basic probability distribution function value, and use the DF diagnosis model as the evidence body. m 2. Input the test data set divided in S1 into the trained LGBM diagnostic model for preliminary fault diagnosis, and use the final output result of the LGBM diagnostic model as the basic probability distribution function value, and use the LGBM diagnostic model as the evidence body. m 3; S4, using the improved DS evidence theory fusion method, first use the Euclidean distance to calculate the support of the three evidence bodies obtained in S3, and then obtain a layer of credibility; S5. Use the multi-core maximum average distance method to calculate the support and double credibility of the three evidence bodies obtained in S3, correct the two credibility to obtain the weight coefficient of each evidence body, and then multiply the weight coefficient by the corresponding evidence body to obtain the weighted evidence body; S6. Perform decision fusion on the three weighted evidence bodies according to the Dempster decision fusion method to obtain a fusion result; use the fault rule corresponding to the maximum basic probability function value to judge the fusion result and output the final fault diagnosis result.

[0005] According to one embodiment of the present invention, in S1, data is collected by using vibration sensors and torque sensors installed at different positions of the same device of the power plant wind turbine, and then multiple time domain feature parameters of the vibration signal and torque signal are respectively extracted to obtain time domain feature vectors to construct a time domain feature data set; Since the magnitudes of features in different time domains vary, Z-score standardization is used to avoid neglecting features with smaller magnitudes and to give them a specific mean and standard deviation. Therefore, the transformed data has the characteristics of a standard normal distribution, which is used to help improve the performance and stability of the model. Z-score standardization is performed through the following steps:

[0006] in: is the original data point is the normalized data point, μ is the mean, σ is the standard deviation; The standardized time domain feature dataset is divided into a training dataset and a test dataset.

[0007] According to an embodiment of the present invention, in S1, the multiple time domain characteristic parameters include: Maximum value:

[0008] RMS:

[0009] Square root amplitude:

[0010] Standard Deviation:

[0011] Peak indicators:

[0012] Margin index:

[0013] Absolute average: .

[0014] According to one embodiment of the present invention, in S2, the SVM model uses a nonlinear mapping function to map the training samples from the input space to a higher-dimensional feature space, and finds an optimal separation hyperplane to distinguish sample points of different categories and maximize the interval difference; the DF model can automatically adjust the number of layers of the deep structure to adapt to a variety of data sets; the LGBM model is an efficient integrated learning algorithm; the training data set in S1 is input into the SVM model to train the model to obtain the SVM diagnostic model; the training data set in S1 is input into the DF model to train the model to obtain the DF diagnostic model; the training data set in S1 is input into the LGBM model to train the model to obtain the LGBM diagnostic model.

[0015] According to an embodiment of the present invention, in S3, the test data set divided in S1 is input into the trained SVM diagnosis model for preliminary fault diagnosis, and the prediction result of the SVM diagnosis model is used as the basic probability distribution function value to construct the evidence body. m 1. Input the test data set divided in S1 into the trained DF diagnosis model for preliminary fault diagnosis, and use the final prediction result of the DF diagnosis model as the basic probability distribution function value to construct the evidence body. m 2. Input the test data set divided in S1 into the trained LGBM diagnostic model for preliminary fault diagnosis, and use the final prediction result of the LGBM diagnostic model as the basic probability distribution function value to construct the evidence body. m 3.

[0016] According to an embodiment of the present invention, in S4, an improved DS evidence theory fusion method is used to calculate the differences of the three evidence bodies obtained in S3, obtain the weight value of each evidence body, and then multiply the weight value by the corresponding evidence body to obtain the weighted evidence body, including the following steps: S41, using the three evidence bodies obtained in S3, calculate the evidence bodies respectively m 1. Body of evidence m 2. Body of evidence m 3Euclidean distance between two pairs; Assume i The first j The basic trust distribution functions of each evidence body are m i and m j , using Euclidean distance To measure the difference between two bodies of evidence, the calculation formula is as follows:

[0017] in: i =1, 2, 3; j =1,2,3, and when i = j hour, =0; represents the inner product of vectors; S42, based on n The difference values ​​between each pair of evidences form an n-row and n-column DM Matrix, since the Euclidean distance between the evidence body and itself is 0, the values ​​on the diagonal of the matrix are all 0, and it is a diagonal matrix; The larger the value, the greater the distance between the vectors, that is, the greater the difference between the evidences, then:

[0018] Calculate the support matrix SM , SM Each element in ,but:

[0019] Obtain support for each body of evidence , the greater the support, the higher the trust in the evidence. m i Support normalization to obtain a level of credibility for each body of evidence Crd 1, and ;

[0020] .

[0021] According to an embodiment of the present invention, in S5, the multi-kernel maximum mean distance method is used to detect whether two distributions are the same, which is an extension of the maximum mean distance method, and improves detection performance and flexibility by combining multiple kernel functions; the maximum mean distance method is a distribution difference measurement method based on the reproducing kernel Hilbert space, by mapping the distribution to the RKHS space and using the mean in the space to compare the two distributions. If the means of the two distributions in the space are the same, they are considered to be the same, including the following steps: S51. For the sample set and , using Gaussian kernel function and polynomial kernel function, defined as follows:

[0022]

[0023] in: is the bandwidth parameter of the kernel function, c is a constant, d is the order of the polynomial, c =1, d =3; The multi-core maximum average distance method is a weighted combination of the two kernel functions, and the formula is as follows:

[0024] Among them: weight and ,satisfy and ;

[0025]

[0026] Combine the Gaussian kernel and the polynomial kernel and adjust the weights β 1 and β 2 to adjust the sensitivity to different features, thereby improving the performance of distribution detection. β 1= β 2=0.5; S52: To comprehensively consider all evidence and give higher weight to reliable evidence, use the formula in S51. Find the relationship between the three bodies of evidence , and constitute MMDM Matrix, while calculating Obtain support to better capture the complex differences between distributions, thereby improving the ability to detect distribution similarities;

[0027]

[0028] To gain support , and double credibility Crd 2, and ;

[0029]

[0030] S53, add a layer of credibility in S42 Crd1 Dual credibility with S52 Crd2 Correct the coefficients to obtain the credibility of the fusion Crd As a body of evidence m i The final weight value of the evidence mi and Crd Multiply them together to get the weighted evidence body ;

[0031] .

[0032] According to an embodiment of the present invention, in S6, the weighted evidence body is conduct n -1 Dempster fusion to get the final probability output value, including the following steps: S61. DS evidence theory is a method for dealing with uncertain information. For an event n A set of mutually exclusive complete hypotheses is called the evidence theory identification framework; its basic probability distribution BPA of m Function satisfies ,and , , and then calculate the subset A The trust function value of , indicating the support set A The degree to which it is true; the likelihood function value , which represents the set A The maximum likelihood that could be true, reflecting the non-exclusion set A The degree of confidence in the hypothesis of set A can be finally obtained, that is, the confidence interval ; Hypothetical wind turbine fault identification framework , F 1 represents the gear crack fault condition in the fault identification framework, F 2 represents the gear missing tooth fault condition, F 3 represents the gear broken tooth fault diagnosis situation; SVM model builds the evidence body m 1. DF model builds evidence body m 2. LGBM model builds evidence body m 3. Take m 1( F 1) Indicates the fault in the SVM evidence F A basic probability assignment of 1; S62. According to Dempster's fusion rule, assume that the three weighted evidence bodies are , , , so two fusions are required to get the final probability output value; first, and The two evidence bodies are combined by rules to obtain a new evidence body function ;

[0033] in: , Fault Identification Framework F Neutron Set B With subset C In case of non-conflicting faults, B ={ F 1}, C ={ F 1}; K is a normalized constant representing the degree of conflict, and the calculation formula is as follows:

[0034] in: Fault Identification Framework F Neutron Set B With subset C In case of conflicting combinations, B ={ F 1}, C ={ F 2, F 3}; will with The two evidence bodies are combined according to the rules to obtain the final evidence fusion result and the fault probability value;

[0035]

[0036] Use the fault rule corresponding to the maximum basic probability function value to judge the fusion result and output the final fault diagnosis result .

[0037] Compared with the prior art, the present invention has the following beneficial effects: (1) Using SVM, DF, and LGBM, three algorithms with good classification performance in the field of fault diagnosis, as evidence, a single algorithm can achieve a high diagnostic accuracy; (2) Fault diagnosis of fan equipment in thermal power plants based on the fusion fault diagnosis model method can better identify all fault types, and the diagnosis accuracy is higher than that of a single algorithm. The effectiveness and robustness of this method are good; (3) By calculating the Euclidean distance between the evidences, we obtain the support and first-level credibility, and then calculate the multi-core maximum average distance between the evidences to obtain the support and second-level credibility; the two credibility are corrected to obtain the final credibility as the weight value. This can fully analyze and consider the deep connection between the evidences, thereby reducing the occurrence of conflicting events and improving the accuracy of diagnosis; (4) It provides an effective means for analyzing the abnormal operating condition detection problem of fans in thermal power plants. This shows that the fusion algorithm has broad application prospects in practical applications, and the diagnostic results are more reasonable and effective, which is of great significance to improving the operating efficiency and safety of fans in power plants. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 It is a flow diagram of the method of the present invention.

[0039] Figure 2 It is the gear torque signal diagram of the three types of faults in the present invention.

[0040] Figure 3 It is a waveform diagram of the torque signal collected by the sensor of the present invention. DETAILED DESCRIPTION

[0041] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0042] In the prior art, fans are important auxiliary equipment in thermal power plants. However, faults often occur during their operation. When a single classifier is used for fault diagnosis, it has low accuracy and randomness, and the traditional DS evidence theory has deficiencies in conflict fusion between evidence. To solve this problem, the present invention proposes a fault diagnosis method for fan equipment in thermal power plants based on a fusion fault diagnosis model. First, a variety of time domain feature vectors are collected from two sensor signals, and then a fusion fault diagnosis model is constructed using three algorithms: support vector machine, deep forest, and lightweight gradient boosting tree. Finally, the Euclidean distance and MK-MMD are calculated using the improved DS evidence theory method, and the fault information of these three models is fused for decision making to reduce the degree of conflict between evidence and achieve comprehensive fault diagnosis of the entire fan system.

[0043] Example 1 like Figure 1 As shown, this embodiment provides a method for diagnosing a fault of a fan device in a thermal power plant based on a fusion fault diagnosis model, comprising the following steps: S1, collect vibration and torque signals of the fan sensor in the power plant, such as Figure 2 and Figure 3 As shown, 200 sampling points are selected and then multiple time domain feature vectors are extracted from the vibration signal and the torque signal to obtain a time domain feature data set and perform a standardized process. The standardized time domain feature data set is randomly divided into a training data set and a test data set; S2, input the training data set in S1 into the SVM model to train the model to obtain the SVM diagnostic model; input the training data set in S1 into the DF model to train the model to obtain the DF diagnostic model; input the training data set in S1 into the LGBM model to train the model to obtain the LGBM diagnostic model; S3: Input the test data set divided in S1 into the trained SVM diagnosis model for preliminary fault diagnosis, use the final output result of the SVM diagnosis model as the basic probability distribution function value, and use the SVM diagnosis model as the evidence body. m 2. Input the test data set divided in S1 into the trained DF diagnosis model for preliminary fault diagnosis, use the final output result of the DF diagnosis model as the basic probability distribution function value, and use the DF diagnosis model as the evidence body. m 2. Input the test data set divided in S1 into the trained LGBM diagnostic model for preliminary fault diagnosis, and use the final output result of the LGBM diagnostic model as the basic probability distribution function value, and use the LGBM diagnostic model as the evidence body. m 3; S4, using the improved DS evidence theory fusion method, first use the Euclidean distance to calculate the support of the three evidence bodies obtained in S3, and then obtain a layer of credibility; S5. Use the multi-core maximum average distance method to calculate the support and double credibility of the three evidence bodies obtained in S3, correct the two credibility to obtain the weight coefficient of each evidence body, and then multiply the weight coefficient by the corresponding evidence body to obtain the weighted evidence body; S6. Perform decision fusion on the three weighted evidence bodies according to the Dempster decision fusion method to obtain a fusion result; use the fault rule corresponding to the maximum basic probability function value to judge the fusion result and output the final fault diagnosis result.

[0044] According to one embodiment of the present invention, in S1, data is collected by using vibration sensors and torque sensors installed at different positions of the same device of the power plant wind turbine, and then multiple time domain feature parameters of the vibration signal and torque signal are respectively extracted to obtain time domain feature vectors to construct a time domain feature data set; Since the magnitudes of features in different time domains vary, Z-score standardization is used to avoid neglecting features with smaller magnitudes and to give them a specific mean and standard deviation. Therefore, the transformed data has the characteristics of a standard normal distribution, which is used to help improve the performance and stability of the model. Z-score standardization is performed through the following steps:

[0045] in: is the original data point is the normalized data point, μ is the mean, σ is the standard deviation; The standardized time domain feature dataset is divided into a training dataset and a test dataset.

[0046] According to an embodiment of the present invention, in S1, the multiple time domain characteristic parameters include: Table 1 Time domain feature description

[0047] According to one embodiment of the present invention, in S2, the SVM model uses a nonlinear mapping function to map the training samples from the input space to a higher-dimensional feature space, and finds an optimal separation hyperplane to distinguish sample points of different categories and maximize the interval difference; the DF model can automatically adjust the number of layers of the deep structure to adapt to a variety of data sets; the LGBM model is an efficient integrated learning algorithm; the training data set in S1 is input into the SVM model to train the model to obtain the SVM diagnostic model; the training data set in S1 is input into the DF model to train the model to obtain the DF diagnostic model; the training data set in S1 is input into the LGBM model to train the model to obtain the LGBM diagnostic model.

[0048] According to an embodiment of the present invention, in S3, the test data set divided in S1 is input into the trained SVM diagnosis model for preliminary fault diagnosis, and the prediction result of the SVM diagnosis model is used as the basic probability distribution function value to construct the evidence body. m 1. Input the test data set divided in S1 into the trained DF diagnosis model for preliminary fault diagnosis, and use the final prediction result of the DF diagnosis model as the basic probability distribution function value to construct the evidence body. m 2. Input the test data set divided in S1 into the trained LGBM diagnostic model for preliminary fault diagnosis, and use the final prediction result of the LGBM diagnostic model as the basic probability distribution function value to construct the evidence body. m 3.

[0049] According to an embodiment of the present invention, in S4, an improved DS evidence theory fusion method is used to calculate the differences of the three evidence bodies obtained in S3, obtain the weight value of each evidence body, and then multiply the weight value by the corresponding evidence body to obtain the weighted evidence body, including the following steps: S41, using the three evidence bodies obtained in S3, calculate the evidence bodies respectively m 1. Body of evidence m 2. Body of evidence m 3Euclidean distance between two pairs; Assume i The first j The basic trust distribution functions of each evidence body are m i and m j , using Euclidean distance To measure the difference between two bodies of evidence, the calculation formula is as follows:

[0050] in: i =1, 2, 3; j =1,2,3, and when i = j hour, =0; represents the inner product of vectors; S42, based on n The difference values ​​between each pair of evidences form an n-row and n-column DM Matrix, since the Euclidean distance between the evidence body and itself is 0, the values ​​on the diagonal of the matrix are all 0, and it is a diagonal matrix; The larger the value, the greater the distance between the vectors, that is, the greater the difference between the evidences, then:

[0051] Calculate the support matrix SM , SM Each element in ,but:

[0052] Obtain support for each body of evidence , the greater the support, the higher the trust in the evidence. m i Support normalization to obtain a level of credibility for each body of evidence Crd 1, and ;

[0053] .

[0054] According to an embodiment of the present invention, in S5, the multi-kernel maximum mean distance method is used to detect whether two distributions are the same, which is an extension of the maximum mean distance method, and improves detection performance and flexibility by combining multiple kernel functions; the maximum mean distance method is a distribution difference measurement method based on the reproducing kernel Hilbert space, by mapping the distribution to the RKHS space and using the mean in the space to compare the two distributions. If the means of the two distributions in the space are the same, they are considered to be the same, including the following steps: S51. For the sample set and , using Gaussian kernel function and polynomial kernel function, defined as follows:

[0055]

[0056] in: is the bandwidth parameter of the kernel function, c is a constant, d is the order of the polynomial, c =1, d =3; The multi-core maximum average distance method is a weighted combination of the two kernel functions, and the formula is as follows:

[0057] Among them: weight and ,satisfy and ;

[0058]

[0059] Combine the Gaussian kernel and the polynomial kernel and adjust the weights β 1 and β 2 to adjust the sensitivity to different features, thereby improving the performance of distribution detection. β 1= β 2=0.5; S52: To comprehensively consider all evidence and give higher weight to reliable evidence, use the formula in S51. Find the relationship between the three bodies of evidence , and constitute MMDM Matrix, while calculating Obtain support to better capture the complex differences between distributions, thereby improving the ability to detect distribution similarities;

[0060]

[0061] To gain support , and double credibility Crd 2, and ;

[0062]

[0063] S53, add a layer of credibility in S42 Crd1 Dual credibility with S52 Crd2 Correct the coefficients to obtain the credibility of the fusion Crd As a body of evidence m i The final weight value of the evidence m i and Crd Multiply them together to get the weighted evidence body ;

[0064] .

[0065] According to an embodiment of the present invention, in S6, the weighted evidence body is conduct n -1 Dempster fusion to get the final probability output value, including the following steps: S61. DS evidence theory is a method for dealing with uncertain information. For an event n A set of mutually exclusive complete hypotheses is called the evidence theory identification framework; its basic probability distribution BPA of m Function satisfies ,and , , and then calculate the subset A The trust function value of , indicating the support set A The degree to which it is true; the likelihood function value , which represents the set A The maximum likelihood that could be true, reflecting the non-exclusion set A The degree of confidence in the hypothesis of set A can be finally obtained, that is, the confidence interval ; Hypothetical wind turbine fault identification framework , F1 represents the gear crack fault condition in the fault identification framework, F 2 represents the gear missing tooth fault condition, F 3 represents the gear broken tooth fault diagnosis situation; SVM model builds the evidence body m 1. DF model builds evidence body m 2. LGBM model builds evidence body m 3. Take m 1( F 1) Indicates the fault in the SVM evidence F A basic probability assignment of 1; S62. According to Dempster's fusion rule, assume that the three weighted evidence bodies are , , , so two fusions are required to get the final probability output value; first, and The two evidence bodies are combined by rules to obtain a new evidence body function ;

[0066] in: , Fault Identification Framework F Neutron Set B With subset C In case of non-conflicting faults, B ={ F 1}, C ={ F 1}; K is a normalized constant representing the degree of conflict, and the calculation formula is as follows:

[0067] in: Fault Identification Framework F Neutron Set B With subset C In case of conflicting combinations, B ={ F 1}, C ={ F 2, F 3}; will with The two evidence bodies are combined according to the rules to obtain the final evidence fusion result and the fault probability value;

[0068]

[0069] Use the fault rule corresponding to the maximum basic probability function value to judge the fusion result and output the final fault diagnosis result .

[0070] Through this improved method, the uncertainty reasoning system has better fusion performance. Even in the extreme case of contradictory evidence, the weight of the contradictory evidence will be very small and thus regarded as unimportant evidence.

[0071] Embodiment 2: In this embodiment, a specific case is given to demonstrate the effect of the present invention. With the help of a fan vibration sensor and a torque sensor, 2000 data of cracks, missing teeth and broken teeth are collected respectively, totaling 6000 data, and the sliding window size is selected as 4 to extract their time domain feature signals (maximum value, root mean square, root square amplitude, standard deviation, peak index, margin index, absolute average value) to obtain a 14-dimensional time domain feature vector, 500 samples of each type of fault type, and finally a time domain data set of 1500 rows and 14 columns is obtained.

[0072] After standardizing the data set, 70% (1500) samples were randomly selected as the training set, and 30% (450) samples were selected as the test set. The training set was used to train the three models, and finally the test set was used to verify the effect of the model.

[0073] Two real cases were selected. In case 1, the fault was a crack. F 1. Case 2 is missing teeth F 2, followed by detailed analysis and interpretation.

[0074] Table 2 Preliminary diagnosis results of case 1

[0075] From Table 2, we can see that using three different diagnostic methods to identify the same crack fault case can obtain different diagnostic results, such as the evidence m 1 and m 3 The fault types diagnosed are F 1, while m 2 were diagnosed as F 3. This leads to conflicting diagnostic results. Even if the initial diagnostic results of the two evidence bodies are the same, the basic probability function values ​​are different, indicating that different algorithms have different methods of analyzing data.

[0076] In order to prove the effectiveness and generalization of the method proposed in this invention, multiple fusion methods are used to fuse the preliminary diagnosis results of Case 1 in Table 2. The fault fusion diagnosis results are shown in Table 3.

[0077] Table 3 Fusion diagnosis results of case 1

[0078] As can be seen from Table 3, the DS evidence theory fusion method fuses the preliminary diagnosis results and judges the fault type as broken tooth. F 3. It contradicts the actual fault situation, indicating that a diagnosis error has occurred. The Murphy method simply sums and averages the basic probability function values ​​of the evidence body, and the fault signal is submerged, making it impossible to determine whether the fault is a crack. F 1 or broken teeth F 3. The Sunquan method improves the conflict factors in the original Dempster fusion rule, but it still cannot correctly determine the fault type. The Dengyong method and the improved DS method can correctly identify the real fault type. F 1, but the improved DS method is F The basic probability function value of 1 is 0.48, which is higher than 0.4 of Dengyong method. The results show that the improved DS method proposed in the present invention can fully consider the correlation and contradiction between evidences, improve the trust of important evidence, and finally realize the correct identification of faults.

[0079] Table 4 Preliminary diagnosis results of case 2

[0080] Table 5 Fusion diagnosis results of case 2

[0081] As can be seen from Table 4, in the preliminary diagnosis results m 1 and m 2 Diagnostic fault is F 2, while m The diagnosis result of 3 is F 3. The diagnosis results are inconsistent, but the actual fault is F 2. Description m 3. Diagnostic error. Again, the DS method still makes a wrong diagnosis, but the other fusion methods can correctly identify the fault type. The improved DS method has a higher confidence level of 0.7 for the correct fault, which is higher than other fusion methods, indicating that the diagnosis result is more reasonable and effective.

[0082] Table 6 Diagnostic accuracy of different fusion methods

[0083] As can be seen from Table 6, a single machine learning algorithm is relatively simple to implement, but its diagnostic results have low credibility and high uncertainty. This may lead to misdiagnosis or missed diagnosis in practical applications, affecting the timeliness and effectiveness of fault handling. In order to solve this problem, this study used a variety of different fusion methods to fuse information for three different classifiers. The results show that the accuracy performance indicators of the algorithm after information fusion have been improved, verifying the practicality and feasibility of information fusion technology in improving classification performance. Furthermore, the improved DS information fusion algorithm proposed in the present invention shows a higher accuracy of 94.44% when compared with the ordinary DS method and the information fusion algorithm of Dengyong et al. This excellent performance is attributed to the adoption of a weighted information fusion strategy, which assigns different weights according to the reliability of information from different sources, so that the fusion results are more accurate and reliable. However, this method also has limitations, and the overall performance of the fusion model depends greatly on the basic classification model. If the fault detection capability of a certain model is weak, the overall accuracy will be reduced.

[0084] In summary, when selecting classifiers for combination, three algorithms with good classification performance in the field of fault diagnosis, namely support vector machine, deep forest and lightweight gradient boosting machine, were selected as evidence bodies. Each algorithm principle has different focuses on data classification, and can process and analyze data from multiple angles. When obtaining weight coefficients, the first credibility is obtained by calculating the Euclidean distance between evidences, and the second credibility is obtained by calculating the multi-core maximum average distance between evidences. The combination of the two can comprehensively consider the internal and external connections between evidences, thereby reducing the occurrence of conflicting events, avoiding misdiagnosis and missed diagnosis, and improving the accuracy of diagnosis.

[0085] Although the present invention has been described in detail by referring to the accompanying drawings and in combination with the preferred embodiments, the present invention is not limited thereto. Without departing from the spirit and essence of the present invention, a person of ordinary skill in the art may make various equivalent modifications or substitutions to the embodiments of the present invention, and these modifications or substitutions shall be within the scope of the present invention. Any person of ordinary skill in the art may easily think of changes or substitutions within the technical scope disclosed by the present invention, and these shall be within the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be subject to the scope of protection of the claims.

Claims

1. A method for diagnosing fan equipment faults in thermal power plants based on a fusion fault diagnosis model, characterized in that: The steps include: S1, collecting vibration signals and torque signals of the fan sensor of the power plant, then extracting multiple time domain feature vectors from the vibration signals and torque signals simultaneously to obtain a time domain feature data set and perform standardization processing, and randomly dividing the standardized time domain feature data set into a training data set and a test data set; S2, input the training data set in S1 into the SVM model to train the model to obtain the SVM diagnostic model; input the training data set in S1 into the DF model to train the model to obtain the DF diagnostic model; input the training data set in S1 into the LGBM model to train the model to obtain the LGBM diagnostic model; S3: Input the test data set divided in S1 into the trained SVM diagnosis model for preliminary fault diagnosis, use the final output result of the SVM diagnosis model as the basic probability distribution function value, and use the SVM diagnosis model as the evidence body. m 2. Input the test data set divided in S1 into the trained DF diagnosis model for preliminary fault diagnosis, use the final output result of the DF diagnosis model as the basic probability distribution function value, and use the DF diagnosis model as the evidence body. m 2. Input the test data set divided in S1 into the trained LGBM diagnostic model for preliminary fault diagnosis, and use the final output result of the LGBM diagnostic model as the basic probability distribution function value, and use the LGBM diagnostic model as the evidence body. m 3; S4, using the improved DS evidence theory fusion method, first use the Euclidean distance to calculate the support of the three evidence bodies obtained in S3, and then obtain a layer of credibility; S5. Use the multi-core maximum average distance method to calculate the support and double credibility of the three evidence bodies obtained in S3, correct the two credibility to obtain the weight coefficient of each evidence body, and then multiply the weight coefficient by the corresponding evidence body to obtain the weighted evidence body; S6. Perform decision fusion on the three weighted evidence bodies according to the Dempster decision fusion method to obtain a fusion result; use the fault rule corresponding to the maximum basic probability function value to judge the fusion result and output the final fault diagnosis result.

2. The method for diagnosing fan equipment faults in thermal power plants based on a fusion fault diagnosis model according to claim 1, characterized in that: In S1, data is collected by using vibration sensors and torque sensors installed at different positions of the same equipment of the power plant fan, and then multiple time domain feature parameters of the vibration signal and torque signal are extracted respectively to obtain time domain feature vectors to construct a time domain feature data set; Since the magnitudes of features in different time domains vary, Z-score standardization is used to avoid neglecting features with smaller magnitudes and to give them a specific mean and standard deviation. Therefore, the transformed data has the characteristics of a standard normal distribution, which is used to help improve the performance and stability of the model. Z-score standardization is performed through the following steps: in: is the original data point is the normalized data point, μ is the mean, σ is the standard deviation; The standardized time domain feature dataset is divided into a training dataset and a test dataset.

3. The method for diagnosing fan faults in thermal power plants based on a fusion fault diagnosis model according to claim 1 or 2, characterized in that: In S1, the multiple time domain characteristic parameters include: Maximum value: RMS: Square root amplitude: Standard Deviation: Peak indicators: Margin index: Absolute average: .

4. The method for diagnosing fan faults in thermal power plants based on a fusion fault diagnosis model according to claim 3, characterized in that: In S2, the SVM model uses a nonlinear mapping function to map the training samples from the input space to a higher-dimensional feature space, and finds an optimal separation hyperplane to distinguish sample points of different categories and maximize the interval difference; the DF model can automatically adjust the number of layers of the deep structure to adapt to a variety of data sets; the LGBM model is an efficient integrated learning algorithm; the training data set in S1 is input into the SVM model to train the model to obtain the SVM diagnostic model; the training data set in S1 is input into the DF model to train the model to obtain the DF diagnostic model; the training data set in S1 is input into the LGBM model to train the model to obtain the LGBM diagnostic model.

5. The method for diagnosing fan faults in thermal power plants based on a fusion fault diagnosis model according to claim 4, characterized in that: In S3, the test data set divided in S1 is input into the trained SVM diagnosis model for preliminary fault diagnosis, and the prediction result of the SVM diagnosis model is used as the basic probability distribution function value to construct the evidence body. m 1. Input the test data set divided in S1 into the trained DF diagnosis model for preliminary fault diagnosis, and use the final prediction result of the DF diagnosis model as the basic probability distribution function value to construct the evidence body. m 2. Input the test data set divided in S1 into the trained LGBM diagnostic model for preliminary fault diagnosis, and use the final prediction result of the LGBM diagnostic model as the basic probability distribution function value to construct the evidence body. m 3.

6. The method for diagnosing fan faults in thermal power plants based on a fusion fault diagnosis model according to claim 5, characterized in that: In S4, the improved DS evidence theory fusion method is used to calculate the differences of the three evidence bodies obtained in S3, obtain the weight value of each evidence body, and then multiply the weight value by the corresponding evidence body to obtain the weighted evidence body, including the following steps: S41, using the three evidence bodies obtained in S3, calculate the evidence bodies respectively m 1. Body of evidence m 2. Body of evidence m 3Euclidean distance between two pairs; Assume i The first j The basic trust distribution functions of each evidence body are m i and m j , using Euclidean distance To measure the difference between two bodies of evidence, the calculation formula is as follows: in: i =1, 2, 3; j =1,2,3, and when i = j hour, =0; represents the inner product of vectors; S42, based on n The difference values ​​between each pair of evidences form an n-row and n-column DM Matrix, since the Euclidean distance between the evidence body and itself is 0, the values ​​on the diagonal of the matrix are all 0, and it is a diagonal matrix; The larger the value, the greater the distance between the vectors, that is, the greater the difference between the evidences, then: Calculate the support matrix SM , SM Each element in ,but: Obtain support for each body of evidence , the greater the support, the higher the trust in the evidence; m i Normalize the support to get the credibility of each evidence Crd 1, and ; 。 7. The method for diagnosing fan faults in thermal power plants based on a fusion fault diagnosis model according to claim 6, characterized in that: In S5, the multi-kernel maximum mean distance method is used to detect whether two distributions are the same. It is an extension of the maximum mean distance method, which improves detection performance and flexibility by combining multiple kernel functions. The maximum mean distance method is a distribution difference measurement method based on the reproducing kernel Hilbert space. It maps the distribution to the RKHS space and uses the mean in the space to compare the two distributions. If the means of the two distributions in the space are the same, they are considered to be the same, including the following steps: S51. For the sample set and , using Gaussian kernel function and polynomial kernel function, defined as follows: in: is the bandwidth parameter of the kernel function, c is a constant, d is the order of the polynomial, c =1, d =3; The multi-core maximum average distance method is a weighted combination of the two kernel functions, and the formula is as follows: Among them: weight and ,satisfy and ; Combine the Gaussian kernel and the polynomial kernel and adjust the weights β 1 and β 2 to adjust the sensitivity to different features, thereby improving the performance of distribution detection. β 1= β 2=0.5; S52: To comprehensively consider all evidence and give higher weight to reliable evidence, use the formula in S51. Find the relationship between the three bodies of evidence , and constitute MMDM Matrix, while calculating Obtain support to better capture the complex differences between distributions, thereby improving the ability to detect distribution similarities; To gain support , and double credibility Crd 2, and ; S53, add a layer of credibility in S42 Crd1 Dual credibility with S52 Crd2 Correct the coefficients to obtain the credibility of the fusion Crd As a body of evidence m i The final weight value of the evidence m i and Crd Multiply to get the weighted evidence body ; 。 8. The method for diagnosing fan faults in thermal power plants based on a fusion fault diagnosis model according to claim 7, characterized in that: In S6, the weighted evidence body conduct n -1 Dempster fusion to get the final probability output value, including the following steps: S61. DS evidence theory is a method for dealing with uncertain information. For an event n a set of mutually exclusive complete hypotheses, called the evidence-theoretic identification framework; Its basic probability distribution BPA of m Function satisfies ,and , , and then calculate the subset A The trust function value of , indicating the support set A The degree to which it is true; the likelihood function value , which represents the set A The maximum likelihood that could be true, reflecting the non-exclusion set A The degree of confidence in the hypothesis of set A can be finally obtained, that is, the confidence interval ; Hypothetical wind turbine fault identification framework , F 1 represents the gear crack fault condition in the fault identification framework, F 2 represents the gear missing tooth fault condition, F 3 represents the gear broken tooth fault diagnosis situation; SVM model builds the evidence body m 1. DF model builds evidence body m 2. LGBM model builds evidence body m 3. Take m 1( F 1) Indicates the fault in the SVM evidence F A basic probability assignment of 1; S62. According to Dempster's fusion rule, assume that the three weighted evidence bodies are , , , so two fusions are required to get the final probability output value; first, and The two evidence bodies are combined by rules to obtain a new evidence body function ; in: , Fault Identification Framework F Neutron Set B With subset C In case of non-conflicting faults, B ={ F 1}, C ={ F 1}; K is a normalized constant representing the degree of conflict, and the calculation formula is as follows: in: Fault Identification Framework F Neutron Set B With subset C In case of conflicting combinations, B ={ F 1}, C ={ F 2, F 3}; will with The two evidence bodies are combined according to the rules to obtain the final evidence fusion result and the fault probability value; Use the fault rule corresponding to the maximum basic probability function value to judge the fusion result and output the final fault diagnosis result .

Citation Information

Patent Citations

  • Driving motor fault diagnosis model construction method based on intra-class feature transfer learning and multi-source information fusion

    CN112036301A

  • Induced draft fan fault diagnosis method, system and device based on PNN-WNN and improved D-S

    CN115563578A

  • Artificial intelligence rolling bearing fault diagnosis method based on information fusion

    CN116484310A

  • Non-intrusive series arc detection method based on multi-classifier fusion

    CN117235659A

Cited By

  • Cloud edge collaboration method based on multi-source heterogeneous information fusion

    CN120766078A

  • Health monitoring system and method for water diversion engineering optical fiber network

    CN120915375A

  • Flexible straight valve capacitor fault diagnosis method and system

    CN121256519A

  • Multi-agent-based power grid fault traceability analysis method and system

    CN121502530A

  • Multi-source data fusion fault diagnosis method and system based on DS evidence theory

    WO2026067893A1