A vision-based online life prediction method for high-temperature components

Abstract

The present invention relates to a vision-based online life prediction method for high-temperature components, comprising the following steps: S1, obtaining a temperature distribution image of a high-temperature component, and preprocessing the temperature distribution image; S2, extracting features from the temperature distribution image, detecting defect data of a coating sample, and inputting the feature data into a life prediction model; S3, obtaining a first remaining life estimate value through a high-temperature component health assessment method based on a physical model, and obtaining a second remaining life estimate value through a high-temperature component health assessment method based on machine learning; S4, calculating the weight values ​​of the first remaining life estimate value and the second remaining life estimate value through a data fusion method based on information entropy; S5, calculating the remaining life value of the high-temperature component according to the corresponding weight value, and obtaining a health assessment result of the high-temperature component. Compared with the prior art, the present invention has the advantages of improving the accuracy of the life prediction result of the high-temperature component, ensuring the safe and efficient operation of the gas turbine, and the like.

Description

Technical Field

[0001] The invention relates to the technical field of gas turbines, and in particular to a vision-based online life prediction method for high-temperature components. Background Art

[0002] In order to improve the efficiency of gas turbines, the gas temperature continues to increase. At present, the gas temperature is much higher than the heat resistance limit of high-temperature component materials such as gas turbine blades, and thermal protection is required by spraying thermal barrier coatings. The high-temperature components (blades, etc.) of gas turbines can significantly extend their service life by coating thermal barrier coatings. Therefore, real-time online health assessment of thermal barrier coatings and accurate prediction of their remaining life have become necessary means to ensure the safe operation of gas turbines, which is of great significance to the normal operation of gas turbines.

[0003] The current methods for detecting the status of high-temperature components of gas turbines are mainly divided into two categories: offline detection and online detection. Among them, offline monitoring requires the gas turbine to be shut down for detection, and the most common method is visual inspection by maintenance personnel. However, the offline detection cycle is long and the efficiency is low. Frequent shutdowns will inevitably result in large economic losses. The online detection method can ensure that the status of high-temperature components can be judged while the gas turbine is operating normally. It not only has high economic benefits, but also develops suitable online health assessment and remaining life prediction methods and devices for high-temperature components, which will effectively guide the normal operation of the gas turbine. Online detection methods usually require real-time data support. Introducing visual algorithms to analyze real-time temperature distribution images will be an effective way.

[0004] Thermal barrier coating is a typical multi-layer porous medium with a relatively complex microstructure. The thermal and mechanical properties of each layer vary greatly, and the internal stress is complex. In addition, thermal barrier coatings are usually used at high temperatures and are susceptible to corrosion, erosion, etc., and the service environment is relatively harsh. The complex internal structure and harsh external conditions make it inevitable for thermal barrier coatings to fail during service. The failure modes of thermal barrier coatings mainly include oxidation, sintering, high-temperature corrosion, particle erosion, crack propagation, and spalling. Among them, sintering reduces the porosity of the coating, and the initiation and development of cracks can easily cause spalling and fracture of the coating, which is the most serious problem in coating failure. Therefore, in order to more accurately predict the life of high-temperature components of gas turbines, it is necessary to detect the internal porosity and cracks of the coating. Summary of the invention

[0005] The purpose of the present invention is to provide a vision-based online life prediction method for high-temperature components in order to overcome the defects of the above-mentioned prior art. The porosity and crack length data of the thermal barrier coating of the high-temperature components are obtained based on the visual algorithm, and the health of the high-temperature components is evaluated based on the physical model, machine learning algorithm and information entropy fusion algorithm, so as to improve the accuracy of the life prediction results of the high-temperature components.

[0006] The purpose of the present invention can be achieved by the following technical solutions:

[0007] A vision-based online life prediction method for high-temperature components comprises the following steps:

[0008] S1, obtaining a temperature distribution image of a high-temperature component, and preprocessing the temperature distribution image;

[0009] S2, extracting features from the preprocessed temperature distribution image, detecting defect data of the coating sample of the current high-temperature component, and inputting the defect data into the life prediction model;

[0010] S3. Obtaining a first estimated remaining life of the high-temperature component by a high-temperature component health assessment method based on a physical model, and obtaining a second estimated remaining life of the high-temperature component by a high-temperature component health assessment method based on machine learning;

[0011] S4. Calculate the weight value of the first estimated remaining life value and the weight value of the second estimated remaining life value by a data fusion method based on information entropy;

[0012] S5. Calculate the remaining life value of the high-temperature component according to the weight value of the first remaining life estimate value and the second remaining life estimate value to obtain a health assessment result of the high-temperature component.

[0013] The temperature distribution image of the high-temperature component is obtained by an infrared thermal imager.

[0014] The defect data of the coating samples includes porosity and crack information.

[0015] The physical model in step S3 is a coating life prediction model for crack length established by taking interface oxidation as the only factor controlling the coating life.

[0016] The machine learning method in step S3 is a relevance vector machine (RVM) method based on sparse Bayesian learning.

[0017] Furthermore, the process of calculating the second estimated remaining life in step S3 is as follows:

[0018] S31, constructing an input vector set and a target vector set respectively according to the input temperature distribution image data of the high-temperature component;

[0019] S32, using a radial basis function as a kernel function, optimizing the parameters of the correlation vector machine based on a five-fold cross validation method, and obtaining the parameters of the correlation vector machine with the smallest mean square error value as the optimal parameters;

[0020] S33, establishing a correlation vector machine regression model according to the optimal parameters to obtain the remaining life value under the current temperature distribution image;

[0021] S34, calculating the average remaining lifespan within a period of time to replace the remaining lifespan value in the transition segment, and outputting the final remaining lifespan value as the second remaining lifespan estimation value.

[0022] Further, the data of the input vector set includes porosity and crack length, and the data of the target vector set includes the remaining life value of the high-temperature component.

[0023] The calculation formula of the remaining life value of the high temperature component is as follows:

[0024] RUL=w m rul m +w r rul r

[0025] Among them, RUL is the remaining life value of high temperature components, rul m is the first estimated remaining life, w m is the weight value of the first estimated remaining life, rul r is the second estimated remaining life, w r is the weight value of the second remaining life estimate.

[0026] Furthermore, the calculation formula of the weight value of the first estimated remaining life value and the weight value of the second estimated remaining life value is as follows:

[0027]

[0028] w m +w r =1

[0029] Among them, d m is the coefficient of variation based on the physical model, d r is the coefficient of variation based on machine learning. Further, the calculation formula of the coefficient of variation based on the physical model is as follows:

[0030] d m =1-h m

[0031]

[0032]

[0033]

[0034] Among them, h m is the information entropy based on the physical model, k h is a constant greater than zero, and h m ≥0, p mt is the normalized error based on the physical model, e mt is the relative error of the prediction result based on the physical model, t=1,2,…,N, rul t is the true value of the remaining life of the coating, rul mt is the predicted value based on the physical model.

[0035] Furthermore, the calculation formula of the coefficient of variation based on machine learning is as follows:

[0036] d r =1-h r

[0037]

[0038]

[0039]

[0040] Among them, h r is the information entropy based on machine learning, p rt is the normalized error based on machine learning, e rt is the relative error of the prediction results based on machine learning, rul rt is the predicted value based on machine learning.

[0041] Compared with the prior art, the present invention has the following beneficial effects:

[0042] Based on the temperature distribution image of the high-temperature component, the present invention inputs the defect data of the coating sample into the life prediction model, obtains the first remaining life estimate and the second remaining life estimate through the physical model and machine learning respectively, and finally calculates the remaining life value of the high-temperature component through the data fusion method based on information entropy, and obtains the health assessment result of the high-temperature component, which effectively improves the accuracy of the life prediction result of the high-temperature component and ensures the safe and efficient operation of the gas turbine. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 It is a schematic diagram of the process of the present invention;

[0044] Figure 2It is a schematic diagram of the prediction result and error rate of the remaining life of the coating based on the physical model of the present invention;

[0045] Figure 3 Schematic diagram of the model structure of RVM of the present invention;

[0046] Figure 4 It is a schematic diagram of the coating remaining life prediction results and error rate based on machine learning in the present invention. DETAILED DESCRIPTION

[0047] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.

[0048] Example

[0049] like Figure 1 As shown, a vision-based online life prediction method for high-temperature components can more accurately analyze and judge the short-term and long-term health conditions of high-temperature components, thereby providing data support for the control strategy of the gas turbine, which specifically includes the following steps:

[0050] S1, obtaining a temperature distribution image of a high-temperature component and preprocessing the temperature distribution image;

[0051] S2, extracting features from the preprocessed temperature distribution image, detecting defect data of the coating sample of the current high-temperature component, and inputting the defect data into the life prediction model;

[0052] S3. Obtaining a first estimated remaining life of the high-temperature component by a high-temperature component health assessment method based on a physical model, and obtaining a second estimated remaining life of the high-temperature component by a high-temperature component health assessment method based on machine learning;

[0053] S4. Calculate the weight value of the first estimated remaining life value and the weight value of the second estimated remaining life value by a data fusion method based on information entropy;

[0054] S5. Calculate the remaining life value of the high-temperature component according to the weight value of the first remaining life estimate value and the second remaining life estimate value to obtain a health assessment result of the high-temperature component.

[0055] The temperature distribution image of high-temperature components is obtained by an infrared thermal imager. The infrared thermal imager receives the infrared radiation of the object being measured and converts it into electrical signals of varying strengths. After image signal processing, it obtains an infrared thermal image corresponding to the temperature field distribution of the object being measured. It has high accuracy and a large measurement area.

[0056] Defect data of coating samples include porosity and crack information.

[0057] The physical model in step S3 is a coating life prediction model for crack length established by taking interface oxidation as the only factor controlling the coating life.

[0058] In this embodiment, the physical model in step S3 assumes that as the number of thermal cycles increases, the residual stress in the ceramic layer continues to increase, and the accumulated residual stress causes cracks to initiate in the ceramic layer and gradually develop, eventually leading to coating failure. Under this model, the increase in crack length will cause an exponential increase in stress inside the coating. The longer the crack length, the more significant the impact on the coating life. Figure 2 The figure shows the life prediction results and error rate of the coating under this method. Under all detection conditions, the prediction error rate is less than 1%, and the detection accuracy is significantly improved when the crack length is higher than 220μm.

[0059] The machine learning method in step S3 is a method based on sparse Bayesian learning of relevance vector machine (RVM). RVM can greatly reduce the time consumption while ensuring the prediction accuracy, and this method can obtain probability output, and can better characterize the uncertainty of the prediction result, so it is more suitable for application in remaining life prediction. As a supervised learning algorithm, the basic principle of RVM is to establish a mapping relationship between the input feature vector and the life prediction value through training data, so as to obtain the corresponding target value according to the new input sample. The specific algorithm is as follows:

[0060] For a given sample set, x is the input feature vector and t is the corresponding life prediction value. Using a model structure similar to SVM, y(x) can be expressed as the sum of the product of the kernel function and its corresponding weight coefficient w:

[0061]

[0062] Among them, w is the weight vector, K(x,x i ) is the kernel function, is the kernel matrix. Assume that t i It follows a normal distribution with a mean of y(x i ,w);

[0063] In the relevant vector machine, a prerequisite is added to w so that w obeys the standard normal distribution to avoid overfitting problems:

[0064]

[0065] Among them, μ=[μ0,μ1,μ2,…,μ N ] T is an N+1-dimensional hyperparameter vector that follows a Gamma distribution. Apply the maximum likelihood method to it:

[0066]

[0067] Find the pair σ 2 The solution with partial derivative equal to 0:

[0068]

[0069]

[0070]

[0071] Among them, ∑i,i is the i-th diagonal element, A=diag(μ0,μ1,μ2,…,μ N ).

[0072] The process of calculating the second estimated remaining life in step S3 is as follows:

[0073] S31, constructing an input vector set and a target vector set respectively according to the input temperature distribution image data of the high-temperature component;

[0074] S32, using a radial basis function as a kernel function, optimizing the parameters of the correlation vector machine based on a five-fold cross validation method, and obtaining the parameters of the correlation vector machine with the smallest mean square error value as the optimal parameters;

[0075] S33, establishing a correlation vector machine regression model according to the optimal parameters to obtain the remaining life value under the current temperature distribution image;

[0076] S34, calculating the average remaining lifespan within a period of time to replace the remaining lifespan value in the transition segment, and outputting the final remaining lifespan value as the second remaining lifespan estimation value.

[0077] The data of the input vector set include porosity and crack length, and the data of the target vector set include the remaining life value of the high-temperature component.

[0078] The final RVM model structure diagram is as follows Figure 3 As shown. Repeated iterations until the convergence criteria are met can complete the RVM training. Figure 4 The RVM-based coating remaining life prediction results are presented. Compared with the prediction method based on the physical model, the prediction results have lower errors and smaller fluctuation ranges.

[0079] However, due to the complex correspondence between condition monitoring data and the remaining life of the coating, it is difficult to specifically evaluate the performance of different prediction methods under different specific working conditions. Therefore, it is necessary to fuse the prediction results of multiple methods. Compared with a single prediction method, the fusion method will effectively improve the accuracy and robustness of the prediction results.

[0080] The remaining life value of high temperature components is calculated as follows:

[0081] RUL=w m rul m +w r rul r

[0082] Among them, RUL is the remaining life value of high temperature components, rul m is the first estimated remaining life, w m is the weight value of the first estimated remaining life, rul r is the second estimated remaining life, w r is the weight value of the second remaining life estimate.

[0083] The weight value of the first remaining life estimate and the weight value of the second remaining life estimate are calculated as follows:

[0084]

[0085] w m +w r =1

[0086] Among them, d m is the coefficient of variation based on the physical model, d r is the coefficient of variation based on machine learning. The calculation formula of the coefficient of variation based on the physical model is as follows:

[0087] d m =1-h m

[0088]

[0089]

[0090]

[0091] Among them, h m is the information entropy based on the physical model, k h is a constant greater than zero, and h m ≥0, p mt is the normalized error based on the physical model, e mt is the relative error of the prediction result based on the physical model, t=1,2,…,N, rul t is the true value of the remaining life of the coating, rul mt is the predicted value based on the physical model.

[0092] The calculation formula of the coefficient of variation based on machine learning is as follows:

[0093] d r =1-h r

[0094]

[0095]

[0096]

[0097] Among them, h r is the information entropy based on machine learning, p rt is the normalized error based on machine learning, e rt is the relative error of the prediction results based on machine learning, rul rt is the predicted value based on machine learning.

[0098] In the specific implementation, the physical model, RVM model and fusion model are used to predict the coating life, where the crack length range is 100μm-400μm and the porosity range is 10%-25%. Table 1 shows the statistical results of life prediction using a single method and a fusion algorithm, as shown in Table 1:

[0099] Table 1 Lifespan prediction statistical results

[0100]

[0101] It can be seen from Table 1 that the prediction of the fusion method based on information entropy is better than that of the single prediction model in terms of the average error and root mean square error of the prediction results, and has the best comprehensive performance.

[0102] In addition, it should be noted that the specific embodiments described in this specification may be named differently, and the above contents described in this specification are merely examples of the structure of the present invention. Any equivalent changes or simple changes made based on the structure, features and principles of the present invention are included in the protection scope of the present invention. Those skilled in the art of the present invention may make various modifications or supplements to the specific examples described or adopt similar methods, as long as they do not deviate from the structure of the present invention or exceed the scope defined by the claims, they should all fall within the protection scope of the present invention.

Claims

1. A vision-based online life prediction method for high-temperature components, characterized in that: The specific steps include: S1, obtaining a temperature distribution image of a high-temperature component, and preprocessing the temperature distribution image; S2, extracting features from the preprocessed temperature distribution image, detecting defect data of the coating sample of the current high-temperature component, and inputting the defect data into the life prediction model; S3. Obtaining a first estimated remaining life of the high-temperature component by a high-temperature component health assessment method based on a physical model, and obtaining a second estimated remaining life of the high-temperature component by a high-temperature component health assessment method based on machine learning; S4. Calculate the weight value of the first estimated remaining life value and the weight value of the second estimated remaining life value by a data fusion method based on information entropy; S5. Calculate the remaining life value of the high-temperature component according to the weight value of the first remaining life estimate value and the second remaining life estimate value to obtain a health assessment result of the high-temperature component; The machine learning method in step S3 is a method of a relevance vector machine based on sparse Bayesian learning; The process of calculating the second estimated remaining life in step S3 is as follows: S31, constructing an input vector set and a target vector set respectively according to the input temperature distribution image data of the high-temperature component; S32, using a radial basis function as a kernel function, optimizing the parameters of the correlation vector machine based on a five-fold cross validation method, and obtaining the parameters of the correlation vector machine with the smallest mean square error value as the optimal parameters; S33, establishing a correlation vector machine regression model according to the optimal parameters to obtain the remaining life value under the current temperature distribution image; S34, calculating the mean value of the remaining life within a period of time to replace the remaining life value in the transition segment, and outputting the final remaining life value as the second remaining life estimate; The calculation formula of the remaining life value of the high temperature component is as follows: RUL=w m NUMBER m +w r NUMBER r Among them, RUL is the remaining life value of high temperature components, rul m is the first estimated remaining life, w m is the weight value of the first estimated remaining life, rul r is the second estimated remaining life, w r is the weight value of the second estimated remaining life; The calculation formula of the weight value of the first estimated remaining life value and the weight value of the second estimated remaining life value is as follows: In m +in r =1 Among them, d m is the coefficient of variation based on the physical model, d r is the coefficient of variation based on machine learning.

2. The method for online life prediction of high-temperature components based on vision according to claim 1, characterized in that: The defect data of the coating samples includes porosity and crack information.

3. The method for online life prediction of high-temperature components based on vision according to claim 1, characterized in that: The physical model in step S3 is a coating life prediction model for crack length established by taking interface oxidation as the only factor controlling the coating life.

4. The method for online life prediction of high-temperature components based on vision according to claim 1, characterized in that: The data of the input vector set include porosity and crack length, and the data of the target vector set include the remaining life value of the high-temperature component.

5. The method for online life prediction of high temperature components based on vision according to claim 1, characterized in that: The calculation formula of the coefficient of variation based on the physical model is as follows: d m =1-h m Among them, h m is the information entropy based on the physical model, k h is a constant greater than zero, and h m ≥0, p mt is the normalized error based on the physical model, e mt is the relative error of the prediction result based on the physical model, t=1,2,…,N, rul t is the true value of the remaining life of the coating, rul mt is the predicted value based on the physical model.

6. The method for online life prediction of high temperature components based on vision according to claim 5, characterized in that: The calculation formula of the coefficient of variation based on machine learning is as follows: d r =1-h r Among them, h r is the information entropy based on machine learning, p rt is the normalized error based on machine learning, e rt is the relative error of the prediction results based on machine learning, rul rt is the predicted value based on machine learning.

Images