Power plant wear fault diagnosis method and system based on ferrographic image analysis

Through a two-level classification algorithm based on iron spectrum image analysis, combined with SVM and SDAE models, Jaya and Bayesian optimization algorithms are used to solve the problem of large data volume and low accuracy in the diagnosis of wear faults of power devices, and efficient and accurate Abrasive particle identification and fault diagnosis provide interpretable results.

CN114565798BActive Publication Date: 2024-12-03CHINA NORTH VEHICLE RES INST
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202210231033.1
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-03-10
Publication Date
2024-12-03
Estimated Expiration
2042-03-10

AI Technical Summary

Technical Problem

The existing iron spectrum analysis technology requires large data volume, low accuracy and lack of interpretability in the diagnosis of wear faults of power plants, making it difficult to effectively monitor and diagnose the wear status of power plants.

Method used

Using a two-level classification algorithm based on frost spectrum image analysis, abrasive type is constructed through the combination of image preprocessing, connectivity domain feature extraction, support vector machine (SVM) multi-classification model optimization and stacked noise reduction automatic encoder (SDAE). Diagnostic model with the type of wear failure of power plants, optimize model parameters using Jaya algorithm and Bayesian optimization algorithm.

Benefits of technology

It improves the accuracy of abrasive particle recognition and the accuracy of wear fault diagnosis of power equipment, provides interpretable diagnostic results, reduces algorithm complexity and improves calculation speed, and achieves fast and accurate abrasive particle recognition and fault diagnosis.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114565798B_ABST
    Figure CN114565798B_ABST
Patent Text Reader

Abstract

The present invention belongs to the technical field of power plant fault diagnosis, and specifically relates to a power plant wear fault diagnosis method and system based on iron spectrum image analysis, including: obtaining original iron spectrum images, image preprocessing, wear feature extraction, wear type identification, wear fault diagnosis, and displaying fault diagnosis results. In the wear type identification model, the Jaya algorithm is used to optimize the parameters of the support vector machine classification model to achieve fast and accurate identification of the wear type; in the wear fault classification, the Bayesian optimization algorithm is used to optimize the stacked denoising autoencoder structure. The optimized SDAE model reconstructs the fault features and can better reflect the key feature information. Therefore, the reconstructed features are input into the Softmax classifier to further improve the accuracy of the diagnosis. This technical solution overcomes the shortcomings of slow speed, low accuracy, and poor interpretability encountered in the implementation of traditional power plant fault diagnosis, and provides effective guarantee for the stable operation of the power plant.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of power plant fault diagnosis, and in particular relates to a power plant wear fault diagnosis method and system based on ferrographic image analysis. Background Art

[0002] The mechanical parts of the power unit will wear out after running for a long time. When the wear exceeds a certain limit, it will cause hidden dangers of failure. By performing iron spectrum analysis on the wear particles in the oil of the power unit and analyzing the wear particle image characteristics, the wear failure of the power unit can be monitored and the fault can be quickly diagnosed, thereby reducing economic losses.

[0003] Traditional ferrography analysis generally uses magnetic force to separate metal particles from oil, and judges the wear severity of the power unit based on the number of particles of different sizes. However, this distinction lacks strict physical meaning, and a large number of statistical experiments are required to build a trend line that can reflect the wear status of the power unit. With the development of microscopic imaging technology and image recognition technology, ferrography fault diagnosis has entered the microscopic and intelligent stage, using the microscopic characteristics of the acquired wear particle images such as material, size, and quantity to analyze the wear status of the power unit. In this process, some researchers conduct fault diagnosis based on human vision, which strongly relies on personal experience technology, has low recognition efficiency, and cannot form large-scale applications. Another part of researchers applies deep learning to ferrography analysis to break through the errors of the human eye. However, the deep learning model is a black box model, lacks analysis of physical meaning, and is difficult to explain the lack of accuracy from a physical level. Summary of the invention

[0004] 1. Technical issues to be resolved

[0005] The technical problem to be solved by the present invention is: how to provide a method and system for diagnosing wear faults of power units based on ferrographic image analysis, and use a two-level classification algorithm to construct a diagnostic algorithm of ferrographic image features-abrasive type-power unit wear fault type, so as to effectively solve the problems of large data volume required, low accuracy and poor interpretability when the existing ferrographic analysis technology is used to diagnose wear faults of power units.

[0006] (II) Technical solution

[0007] In order to solve the above technical problems, the present invention provides a method for diagnosing wear faults of power plants based on ferrographic image analysis, the method comprising the following steps:

[0008] Step 1, capturing an original ferrographic image using an imaging device at a predetermined period;

[0009] Step 2, performing image preprocessing on the original ferrographic image to obtain a processed binary image;

[0010] Step 3, marking the connected domain of the binary ferrographic image, and extracting the geometric, morphological and other attribute parameters of the connected domain as ferrographic wear particle features;

[0011] Step 4, training a wear particle recognition model according to the ferrographic wear particle characteristics, wherein the input of the wear particle recognition model is the ferrographic wear particle characteristics, the output of the wear particle recognition model is the wear particle category, the wear particle recognition model is an SVM multi-classification model optimized by an algorithm, the SVM multi-classification model adopts a combination of multiple SVMs, and a sub-SVM is designed between any two categories of wear particle samples. The training set of the sub-SVM adopts a pairwise combination method, and the Jaya algorithm is used to optimize the parameters of each sub-SVM in the SVM multi-classification model;

[0012] Step 5, training the deep network model optimized by the algorithm, reconstructing the fault characteristics of the power unit, classifying the reconstructed fault characteristics of the power unit using the Softmax classifier, and finally outputting the type of wear fault of the power unit; the fault characteristics of the power unit are the categories of abrasive particles obtained based on the abrasive particle recognition model, and the number of abrasive particles of each category and / or the proportion of abrasive particles of each category obtained by counting the abrasive particle categories; the deep network model is a stacked denoising autoencoder, and the parameters of the stacked denoising autoencoder are optimized using the Bayesian optimization algorithm;

[0013] Step 6, displaying the iron spectrum images of each stage, the number of each type of wear particles, and diagnosing the type of wear fault obtained.

[0014] Wherein, in step 2, the image preprocessing operations include: grayscale, binarization, and denoising.

[0015] Among them, in the step 3, before obtaining the ferrographic abrasive features, the connected domain is first marked, the boundary rectangle of the connected domain is obtained, and the relevant attributes of the connected domain are calculated as the ferrographic abrasive features, and the ferrographic abrasive features include: the area, perimeter, circularity, rectangularity, aspect ratio of the connected domain, the major axis length, minor axis length, eccentricity of the ellipse with the same second-order center distance as the boundary rectangle, the diameter of a circle of equal area to the connected domain, and the pixel ratio of the connected domain to its minimum convex polygon.

[0016] Wherein, in step 4, the classification decision function formula of the SVM multi-classification model is:

[0017]

[0018] in, σ is the Gaussian kernel function parameter; x is the function independent variable; x i is the characteristic vector of the i-th sample wear particle; y iis the category label of the i-th sample wear particle; m is the number of samples in the training set; α i is the Lagrange multiplier; α i Is the constraint condition: 0≤α i ≤γ; γ is the regularization parameter; b is the bias.

[0019] Wherein, in step 4, the Jaya algorithm is used to optimize the parameter combination (γ, σ) of each sub-SVM in the SVM multi-classification model, which specifically includes the following steps:

[0020] Step 4.1, the Jaya algorithm is designed to have a population size of 30 to 40, the algorithm is terminated when the maximum number of iterations reaches 1000 to 2000, the number of algorithm design variables is 2, and the ranges of the regularization parameter γ and the Gaussian kernel function parameter σ are [0.1, 100] and [0.1, 500], respectively;

[0021] Step 4.2, select E RMSE As the objective function of the Jaya algorithm, its formula is as follows:

[0022]

[0023] m is the number of wear particle samples; y i is the actual label of the i-th sample wear particle; y′ i is the predicted label of the wear particle of the i-th sample;

[0024] Step 4.3, a set of vectors (γ, σ) is randomly generated within the range of the optimization parameters, and a set of (γ, σ) is assigned to each individual in the Jaya algorithm population;

[0025] Step 4.4, based on the training results of SVM, calculate the E of each individual in the population RMSE The objective function value is used to obtain the optimal solution E for the current population. RMSEmin and the worst solution E RMSEmax ;

[0026] Step 4.5, according to the obtained optimal solution E RMSEmin and the worst solution E RMSEmax Correct the current solution E of the objective function during the iteration RMSE , as follows:

[0027] E′ j,k,i =E j,k,i +r1 j,i (E j,min,i -|E RMSE |)-r2 j,i (E j,max,i -|E RMSE |)

[0028] Among them, E j,k,i is the solution corresponding to the jth variable of the kth candidate variable in the i-th iteration; E′ j,k,i is the modified solution of the jth variable of the kth candidate variable in the i-th iteration; E j,min,i is the solution of the jth variable of the optimal option in the i-th iteration; E j,max,i is the solution of the jth variable of the worst option in the i-th iteration; r1 j,i 、r2 j,i is the random number corresponding to the jth variable at the i-th iteration, and its value range is [0, 1];

[0029] Step 4.6, if the corrected solution E′ RMSE Better than the current solution E RMSE , then retain the modified solution E′ RMSE , otherwise keep the current solution; all the retained solutions are used as the input of the next iteration;

[0030] Step 4.7: Determine whether the maximum number of iterations has been reached. If not, go to step 4.4. Otherwise, output the optimal solution, i.e., E RMSE Minimum value, and output the corresponding optimal parameter combination (γ, σ).

[0031] Wherein, the step 5 includes the following steps:

[0032] Step 5.1, performing wear particle recognition on the processed binary image, and obtaining the number of each type of wear particles and / or the proportion of each type of wear particles, that is, the fault characteristics of the power device;

[0033] Step 5.2, in order to make the learned stacked denoising autoencoder more robust, the power unit fault feature is subjected to noise addition and normalization processing;

[0034] Step 5.3, initialize the stacked denoising autoencoder, and train it layer by layer with the goal of minimizing the reconstruction error. In this process, the Gaussian process is used to continuously optimize the combination of the number of nodes in the hidden layer of the initialized stacked denoising autoencoder (SDAE) and train the stacked denoising autoencoder until the reconstruction error obtained under a certain combination of the number of nodes meets the requirements, then stop optimizing the stacked denoising autoencoder;

[0035] Step 5.4, putting the power unit fault feature into the trained stacked denoising autoencoder to obtain the reconstructed feature of the power unit fault;

[0036] Step 5.5, classify the reconstructed features using a Softmax classifier, and the classification result is the wear fault type of the power unit.

[0037] Wherein, in step 5.3, the parameters of the stacked denoising autoencoder are optimized using a Bayesian optimization algorithm, specifically including the following:

[0038] Step 5.3.1, initialize the node number combination, the node number combination is expressed as: g = [m 1 ,m 2 ,m 3 ]; where m 1 Indicates the number of nodes in the first hidden layer, m 2 Indicates the number of nodes in the second hidden layer, m 3 Indicates the number of nodes in the third hidden layer, the range of the number of nodes is [50, 500];

[0039] Step 5.3.2, optimize the node number combination, the optimization function is:

[0040] g represents the parameter combination to be optimized; f(g) represents the reconstruction error corresponding to the parameter combination; χ is the set of g;

[0041] Using the Gaussian process, the t-th Gaussian prediction distribution is:

[0042]

[0043]

[0044] g t represents the number of nodes in the tth hidden layer; μ(g t ) is g t The mean function of

[0045] k t =[k(g1,g t ),k(g2,g t ),…,k(g t-1 ,g t )]

[0046] k t T k t The transpose of; k(.) represents the covariance function; K is a semidefinite matrix; I is the identity matrix; h=[h1,h2,…,h t ] is the observation sequence of the Gaussian process; h t =f(g t )+ε t is the observed value of the Gaussian process; f(g t ) is the reconstruction error corresponding to the parameter combination; ε t is an independent and identically distributed Gaussian noise, which is a random variable; σ n is t The variance of

[0047] Step 5.3.3, obtain g through Gaussian process t After the prediction distribution, a collection function is selected to construct a utility function from the posterior distribution of the model to determine the next parameter combination point and the new parameter combination g t+1 Input to the stacked denoising autoencoder and train with the reconstruction error between the input signal and the reconstructed output signal as the model output;

[0048] Step 5.3.4, if the reconstruction error f(g t+1 ) meets the target requirements, the algorithm execution is terminated and exited, and the corresponding hidden layer node number combination g is output t+1 ; If f(g t+1 ) does not meet the requirements, then (g t+1 ,f(g t+1 )) is input into the Gaussian model for correction and step 5.3.2 is repeated until the required accuracy of the experiment is met.

[0049] Among them, the abrasive categories include: normal abrasives, strict lubrication abrasives, cutting abrasives, fatigue abrasives, flaky abrasives, spherical abrasives, and oxide abrasives; the wear failure types include: normal wear, adhesive wear, cutting wear, fatigue wear, and corrosive wear.

[0050] In addition, the present invention also provides a system for executing the power device wear fault diagnosis method based on ferrographic image analysis, comprising:

[0051] An image acquisition module 1, wherein the image acquisition module 1 repeatedly captures images at a predetermined period and sends the captured images to a control unit 2;

[0052] An image processing module 3, the image processing module 3 is used to obtain the image captured by the image acquisition module from the control unit 2, and perform grayscale, binarization and denoising processing on the image;

[0053] A feature extraction module 4, which is used to extract relevant attributes of the connected domain of the binary image after image processing as the ferrographic wear particle features;

[0054] A wear particle identification module 5, the wear particle identification module 5 is used to identify the wear particle category by using an algorithm-optimized SVM multi-classification model, the input of the algorithm-optimized SVM multi-classification model is the extracted iron spectrum wear particle features, the output of the algorithm-optimized SVM multi-classification model is the wear particle category, the SVM multi-classification model uses a combination of multiple SVMs, designs a sub-SVM between any two categories of wear particle samples, and uses the Jaya algorithm to optimize the parameters of each sub-SVM in the SVM multi-classification model;

[0055] A fault diagnosis module 6, wherein the fault diagnosis module 6 is used to reconstruct the fault characteristics of the power device and perform wear fault diagnosis on the power device; the fault characteristics of the power device are the number of each category of abrasive particles and / or the proportion of each category of abrasive particles in the iron spectrum image, and the fault characteristics of the power device are used as input, and the deep network model optimized by the algorithm is first used to reconstruct the fault characteristics of the power device, and then the reconstructed power device fault characteristics are classified using a Softmax classifier, and finally the wear fault type of the power device is output; the deep network model is a stacked denoising autoencoder, and the parameters of the stacked denoising autoencoder are optimized using a Bayesian optimization algorithm;

[0056] The visualization module 7 is used to display images, the number of each type of wear particles and the wear fault type obtained by diagnosis.

[0057] In addition, the present invention also provides a computer-readable storage medium, in which computer-executable instructions are stored, and the computer-executable instructions are used to execute the power device wear fault diagnosis method.

[0058] (III) Beneficial effects

[0059] In view of the problems of the prior art, the present invention first adopts the support vector machine classification model optimized by the Jaya algorithm to identify the iron spectrum abrasive particles, and uses a limited number of samples to construct the relationship between the characteristics of the abrasive particle size, morphology and the type of abrasive particles from a physical level, and then establishes a secondary classification correspondence between the wear fault characteristics such as the type, quantity and proportion of abrasive particles and the type of wear fault of the power unit. The process uses the Bayesian optimization algorithm to optimize the wear fault characteristics, improves the quality of the characteristic information, and further improves the accuracy of fault diagnosis. In the overall process, the present invention applies the iron spectrum image recognition results of the abrasive particles to the power unit fault diagnosis, so that the fault diagnosis results are interpretable, which is conducive to subsequent maintenance and early warning.

[0060] Compared with the prior art, the present invention has the following beneficial effects:

[0061] (1) The power plant fault diagnosis method and system based on ferrographic image analysis of the present invention extracts characteristic information of ferrographic wear particles and adopts the SVM classification model optimized by the Jaya algorithm to identify and classify the wear particles. The use of the SVM multi-classification model can avoid the limitation of small sample data volume. The regularization parameter γ and the Gaussian kernel function parameter σ in the model have a great influence on the prediction accuracy. The regularization parameter γ and the Gaussian kernel function parameter σ are optimized by the Jaya algorithm, which can achieve the optimization of the SVM parameters under the condition of low algorithm complexity, improve the calculation speed and thus improve the diagnosis speed; multi-classification The training set of the sub-classifiers of the model adopts a pairwise combination method, which can improve the classification accuracy of multiple wear categories of power devices compared with the existing technology, improve the classification accuracy, and realize fast and accurate wear identification; using 10 image recognition parameters such as the area, perimeter, circularity, rectangularity, aspect ratio of the connected domain, the major axis length, minor axis length, eccentricity of the ellipse with the same second-order center distance as the boundary rectangle, the diameter of the circle of equal area with the connected domain, and the pixel ratio of the connected domain to its minimum convex polygon, the relationship between the iron spectrum image characteristics and the wear type is constructed from the microscopic physical level, thereby improving the interpretability of the algorithm.

[0062] (2) The power plant fault diagnosis method and system based on iron spectrum image analysis of the present invention adopts the Bayesian algorithm (BO) to tune the parameters of the stacked denoising autoencoder (SDAE) according to the number of each category of abrasive particles and / or the proportion of each category of abrasive particles, so as to reconstruct the power plant fault characteristics. The stacked denoising autoencoder (SDAE) is used to learn through a series of nonlinear mappings to better reflect the key information of the fault characteristics. The reconstructed characteristics are classified by a classifier to diagnose the type of power plant wear fault, thereby further improving the accuracy of power plant wear fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 It is an overall implementation block diagram of the present invention. Figure 2 It is the basic flow chart of abrasive type identification of the present invention. Figure 3 It is the basic flow chart of wear fault diagnosis of the present invention. Figure 4 It is a fault diagnosis result diagram of the present invention. Figure 5 Schematic diagram of the wear fault diagnosis system of the present invention. Figure 6 It is the initialization interface after the visualization module of the present invention logs in successfully. Figure 7 It is a diagram of image operation selection of the visualization module of the present invention. Figure 8 This is an image preprocessing effect diagram of the visualization module of the present invention. Fig. 9 This is a diagram showing the effect of the visualization module for identifying the abrasive type of the present invention. Fig.10 This is a diagram showing the fault diagnosis effect of the visualization module of the present invention. DETAILED DESCRIPTION

[0064] In order to make the purpose, content, and advantages of the present invention more clear, the specific implementation methods of the present invention are further described in detail below in conjunction with the accompanying drawings and examples.

[0065] Example 1

[0066] In order to solve the problems in the prior art, the present invention provides a method for diagnosing wear faults of a power plant based on ferrographic image analysis, and the method comprises the following steps:

[0067] Step 1: capturing an original ferrographic image using an imaging device at a predetermined period;

[0068] Step 2, performing image preprocessing on the original ferrographic image to obtain a processed binary image;

[0069] Step 3, selecting the relevant attributes of the connected domain of the processed binary image as the ferrographic wear particle features;

[0070] Step 4, training a wear particle recognition model according to the ferrographic wear particle characteristics, wherein the input of the wear particle recognition model is the ferrographic wear particle characteristics, the output of the wear particle recognition model is the wear particle category, the wear particle recognition model is an SVM multi-classification model optimized by an algorithm, the SVM multi-classification model adopts a combination of multiple SVMs, and a sub-SVM is designed between any two categories of wear particle samples. The training set of the sub-SVM adopts a pairwise combination method, and the Jaya algorithm is used to optimize the parameters of each sub-SVM in the SVM multi-classification model;

[0071] Step 5, training the deep network model optimized by the algorithm, reconstructing the fault characteristics of the power unit, classifying the reconstructed fault characteristics of the power unit using the Softmax classifier, and finally outputting the type of wear fault of the power unit; the fault characteristics of the power unit are the abrasive particle categories obtained based on the abrasive particle recognition model, and the number of abrasive particles of each category and / or the proportion of abrasive particles of each category are obtained by counting the abrasive particle categories; the deep network model is a stacked denoising autoencoder (SDAE), and the parameters of the stacked denoising autoencoder (SDAE) are optimized using the Bayesian optimization algorithm (BO);

[0072] Step 6, displaying the iron spectrum images of each stage, the number of each type of wear particles, and the type of wear fault diagnosed.

[0073] Wherein, in the process of capturing the original ferrogram image in step 1, the original ferrogram image of the ferrogram spectrum sheet in the oil wear particle device is captured by an imaging device, the imaging device is a sensor with a built-in shooting probe or a laser scanning confocal microscope, etc., the oil wear particle device includes a ferrogram spectrum sheet, a driving device and a control device; wherein, the ferrogram spectrum sheet is driven by the driving device, and the control device controls the ferrogram spectrum sheet by controlling the driving device; the ferrogram spectrum sheet image is obtained by the OLVF online visual ferrogram.

[0074] Wherein, in step 2, the image preprocessing operations include: grayscale, binarization, and denoising. Wherein, grayscale is to grayscale the iron spectrum image and convert it into a single-channel grayscale image; binarization is to threshold segment the grayscale image using the between-class variance method (OTSU) to obtain a binary image; denoising is to denoise the image using, for example, the total variation (TV) algorithm.

[0075] Wherein, in step 3, before obtaining the iron spectrum abrasive grain features, the connected domain is first marked, the boundary rectangle of the connected domain is obtained, and the related attributes of the connected domain are calculated as the iron spectrum abrasive grain features, and the iron spectrum abrasive grain features include: the area, perimeter, circularity, rectangularity, aspect ratio of the connected domain, the major axis length, minor axis length, eccentricity of the ellipse with the same second-order center distance as the boundary rectangle, the diameter of the circle with the same area as the connected domain, and the pixel ratio of the connected domain to its minimum convex polygon (i.e., 10 types of abrasive grain features). This step can use the calculation function in the mathematical software tool or programming language.

[0076] Among them, in step 4, support vector machine (SVM) is a widely used classification model. The wear fault diagnosis of power plant is limited by the small sample of oil analysis data. The SVM multi-classification model can use small samples for training and learning to solve the pattern recognition problem in multidimensional vector space. In addition, the regularization parameter γ and kernel function parameters in the model have a great influence on the prediction accuracy, so the appropriate parameter selection plays a vital role in the performance of the prediction model. Jaya has fewer control parameters, which can achieve the optimization of SVM parameters under low algorithm complexity, improve the classification accuracy, and achieve fast and accurate wear particle identification.

[0077] Furthermore, in step 4, the abrasive particle identification model adopts a SVM multi-classification model, the model input is the abrasive particle characteristic parameters, and the model output is the category of the abrasive particles; the categories of the abrasive particles include the following 7 categories: normal abrasive particles, strict lubrication abrasive particles, cutting abrasive particles, fatigue abrasive particles, flaky abrasive particles, spherical abrasive particles, and oxide abrasive particles.

[0078] Furthermore, the multi-classification model needs to use multiple SVM combinations, and a sub-SVM is designed between any two categories of abrasive samples. A total of 21 SVM sub-classifiers are required. The training sets of the sub-classifiers are combined in pairs. Compared with the existing technology, this combination method can improve the classification accuracy of multiple abrasive categories in power devices.

[0079] Among them, the classification decision function formula of the SVM multi-classification model is:

[0080]

[0081] in, σ is the Gaussian kernel function parameter; x is the function independent variable; x i is the characteristic vector of the i-th sample wear particle; y i is the category label of the i-th sample wear particle; m is the number of samples in the training set; α i is the Lagrange multiplier; α i Is the constraint condition: 0≤α i ≤γ; γ is the regularization parameter; b is the bias.

[0082] Wherein, in step 4, the Jaya algorithm is used to optimize the parameter combination (γ, σ) of each sub-SVM in the SVM multi-classification model, which specifically includes the following steps:

[0083] Step 4.1, the Jaya algorithm is designed to have a population size of 30 to 40, the algorithm is terminated when the maximum number of iterations reaches 1000 to 2000, the number of algorithm design variables is 2, and the ranges of the regularization parameter γ and the Gaussian kernel function parameter σ are [0.1, 100] and [0.1, 500], respectively;

[0084] Step 4.2, select E RMSE As the objective function of the Jaya algorithm, its formula is as follows:

[0085]

[0086] m is the number of wear particle samples; y i is the actual label of the i-th sample wear particle; y′ i is the predicted label of the wear particle of the i-th sample;

[0087] Step 4.3, a set of vectors (γ, σ) is randomly generated within the range of the optimization parameters, and a set of (γ, σ) is assigned to each individual in the Jaya algorithm population;

[0088] Step 4.4, based on the training results of SVM, calculate the E of each individual in the population RMSE The objective function value is used to obtain the optimal solution E for the current population.RMSEmin and the worst solution E RMSEmax ;

[0089] Step 4.5, according to the obtained optimal solution E RMSEmin and the worst solution E RMSEmax Correct the current solution E of the objective function during the iteration RMSE , as follows:

[0090] E′ j,k,i =E j,k,i +r1 j,i (E j,min,i -|E RMSE |)-r2 j,i (E j,max,i -|E RMSE |)

[0091] Among them, E j,k,i is the solution corresponding to the jth variable of the kth candidate variable in the i-th iteration; E′ j,k,i is the modified solution of the jth variable of the kth candidate variable in the i-th iteration; E j,min,i is the solution of the jth variable of the optimal option in the i-th iteration; E j,max,i is the solution of the jth variable of the worst option in the i-th iteration; r1 j,i 、r2 j,i is the random number corresponding to the jth variable at the i-th iteration, and its value range is [0, 1];

[0092] Step 4.6, if the corrected solution E′ RMSE Better than the current solution E RMSE , then retain the modified solution E′ RMSE , otherwise keep the current solution; all the retained solutions are used as the input of the next iteration;

[0093] Step 4.7: Determine whether the maximum number of iterations has been reached. If not, go to step 4.4. Otherwise, output the optimal solution, i.e., E RMSE Minimum value, and output the corresponding optimal parameter combination (γ, σ).

[0094] Wherein, the step 5 comprises the following steps:

[0095] Step 5.1, performing wear particle recognition on the processed binary image, and obtaining the number of each type of wear particles and / or the proportion of each type of wear particles, that is, the fault characteristics of the power device;

[0096] Step 5.2, in order to make the learned stacked denoising autoencoder (SDAE) more robust, the power unit fault feature is denoised and normalized;

[0097] Step 5.3, initialize the stacked denoising autoencoder (SDAE), and train it layer by layer with the goal of minimizing the reconstruction error. In this process, the Gaussian process is used to continuously optimize the combination of the number of nodes in the hidden layer of the initialized stacked denoising autoencoder (SDAE) and train the stacked denoising autoencoder (SDAE) until the reconstruction error obtained under a certain combination of the number of nodes meets the requirements, then stop optimizing the stacked denoising autoencoder (SDAE);

[0098] Step 5.4, putting the power unit fault feature into the trained stacked denoising autoencoder (SDAE) to obtain the reconstructed feature of the power unit fault;

[0099] Step 5.5, classify the reconstructed features using a Softmax classifier, and the classification result is the wear fault type of the power unit.

[0100] Since there are multicollinear relationships between the above fault features, direct classification will affect the diagnosis accuracy, so the fault features need to be reconstructed. Stacked denoising autoencoder (SDAE) is a variant of deep autoencoder network. It replaces autoencoder (SAE) with denoising autoencoder (DAE). By stacking multiple layers of DAE, stacked denoising autoencoder (SDAE) is established. It can learn more abstract and higher-level features through a series of nonlinear mappings, so as to more effectively represent key information. The features reconstructed by stacked denoising autoencoder (SDAE) help Softmax to classify faults, thereby improving the accuracy of wear fault diagnosis. Stacked denoising autoencoder (SDAE) is very dependent on the number of hidden layer nodes in the training process. The number of nodes that can be selected in each hidden layer is very large, and the number of nodes in each hidden layer can affect each other. Therefore, it is theoretically difficult to obtain the optimal hidden layer node combination scheme of stacked denoising autoencoder (SDAE) by adjusting the number of hidden layer nodes only by prior experience. However, the reconstruction effect of the stacked denoising autoencoder (SDAE) depends largely on the number of hidden layers and the number of hidden layer nodes. In order to reduce the complexity of the algorithm, the present invention sets the number of hidden layers to a fixed value, and the range of the number of hidden layer nodes of the stacked denoising autoencoder (SDAE) is [50, 500].

[0101] Among them, in the step 5.3, the Bayesian optimization algorithm (BO) is used to optimize the parameters of the stacked denoising autoencoder (SDAE), that is, the number of nodes, and specifically includes the following steps:

[0102] Step 5.3.1, initialize the node number combination, the node number combination is expressed as: g = [m 1 ,m 2 ,m 3 ]; where m 1Indicates the number of nodes in the first hidden layer, m 2 Indicates the number of nodes in the second hidden layer, m 3 Indicates the number of nodes in the third hidden layer, the range of the number of nodes is [50, 500];

[0103] Step 5.3.2, optimize the node number combination, the optimization function is:

[0104] g represents the parameter combination to be optimized; f(g) represents the reconstruction error corresponding to the parameter combination; χ is the set of g;

[0105] Using the Gaussian process, the t-th Gaussian prediction distribution is:

[0106]

[0107]

[0108] g t represents the number of nodes in the tth hidden layer; μ(g t ) is g t The mean function of

[0109] k t =[k(g1,g t ),k(g2,g t ),…,k(g t-1 ,g t )]

[0110] k t T k t The transpose of; k(.) represents the covariance function; K is a semidefinite matrix; I is the identity matrix; h=[h1,h2,…,h t ] is the observation sequence of the Gaussian process; h t =f(g t )+ε t is the observed value of the Gaussian process; f(g t ) is the reconstruction error corresponding to the parameter combination; ε t is an independent and identically distributed Gaussian noise, which is a random variable; σ n is t The variance of

[0111] Step 5.3.3, obtain g through Gaussian process t After the prediction distribution, a collection function is selected to construct a utility function from the posterior distribution of the model to determine the next parameter combination point and the new parameter combination g t+1 Input to the stacked denoising autoencoder and train with the reconstruction error between the input signal and the reconstructed output signal as the model output;

[0112] Step 5.3.4, if the reconstruction error f(g t+1 ) meets the target requirements, the algorithm execution is terminated and exited, and the corresponding hidden layer node number combination g is output t+1 ; If f(g t+1 ) does not meet the requirements, then (g t+1 ,f(g t+1 )) is input into the Gaussian model for correction and step 5.3.2 is repeated until the required accuracy of the experiment is met.

[0113] In the actual wear working conditions of power units, multiple types of wear often exist at the same time and affect each other. Due to the different working conditions under which each type of wear occurs, as the wear state changes, the priority of different types of wear varies, and the proportion of different types of abrasive particles is also different. By statistically analyzing the wear failure data of power units, the wear type with the highest failure rate of power units is selected, and the range and / or proportion of abrasive particles corresponding to different wear categories for each type of power unit wear failure are obtained. Considering the wear mechanism of the main moving parts in the work of the power unit, the types of wear failures are divided into normal wear, adhesive wear, cutting wear, fatigue wear and corrosive wear. Among them, different types of wear produce different abrasive particles. Normal wear is mainly normal abrasive particles, adhesive wear is mainly strict lubrication abrasive particles, cutting wear is mainly cutting abrasive particles, fatigue wear is mainly fatigue abrasive particles, flaky abrasive particles and spherical abrasive particles, and corrosive wear is mainly oxide abrasive particles.

[0114] Among them, the abrasive categories include: normal abrasives, strict lubrication abrasives, cutting abrasives, fatigue abrasives, flaky abrasives, spherical abrasives, and oxide abrasives; the wear failure types include: normal wear, adhesive wear, cutting wear, fatigue wear, and corrosive wear.

[0115] In addition, the present invention also provides a system for executing the power plant wear fault diagnosis method based on ferrographic image analysis, which comprises:

[0116] An image acquisition module 1, wherein the image acquisition module 1 repeatedly captures images at a predetermined period, and sends the captured images to a control unit 2, and uses a sensor built-in shooting probe or a laser scanning confocal microscope to capture the original ferrogram image of the ferrogram spectrum in the oil wear device; wherein the oil wear device includes a ferrogram spectrum, a driving device and a control device; wherein the ferrogram spectrum is driven by the driving device, and the control device controls the ferrogram spectrum by controlling the driving device; the ferrogram spectrum image can also be obtained by an OLVF online visual ferrogram;

[0117] An image processing module 3, the image processing module 3 is used to obtain the image captured by the image acquisition module from the control unit 2, and perform grayscale, binarization and denoising processing on the image;

[0118] A feature extraction module 4 is used to extract relevant attributes of the connected domain of the binary image after image processing as the iron spectrum wear particle features; the iron spectrum wear particle features include: the area, perimeter, circularity, rectangularity, aspect ratio of the connected domain, the major axis length, minor axis length, eccentricity of the ellipse with the same second-order center distance as the boundary rectangle, the diameter of the circle with the same area as the connected domain, and the pixel ratio of the connected domain to its minimum convex polygon;

[0119] A wear particle identification module 5, the wear particle identification module 5 is used to identify the wear particle category by using an algorithm-optimized SVM multi-classification model, the input of the algorithm-optimized SVM multi-classification model is the extracted iron spectrum wear particle features, the output of the algorithm-optimized SVM multi-classification model is the wear particle category, the SVM multi-classification model uses a combination of multiple SVMs, designs a sub-SVM between any two categories of wear particle samples, and uses the Jaya algorithm to optimize the parameters of each sub-SVM in the SVM multi-classification model;

[0120] A fault diagnosis module 6, wherein the fault diagnosis module 6 is used to reconstruct the fault characteristics of the power device and perform wear fault diagnosis on the power device; the fault characteristics of the power device are the number of each category of abrasive particles and / or the proportion of each category of abrasive particles in the iron spectrum image, and the fault characteristics of the power device are used as input, and the deep network model optimized by the algorithm is first used to reconstruct the fault characteristics of the power device, and then the reconstructed power device fault characteristics are classified by using a Softmax classifier, and finally the wear fault type of the power device is output; the deep network model is a stacked denoising autoencoder (SDAE), and the Bayesian optimization algorithm (BO) is used to optimize the parameters of the stacked denoising autoencoder (SDAE);

[0121] A visualization module (7), wherein the visualization module (7) is used to display images, the number of each type of wear particles, and the type of wear fault obtained through diagnosis.

[0122] In addition, the present invention also provides a computer-readable storage medium, in which computer-executable instructions are stored, and the computer-executable instructions are used to execute the power device wear fault diagnosis method.

[0123] Example 1

[0124] Figure 1This is the overall implementation block diagram of this embodiment, in which the input is the original iron spectrum image, and the output is the wear fault type of the power unit. The overall process of this embodiment includes: obtaining iron spectrum images, image preprocessing, abrasive feature extraction, abrasive type identification, wear fault diagnosis, and displaying fault diagnosis results.

[0125] The present embodiment will be further described below in conjunction with the accompanying drawings and embodiments.

[0126] Step 1: Obtain the original iron spectrum image.

[0127] The original ferrographic image of the ferrographic spectrum sheet in the oil wear particle device is repeatedly captured by using a sensor built-in shooting probe or a laser scanning confocal microscope at a predetermined period, and the captured image is sent to the control unit.

[0128] Among them, the oil abrasive device includes a ferroscope spectrum plate, a driving device and a control device; the ferroscope spectrum plate is a glass substrate on which ferromagnetic particles are deposited after magnetic separation, and the control device mainly realizes the movement, focusing imaging, and imaging of the ferroscope spectrum plate through the driving device. The ferroscope image can be used offline after scanning and imaging storage, and can also be viewed online through OLVF.

[0129] Step 2: Image preprocessing.

[0130] The iron spectrum image is converted into a single-channel grayscale image using grayscale technology, and the grayscale image is threshold segmented using the between-class variance method (OTSU) to obtain a binary image. Furthermore, the total variation (TV) algorithm is used to denoise the image.

[0131] Step 3: Extraction of wear particle features.

[0132] First, the connected domain is marked and the bounding rectangle of the connected domain is obtained. The related attributes of the connected domain are calculated as the abrasive particle features, including: the area, perimeter, circularity, rectangularity, aspect ratio of the connected domain, the major axis length, minor axis length, eccentricity of the ellipse with the same second-order center distance as the bounding rectangle, the diameter of the circle with the same area as the connected domain, and the pixel ratio of the connected domain to its minimum convex polygon (i.e., 10 abrasive particle features). This step can use the calculation function in mathematical software tools or programming languages.

[0133] Step 4: Identify the abrasive type.

[0134] According to the internal wear mechanism of the power unit and the morphological differences of the generated abrasive particles, the abrasive particles are divided into the following 7 categories: normal abrasive particles, severe lubrication abrasive particles, cutting abrasive particles, fatigue abrasive particles, flaky abrasive particles, spherical abrasive particles, and oxide abrasive particles.

[0135] Experimental samples: m typical abrasive samples, m value range is [60, 100]. Marking experimental samples:

[0136] T1={(x1,y1),(x2,y2),…,(x m ,y m )}

[0137] where x i =(x i1 ,x i2 ,…,x i10 ) T , (i=1,2,…,m;j=1,2,…,10) is the jth feature of the i-th sample (including 10 types of wear particle features in total), so y i ∈{1,-1}(i=1,2,…,m) represents the i-th sample category label.

[0138] Figure 2 This is the basic flow chart of the abrasive type identification of the present invention, which is described in detail below.

[0139] Step 4.1: Use the OVOSVMs method to divide the sub-training set.

[0140] Single SVM classifier only supports binary classification. For multi-target types, multiple SVM combinations are required. The "SVM one-to-one (OVOSVMs)" method is adopted, that is, a sub-SVM is designed between any two types of abrasive samples. A total of 21 SVM sub-classifiers are required for the 7 types of abrasives, and the training sets of the sub-classifiers are combined in pairs. For the convenience of representation, the abrasive categories are marked, A: normal abrasive; B: strict lubrication abrasive; C: cutting abrasive; D: fatigue abrasive; E: flaky abrasive; F: spherical abrasive; G: oxide abrasive. The training sets of the 21 sub-classifiers are (A, B), (A, C), (A, D), (A, E), (A, F), (A, G), (B, C), (B, D), (B, E), (B, F), (B, G), (C, D), (C, E), (C, F), (C, G), (D, E), (D, F), (D, G), (E, F), (E, G), and (F, G).

[0141] Step 4.2, data preprocessing.

[0142] 1) Remove data noise: Use Gaussian low-pass filter to remove data noise;

[0143] 2) Feature normalization: Since the scales of the features of different types of abrasive particles are different, normalization is required. The normalization formula is as follows:

[0144]

[0145] In the formula, x′ is the normalized feature, x is the original feature, μ is the average value of the original feature, and σ is the standard deviation of the original feature;

[0146] The input and output of the wear particle category must be digital. The wear particle category in each training set needs to be represented by a category label (1, -1), where 1 represents a positive sample and -1 represents a negative sample. For example, when training a subclassifier with a training set of (A, B), category A (i.e., "normal wear particle") is encoded as 1, and category B (i.e., "strictly smooth wear particle") is encoded as -1.

[0147] Step 4.3, build a classification model.

[0148] The classification decision function formula of SVM is:

[0149]

[0150] in, σ is the Gaussian kernel function parameter; x is the function independent variable; x i is the characteristic vector of the i-th sample wear particle; y i is the category label of the i-th sample wear particle; m is the number of samples in the training set; α i is the Lagrange multiplier; α i Is the constraint condition: 0≤α i ≤γ; γ is the regularization parameter; b is the bias.

[0151] Step 4.4, parameter optimization.

[0152] Step 4.4.1, initialize Jaya-SVM parameters.

[0153] The number of design variables of the Jaya algorithm is 2: the population size is 30 (usually set to 30-40), and the condition for the algorithm to terminate is to reach the maximum number of iterations 1000 (usually set to 1000-2000). The ranges of the regularization parameter γ and the kernel function parameter σ in the model are [0.1, 100] and [0.1, 500] respectively.

[0154] Step 4.4.2, select E RMSE As the objective function of the Jaya algorithm, its formula is as follows:

[0155]

[0156] m is the number of wear particle samples; y i is the actual label of the i-th sample wear particle; y i ′ is the predicted label of the wear particle of the i-th sample;

[0157] A set of vectors (γ, σ) is randomly generated within the range of optimization parameters, and a set of (γ, σ) is assigned to each individual in the Jaya algorithm population.

[0158] Step 4.4.3, based on the training results of SVM, calculate the E of each individual in the population RMSE The objective function value is used to obtain the optimal solution E for the current population. RMSEmin and the worst solution E RMSEmax ;

[0159] Step 4.4.4, according to the obtained optimal solution E RMSEmin and the worst solution E RMSEmax Correct the current solution E of the objective function during the iteration RMSE , as follows:

[0160] E′ j,k,i =E j,k,i +r1 j,i (E j,min,i -|E RMSE |)-r2 j,i (E j,max,i -|E RMSE |)

[0161] Among them, E j,k,i is the solution corresponding to the jth variable of the kth candidate variable in the i-th iteration; E j ' ,k,i is the modified solution of the jth variable of the kth candidate variable in the i-th iteration; E j,min,i is the solution of the jth variable of the optimal option in the i-th iteration; E j,max,i is the solution of the jth variable of the worst option in the i-th iteration; r1 j,i 、r2 j,i is the random number corresponding to the jth variable at the i-th iteration, and its value range is [0, 1];

[0162] Step 4.4.5, if the corrected solution E′ RMSE Better than the current solution E RMSE , then retain the modified solution E′ RMSE , otherwise keep the current solution; all the retained solutions are used as the input of the next iteration;

[0163] Step 4.4.6, determine whether the maximum number of iterations has been reached. If not, go to step 4.4. Otherwise, output the optimal solution, i.e., E RMSE Minimum value, and output the corresponding optimal parameter combination (γ, σ).

[0164] Step 4.5, build a multi-classification model.

[0165] After all sub-classifiers are trained, a voting strategy is used for classification, and the category with the most votes is the sample category:

[0166] 4.5.1, put the test sample into the trained sub-classifier for classification. If the classification result is a certain category, the number of votes for that category will be increased by 1;

[0167] 4.5.2, repeat 4.5.1 until the test sample has been tested in all sub-classifiers;

[0168] 4.5.3. Count the votes. The category with the highest number of votes is the final classification result of the test sample.

[0169] Step 5, wear fault diagnosis.

[0170] The wear failure types of diesel power units are divided into five types: normal wear, adhesive wear, cutting wear, fatigue wear and corrosive wear, which correspond to seven types of abrasive particles respectively: normal wear is mainly normal abrasive particles, adhesive wear is mainly severe abrasive particles, cutting wear is mainly cutting abrasive particles, fatigue wear is mainly fatigue abrasive particles, flaky abrasive particles and spherical abrasive particles, and corrosive wear is mainly oxide abrasive particles.

[0171] Considering the multicollinear relationship between the iron spectrum abrasive features and the wear fault types, the stacked denoising autoencoder (SDAE) is used to reconstruct the fault features. The number of hidden layers is set to a fixed value, and the Bayesian optimization algorithm BO is used to optimize the combination of the number of hidden layer nodes of the stacked denoising autoencoder (SDAE); the range of the number of hidden layer nodes of the stacked denoising autoencoder SDAE is [50, 500].

[0172] Figure 3 The flowchart of wear fault diagnosis generally includes the following steps: (1) by performing wear particle identification on the processed binary image, the number of each category of wear particles and / or the proportion of each category of wear particles are obtained by statistics, that is, the power unit fault characteristics; (2) in order to make the learned stacked denoising autoencoder (SDAE) more robust, the fault characteristics are denoised and normalized; (3) the stacked denoising autoencoder (SDAE) is initialized and trained layer by layer with the goal of minimizing the reconstruction error. In this process, the Gaussian process is used to continuously optimize the number of nodes in the hidden layer of the stacked denoising autoencoder (SDAE) and train the stacked denoising autoencoder (SDAE) until the reconstruction error obtained under a certain number of nodes meets the requirements, then the optimization of the stacked denoising autoencoder (SDAE) is stopped; (4) the power unit fault characteristics are placed in the trained stacked denoising autoencoder (SDAE) to obtain the reconstruction characteristics of the power unit fault; (5) the reconstructed characteristics are classified using the Softmax classifier, and the classification result is the type of power unit wear fault.

[0173] The following is a detailed description:

[0174] Step 5.1, obtain characteristic parameters.

[0175] Jaya-SVM was used to identify wear particles in m2 iron spectrum images, and the value range of m2 was [50, 100]. The number of each type of wear particles in each group of results was counted, and the number of each type of wear particles was used as the first characteristic parameter to obtain m2 groups of characteristic parameters: T1 = [d(1), d(2), ..., d(7)], where d(i) represents the number of the i-th type of wear particles (including 7 types of wear particles in total).

[0176] In addition, the proportion of each type of abrasive can be calculated based on the statistical number of each type of abrasive, and the proportion of each type of abrasive can be used as the second characteristic parameter to obtain the m2 group of characteristic parameters: T2 = [p(1), p(2), ..., p(7)], where p(i) represents the proportion of the i-th type of abrasive.

[0177] Step 5.2, data preprocessing.

[0178] In order to prevent the stacked denoising autoencoder from overfitting, noise is added to the input signal to make the learned encoder more robust. In this example, Gaussian noise is added:

[0179]

[0180] ε~N(0,σ 2 )

[0181] In the formula, is the signal after adding noise, z is the original fault feature signal, ε is Gaussian noise, obeys normal distribution, and σ is the variance of ε.

[0182] It should be noted that the input data needs to be denoised only during training, and no denoising is required after training is completed.

[0183] Since the scales of the features of different wear faults are different, normalization is required. The normalization formula is as follows: In the formula, is the normalized feature, is the signal after adding noise, μ′ is The average value of The standard deviation of .

[0184] Step 5.3, establish the SDAE network.

[0185] In order to reduce the complexity of the algorithm, the present invention sets the number of hidden layers of SDAE to 3 and randomly generates initialization points according to the range of node numbers.

[0186] Step 5.3.1, initialize the node number combination.

[0187] The node number combination is expressed as: g = [m 1 ,m 2 ,m 3 ];

[0188] Among them, m 1 Indicates the number of nodes in the first hidden layer, m 2 Indicates the number of nodes in the second hidden layer, m 3 represents the number of nodes in the third hidden layer, and the range of the number of nodes is [50, 500]. The initialization parameters and experimental samples are input into SDAE, and the reconstruction error between the input signal and the reconstructed output signal is used as the model output, which is corrected to make the model close to the real distribution of the function.

[0189] Step 5.3.2, optimize the node number combination.

[0190] The optimization function is:

[0191] g represents the parameter combination to be optimized; f(g) represents the reconstruction error corresponding to the parameter combination; χ is the set of g;

[0192] Using the Gaussian process, the t-th Gaussian prediction distribution is:

[0193]

[0194]

[0195] g t represents the number of nodes in the tth hidden layer; μ(g t ) is g t The mean function of

[0196] k t =[k(g1,g t ),k(g2,g t ),…,k(g t-1 ,g t )]

[0197] k t The transpose of; k(.) represents the covariance function; K is a semidefinite matrix; I is the identity matrix; h=[h1,h2,…,h t ] is the observation sequence of the Gaussian process; h t =f(g t )+εt is the observed value of the Gaussian process; f(g t ) is the reconstruction error corresponding to the parameter combination; ε t is an independent and identically distributed Gaussian noise, which is a random variable; σ n is t The variance of

[0198] Get g through Gaussian process t After the prediction distribution is obtained, it is necessary to select an acquisition function to construct a utility function from the posterior distribution of the model to determine the next parameter combination point. There are many forms of acquisition functions in the prior art, the most commonly used of which is Expected Improvement (EI). The main idea of ​​EI is to search for the next hyperparameter set so that the expectation of the objective function is better than the expectation corresponding to the current optimal hyperparameter set.

[0199] The present invention selects Expected improvement as the acquisition function, and the Gaussian process update formula is as follows:

[0200] α(g)=[μ(g)-q + ]ψ(z)+σ(g)φ(z)

[0201]

[0202] Where: g represents the parameter combination to be optimized, χ is the set of g, z = (μ(g)-q + ) / σ(g), α(g) represents the utility function, μ(g) represents the mean, σ(g) is the standard deviation, and q + is the current maximum utility function value, ψ(z) represents the cumulative probability function of the Gaussian distribution, φ(z) represents the probability density function of the Gaussian distribution, and g t+1 represents the parameter combination evaluated t+1 times.

[0203] The new parameter combination g t+1 The input is fed into SDAE, and the reconstruction error between the input signal and the reconstructed output signal is used as the model output for training, so that the Gaussian model can approach the true distribution of the objective function (here the objective function is the reconstruction error) more quickly and accurately than other combinations of the candidate set χ.

[0204] Step 5.3.3, if the reconstruction error f(g t+1 ) meets the target requirements, the algorithm execution is terminated and exited, and the corresponding hidden layer node number combination g is output t+1 .

[0205] Step 5.3.4, if f(g t+1 ) does not meet the requirements, then (g t+1 ,f(gt+1 )) is input into the Gaussian model, the Gaussian model is modified, and step 5.3.2 is repeated until the required accuracy of the experiment is met.

[0206] Step 5.4, reconstruct the characteristic signal.

[0207] The characteristic signal obtained in step 4.2 (i.e., the preprocessed power unit fault characteristics) is added to the trained SDAE model for reconstruction.

[0208] Step 5.5, wear fault diagnosis.

[0209] The reconstructed feature signal was classified using the Softmax classifier. There was one group of classification errors for each of the normal wear, fatigue wear, cutting wear, and corrosion wear types, and two groups of classification errors for the adhesive wear type. The overall accuracy rate was 92.50%.

[0210] Step 6: Display the fault diagnosis results.

[0211] The wear fault diagnosis results after reconstruction are as follows: Figure 4 As shown in the figure, type labels 1 to 5 represent normal wear, adhesive wear, cutting wear, fatigue wear, and corrosive wear, respectively.

[0212] Example 2

[0213] Figure 5 Schematic diagram of the wear fault diagnosis system of the present invention.

[0214] The power plant fault diagnosis system based on ferrographic image analysis comprises:

[0215] Image acquisition module 1, image acquisition module 1 repeatedly captures images in a predetermined period, and sends the captured images to control unit 2, using a sensor built-in shooting probe or a laser scanning confocal microscope to capture the original ferrogram image of the ferrogram spectrum in the oil wear device; wherein the oil wear device includes a ferrogram spectrum, a driving device and a control device; wherein the ferrogram spectrum is driven by the driving device, and the control device controls the ferrogram spectrum by controlling the driving device; the ferrogram spectrum image can also be obtained through the OLVF online visual ferrogram.

[0216] The image processing module 3 obtains the image captured by the image acquisition module from the control unit 2, and performs grayscale, binarization and denoising processing on the image.

[0217] Feature extraction module 4, feature extraction module 4 extracts relevant attributes of the connected domain of the binary image after image processing as the ferrographic wear particle features; the ferrographic wear particle features include: the area, perimeter, circularity, rectangularity, aspect ratio of the connected domain, the major axis length, minor axis length, eccentricity of the ellipse with the same second-order center distance as the boundary rectangle, the diameter of the circle with the same area as the connected domain, and the pixel ratio of the connected domain to its minimum convex polygon.

[0218] Wear particle identification module 5, wear particle identification module 5 uses an algorithm-optimized SVM multi-classification model to identify wear particle categories, the input of the algorithm-optimized SVM multi-classification model is the extracted iron spectrum wear particle features, the output of the algorithm-optimized SVM multi-classification model is the wear particle category, the SVM multi-classification model uses a combination of multiple SVMs, a sub-SVM is designed between any two categories of wear particle samples, and the Jaya algorithm is used to optimize the parameters of each sub-SVM in the SVM multi-classification model;

[0219] Fault diagnosis module 6, fault diagnosis module 6 reconstructs the fault characteristics of the power unit and performs wear fault diagnosis on the power unit; the fault characteristics of the power unit are the number of each category of abrasive particles and / or the proportion of each category of abrasive particles in the iron spectrum image, these fault characteristics are used as input, the fault characteristics are first reconstructed using a deep network model optimized by an algorithm, and then the reconstructed fault characteristics are classified using a Softmax classifier, and finally the wear fault type of the power unit is output, the deep network model is a stacked denoising autoencoder (SDAE), and the parameters of the stacked denoising autoencoder (SDAE) are optimized using a Bayesian optimization algorithm (BO);

[0220] Visualization module 7, visualization module 7 is used to display images, the number of each category of abrasive particles and the type of wear faults obtained by diagnosis. The displayed images include the original iron spectrum image and the images of connected domains, edges and abrasive particle marks processed by the image processing module.

[0221] A control unit 2, wherein the control unit 2 is connected to a shooting probe or a laser scanning confocal microscope in the image acquisition module 1, and the control unit includes a central processing unit 20 (CPU) and a memory 21, wherein the memory 21 is used to store one or more programs. When the one or more programs are executed by the central processing unit 20 (CPU), the central processing unit 20 (CPU) implements the power unit wear fault diagnosis method proposed in the present invention to diagnose the power unit wear fault.

[0222] The power unit wear fault diagnosis system is used in an operational and visual platform. The operation through the platform can realize the integration of image processing, feature extraction, wear particle identification, and fault diagnosis. Specifically, the total number of wear particles, the number of various types of wear particles, the type of wear fault and other information can be obtained through the original iron spectrum image captured in the oil. The specific operation is introduced as follows.

[0223] Login: Enter your account and password on the login page and click the confirmation button to log in. Click "Remember Password" to confirm the login next time you open the page. The login interface is as follows: Figure 6 shown.

[0224] Image processing: After entering the visualization module initialization interface, first select the wear particle image to be analyzed. Click the "Select Image" button. Figure 7 As shown, open the image selection dialog box, find the address where the image is stored, select the image, and click "Open". The selected initialization image will be displayed in the image display area. After selecting the image to be processed, click the "Image Processing" button. According to the results of the image processing, the processed connected domain, edge and wear particle mark images are displayed in the image display area; at the same time, the total number of wear particles in the image is displayed in the data display area, as shown in the figure below. Figure 8 shown.

[0225] Wear particle identification: After the image processing is completed, the wear particles are identified and classified according to the parameter characteristics of each wear particle extracted from the wear particle mark image. Click the "Wear Particle Identification" button to classify each wear particle. The identification results are as follows: Fig. 9 .

[0226] Fault diagnosis: After the abrasive particles are successfully identified, click the "Fault diagnosis" button. According to the type and quantity of the identified abrasive particles and / or the proportion of each type of abrasive particles, the type of wear fault can be determined, such as Fig.10 shown.

[0227] Clear and exit: After completing a complete analysis process, you can choose to continue analyzing and processing the next image. Before selecting, click the "Clear" button, and the visualization module is in the initialization state.

[0228] Furthermore, an embodiment of the present application also provides a computer-readable storage medium, in which computer-executable instructions are stored, and the computer-executable instructions are used to execute the power device wear fault diagnosis method disclosed in any one of the above-mentioned present inventions.

[0229] A person skilled in the art can understand that all or part of the steps in the various methods of the above embodiments can be completed by instructing related hardware through a program, and the program can be stored in a computer-readable storage medium, which can include: ROM, RAM, disk or CD, etc.

[0230] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A method for diagnosing wear faults of power plants based on ferrographic image analysis, characterized in that: The method comprises the following steps: Step 1, capturing an original ferrographic image using an imaging device at a predetermined period; Step 2, performing image preprocessing on the original ferrographic image to obtain a processed binary image; Step 3, marking the connected domain of the binary ferrographic image, and selecting the relevant attributes of the connected domain of the processed binary image, including the geometric and morphological attribute parameters of the connected domain, as the ferrographic wear particle features; Step 4, training a wear particle recognition model according to the ferrographic wear particle characteristics, wherein the input of the wear particle recognition model is the ferrographic wear particle characteristics, the output of the wear particle recognition model is the wear particle category, the wear particle recognition model is an SVM multi-classification model optimized by an algorithm, the SVM multi-classification model adopts a combination of multiple SVMs, and a sub-SVM is designed between any two categories of wear particle samples. The training set of the sub-SVM adopts a pairwise combination method, and the Jaya algorithm is used to optimize the parameters of each sub-SVM in the SVM multi-classification model; Step 5, training the deep network model optimized by the algorithm, reconstructing the fault characteristics of the power unit, classifying the reconstructed fault characteristics of the power unit using the Softmax classifier, and finally outputting the type of wear fault of the power unit; the fault characteristics of the power unit are the categories of abrasive particles obtained based on the abrasive particle recognition model, and the number of abrasive particles of each category and / or the proportion of abrasive particles of each category obtained by counting the abrasive particle categories; the deep network model is a stacked denoising autoencoder, and the parameters of the stacked denoising autoencoder are optimized using the Bayesian optimization algorithm; Step 6, displaying the iron spectrum images of each stage, the number of each type of wear particles, and the wear fault types obtained by diagnosis; Wherein, in step 4, the classification decision function formula of the SVM multi-classification model is: in, σ is the Gaussian kernel function parameter; x is the function independent variable; x i is the characteristic vector of the i-th sample wear particle; y i is the category label of the i-th sample wear particle; m is the number of samples in the training set; α i is the Lagrange multiplier; α i Is the constraint condition: 0≤α i ≤γ; γ is the regularization parameter; b is the bias; Wherein, in step 4, the Jaya algorithm is used to optimize the parameter combination (γ, σ) of each sub-SVM in the SVM multi-classification model, which specifically includes the following steps: Step 4.1, the Jaya algorithm is designed to have a population size of 30 to 40, the algorithm is terminated when the maximum number of iterations reaches 1000 to 2000, the number of algorithm design variables is 2, and the ranges of the regularization parameter γ and the Gaussian kernel function parameter σ are [0.1, 100] and [0.1, 500], respectively; Step 4.2, select E RMSE As the objective function of the Jaya algorithm, its formula is as follows: m is the number of wear particle samples; y i is the actual label of the i-th sample wear particle; y i ′ is the predicted label of the wear particle of the i-th sample; Step 4.3, a set of vectors (γ, σ) is randomly generated within the range of the optimization parameters, and a set of (γ, σ) is assigned to each individual in the Jaya algorithm population; Step 4.4, based on the training results of SVM, calculate the E of each individual in the population RMSE The objective function value is used to obtain the optimal solution E for the current population. RMSEmin and the worst solution E RMSEmax ; Step 4.5, according to the obtained optimal solution E RMSEmin and the worst solution E RMSEmax Correct the current solution E of the objective function during the iteration RMSE , as follows: E′ j,k,i =E j,k,i +r1 j,i (E j,min,i -|E RMSE |)-r2 j,i (E j,max,i -|E RMSE |) Among them, E j,k,i is the solution corresponding to the jth variable of the kth candidate variable in the i-th iteration; E j ' ,k,i is the modified solution of the jth variable of the kth candidate variable in the i-th iteration; E j,min,i is the solution of the jth variable of the optimal option in the i-th iteration; E j,max,i is the solution of the jth variable of the worst option in the i-th iteration; r1 j,i 、r2 j,i is the random number corresponding to the jth variable at the i-th iteration, and its value range is [0, 1]; Step 4.6, if the corrected solution E′ RMSE Better than the current solution E RMSE , then retain the modified solution E′ RMSE , otherwise keep the current solution; all the retained solutions are used as the input of the next iteration; Step 4.7: Determine whether the maximum number of iterations has been reached. If not, go to step 4.

4. Otherwise, output the optimal solution, i.e., E RMSE Minimum value, and output the corresponding optimal parameter combination (γ, σ).

2. The method for diagnosing wear faults of power plants based on ferrographic image analysis according to claim 1, characterized in that: In step 2, the image preprocessing operations include: grayscale, binarization, and denoising.

3. The method for diagnosing wear faults of power plants based on ferrographic image analysis according to claim 1, characterized in that: In the step 3, before obtaining the ferrographic wear particle features, the connected domain is first marked, the boundary rectangle of the connected domain is obtained, and the related attributes of the connected domain are calculated as the ferrographic wear particle features; The iron spectrum abrasive features include: the area, perimeter, circularity, rectangularity, aspect ratio of the connected domain, the major axis length, minor axis length, eccentricity of the ellipse with the same second-order center distance as the boundary rectangle, the diameter of a circle with the same area as the connected domain, and the pixel ratio of the connected domain to its minimum convex polygon.

4. The method for diagnosing wear faults of power plants based on ferrographic image analysis according to claim 1, characterized in that: The step 5 comprises the following steps: Step 5.1, performing wear particle recognition on the processed binary image, and obtaining the number of each type of wear particles and / or the proportion of each type of wear particles, that is, the fault characteristics of the power device; Step 5.2, in order to make the learned stacked denoising autoencoder more robust, the power unit fault feature is subjected to noise addition and normalization processing; Step 5.3, initialize the stacked denoising autoencoder, and train it layer by layer with the goal of minimizing the reconstruction error. In this process, the Gaussian process is used to continuously optimize the combination of the number of nodes in the hidden layer of the initialized stacked denoising autoencoder and train the stacked denoising autoencoder until the reconstruction error obtained under a certain combination of the number of nodes meets the requirement, then stop optimizing the stacked denoising autoencoder; Step 5.4, putting the power unit fault feature into the trained stacked denoising autoencoder to obtain the reconstructed feature of the power unit fault; Step 5.5, classify the reconstructed features using a Softmax classifier, and the classification result is the wear fault type of the power unit.

5. The method for diagnosing wear faults of power plants based on ferrographic image analysis according to claim 4, characterized in that: In step 5.3, the Bayesian optimization algorithm is used to optimize the parameters of the stacked denoising autoencoder, that is, the number of nodes, which specifically includes the following steps: Step 5.3.1, initialize the node number combination, the node number combination is expressed as: g = [m 1 ,m 2 ,m 3 ]; where m 1 Indicates the number of nodes in the first hidden layer, m 2 Indicates the number of nodes in the second hidden layer, m 3 Indicates the number of nodes in the third hidden layer, the range of the number of nodes is [50, 500]; Step 5.3.2, optimize the node number combination, the optimization function is: g represents the parameter combination to be optimized; f(g) represents the reconstruction error corresponding to the parameter combination; χ is the set of g; Using the Gaussian process, the t-th Gaussian prediction distribution is: g t represents the number of nodes in the tth hidden layer; μ(g t ) is g t The mean function of k t =[k(g1,g t ),k(g2,g t ),…,k(g t-1 ,g t )] k t The transpose of; k(.) represents the covariance function; K is a semidefinite matrix; I is the identity matrix; h=[h1,h2,…,h t ] is the observation sequence of the Gaussian process; h t =f(g t )+ε t is the observed value of the Gaussian process; f(g t ) is the reconstruction error corresponding to the parameter combination; ε t is an independent and identically distributed Gaussian noise, which is a random variable; σ n is t The variance of Step 5.3.3, obtain g through Gaussian process t After the prediction distribution, a collection function is selected to construct a utility function from the posterior distribution of the model to determine the next parameter combination point and the new parameter combination g t+1 Input to the stacked denoising autoencoder and train with the reconstruction error between the input signal and the reconstructed output signal as the model output; Step 5.3.4, if the reconstruction error f(g t+1 ) meets the target requirements, the algorithm execution is terminated and exited, and the corresponding hidden layer node number combination g is output t+1 ; If f(g t+1 ) does not meet the requirements, then (g t+1 ,f(g t+1 )) is input into the Gaussian model for correction and step 5.3.2 is repeated until the required accuracy of the experiment is met.

6. The method for diagnosing wear faults of power plants based on ferrographic image analysis according to claim 1, characterized in that: The abrasive grain categories include: normal abrasive grains, strict lubrication abrasive grains, cutting abrasive grains, fatigue abrasive grains, flaky abrasive grains, spherical abrasive grains, and oxide abrasive grains; the wear failure types include: normal wear, adhesive wear, cutting wear, fatigue wear, and corrosive wear.

7. A system for executing the method for diagnosing wear faults of power plants based on ferrographic image analysis as claimed in any one of claims 1 to 6, characterized in that: include: An image acquisition module (1), the image acquisition module (1) being used to repeatedly capture images at a predetermined period and send the captured images to a control unit (2); An image processing module (3), the image processing module (3) being used to obtain the image captured by the image acquisition module from the control unit (2), and to perform grayscale, binarization and denoising processing on the image; A feature extraction module (4), wherein the feature extraction module (4) is used to extract relevant attributes of a connected domain of a binary image after image processing as a ferrographic wear particle feature; an abrasive particle identification module (5), wherein the abrasive particle identification module (5) is used to identify the abrasive particle category by using an algorithm-optimized SVM multi-classification model, wherein the input of the algorithm-optimized SVM multi-classification model is the extracted iron spectrum abrasive particle feature, and the output of the algorithm-optimized SVM multi-classification model is the abrasive particle category, wherein the SVM multi-classification model uses a combination of multiple SVMs, designs a sub-SVM between any two categories of abrasive particle samples, and uses the Jaya algorithm to optimize the parameters of each sub-SVM in the SVM multi-classification model; A fault diagnosis module (6), wherein the fault diagnosis module (6) is used to reconstruct a power device fault feature and perform wear fault diagnosis on the power device; the power device fault feature is the number of each category of abrasive particles and / or the proportion of each category of abrasive particles in the iron spectrum image, and the power device fault feature is used as an input, and the power device fault feature is first reconstructed using a deep network model optimized by an algorithm, and then the reconstructed power device fault feature is classified using a Softmax classifier, and finally the power device wear fault type is output; the deep network model is a stacked denoising autoencoder, and the parameters of the stacked denoising autoencoder are optimized using a Bayesian optimization algorithm; A visualization module (7), wherein the visualization module (7) is used to display images, the number of each type of wear particles, and the type of wear fault obtained through diagnosis.

8. A computer-readable storage medium, characterized in that: The storage medium stores computer executable instructions, and the computer executable instructions are used to execute the power unit wear fault diagnosis method as described in any one of claims 1 to 6.