A bearing fault diagnosis method based on multi-granularity scanning and catboost model

Abstract

The application discloses a bearing fault diagnosis method based on multi-granularity scanning and a catboost model, and comprises the following steps: pre-processing a bearing vibration signal, normalizing the vibration signal, and storing the vibration signal into a two-dimensional gray picture; changing the structure of a deep forest, removing a cascade forest in the deep forest, and retaining a multi-granularity scanning part; extracting features in the gray picture by using the multi-granularity scanning; using a catboost to replace the cascade forest part in the deep forest, constructing a multi-granularity scanning plus catboost fault diagnosis model, inputting features extracted by using the multi-granularity scanning in an upper layer, and using the catboost model to perform fault diagnosis; training the catboost model by using a training set; determining a verification index; and testing the accuracy of the model by using a test set; the application has higher accuracy than a convolutional neural network and a deep forest, and improves operation efficiency; the application reduces the demand for optimization of extensive hyperparameters, reduces the risk of overfitting, and makes the model have stronger robustness and generalization.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of mechanical bearing fault diagnosis and prognostics health management (PHM) artificial intelligence, and particularly relates to a bearing fault diagnosis method based on multi-granularity scanning and a catboost model. BACKGROUND

[0002] Rotary machines, as transmission devices in mechanical systems, are widely used in aviation machinery, agricultural machinery and modern machine tools, and play an important role in national economic production. As an important part of mechanical systems, rotary equipment usually needs to operate in high pressure, high speed, heavy load and other environments, which greatly increases the probability of problems occurring in rotary machines.

[0003] As one of the most critical parts in rotary machines, bearings account for about 30% of rotary machine failures. The detection of bearing components is the most important for the detection of the entire electromechanical device. Due to the large scale of modern electromechanical equipment, the detection points are many, the sensor detection frequency is high, and the equipment working time is long. The detection system data is developing towards large and giant, and the data types are developing towards diversification and multi-frequency. The original expert system, which relies solely on manpower for detection, cannot meet the requirements, and there is an urgent need for automatic diagnosis algorithms.

[0004] In recent years, fault diagnosis algorithms can be divided into model-based methods and data-driven methods. The data-driven method conforms to the current big data era and has been paid attention by many engineers. Common data-driven fault diagnosis techniques include support vector machine (SVM), artificial neural network (ANN), k-nearest neighbor (KNN), etc. For fault diagnosis based on a large amount of data, the model parameters of the traditional intelligent diagnosis method are limited, the model representation learning ability is weak, and the fault diagnosis ability is limited. Deep neural network (Deep Learning Network) has strong representation learning ability and has been successfully applied in machine vision, medical insurance and aerospace. In recent years, domestic and foreign scholars have also applied deep learning in the field of fault diagnosis. However, deep learning has too many hyperparameters, and the learning ability is seriously dependent on the adjustment of the parameters. The performance of the neural network under the influence of different parameters is quite different. The most representative CNN adopts the strategy of "weight sharing" in order to save the training cost. In summary, although CNN has strong representation learning ability, its performance depends on the selection of hyperparameters. In order to overcome the shortcomings of CNN and traditional machine learning, deep forest based on decision tree integration is proposed in recent years. Compared with deep network, deep forest has fewer parameters and is suitable for small sample learning. Compared with a large number of parameters in CNN, deep forest can be explained better than CNN under the support of gini coefficient and information gain.

[0005] Since the original mechanical vibration signal is one-dimensional data, the current popular signal preprocessing methods are fast Fourier transform (Fast Fourier Transform, FFT), wavelet transform (Wavelet Transform, WT) and empirical mode decomposition (Empirical Mode Decomposition, EMD), etc. These methods need a lot of expert experience, and through human perception, overfitting is easy to occur, and the prediction accuracy of a single predictor is poor, the operation efficiency of the prediction method is low, and the prediction effect on small samples is poor. SUMMARY

[0006] The purpose of the application is to overcome the deficiencies in the prior art, provide a bearing fault diagnosis method based on multi-granularity scanning and catboost model, which can obtain higher accuracy than convolutional neural network and deep forest, improve operation efficiency, reduce the need for optimization of extensive hyperparameters, reduce the risk of overfitting, and make the model have stronger robustness and generalization.

[0007] Technical solution: To achieve the above purpose, the technical solution adopted by the application is:

[0008] A bearing fault diagnosis method based on multi-granularity scanning and catboost model, comprising the following steps:

[0009] Step a, preprocessing the bearing vibration signal, normalizing the bearing vibration signal, and storing it as a two-dimensional gray picture of the bearing vibration signal;

[0010] Step b, changing the structure of the deep forest, removing the cascade forest in the deep forest, and retaining the multi-granularity scanning structure part;

[0011] Step c, using the multi-granularity scanning structure to extract the features of the two-dimensional gray picture of the bearing vibration signal and the bearing vibration signal to obtain a feature vector, splicing and outputting the extracted feature vector, specifically including the following steps:

[0012] Step c1, the multi-granularity scanning structure first divides the input original feature data by using a sliding window to obtain Q feature vectors, and the formula is:

[0013] Q=(J-K) / λ+1,

[0014] Wherein, K is the length and width of the sliding window;

[0015] J is the dimension of the original feature data;

[0016] λ refers to the set sliding step;

[0017] Step c2, input the obtained feature vector into the random forest and the completely random forest, and the two forests respectively obtain Q x-dimensional class vectors, and splice to obtain a 2·Q·x-dimensional vector output;

[0018] Step c3, splice the class vectors obtained by the two forests in step c2 as the output of multi-granularity scanning, extract the features in the gray picture, splice the feature vectors extracted from the one-dimensional original vibration signal and the feature vectors extracted from the two-dimensional gray picture to obtain a 4·Q·x-dimensional vector, and output it;

[0019] Step d, use catboost structure instead of cascade forest part in deep forest, construct multi-granularity scanning plus catboost fault diagnosis model, input the features extracted by multi-granularity scanning in step c, and use catboost structure to do fault diagnosis;

[0020] Step e, training catboost structure using training set, building 4 catboost structures as a Cascatboost cascade layer, each Cascatboost cascade layer obtains a decision coefficient R while obtaining a prediction result 2 , until the average of 4 decision coefficients no longer improves, the Cascatboost cascade layer stops expanding, and the average of 4 prediction results is output as the final prediction result, wherein:

[0021] In the Cascatboost cascade layer, the decision coefficient R 2 is used as an index to judge whether the cascade continues to expand, and the calculation formula of the decision coefficient R 2 is:

[0022]

[0023] In the formula, y i is the true value;

[0024] is the average of the sum of y i ;

[0025] is the predicted value;

[0026] n is the total number of data points, that is, the sample number;

[0027] The closer the value of the decision coefficient R 2 is to 1, the stronger the explanatory ability of the independent variable to the dependent variable, and the better the fitting degree of the model to the data; when the Cascatboost cascade layer expands a new Cascatboost cascade layer i, it is judged whether R is greater than 0, if greater than 0, the Cascatboost cascade layer continues to expand;

[0028] Step f, determine the verification index;

[0029] Step g, use the test set to test the accuracy of the model, and obtain the bearing fault judgment result.

[0030] As a preferred embodiment of the present application: the step a includes the following sub-steps:

[0031] Step a1, obtain an MxM picture, take a sequence with a length of M 2 on the original vibration signal, the pixel value range of the picture is 0-255, and the value M is within the boundary value of the picture pixel, and the normalization formula of a single pixel value N(i) is

[0032] Wherein, L(i) represents the value of the vibration signal on the sequence, and L(i), i∈(0, M 2 ];

[0033] M represents the length of the picture;

[0034] MIN(L) represents the minimum vibration signal value;

[0035] MAX(L) represents the maximum vibration signal value;

[0036] Step a2, the vibration signal is converted into a picture, using P(j, k), j, k∈(0, M 2 ] represents the pixel value on the picture, and the formula of the pixel value is

[0037] Wherein, j represents the horizontal coordinate of the pixel value;

[0038] k represents the vertical coordinate of the pixel value.

[0039] As a preferred embodiment of the application: in step f, the verification standard index adopts accuracy rate Accuracy and precision rate Precision, the accuracy rate is the proportion of all correct predictions, and the precision rate is the ratio of all samples actually positive to those predicted as positive, and the formula is:

[0040] Accuracy rate

[0041] Precision rate

[0042] Wherein, TP, the number of correctly classified positive examples, that is, the number of instances, that is, the number of samples, actually positive and classified as positive by the classifier;

[0043] FP, the number of incorrectly classified positive examples, that is, the number of instances, that is, the number of samples, actually negative and classified as positive by the classifier;

[0044] FN, the number of incorrectly classified negative examples, that is, the number of instances, that is, the number of samples, actually positive and classified as negative by the classifier;

[0045] TN, the number of correctly classified negative examples, that is, the number of instances, that is, the number of samples, actually negative and classified as negative by the classifier.

[0046] Compared with the prior art, the application has the following beneficial effects:

[0047] (1) The vibration signal is converted into a two-dimensional gray picture by normalization, multi-granularity scanning is used for feature learning, and multi-granularity scanning is used for feature learning of the one-dimensional original vibration signal, the extracted feature vectors are spliced, the problem of data loss that may occur in data processing is effectively solved, and the cascaded feature learning ability is strengthened.

[0048] (2) The method of the application automatically determines the cascade by setting the cascade determination index, the model is simple, does not need extensive parameter adjustment, is easy to use, can automatically learn features, has intelligence, the model is simple, is easy to understand and promote for operators, and is easy to use.

[0049] (3) The current cascade forest can only run on CPU, and the operation efficiency is low; catboost can run on GPU, greatly reducing the running time; the deep forest algorithm is similar to the deep neural network structure, adopts a cascade structure to process the original data layer by layer, each cascade layer takes the output of the previous layer as input, and outputs the processed feature information to the next layer; the structure enhances the representation ability of the algorithm, is conducive to improving the prediction accuracy, improving the operation efficiency, reducing the operation time, and can obtain higher accuracy and precision.

[0050] (4) The method of the application strengthens the multi-granularity scanning of high-dimensional input data, and the number of cascade layers is automatically determined by the determination coefficient R 2 , without complex manual parameter adjustment, increasing the efficiency of the entire model.

[0051] (5) The application has a determination coefficient R 2 to automatically determine whether the cascade needs to be expanded, so as to reduce the need for hyperparameter tuning, reduce the risk of overfitting, and make the model have stronger robustness and generalization. BRIEF DESCRIPTION OF DRAWINGS

[0052] Figure 1 is a signal conversion picture schematic diagram;

[0053] Figure 2 is a conversion picture of 9 kinds of faults of CWRU data set;

[0054] Figure 3 is a multi-granularity scanning structure schematic diagram;

[0055] Figure 4 is a part of Cascatboost cascade model structure schematic diagram;

[0056] Figure 5 is a part of Cascatboost cascade model structure schematic diagram;

[0057] Figure 6 is a structural diagram of a multi-granularity scanning plus catboost cascade model;

[0058] Figure 7 is a bearing fault diagnosis flowchart based on multi-granularity scanning and catboost model based on signal conversion picture. DETAILED DESCRIPTION

[0059] The application will be further described below in conjunction with the drawings:

[0060] In deep neural networks, the key to processing image data is the spatial relationship between original pixels, while the key to processing sequence data is the sequence relationship of original data. Based on the above characteristics, a multi-granularity scanning process is used to enhance the cascade, and then a catboost structure is used, which are combined into a Cascatboost cascade layer to diagnose bearing faults. The application can achieve higher accuracy than a single CNN and deep forest, has stronger representation learning ability, lower time consumption, stronger robustness and generalization, and effectively solves the problem of data loss that may occur during data processing in CNN. The multi-granularity scanning structure is as shown in Figure 3 , Figure 4 The catboost structure is as shown in Figure 5 The structure of the cascade model of multi-granularity scanning plus catboost is as shown in Figure 6 The entire application flowchart is as shown in Figure 7 .

[0061] In order to specifically illustrate the method of the application, the application analyzes the Case Western Reserve University data set. This data set uses electric spark technology to simulate fault generation and records the acceleration signals of the equipment under different loads when faults occur and when the equipment is normal. The faults are divided into bearing outer ring fault (Outer Race Fault, OF), bearing inner ring fault (Inner Race Fault, IF) and rolling fault (Rolling Fault), and the fault diameters are 0.007 inches, 0.004 inches and 0.021 inches. The following Figure 2 is a visualization grayscale diagram of the first 80x80 signals under 0hp load for 9 types of faults. The proportions of the training set, the validation set and the test set are 0.5, 0.25 and 0.25. A total of 5 data sets are prepared. Data sets A, B, C, D and E are samples under loads of 0hp, 1hp, 2hp and 3hp, respectively. Data set E is the collection of data sets A, B, C and D. The number of the first four samples is 1160, and the sliding size is 100. The data set enhancement data technology is sliding block offset, which ensures the diversity of the samples.

[0062] The application provides a bearing fault diagnosis method based on multi-granularity scanning and a catboost model, the model has better accuracy, and the processing of signals does not require expert experience. Figure 1 The specific steps are as follows.

[0063] In step a, the bearing vibration signal is preprocessed, the bearing vibration signal is normalized, and the bearing vibration signal is stored as a two-dimensional gray picture.

[0064] The step a includes the following substeps.

[0065] In step a1, an M*M picture is obtained, an M 2 long sequence is taken on the original vibration signal, the pixel value range of the picture is 0-255, the value M is within the boundary value of the picture pixel, and the normalization formula of the single pixel value N(i) is

[0066] Wherein, L(i) represents the value of the vibration signal on the sequence, and L(i), i∈(0, M 2 ];

[0067] M represents the picture length.

[0068] MIN(L) represents the minimum vibration signal value.

[0069] MAX(L) represents the maximum vibration signal value.

[0070] In step a2, the vibration signal is converted into a picture, P(j, k), j, k∈(0, M 2 ] represents the pixel value on the picture, and the formula of the pixel value is

[0071] Wherein, j represents the horizontal coordinate of the pixel value.

[0072] k represents the vertical coordinate of the pixel value.

[0073] In step b, the structure of the deep forest is changed, the cascade forest in the deep forest is removed, and the multi-granularity scanning structure part is reserved.

[0074] The multi-granularity scanning structure is used to extract the feature vectors from the two-dimensional gray picture of the bearing vibration signal and the features in the bearing vibration signal, the extracted feature vectors are spliced and output, and the specific steps include the following steps.

[0075] In step c1, the multi-granularity scanning structure first divides the input original feature data by using a sliding window to obtain Q feature vectors, Q=(J-K) / λ+1, wherein K is the length and width of the sliding window, J is the dimension of the original feature data, and λ refers to the set sliding step.

[0076] Step c2: Input the obtained feature vector into the random forest and the fully random forest. The two forests will each produce Q x-dimensional class vectors. After concatenation, a 2·Q·x-dimensional vector will be output.

[0077] Step c3: The class vectors obtained from the two types of forests in step c2 are concatenated as the output of multi-granularity scanning. Features in the grayscale image are extracted. The feature vectors extracted from the one-dimensional original vibration signal and the feature vectors extracted from the two-dimensional grayscale image are concatenated to obtain a 4·Q·x-dimensional vector, which is then output.

[0078] Step d: Replace the cascaded forest part in the deep forest with the catboost structure to build a multi-granularity scan plus catboost fault diagnosis model. Input the features extracted by the multi-granularity scan used in step c and use the catboost structure to perform fault diagnosis.

[0079] Step e: Train the catboost structure using the training set, constructing four catboost structures as a Cascatboost cascade layer. Each Cascatboost cascade layer obtains a coefficient of determination R^2 while acquiring the prediction result. The Cascatboost cascade layer stops expanding when the mean of the four coefficients of determination no longer increases, and outputs the mean of the four prediction results as the final prediction result, where:

[0080] In the Cascatboost cascade layer, the coefficient of determination R is used. 2 As an indicator for determining whether a cascade should continue to expand, R 2 Calculation method:

[0081]

[0082] In the formula: y i The actual value; For y i The average of the sums; This is the predicted value. R 2 The closer the value of is to 1, the stronger the explanatory power of the independent variable on the dependent variable, proving that the model fits the data well; when the Cascatboost cascade layer is expanded to a new Cascatboost cascade layer i, the judgment is... If the value is greater than 0, the Cascatboost cascade layer continues to expand.

[0083] Step f, determine the validation metrics; the validation criteria are accuracy and precision. Accuracy is the proportion of all correctly predicted predictions, and precision is the ratio of all samples predicted as positive to those that were actually positive. The formula is:

[0084] Accuracy

[0085] Precision

[0086] TP, the number of correctly classified positive examples, i.e. the number of instances, i.e. samples, that are actually positive and classified as positive by the classifier;

[0087] FP, the number of incorrectly classified positive examples, i.e. the number of instances, i.e. samples, that are actually negative and classified as positive by the classifier;

[0088] FN, the number of incorrectly classified negative examples, i.e. the number of instances, i.e. samples, that are actually positive and classified as negative by the classifier;

[0089] TN, the number of correctly classified negative examples, i.e. the number of instances, i.e. samples, that are actually negative and classified as negative by the classifier.

[0090] Step g, using the test set to test the accuracy of the model, and obtaining the bearing fault judgment result.

[0091] The above only describes the preferred embodiments of the present application, and it should be noted that for those skilled in the art, without departing from the principles of the present application, a number of improvements and refinements can be made, and these improvements and refinements should also be considered within the scope of protection of the present application.

Claims

1. A bearing fault diagnosis method based on multi-granularity scanning and the CatBoost model, characterized in that: Includes the following steps: Step a: Preprocess the bearing vibration signal by normalizing it and storing it as a two-dimensional grayscale image of the bearing vibration signal. Step b: Modify the structure of the deep forest, remove the cascaded forest in the deep forest, and retain the multi-granularity scan structure part; Step c involves using a multi-granularity scanning structure to extract feature vectors from the two-dimensional grayscale image of the bearing vibration signal and the features in the bearing vibration signal. The extracted feature vectors are then concatenated and output. This process includes the following steps: Step c1: The multi-granularity scanning structure first uses a sliding window to segment the input raw feature data, obtaining Q feature vectors, as shown in the formula: Q = (J - K) / λ + 1, Where K represents the length and width of the sliding window; J represents the dimension of the original feature data; λ refers to the set sliding step size; Step c2: Input the obtained feature vector into the random forest and the fully random forest. The two forests will each produce Q x-dimensional class vectors. After concatenation, a 2·Q·x-dimensional vector will be output. Step c3: The class vectors obtained from the two types of forests in step c2 are concatenated as the output of multi-granularity scanning. Features in the grayscale image are extracted. The feature vectors extracted from the one-dimensional original vibration signal and the feature vectors extracted from the two-dimensional grayscale image are concatenated to obtain a 4·Q·x-dimensional vector, which is then output. Step d: Replace the cascaded forest part in the deep forest with the catboost structure to build a multi-granularity scan plus catboost fault diagnosis model. Input the features extracted by the multi-granularity scan used in step c and use the catboost structure to perform fault diagnosis. Step e: Train the catboost structure using the training set, constructing four catboost structures as a Cascatboost cascade layer. Each Cascatboost cascade layer obtains a coefficient of determination R along with the prediction result. 2 The Cascatboost cascade layer stops expanding when the mean of the four determination coefficients no longer increases, and outputs the mean of the four prediction results as the final prediction result, where: In the Cascatboost cascade layer, the coefficient of determination R is used. 2 The coefficient of determination R is used as an indicator to determine whether a cascade will continue to expand. 2 The calculation formula is: In the formula: y i The actual value; For y i The average of the sums; This is a predicted value; n is the total number of data points; Coefficient of determination R 2 The closer the value of is to 1, the stronger the explanatory power of the independent variable on the dependent variable, proving that the model fits the data well; when the Cascatboost cascade layer is expanded to a new Cascatboost cascade layer i, the judgment is... If the value is greater than 0, the Cascatboost cascade layer continues to expand. Step f: Determine the validation metrics; Step g: Use the test set to test the accuracy of the model and obtain the bearing fault diagnosis results.

2. The bearing fault diagnosis method based on multi-granularity scanning and the CatBoost model according to claim 1, characterized in that: Step a includes the following sub-steps: Step a1: Obtain an M×M image, and extract a segment of length M from the original vibration signal. 2 The sequence of images has pixel values ​​ranging from 0 to 255, with the value M falling within the boundary values ​​of the image pixels. The formula for normalizing a single pixel value N(i) is: Where L(i) represents the value of the vibration signal on the sequence, and L(i), i∈(0,M) 2 ]; M represents the image length; MIN(L) represents the minimum vibration signal value; MAX(L) represents the maximum vibration signal value; Step a2: Convert the vibration signal into an image, using P(j,k), j,k∈(0,M) 2 ] represents the pixel value in the image. The formula for the pixel value is: Where j represents the x-coordinate of the pixel value; k represents the vertical coordinate of the pixel value.

3. The bearing fault diagnosis method based on multi-granularity scanning and the CatBoost model according to claim 1, characterized in that: In step f, the validation criteria are accuracy and precision. Accuracy is the proportion of all correctly predicted predictions, and precision is the ratio of all samples predicted as positive to those that were actually positive. The formula is as follows: accuracy Accuracy Wherein, TP is the number of instances that are correctly classified as positive, that is, the number of instances that are actually positive and are classified as positive by the classifier, i.e., the number of samples; FP, the number of instances incorrectly classified as positive, that is, the number of instances that are actually negative but were classified as positive by the classifier, i.e. the number of samples; FN, the number of instances incorrectly classified as negative, that is, the number of instances that are actually positive but were classified as negative by the classifier, i.e. the number of samples; TN, the number of instances correctly classified as negative, that is, the number of instances that are actually negative and classified as negative by the classifier, i.e., the number of samples.

Images