A bearing state monitoring method based on a signal-based open set incremental identification model

Abstract

The application relates to the technical field of classification recognition, in particular to a bearing state monitoring method based on an open set incremental recognition model of signals, which comprises the following steps: pretreating and extracting features of collected bearing data; adopting a PCA algorithm to reduce dimensions of the feature data processed in step one according to eigenvalues, so as to obtain feature components after dimension reduction; establishing an open set incremental model, recognizing bearing states, and detecting abnormal data. The application proposes that the incremental open set model is applied to bearing vibration signals in industry, when new bearing data is available, an incremental learning method is adopted, new knowledge is used to update an existing model, and an open set recognition algorithm provides a rejection option for a classifier, so as to recognize new untrained target types.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of classification recognition, in particular to a bearing state monitoring method based on a signal open set incremental recognition model. BACKGROUND

[0002] Machines in Industry 4.0 are becoming more and more complex, which makes their vibration signal analysis a challenging task. Bearings are recognized as key components in industrial machines, therefore, mastering their vibration information can enhance the diagnostic process. The vibration signal of a defective bearing operating under stationary conditions can be considered as an amplitude modulated waveform. This makes envelope analysis one of the most effective methods for bearing health monitoring under stationary conditions.

[0003] In the document Order-Based Identification of Bearing Defects under Variable Speed Condition, three frequency domain features SPRO, SPRI and SPRR are extracted from vibration data, and a multi-kernel support vector machine (MSVM) classification model is established, which can effectively identify the bearing state; but the data in industry is changing day by day, and new fault signals may also appear, which may cause the original model to be misrecognized, the model proposed in the document cannot effectively identify new fault categories, and needs to be retrained, which will require a large amount of memory time. SUMMARY

[0004] In view of the deficiencies of the existing algorithm, the application proposes an incremental open set model applied to the bearing vibration signal in industry. When new bearing data is obtained, an incremental learning method is used to update the existing model with new knowledge. The open set recognition algorithm provides a rejection option for the classifier to identify new untrained target types.

[0005] The technical scheme adopted by the application is: a bearing state monitoring method based on a signal open set incremental recognition model, comprising the following steps:

[0006] Step 1: Preprocessing and feature extraction of collected bearing data;

[0007] Further, the preprocessing is to divide the bearing vibration data and the rotating speed data into 2-second windows;

[0008] Further, the feature extraction includes mean, standard deviation, mode, maximum value, minimum value, skewness and kurtosis of the segmented data.

[0009] Step 2: using the PCA algorithm to reduce the dimension of the feature data processed in step 1 according to the characteristic value, to obtain the reduced feature components;

[0010] Further, it specifically includes:

[0011] S21, set n rows of d-dimensional data, the processed feature data is arranged in n rows and d columns to form a matrix X; each column of X is zero-meaned;

[0012] S22, eigenvalues and corresponding eigenvectors of the covariance matrix are calculated; the eigenvectors are arranged in rows according to the corresponding eigenvalues from large to small to form a matrix, and the first k rows are taken to form a matrix P; wherein Y=PX is the data after dimensionality reduction to k dimensions.

[0013] Step three, an open set incremental model is established to identify the bearing state and detect abnormal data;

[0014] Further, it specifically comprises:

[0015] First, training starts from K initial classes, wherein K is the number of categories of known classes, an initial model M0 is generated through extreme value theory, bearing state data of test samples are classified into K classes, and data of the remaining classes is effectively rejected into unknown classes;

[0016] Secondly, a new class is learned, the model M0 is updated, and the updated model classifies the bearing state data of the test samples into K+1 classes, and the data of the remaining unlearned classes is effectively rejected into unknown classes;

[0017] Finally, unknown class data is continuously added and identified.

[0018] The beneficial effects of the present application are:

[0019] 1. The open set incremental model of the present application has better recognition effect on new class samples than existing methods, and only needs to increase new data each time after the model is generated, without the need to retrain the model;

[0020] 2. The signal data in the bearing and other industries has good classification and recognition effect, effectively reduces the cost of model training, and has good effect in identifying new class data, and has good application prospect in human-computer integrated intelligent systems. BRIEF DESCRIPTION OF DRAWINGS

[0021] Figure 1 is a bearing state monitoring method flow chart of the open set incremental recognition model based on signals of the present application;

[0022] Figure 2 is a data dimensionality reduction corresponding Pareto chart of the present application;

[0023] Figure 3 is an F1 index chart in the open set incremental scenario of the present application. DETAILED DESCRIPTION

[0024] The application will be further described below with reference to the accompanying drawings and examples, which are simplified schematic diagrams and only schematically show the basic structure of the application, and thus only show the components related to the application.

[0025] The data set "Bearing Vibration Data under Time-varying Rotational Speed Conditions" contains vibration signals collected from bearings of different health conditions under time-varying rotational speed conditions, with a total of 60 data sets; for each data set, there are two experimental settings: bearing health condition and variable speed condition. The health conditions of the bearings include (i) healthy, (ii) inner ring defect failure, (iii) outer ring defect failure, (iv) failure with ball defect, and (v) failure with comprehensive defects of inner ring, outer ring and ball; the running speed conditions are: (i) increasing speed, (ii) decreasing speed, (iii) increasing then decreasing speed, and (iv) decreasing then increasing speed. Therefore, there are 20 different cases; to ensure the authenticity of the data, 3 experiments are collected for each experimental setting, and the result is a total of 60 data sets. Each data set contains two channels: "Channel_1" is the vibration data measured by the accelerometer, and "Channel_2" is the rotational speed data measured by the encoder. All these data are sampled at a frequency of 200,000 Hz, with a sampling duration of 10 seconds. The CPR (Cycles Per Revolution) of the encoder is 1024.

[0026] As shown in Figure 1 A bearing condition monitoring method based on a signal-based open set incremental identification model includes the following steps:

[0027] Step one, pre-processing and feature extraction of bearing data;

[0028] Further, the extracted features include: mean, standard deviation, mode, maximum value, minimum value, skewness and kurtosis;

[0029] In this embodiment, for the bearing vibration signal, a window size of 2S is selected, a total of 300 samples are generated, of which 150 groups are test samples and 150 groups are training samples, according to the vibration data measured by the accelerometer and the rotational speed data measured by the encoder, 14 groups of features are extracted, including mean, standard deviation, mode, maximum value, minimum value, skewness and kurtosis.

[0030] The health conditions of the bearings are shown in Table 1, which have five states.

[0031] Table 1 Health conditions of bearings and abbreviations

[0032]

[0033] Step two, feature dimension reduction using PCA algorithm;

[0034] PCA is used for feature dimensionality reduction. PCA is a common data analysis method, often used for dimensionality reduction of high-dimensional data, and can be used to extract the main feature components of the data.

[0035] The specific steps are as follows:

[0036] Given n rows and d dimensions of data, the original feature data is arranged into an n-row, d-column matrix X; each column of X (representing an attribute) is zero-meaned, that is, the mean of that column is subtracted;

[0037] Next, calculate the covariance matrix. Let y be zero-mean data of X; find the eigenvalues ​​and corresponding eigenvectors of the covariance matrix; arrange the eigenvectors into a matrix by row according to the corresponding eigenvalues ​​from largest to smallest, and take the first k rows to form matrix P; where Y = PX is the data after dimensionality reduction to k dimensions, and the dimension of Y is n*k.

[0038] In this embodiment, PCA dimensionality reduction is used to reduce the original 14 sets of features to 5 sets of features, such as... Figure 2 The diagram shows the corresponding Pareto plot. It can be seen from the plot that these five features account for 90% of the total contribution (feature values). Among them, feature F14 accounts for 36% of the total contribution. Therefore, F14, F11, F5, F8, and F12 are selected as the final features. F1-F7 represent the mean, standard deviation, mode, maximum, minimum, skewness, and kurtosis of the vibration data, respectively. F8-F14 represent the mean, standard deviation, mode, maximum, minimum, skewness, and kurtosis of the rotational speed data, respectively.

[0039] Step 3: Establishing the open set incremental model;

[0040] Incremental Scenario: During training, an initial model M0 is generated from K initial classes, where K is the number of known classes. In this embodiment, K=2. The initial model M0 is generated using Extreme Value Theory (EVM). When a new batch of samples arrives, the test set samples are classified to identify K classes, and the new classes are rejected as "unknown". When a new batch of samples arrives, after labeling the new classes as (K+1) to (K+S), where S is the number of classes learned each time (in this embodiment, S=1), incremental learning is performed to update the initial model M0 to adapt to new samples of existing classes and new classes. The test samples are classified into classes 1 to (K+S), and the new classes are rejected as "unknown". Therefore, unknown classes can be continuously identified and added.

[0041] Open set model: According to the central limit theorem, regardless of the distribution of a random variable X, the mean of each sample of random variables follows a normal distribution; similarly, extreme value theory states that regardless of the distribution of a random variable X, the maximum value of each sample of random variables follows an extreme value distribution F; let z = -mij where m ij is half of the nearest distance of the sample point of class A i to all sample points of class A j , so the minimum half distance is considered to obey the Weibull distribution; let y i be the i-th row data in Y, for sample y i , the probability density function of x of class C

[0042]

[0043] where ||y i -x|| is the distance of x to sample y i , k i and λ i are Weibull shape and scale parameters respectively; k i and λ i are obtained by fitting to the minimum distance; after obtaining the parameters of all samples, the probability that x is not more than the margin can be expressed as:

[0044] ψ i = 1-F(||x-y i ||; k i , λ i ) (2)

[0045] where F is the distribution function of f, the probability that the test sample x is associated with class C l is:

[0046]

[0047] Therefore, the decision function for classifying K known classes and identifying a new class of "unknown" is defined as follows, and the output label formula is:

[0048]

[0049] where y* is the output label, δ is the probability rejection threshold, and K is the number of classes of known classes.

[0050] In this embodiment, class samples are gradually added to simulate the incremental open set scene;

[0051] The specific operation is as follows: 1) select 2 class samples in the training samples for training to generate model M0, and identify the health status of 5 classes of bearings in the test samples (the remaining 3 classes are identified as unknown); 2) continue to train the third class samples of the training samples and the model MO to generate the model M1, which can identify the health status of 3 classes of bearings in the test samples, and the remaining 2 classes are identified as unknown; 3) in this way, new class data is continuously added to generate a model to identify the data of the test samples.

[0052] The experiments are shown in Tables 2-5, which are incremental results:

[0053] Table 2: The H, ORD data are trained to generate a model M0, and five types of bearing health states in the test sample are identified. As can be seen from the table, under the model, the other three health states will be misclassified as H, ORD; the IRD data are trained on the basis of the M0 model to obtain a model M1, and as can be seen from Table 3, in the model, H, ORD and IRD can be basically classified, and the remaining two health states can also be well distinguished. The BD data are trained on the basis of the M1 model to obtain a model M2, and as can be seen from Table 4, in the model M2, H, ORD, IRD and BD can be basically classified, and the bearing state CD that is not in the training can also be well distinguished from the four types. It can be seen that the model does not lose the feature information of the known categories, and can also distinguish the unknown categories. The CD data are trained on the basis of the M2 model to obtain a model M3, and as can be seen from Table 5, in the model, the five types of bearing health states H, ORD, IRD, BD and CD can be basically classified.

[0054] Table 2: Classification results of the model M0 trained by H, ORD data

[0055]

[0056] Table 3: Classification results of the model M1 trained by adding IRD data on the basis of the M0 model

[0057]

[0058] Table 4: Classification results of the model M2 trained by adding BD data on the basis of the M1 model

[0059]

[0060] Table 5: Classification results of the model M3 trained by adding CD data on the basis of the M2 model

[0061]

[0062] As Figure 3As shown, it is the complete process of increment, and it can be seen that the F1 evaluation index is rising with the increasing of the class data; in this embodiment, the present application is compared with the document "Order-Based Identification of Bearing Defects under Variable Speed Condition" (referred to as comparative document 1), and the results are shown in Tables 6 and 7; the precision of BD and CD of the present application is higher than that of comparative document 1, and the recall of ORD, IRD, BD and CD is higher than that of comparative document 1; as shown in Table 8, the accuracy of the method proposed in this example is 81.3%, which is higher than 79% of comparative document 1.

[0063] Table 6 Comparison of precision between the method of the present application and comparative document 1

[0064]

[0065] Table 7 Comparison of recall between the method of the present application and comparative document 1

[0066]

[0067] Table 8 Comparison of overall accuracy between the method of the present application and comparative document 1

[0068]

[0069] Based on the above ideal embodiments according to the present application, through the above description, relevant personnel can make various changes and modifications without deviating from the technical idea of the present application. The technical scope of the present application is not limited to the content of the specification, and must be determined according to the scope of the claims.

Claims

1. A bearing condition monitoring method based on an open-set incremental identification model of signals, characterized in that, Includes the following steps: Step 1: Preprocess and extract features from the collected bearing data; Step 2: Use the PCA algorithm to reduce the dimensionality of the feature data processed in Step 1 based on the eigenvalues, and obtain the dimensionality-reduced feature components. Step 3: Establish an open set incremental model to identify bearing status and detect abnormal data; Step three specifically includes: First, training from K Starting with an initial class K Given the number of classes, an initial model is generated using extreme value theory to classify the bearing state data of the test samples into... K Data from one class is validly rejected and classified as an unknown class; Secondly, a new class is learned to update the initial model, enabling the updated initial model to classify the bearing state data of the test samples into... K +1 class, and the remaining data from unlearned classes are effectively rejected as unknown classes; Finally, continuously add data of unknown categories and perform identification; Extreme value theory includes: Let ,in, Is class Sample points to class Half the nearest distance of all sample points; let y i It is the first in Y i Row data, for samples y i Category x The probability density function is derived based on the Weibull distribution, as shown in Equation 1: (1) in, yes x With sample y i distance, k i and λ i These are the Weibull shape and scale parameters, respectively; after obtaining the parameters for all samples, x The probability of not exceeding the margin is expressed as: (2) in, F yes f The distribution function of the test sample x With class C l The probability of association is: (3) use K The decision function for classifying known classes and identifying unknown classes, with the output label formula as follows: (4) in, y Here, δ is the output label, and δ is the probability rejection threshold. K The number of known classes.

2. The bearing condition monitoring method based on the open set incremental identification model of the signal according to claim 1, characterized in that: Preprocessing involves dividing the bearing vibration data and rotational speed data into 2-second windows.

3. The bearing condition monitoring method based on the open set incremental identification model of the signal according to claim 1, characterized in that: Feature extraction includes calculating the mean, standard deviation, mode, maximum, minimum, skewness, and kurtosis of the segmented data.

4. The bearing condition monitoring method based on the open set incremental identification model of the signal according to claim 1, characterized in that, Step two specifically includes: S21. Given n rows and d dimensions of data, arrange the processed feature data into an n-row, d-column matrix X; then perform zero-mean normalization on each column of X. S22. Find the eigenvalues ​​and corresponding eigenvectors of the covariance matrix; arrange the eigenvectors into a matrix by row according to the corresponding eigenvalues ​​from largest to smallest, and take the first k rows to form matrix P; where Y=PX is the data after dimensionality reduction to k dimensions.

Images