A bearing fault diagnosis method and system based on multi-information fusion

Abstract

This invention provides a bearing fault diagnosis method and system based on multi-information fusion. Step 1: Acquire time-varying vibration signals from the bearing. Calculate the number of sample points collected per revolution of the bearing based on the bearing speed and sampling frequency. Divide the acquired time-varying vibration signals into multiple samples according to the calculated number of sample points. Divide the samples into training and test datasets according to a certain ratio and add corresponding category labels. Step 2: Construct a signal feature extraction layer that includes extracted data features, impact features, time-domain features, and frequency-domain features. Step 3: Use an ELM classifier model to randomly map the data features from the experience pool and output the results. Step 4: Train the output weights of the bearing fault diagnosis model using the training dataset and use the trained weights for fault diagnosis on the test dataset. This invention can extract and effectively fuse signal features with multiple different attributes, facilitating bearing fault diagnosis.

Description

Technical Field

[0001] This invention relates to the field of bearing fault diagnosis, and in particular to a bearing fault diagnosis method based on multi-information fusion. Background Technology

[0002] Rotating equipment is used in all aspects of social life and production, and is an integral part of human life. However, bearings are one of the most critical components in rotating equipment. Bearing failure not only increases the vibration and noise of the rotating equipment but also affects its normal operation, and can even cause irreparable losses. Therefore, how to identify bearing faults and their types in a timely and effective manner is a challenge currently faced by professionals in this field. As the diagnostic accuracy of existing intelligent fault diagnosis algorithms improves, the computational resources and time requirements also continuously increase. Summary of the Invention

[0003] In view of this, the purpose of the present invention is to provide a bearing fault diagnosis method based on multi-information fusion with high diagnostic accuracy.

[0004] To achieve the above objectives, the present invention employs the following technical solution:

[0005] A bearing fault diagnosis method based on multi-information fusion, characterized by comprising the following steps:

[0006] Step 1: Collect the time vibration signal of the bearing. Calculate the number of sample points collected per revolution of the bearing based on the bearing speed and sampling frequency. Divide the collected time vibration signal into multiple samples according to the calculated number of sample points. Divide the samples into training dataset and test dataset according to a certain ratio and add corresponding category labels.

[0007] Step 2: Construct a signal feature extraction layer that includes extracted data features, impulse features, time-domain features, and frequency-domain features;

[0008] Step 3: Use the ELM classifier model to randomly map the data features from the experience pool and output the results;

[0009] Step 4: Train the output weights of the bearing fault diagnosis model using the training dataset, and use the trained weights on the test dataset for fault diagnosis.

[0010] Furthermore, in step one, the formula for calculating the number of sample points is as follows:

[0011]

[0012] In the formula, n s f is the rotational speed of the bearing. re The sampling frequency;

[0013] The original dataset is divided into multiple samples based on the number of sample points, and then the samples are further divided into training dataset and test dataset according to a certain ratio.

[0014] Furthermore, in step two, the data features extracted by the signal feature extraction layer include the mean, standard deviation, mean absolute error, and bias, as detailed below:

[0015] A. Average value—shows the average level across samples:

[0016]

[0017] In the formula: N is the number of data points contained in a sample; x(n) is the nth element in the sample;

[0018] B. Standard deviation—shows the dispersion of each individual relative to the sample mean:

[0019]

[0020] In the formula: The average value for each sample;

[0021] C. Mean Absolute Error—This shows the magnitude of the actual error in the sample:

[0022]

[0023] D. Skewness—Indicates the degree of asymmetry in the distribution of data in a sample:

[0024]

[0025] In the formula: sσ is the standard deviation of the sample.

[0026] Furthermore, in step two, the impact features extracted by the signal feature extraction layer include root mean square and kurtosis, as detailed below:

[0027] A. Root Mean Square (RMS) – Commonly used for analyzing signal noise:

[0028]

[0029] B. Kurtosis—a numerical signal statistic used to reflect the distribution characteristics of random variables, especially vibration and shock signals of surface damage:

[0030]

[0031] Furthermore, in step two, the temporal features extracted by the signal feature extraction layer include energy and difference, as detailed below:

[0032] A. Energy — used to reflect the energy level of a time-domain signal:

[0033]

[0034] B. Differential – Used to reflect the differential characteristics of time-domain signals:

[0035]

[0036] Furthermore, in step two, the frequency domain features extracted by the signal feature extraction layer include Fast Fourier Transform to reflect the frequency domain features of the sample.

[0037] Furthermore, in step three, the random mapping layer of ELM is first used to perform random mapping to obtain the feature matrix of the sample, the formula of which is as follows:

[0038]

[0039] In the formula: a = [a i1 ...a in ] and b = [b1...b i [ ] are the weight matrix and threshold vector, respectively, which can be obtained through initialization; l is the number of nodes in the hidden layer; h(·) is the activation function of the random mapping layer, and its expression can be represented by, but is not limited to, the following expressions:

[0040] h(x i )=h(a il x i +b l )

[0041] Secondly, the output weights of the training are obtained using the ELM output algorithm, and their expression is as follows:

[0042]

[0043] In the formula: β=[β l1 …β li ] T To output the weights; The inverse of the characteristic matrix; T = [t1, t2, ..., t] n ] T These are the labels for the samples.

[0044] Furthermore, in step four, firstly, the initial input weights and thresholds, as well as the trained output weights, are obtained by training the training dataset. Secondly, the dataset is used for testing, and the output expression of the test is as follows:

[0045] Hβ=O

[0046] In the formula: O=[O1…O j[] represents the output label.

[0047] The present invention also provides a bearing fault diagnosis system based on multi-information fusion, characterized in that it includes a data acquisition layer, an input layer, a first hidden layer, a second hidden layer and an output layer;

[0048] The data acquisition layer is used to acquire the bearing's time vibration signal; the first hidden layer performs a first mapping on the data input from the input layer; the second hidden layer performs a second mapping on the random mapping layer of ELM; and the output layer is used to output fault labels.

[0049] Furthermore, the signal feature extraction layers, which extract data features, impulse features, time-domain features, and frequency-domain features, form the first hidden layer. The calculation formula for each feature is embedded into the first hidden layer as an activation function of a neural node.

[0050] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0051] 1. This invention can extract signal features with multiple different attributes and effectively fuse these features, which facilitates bearing fault diagnosis.

[0052] 2. This invention belongs to a single-feedback fault diagnosis neural network, which does not require repeated iterative updates, has a very fast computation speed, and has very low requirements for computing resources.

[0053] 3. This invention has the ability to mine data and learn autonomously, and has better diagnostic accuracy and efficiency compared with traditional fault diagnosis methods. Attached Figure Description

[0054] Figure 1 This is a flowchart of the calculation process of the present invention;

[0055] Figure 2 This is a network architecture diagram of the present invention;

[0056] Figure 3 This is a visualization of the data features of the test dataset for this invention.

[0057] Figure 4 This is the confusion matrix of the test results of this invention. Detailed Implementation

[0058] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, many other embodiments obtained by those skilled in the art without creative effort are all within the scope of protection of the present invention.

[0059] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0060] This invention relates to a bearing fault diagnosis method and system based on multi-information fusion, primarily used to improve the accuracy and efficiency of rolling bearing fault diagnosis. The system comprises an input layer, two hidden layers, and an output layer, as follows: Figure 2 As shown, Figure 1 The method shown includes the following steps:

[0061] Step 1: Collect the time vibration signal of the bearing. Calculate the number of sample points collected per revolution of the bearing based on the bearing speed and sampling frequency. Divide the collected vibration signal into multiple samples according to the calculated number of sample points. Divide the samples into training dataset and test dataset according to a certain ratio and add corresponding category labels.

[0062] Step 2: Construct a signal feature extraction layer that includes extracted data features, impulse features, time-domain features, and frequency-domain features;

[0063] Step 3: Use the ELM classifier model to randomly map the data features from the experience pool and output the results;

[0064] Step 4: Train the output weights of the bearing fault diagnosis model using the training dataset, and then use the trained weights on the test dataset for fault diagnosis.

[0065] The present invention was validated using a public dataset of rolling bearings shared by Case Western Reserve University. The bearing model was SKF6205, the bearing fault size was 0.007 in, the condition was no-load, the sampling frequency was 12 kHz, four speeds were used (1730, 1750, 1772, and 1797 r / min), and six bearing states were used (normal bearing, inner ring fault, rolling element fault, and outer ring fault at three different locations: three o'clock fault (@3:00), six o'clock fault (@6:00), and twelve o'clock fault (@12:00)). 200 samples were taken for each fault type under each condition, and samples with the same fault were assigned the same label. The relevant information for the training and testing datasets is shown in Table 1.

[0066] Table 1. Detailed information on the training and testing datasets.

[0067]

[0068] In step one, the formula for calculating the number of sample points is as follows:

[0069]

[0070] In the formula, n s f is the rotational speed of the bearing.re The sampling frequency;

[0071] The original dataset is divided into multiple samples based on the number of sample points, and then the samples are further divided into training dataset and test dataset according to a certain ratio.

[0072] Based on a rotational speed of 1797 r / min and a sampling frequency of 12 kHz, it can be calculated that the number of data points collected for one revolution of the bearing is approximately 401.

[0073] Step 2: Extract the following features from the sample using a feature extraction layer:

[0074] (1) Data characteristics:

[0075] A. Average value—shows the average level across samples:

[0076]

[0077] In the formula: N is the number of data points contained in a sample; x(n) is the nth element in the sample.

[0078] B. Standard deviation—shows the dispersion of each individual relative to the sample mean:

[0079]

[0080] In the formula: The average value for each sample.

[0081] C. Mean Absolute Error—This shows the magnitude of the actual error in the sample:

[0082]

[0083] D. Skewness—Indicates the degree of asymmetry in the distribution of data in a sample:

[0084]

[0085] In the formula: sσ is the standard deviation of the sample.

[0086] (2) The impact characteristics mainly include:

[0087] A. Root Mean Square (RMS) – Commonly used for analyzing signal noise:

[0088]

[0089] B. Kurtosis—a numerical signal statistic used to reflect the distribution characteristics of random variables, especially vibration and shock signals of surface damage:

[0090]

[0091] (3) The time-domain features mainly include:

[0092] A. Energy — used to reflect the energy level of a time-domain signal:

[0093]

[0094] B. Differential – Used to reflect the differential characteristics of time-domain signals:

[0095]

[0096] (4) Frequency domain features mainly include Fast Fourier Transform (FFT) to reflect the frequency domain features of the sample:

[0097] Step 3: First, use the random mapping layer of ELM to perform random mapping to obtain the feature matrix of the sample. The formula expression is as follows:

[0098]

[0099] In the formula: a = [a i1 …a in ] and b = [b1...b i [ ] are the weight matrix and threshold vector, respectively, which can be obtained through initialization. h(·) is the activation function of the random mapping layer, and its expression can be represented, but is not limited to, the following expression:

[0100] h(x i )=h(a il x i +b l )

[0101] Secondly, the output weights of the training are obtained using the ELM output algorithm, and their expression is as follows:

[0102]

[0103] In the formula: β=[β l1 …β li ] T To output the weights; The inverse of the characteristic matrix; T = [t1, t2, ..., t] n ] T These are the labels for the samples.

[0104] Step 4: Train the initial input weights and thresholds using the training dataset, and then train the output weights. Test the dataset with the following output expression:

[0105] Hβ=O

[0106] In the formula: O=[O1…O j [] represents the output label.

[0107] Tests showed that the bearing fault diagnosis method of this invention, KELM, ELM, SVM and ANN had test accuracies of 99.25%, 67.83%, 36.77%, 75.00% and 63.50%, respectively, and the total training and testing time of this invention was less than 2 seconds. Figure 3 The diagram illustrates the data features in the original data (a) and hidden layer 2 (b). As can be seen from the diagram, the data features in the original data are not only scattered spatially but also exhibit significant overlap. In hidden layer 2, the features of various data types are clustered together, with a certain distance between them, and only a very small number of features overlap. To more accurately observe the diagnostic results and classification errors of this invention, a confusion matrix is ​​used to describe the diagnostic results. Figure 4 As shown. From Figure 4 It can be seen that the fault diagnosis method of the present invention achieves 100% diagnostic effect for rolling element faults, inner ring faults and outer ring faults @3:00, with diagnostic effect of 99.5% for outer ring faults @6:00 and outer ring faults @12:00, and diagnostic effect of 99% for normal bearings.

[0108] The specific embodiments of the present invention have been described in detail above, but these are merely one example, and the present invention is not limited to the specific embodiments described above. For those skilled in the art, any equivalent modifications and substitutions to the present invention are also within the scope of the present invention. Therefore, all equivalent changes and modifications made without departing from the spirit and scope of the present invention should be covered within the scope of the present invention.

Claims

1. A bearing fault diagnosis method based on multi-information fusion, characterized in that, Includes the following steps: Step 1: Collect the time vibration signal of the bearing. Calculate the number of sample points collected per revolution of the bearing based on the bearing speed and sampling frequency. Divide the collected time vibration signal into multiple samples according to the calculated number of sample points. Divide the samples into training dataset and test dataset according to a certain ratio and add corresponding category labels. Step 2: Construct a signal feature extraction layer that includes extracted data features, impulse features, time-domain features, and frequency-domain features. The data features extracted by the signal feature extraction layer include mean, standard deviation, mean absolute error, and bias, as detailed below: A. Average value—shows the average level across samples: In the formula: The number of data points contained in a sample is called the number of sample points. For the first in the sample One element; B. Standard deviation—shows the dispersion of each individual relative to the sample mean: In the formula: The average value for each sample; C. Mean Absolute Error—This shows the magnitude of the actual error in the sample: D. Skewness—Indicates the degree of asymmetry in the distribution of data in a sample: In the formula: The standard deviation of the sample; The temporal features extracted by the signal feature extraction layer include energy and difference, as detailed below: A. Energy — used to reflect the energy level of a time-domain signal: B. Differential – Used to reflect the differential characteristics of time-domain signals: Step 3: Use the ELM classifier model to randomly map the data features from the experience pool and output the results. First, use the random mapping layer of ELM to perform random mapping to obtain the feature matrix of the samples. The formula expression is as follows: In the formula: and These are the weight matrix and the threshold vector, respectively, which can be obtained through initialization. l It is the number of nodes in the hidden layer; The activation function of the random mapping layer can be expressed, but is not limited to, the following expressions: Secondly, the output weights of the training are obtained using the ELM output algorithm, and their expression is as follows: In the formula: To output the weights; It is the inverse of the characteristic matrix; These are the labels for the samples; Step 4: Train the output weights of the bearing fault diagnosis model using the training dataset, and use the trained weights on the test dataset for fault diagnosis.

2. The bearing fault diagnosis method based on multi-information fusion as described in claim 1, characterized in that: In step one, the formula for calculating the number of sample points is as follows: In the formula, The bearing's rotational speed. The sampling frequency; The original dataset is divided into multiple samples based on the number of sample points, and then the samples are further divided into training dataset and test dataset according to a certain ratio.

3. The bearing fault diagnosis method based on multi-information fusion as described in claim 1, characterized in that: In step two, the impact features extracted by the signal feature extraction layer include root mean square and kurtosis, as detailed below: A. Root mean square (RMS) – commonly used for analyzing signal noise: B. Kurtosis—a numerical signal statistic used to reflect the distribution characteristics of random variables, especially vibration and shock signals of surface damage: 。 4. The bearing fault diagnosis method based on multi-information fusion as described in claim 1, characterized in that: In step two, the frequency domain features extracted by the signal feature extraction layer include Fast Fourier Transform to reflect the frequency domain features of the sample.

5. The bearing fault diagnosis method based on multi-information fusion as described in claim 1, characterized in that: In step four, firstly, the initial input weights and thresholds, as well as the trained output weights, are obtained by training the training dataset. Secondly, the dataset is used for testing, and the output expression of the test is as follows: In the formula: For output labels.

6. A bearing fault diagnosis system based on multi-information fusion, characterized in that, It includes a data acquisition layer, an input layer, a first hidden layer, a second hidden layer, and an output layer; The data acquisition layer is used to acquire the bearing's time vibration signal; the first hidden layer performs a first mapping on the data input from the input layer, the second hidden layer performs a second mapping on the random mapping layer of ELM, and the output layer is used to output fault labels. The bearing fault diagnosis system based on multi-information fusion is used to perform the steps in the bearing fault diagnosis method based on multi-information fusion as described in any one of claims 1-5.

7. The bearing fault diagnosis system based on multi-information fusion as described in claim 6, characterized in that: The extracted signal feature extraction layers, consisting of data features, impulse features, time-domain features, and frequency-domain features, form the first hidden layer. The calculation formula for each feature is embedded into the first hidden layer as an activation function of a neural node.

Images