Gear fault diagnosis method based on local temporal self-similarity and bag-of-words model

Abstract

The present invention belongs to the field of mechanical system fault diagnosis, and is specifically a gear fault diagnosis method based on local temporal self-similarity and a bag-of-words model. The method comprises the following steps: S100~acquire acceleration signals of gears under various states, make the signals into non-overlapping equal-length samples and give state labels; S200~feature extraction based on local temporal self-similarity: use a sliding window to divide the samples into multiple subsequences of length #imgabs0#, construct an LTSS matrix through the subsequences, calculate the gradient of the LTSS matrix and express it as a multi-dimensional feature vector, the sample is represented by the LTSS feature vector #imgabs1# and constitutes the initial sample feature Z; S300~sample feature expression based on the bag-of-words model: construct a feature space T using the initial sample feature Z, cluster the constructed feature space using a clustering algorithm, learn to obtain code words to generate a codebook, and all samples are re-represented by a histogram; S400~mode decision classification: use the histogram constructed in S300 as the input of the classifier for training and testing.

Description

Technical Field

[0001] The invention belongs to the field of mechanical system fault diagnosis, and in particular is a gear fault diagnosis method based on local time series self-similarity and a bag-of-words model. Background Art

[0002] The gear transmission system is the core transmission component of large and complex mechanical equipment such as wind turbines, helicopters, ships, and engineering machinery. Its operating reliability directly restricts the service life of the entire equipment. In the harsh working environment of low speed and heavy load, the tooth surface is subjected to the action of alternating loads, which causes gears to produce damage modes such as broken teeth, root cracks, local peeling of tooth surfaces, pitting, and tooth surface wear, thereby reducing the transmission accuracy, efficiency, and running stability of the system, and even inducing or accelerating other failure forms, seriously affecting the safe, efficient, and stable operation of the equipment. Therefore, it is of great significance to carry out research on fault diagnosis technology of gear transmission systems.

[0003] Fault diagnosis of gear transmission systems based on vibration analysis is one of the current research focuses, which mainly includes several steps such as signal processing, fault feature extraction, fault classification, and fault prediction. Among them, accurate and effective early fault feature recognition is the key link in fault diagnosis. During the operation of the gear, when the gear teeth with local damage mesh with other gear teeth, impact vibration will be generated. Under the effect of the gear rotation motion, the impact will repeat at a certain time interval. Therefore, periodic or quasi-periodic impact is a key indicator of local gear damage.

[0004] In engineering applications, the impact signal caused by the fault is very weak in the early stage of the fault. In addition, the vibration signal components obtained by the actual sensor are extremely complex, and the transmission path, environmental noise and non-stationary operating conditions also make it difficult to accurately extract and identify weak pulses. Existing research mainly focuses on damage feature extraction based on signal processing, using methods such as wavelet transform, singular value decomposition, empirical mode decomposition, spectral kurtosis analysis, etc. to pre-process the original signal to remove noise and enhance fault characteristics. However, improper signal processing methods may delete key fault information or even create more false information. Summary of the invention

[0005] In order to solve the above problems, the present invention provides a gear fault diagnosis method based on local time series self-similarity and bag-of-words model.

[0006] The present invention adopts the following technical scheme: a gear fault diagnosis method based on local time series self-similarity and word bag model, comprising the following steps: S100~acquire acceleration signals of gears under various states, make the signals into non-overlapping equal-length samples and give state labels; S200~feature extraction based on local time series self-similarity: use a sliding window to divide the samples into multiple lengths of The LTSS matrix is ​​constructed through the subsequence, the gradient of the LTSS matrix is ​​calculated and represented as a multidimensional feature vector, and the sample is represented by the LTSS feature vector Represent and constitute the initial sample feature Z; S300~Sample feature expression based on the bag-of-words model: Use the initial sample feature Z to construct the feature space T, use the clustering algorithm to cluster the constructed feature space, learn the coding words to generate the codebook, and all samples are re-represented by the histogram; S400~Pattern decision classification: Use the histogram constructed in S300 as the input of the classifier for training and testing.

[0007] Step S200 includes the following steps: S201~constructing the LTSS matrix through the subsequence; S202~calculating the gradient of the LTSS matrix and expressing it as a multi-dimensional feature vector; S203~samples are represented by the LTSS feature vector Represents and constitutes the initial sample feature Z.

[0008] The specific process of step S201 is: given a sample, select a length A sliding window collects the neighborhood sample signals at time point t within the window ,calculate y ( t ) to obtain the LTSS matrix.

[0009] The specific process of step S202 is: given time t LTSS matrix, calculate the gradient of LTSS matrix; divide the matrix into several sub-matrices of size 3×3, calculate the directional gradient histogram of each sub-matrix based on 8-bin; since LTSS matrix is ​​a symmetric matrix, take the upper right triangular elements and call it the upper right triangular block descriptor; finally, convert the directional gradient histogram vector of each sub-matrix of the upper right triangular block descriptor into m dimensional LTSS feature column vector .

[0010] The specific process of step S203 is: the sliding length is The window is 1, the step size is 1, the whole sample is traversed, and steps S201 and S202 are repeated; the LTSS feature vectors of the samples are merged to form the initial sample feature Z. ,in is the LTSS feature vector extracted from the jth window, and k is the number of windows.

[0011] Step S300 includes the following steps: S301 - constructing a feature space T through all initial sample features Z, S302 - clustering the feature space T using a clustering algorithm to construct a codebook, and S303 - re-representing all samples by a histogram.

[0012] The specific process of step S301 is to construct a feature space T through all initial sample features Z, and its expression is: , where s is the number of samples.

[0013] The specific process of step S302 is to perform K-means clustering on the feature space T, the codewords are defined by the centers of the clustering algorithm, the set of all codewords is called a codebook, and the size of the codebook is determined by the number of clustering centers.

[0014] The specific process of step S303 is that each LTSS feature vector corresponds to a unique codeword. By calculating the Euclidean distance from each LTSS feature vector in the sample to all codewords, the codeword with the smallest distance is found and counted on the codeword. Then, the sample can be represented by a histogram composed of the codebook, which is a histogram feature vector. The vector dimension is determined by the sliding window size, i.e. Sure.

[0015] Different Experiments were conducted with the number of cluster centers to select the combination that achieved the best classification results. In addition, the number of cluster centers was also combined with the Davies-Bouldin index. When the model was finally constructed, Select the best combination result with the cluster center.

[0016] In step S400, after all samples have completed feature extraction in S200 and S300, the samples are represented by the extracted histogram feature vectors, and all sample histogram feature vectors are input into the classifier for training and testing.

[0017] Compared with the prior art, the present invention is beneficial in that it avoids reliance on expert experience and manual feature extraction, reducing the impact of human factors; it realizes end-to-end connection from data collection to fault diagnosis, improving the efficiency of fault diagnosis; compared with other features, the complexity of LTSS features is significantly reduced, and the feature extraction speed is also greatly improved. The LTSS features combined with the bag-of-words model make the extracted features highly robust in terms of noise, have strong generalization to different classifiers, and can achieve good classification accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 is a flow chart of the method of the present invention;

[0019] Figure 2 is a signal sample used in the embodiment of the present invention;

[0020] Figure 3 It is a scatter plot of the fault classification result in the embodiment of the present invention. DETAILED DESCRIPTION

[0021] Figure 1 The figure is a flow chart of a gear fault diagnosis method based on local temporal self-similarity and bag-of-words model. The specific implementation steps of the method are as follows:

[0022] S100: Acquire acceleration signals of the gears in various states, convert the signals into non-overlapping equal-length samples (ie, divide the signals into multiple samples of equal length with no overlap between the samples) and give state labels.

[0023] S200: Use sliding window to divide the sample into multiple The LTSS matrix is ​​constructed through the subsequence, the gradient of the LTSS matrix is ​​calculated and represented as a multidimensional feature vector, and the sample is represented by the LTSS feature vector Represents and constitutes the initial sample feature Z.

[0024] S300~Expression of sample features based on the bag-of-words model; constructing a feature space T through all initial sample features Z; clustering the feature space T using a clustering algorithm to construct a codebook, and re-representing all samples by a histogram.

[0025] S400~Pattern decision classification: The histogram feature vector constructed in S300 is used as the input of the classifier for training and testing.

[0026] Specifically, step S200 includes:

[0027] S201: constructing a LTSS matrix through subsequences.

[0028] Given a sample, choose a length A sliding window collects the neighborhood sample signals at time point t within the window ,calculate y ( t ) to obtain the LTSS matrix. The calculation is as follows:

[0029]

[0030]

[0031] in express Middle i elements and j The Euclidean distance between elements, and n yes The length of , the diagonal calculates the distance between the element and itself, so all are zero, and the LTSS matrix is ​​a symmetric matrix.

[0032] S202: Calculate the LTSS matrix gradient and express it as a multi-dimensional vector.

[0033] Given time t The LTSS matrix is ​​first calculated. Then, the matrix is ​​divided into several sub-matrices of size 3×3, and the directional gradient histogram based on 8-bin is calculated for each sub-matrix. Since the LTSS matrix is ​​a symmetric matrix, the upper right triangular elements can complete the information of the matrix. For the convenience of calculation, the present invention takes the diagonal and upper right triangular sub-matrices for calculation, which are called upper right triangular block descriptors. Finally, the directional gradient histogram vector of each sub-matrix of the upper right triangular block descriptor is converted into an m-dimensional LTSS feature column vector. .

[0034] , c is the column vector of the histogram of directional gradients of the submatrix, where T is the transpose and m is the number of submatrices in the upper right triangular block descriptor.

[0035] S203: Samples are composed of LTSS feature vectors Represents and constitutes the initial sample feature Z.

[0036] The specific process is: the sliding length is The window is 1, the step size is 1, the whole sample is traversed, and steps S201 and S202 are repeated; the LTSS feature vectors of the samples are merged to form the initial sample feature Z. ,in is the LTSS feature vector extracted from the jth window, and k is the number of windows.

[0037] S300: Sample feature expression based on the bag-of-words model: construct a feature space T through all initial sample features Z, cluster the feature space T using a clustering algorithm to construct a codebook, and re-represent all samples by a histogram.

[0038] S301: Reorganize the LTSS feature vectors of all samples to form a feature space T, which is expressed as: , where s is the number of samples.

[0039] S302: Clustering the feature space using a clustering algorithm to construct a codebook.

[0040] K-means clustering is performed on the feature space T. The codewords are defined by the centers of the clustering algorithm. The set of all codewords is called a codebook, and the size of the codebook is determined by the number of clustering centers.

[0041] S302: The samples are re-represented by a histogram.

[0042] Each LTSS feature vector corresponds to a unique codeword, that is, by calculating the Euclidean distance from each LTSS feature vector in the sample to all codewords, the codeword with the smallest distance is found and counted on the codeword. Then the sample can be represented by the histogram composed of the codebook, which is the histogram feature vector. The vector dimension is determined by the sliding window size, i.e. Sure.

[0043] is a multiple of 3, different Experiments were conducted with the number of cluster centers to select the combination that achieved the best classification results. In addition, the number of cluster centers was also combined with the Davies-Bouldin index. When the model was finally constructed, Select the best combination result with the cluster center.

[0044] S400: Pattern decision classification: After all samples have completed S200 and S300 to extract features, the samples are represented by the extracted histogram feature vectors, and the histogram feature vectors of all samples are input into the classifier for training and testing.

[0045] Experimental case:

[0046] In order to verify the effectiveness of the method proposed in the present invention, a fault data set of a multi-stage planetary gear transmission system is used for verification.

[0047] The experimental data are acceleration signals collected under 8 different fault states. The health status types include: normal, single tooth surface peeling of the first-stage sun gear, partial breakage of two teeth of the first-stage sun gear, single tooth surface peeling of the second-stage sun gear, partial breakage of two teeth of the second-stage sun gear, 60% tooth root crack of the second-stage sun gear, partial breakage of a single tooth of the second-stage planetary gear, and partial breakage of a single tooth of the second-stage ring gear.

[0048] Sampling frequency: 5120Hz, speed: 700rpm, torque: 200 Nm.

[0049] Each group of samples consists of samples in 8 different health states, 32 samples for each fault state, and the length of each sample is 2048.

[0050] Table 1

[0051]

[0052] In step S200 Select 4, 7, 10, 13, and 16 respectively; in step S300, the cluster center starts from 45 and the interval is 10, and ends at 145. The combination of the cluster center and the classifier was tested, and the Davies-Bouldin index was combined to obtain the optimal is 7, the cluster center is 125, that is, the codebook size is 125. When building the model, Take 7 and the cluster center as 125.

[0053] After the feature histogram vectors of all samples are extracted, they are input into the SVM classifier for training and testing.

[0054] It should be noted that in order to increase the credibility of the experimental results, all 32 samples were used for training and testing, 31 samples were selected for training each time without duplication, and 1 sample was used for verification; this was performed 32 times in total.

[0055] Experimental results and analysis:

[0056] The following table shows the Euclidean distances between the feature histogram vectors of the selected 9 samples, where x3-1 represents the first sample of the third fault, and so on.

[0057] Table 2

[0058]

[0059] Table 3

[0060]

[0061] In this experiment, the final extracted sample features are feature histogram vectors. As can be seen from Table 2, different health states have large differences in Euclidean distance. As can be seen from the final confusion matrix in Table 3, this method has very good classification results, with a classification accuracy of 99.22%. Compared with traditional methods, this method does not require any professional knowledge when classifying faults, and is superior to many feature extraction methods that require feature comparison, which also shows that the method proposed in the present invention can be widely used.

[0062] The above description is only the preferred embodiment of the present application and is not intended to limit the present application. For those skilled in the art, the present application may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A gear fault diagnosis method based on local temporal self-similarity and bag-of-words model, characterized by: The following steps are included: S100~acquire acceleration signals of gears in various states, convert the signals into non-overlapping equal-length samples and assign state labels; S200~Feature extraction based on local time series self-similarity: Use sliding windows to divide samples into multiple The subsequence is constructed by subsequence, the gradient of LTSS matrix is ​​calculated and expressed as a multi-dimensional feature vector, and the sample is represented by LTSS feature vector Represents and constitutes the initial sample feature Z; Step S200 includes the following steps: S201~Construct LTSS matrix through subsequences; The specific process of step S201 is: given a sample, select a length A sliding window collects the neighborhood sample signals at time point t within the window ,calculate y ( t ) to obtain the LTSS matrix; S202~calculate the gradient of the LTSS matrix and express it as a multi-dimensional feature vector; The specific process of step S202 is: given time t LTSS matrix, calculate the gradient of LTSS matrix; divide the matrix into several sub-matrices of size 3×3, calculate the directional gradient histogram based on 8-bin for each sub-matrix; take the upper right triangular element, which is called the upper right triangular block descriptor; finally, convert the directional gradient histogram vector of each sub-matrix of the upper right triangular block descriptor into m dimensional LTSS feature column vector , , c is the column vector of the histogram of oriented gradients of the submatrix, where T is the transpose and m is the number of submatrices in the upper right triangular block descriptor; S203~Samples are composed of LTSS feature vectors Represents and constitutes the initial sample feature Z; The specific process of step S203 is: the sliding length is The window is 1, and the step length is 1. The entire sample is traversed and steps S201 and S202 are repeated; the LTSS feature vectors of the samples are combined to form the initial sample feature Z; ,in is the LTSS feature vector extracted from the jth window, k is the number of windows; S300~ Sample feature expression based on the bag-of-words model: construct feature space T using initial sample feature Z, cluster the constructed feature space using a clustering algorithm, learn to obtain code words to generate a codebook, and re-represent all samples by a histogram; S400~Pattern decision classification: The histogram constructed in S300 is used as the input of the classifier for training and testing.

2. The gear fault diagnosis method based on local temporal self-similarity and bag-of-words model according to claim 1 is characterized in that: The step S300 includes the following steps: S301~Construct feature space T from initial sample feature Z, and its expression is: , where s is the number of samples; S302~Using K-means clustering algorithm to cluster the feature space T and construct a codebook; S303~All samples are represented by histograms again.

3. The gear fault diagnosis method based on local temporal self-similarity and bag-of-words model according to claim 2 is characterized in that: The specific process of step S302 is to perform K-means clustering on the feature space T, the codewords are defined by the centers of the clustering algorithm, the set of all codewords is called a codebook, and the size of the codebook is determined by the number of clustering centers.

4. The gear fault diagnosis method based on local temporal self-similarity and bag-of-words model according to claim 2 is characterized in that: Each LTSS feature vector corresponds to a unique codeword. By calculating the Euclidean distance from each LTSS feature vector in the sample to all codewords, the codeword with the smallest distance is found and counted on the codeword. The sample is represented by the histogram formed by the codebook, which is the histogram feature vector. The vector dimension is determined by the sliding window size, i.e. Sure.

5. The gear fault diagnosis method based on local temporal self-similarity and bag-of-words model according to claim 1 or 4, characterized in that: Different Experiments were conducted with the number of cluster centers to select the combination that achieved the best classification results. In addition, the number of cluster centers was also combined with the Davies-Bouldin index. When the model was finally constructed, Select the best combination result with the cluster center.

6. The gear fault diagnosis method based on local temporal self-similarity and bag-of-words model according to claim 1 is characterized in that: In the step S400, after all samples have completed the feature extraction in S200 and S300, the samples are represented by the extracted histogram feature vectors, and the histogram feature vectors of all samples are input into the classifier for training and testing.

Images