Bearing fault classification method based on improved convolutional neural network

Abstract

The invention provides a bearing fault classification method based on an improved convolutional neural network, and the method comprises the steps: collecting a vibration signal in an operation process according to a vibration sensor, and converting an original vibration signal of a bearing into an envelope spectrum signal through processing; frequency domain features are extracted from the envelope spectrum signal, and these features are used to describe frequency domain characteristics of the signal. Then, the extracted features are subjected to data enhancement, and the data enhancement can be realized through translation, rotation and zooming methods. Secondly, optimizing by using a particle swarm algorithm and replacing a full connection layer with a random forest, and adjusting the structure of the convolutional neural network; and processing the enhanced data by using the improved convolutional neural network. And finally, obtaining a classification result of the bearing fault type.

Description

Technical Field

[0001] The present invention belongs to the technical field of bearing fault diagnosis, and particularly relates to a method for bearing fault classification based on an improved convolutional neural network. Background Technique

[0002] With the progress of modern industry, industrial equipment has become more intelligent and complex. As a supporting component in industry, the performance of bearings directly affects the reliability of equipment. Their failures may lead to huge economic losses and safety problems. Therefore, the research on reliable fault classification of bearings is of great significance.

[0003] In previous research on bearing fault diagnosis based on convolutional neural networks, since both the training set and the test set come from the same bearing test bench under the same working conditions. If the data in the test set has not appeared during the training process, for bearing data from different workbenches, the existing bearing fault classification methods are inaccurate, resulting in low accuracy of bearing detection.

[0004] For the above reasons, it is necessary to modify the structure of the convolutional neural network, for example, replacing the fully connected layer of the convolution with a random forest model. Summary of the Invention

[0005] Aiming at the deficiencies of the prior art, the present invention provides a method for bearing fault classification based on an improved convolutional neural network, which can solve the problem that the bearing detection accuracy decreases due to the change of working conditions in the prior art.

[0006] Technical Solution: The object of the present invention can be achieved by the following technical solutions: A method for bearing fault classification based on an improved convolutional neural network, comprising the following steps:

[0007] Step 1: Collect the original vibration signal of the bearing and convert it into an envelope spectrum signal through preprocessing;

[0008] Step 2: Extract frequency-domain features, construct feature vectors, and input them into the convolutional neural network after data augmentation;

[0009] Step 3: Preset a set of discrete parameter values, and within the preset set of discrete parameter values, optimize through the particle swarm optimization algorithm, and replace the fully connected layer in the convolutional neural network with a random forest model;

[0010] Step 4: Classify the bearing fault types with the convolutional neural network after the structure change.

[0011] Preferably, in Step 1, the process of converting the preprocessing into an envelope spectrum signal includes:

[0012] Step 1-1: Perform empirical mode decomposition on the collected vibration signal, and use the Hilbert transform on the decomposed signal to capture the signal envelope;

[0013] Step 1-2: Obtain the envelope spectrum of the input signal by performing a frequency transformation on the signal envelope obtained in Step 1-1.

[0014] Preferably, in Step 1-1, the empirical mode decomposition is as follows: The input signal is decomposed into a finite number of small components, namely intrinsic mode functions, which are a complete orthogonal basis for the initial segment of the signal.

[0015] Preferably, in Step 1-1, the process of using the Hilbert transform to capture the signal envelope for the decomposed signal includes:

[0016] First, obtain the corresponding analytic signal based on the obtained intrinsic mode function components

[0017] z i (t) = c i (t) + jH[c i (t)]

[0018] where c i (t) is the i-th component of the intrinsic mode function; H[c i (t)] is the Hilbert transform, and its expression is:

[0019]

[0020] Then, obtain the envelope of the initial vibration signal by calculating the absolute value of the corresponding analytic signal, and its expression is:

[0021]

[0022] Preferably, in Step 1-2, the frequency transformation has the following expression:

[0023]

[0024] The envelope spectrum A(f) shows the frequency components in the envelope.

[0025] Preferably, in Step 2, the expression for extracting the frequency domain features is:

[0026]

[0027] where f is the rotational frequency of the shaft, and n, d, D, and θ are the geometric parameters of the bearing.

[0028] Preferably, in Step 2, the data augmentation method includes:

[0029] 1. Move the envelope spectrum signal a certain distance in the up-down or left-right direction;

[0030] p′(x′,y′) = p(x + m,y + n)

[0031] Where p(x,y) is a point in a two-dimensional space, m and n respectively represent the distances moved along the x-axis and y-axis, and p′(x′,y′) is the new point after the movement.

[0032] 2. Rotate the envelope spectrum signal clockwise or counterclockwise around the origin by a certain angle;

[0033]

[0034] Where θ represents the rotation angle

[0035] 3. Scale the envelope spectrum signal by multiplying it by a scaling factor k;

[0036] When 0 < k < 1:

[0037]

[0038] The envelope spectrum signal is amplified;

[0039] When k > 1:

[0040] x′ ki = x′ ki+1 = … = x′ ki+k-1 = x i

[0041] y′ ki = y′ ki+1 = … = y′ ki+k-1 = ky i

[0042] The envelope spectrum signal is reduced.

[0043] Preferably, in step 3, the expression of the particle swarm optimization algorithm is:

[0044]

[0045] Where k represents the number of iterations, ν ij represents the velocity of the i-th particle in the j-th dimension, x ij represents the position of the i-th particle in the j-th dimension, p best is an i×j matrix, is the best position of the i-th particle in the j-th dimension, g best is a j-dimensional vector, is the global optimal position of all particles in the j-th dimension, ω is the inertia coefficient, c1 and c2 are learning factors, and r1 and r2 are random numbers generated uniformly from the range [0,1].

[0046] Preferably, in step 3, the neural network structure includes: two sets of convolutional layers and pooling layers, a flattened convolutional layer, two fully connected layers, and a softmax classifier.

[0047] The softmax layer as a classifier is a core component of the CNN classification task. Through probabilistic output and cross-entropy loss optimization, it realizes efficient multi-class decision-making.

[0048] The softmax output is:

[0049]

[0050] In the softmax layer, the process of probabilistic output includes:

[0051] Converting the raw scores output by the last layer (fully connected layer) of the neural network into a probability distribution, such that the probability value of each class is between 0 and 1, and the sum of the probabilities of all classes is 1:

[0052]

[0053] In the formula, e zi For each score z i Exponentiating it aims to amplify the influence of high scores, where K is the number of classes.

[0054] Preferably, in step 3, the random forest, which is a classifier that uses multiple decision trees to train and predict samples, has the following classification process:

[0055] First, randomly draw N samples with replacement from the original training set (containing N samples) to construct a training subset for a decision tree. Then, for each decision tree, randomly select m candidate features from all features, and repeat to construct multiple decision trees to form a "forest". Finally, each decision tree votes on the class of the sample, and finally select the class with the most votes and take the average of the classification results of all decision trees.

[0056] Beneficial effects:

[0057] Compared with the prior art, the present invention has at least the following technical effects:

[0058] 1. By using the envelope spectrum signal instead of the original vibration signal, the present invention filters out the high-frequency noise in the signal and retains the fault characteristics of the bearing.

[0059] 2. Using a random forest instead of a fully connected layer, through multiple decision trees and random feature sampling, it has the ability to resist overfitting, can automatically identify key features, and reduce the interference of redundant features. Description of the Drawings

[0060] Figure 1 This is the overall framework diagram of the present invention;

[0061] Figure 2 This is the fault diagnosis flow chart of the present invention;

[0062] Figure 3 This is the schematic diagram of the random forest model of the present invention;

[0063] Figure 4 This is the structure diagram of the improved convolutional neural network of the present invention;

[0064] Figure 5 This is the diagnostic accuracy table of the present invention from the same workbench;

[0065] Figure 6 This is the diagnostic accuracy table of the present invention from different workbenches. Detailed implementation manners

[0066] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0067] As Figure 1 shown, it is the overall framework diagram of a method for bearing fault classification based on an improved convolutional neural network provided by the present invention. First, vibration signals during operation are collected by vibration sensors, and the original vibration signals are converted into envelope spectrum signals through processing; frequency domain features are extracted from the envelope spectrum signals, and these features are used to describe the frequency domain characteristics of the signals. Then, data augmentation is performed on the extracted features, which can be achieved by methods such as translation, rotation, and scaling. Secondly, the particle swarm optimization algorithm is used to optimize and the random forest is used to replace the fully connected layer to adjust the structure of the convolutional neural network. The improved convolutional neural network is used to process the augmented data. Finally, the classification results of the bearing fault types are obtained, and the specific implementation steps are as follows:

[0068] Step 1: Collect the original vibration signals and convert them into envelope spectrum signals through processing;

[0069] According to the vibration signals collected by vibration sensors during operation, the original vibration signals are converted into envelope spectrum signals through processing, and the extraction process is as follows:

[0070] First, the corresponding analytic signal is obtained according to the obtained intrinsic mode function components

[0071] z i (t) = c i (t) + jH[c i (t)]

[0072] In the formula, c i(t) is the i-th component of the intrinsic mode function; j represents the imaginary unit, and H[c i (t)] is the Hilbert transform, and its expression is:

[0073]

[0074] where c i (t) is the i-th component of the intrinsic mode function, and τ is the integration variable in the integration operation process.

[0075] Then, the envelope of the initial vibration signal is obtained by calculating the absolute value of the corresponding analytic signal, and its expression is:

[0076]

[0077] The envelope spectrum signal can better reflect the frequency characteristics of the signal and is helpful for subsequent frequency-domain feature extraction.

[0078] Finally, the frequency spectrum of the original vibration signal after frequency transformation, and its expression is:

[0079]

[0080] Step 2: Extract frequency-domain features, construct a feature vector, and input it into the convolutional neural network after data augmentation;

[0081] Since in the frequency domain, when a bearing has a defect, a peak will appear at the corresponding fault characteristic frequency, and these fault characteristic frequencies contain a large amount of bearing state information.

[0082] The expression for extracting frequency-domain features is:

[0083]

[0084] where f is the rotational frequency of the shaft, and n, d, D, θ are the geometric parameters of the bearing.

[0085] Step 3: Optimize within the preset set of discrete parameter values through the particle swarm optimization algorithm, and replace the fully connected layer with a random forest model;

[0086] Step 4: Classify the bearing fault types using the convolutional neural network with the changed structure.

[0087] Fault diagnosis process: As Figure 2 shown is the specific process of fault diagnosis, which is described as follows: Input the original signal and set the target number of new samples to be generated. Subsequently, enter the loop generation stage: In each iteration, first randomly select an operator A from A i , and then according to A iSpecific parameter combinations are randomly generated within the corresponding preset parameter value range, and then the original signal is processed using this operator and the parameters to generate an enhanced new signal. This process is continuously repeated until the number of generated new signals reaches the preset target value, and finally the process ends and all enhanced data is output.

[0088] The methods of data enhancement include:

[0089] 1. Move the envelope spectrum signal a certain distance in the up-down or left-right direction;

[0090] p′(x′,y′) = p(x + m, y + n)

[0091] Where p(x, y) is a point in the two-dimensional space, m and n respectively represent the distances moved along the x-axis and y-axis, and p′(x′,y′) is the new point after movement.

[0092] 2. Rotate the envelope spectrum signal a certain angle clockwise or counterclockwise around the origin;

[0093]

[0094] Where θ represents the rotation angle

[0095] 3. Scale the envelope spectrum signal by multiplying it with a scaling factor k;

[0096] When 0 < k < 1:

[0097]

[0098] The envelope spectrum signal is amplified;

[0099] When k > 1:

[0100] x′ ki = x′ ki+1 = … = x′ ki+k-1 = x i

[0101] y′ ki = y′ ki+1 = … = y′ ki+k-1 = ky i

[0102] The envelope spectrum signal is reduced.

[0103] In the present invention, the particle swarm optimization algorithm and random forest are used to replace the fully connected layer to change the structure of the convolutional neural network, so as to improve the performance of the convolutional neural network in dealing with bearing test set data under different working conditions.

[0104] The expression of the particle swarm optimization algorithm is:

[0105]

[0106] In the formula, k represents the number of iterations, ν ij represents the velocity of the i-th particle in the j-th dimension, x ij represents the position of the i-th particle in the j-th dimension, p best is an i×j matrix, is the best position of the i-th particle in the j-th dimension, g best is a j-dimensional vector, is the global optimal position of all particles in the j-th dimension, ω is the inertia coefficient, c1 and c2 are learning factors, and r1 and r2 are random numbers generated from a uniform distribution within the range [0,1].

[0107] The neural network structure includes: two sets of convolutional layers and pooling layers, a flattened convolutional layer, two fully connected layers, and a softmax classifier. The structure diagram of the improved convolutional neural network is as shown in Figure 4 shown, where the fully connected layer is replaced by a random forest model.

[0108] Among them, the softmax layer as a classifier is the core component of the CNN classification task. Through probabilistic output and cross-entropy loss optimization, efficient multi-class decision-making is achieved.

[0109] The softmax output is:

[0110]

[0111] In the softmax layer, the process of probabilistic output is:

[0112] The original scores output by the last layer (fully connected layer) of the neural network are converted into a probability distribution, so that the probability values of each class are between 0 and 1, and the sum of the probabilities of all classes is 1:

[0113]

[0114] In the formula, e zi For each score z i is exponentiated, aiming to amplify the influence of high scores. K is the number of classes

[0115] The schematic diagram of the random forest model is as shown in Figure 3 shown. The random forest is a classifier that uses multiple decision trees to train and predict samples. Its classification process is:

[0116] First, randomly draw N samples with replacement from the original training set (containing N samples) to construct a training subset for a decision tree. Then, for each decision tree, randomly select m candidate features from all features, and repeat to construct multiple decision trees to form a "forest". Finally, each decision tree votes on the class of the sample, and ultimately selects the class with the most votes, and takes the average of the classification results of all decision trees.

[0117] To accurately evaluate the performance of the proposed improved convolutional neural network model, there are 5 types of bearing fault categories set in the experiment: rolling element fault (a 3-mm spalling pit on the bearing ball), inner race fault (a 2-mm crack on the inner race), outer race fault (a 2-mm crack on the outer race), compound fault (2-mm cracks on both the inner and outer races), and cage fault (the bearing cage is damaged). The diagnostic accuracies of the improved and conventional convolutional neural networks are compared respectively on the same workbench and different workbenches.

[0118] The diagnostic accuracies of the training set and the test set from the same workbench and from different workbenches are as Figure 5 and Figure 6 shown. It can be seen from the data in the figure that the bearing fault classification scheme of the present invention has higher accuracy than the existing conventional bearing fault classification methods. Moreover, when the data comes from different workbenches, the conventional fault classification methods are extremely severely affected, which seriously does not meet the actual classification requirements.

[0119] In summary, the present invention uses the envelope spectrum signal to replace the original vibration signal, filters out the high-frequency noise in the signal, and retains the fault characteristics of the bearing. And it uses the random forest to replace the fully connected layer. Through multiple decision trees and random feature sampling, it has the ability to resist overfitting, can automatically identify key features, and reduce the interference of redundant features.

[0120] Although the specific implementation manners of the present invention are described above, those skilled in the art should understand that these are only examples, and various changes or modifications can be made to these implementation manners without departing from the principles and essences of the present invention. The scope of the present invention is only defined by the appended claims.

Claims

1. A method for bearing fault classification based on an improved convolutional neural network, characterized in that It includes the following steps: Step 1: Collect the original vibration signal of the bearing and convert it into an envelope spectrum signal through preprocessing; Step 2: Extract frequency-domain features, construct feature vectors, and input them into a convolutional neural network after data augmentation; Step 3: Preset a set of discrete parameter values, optimize within the preset set of discrete parameter values through the particle swarm optimization algorithm, and replace the fully connected layer in the convolutional neural network with a random forest model; Step 4: Classify the bearing fault types using the convolutional neural network with the changed fully connected layer structure.

2. The method for bearing fault classification based on an improved convolutional neural network according to claim 1, wherein, In Step 1, the preprocessing is to convert the original vibration signal of the bearing into an envelope spectrum signal, and the preprocessing process includes: Step 1-1: Decompose the collected original vibration signal through empirical mode decomposition, and use the Hilbert transform for the decomposed signal to capture the signal envelope; Step 1-2: Obtain the envelope spectrum of the input signal by performing a frequency transformation on the signal envelope obtained in Step 1-1.

3. The method for bearing fault classification based on an improved convolutional neural network according to claim 2, characterized in that The empirical mode decomposition described in Step 1-1 is: Decompose the input signal into a finite number of small components, namely intrinsic mode functions, which are the complete orthogonal basis of the initial segment of this signal.

4. A method for bearing fault classification based on an improved convolutional neural network according to claim 2, characterized in that, In Step 1-1, the process of using the Hilbert transform for the decomposed signal to capture the signal envelope includes: First, obtain the corresponding analytic signal according to the obtained intrinsic mode function components: z i z(t) = c i z(t) + jH[c i (t)] where c i (t) is the i-th component of the intrinsic mode function; j represents the imaginary unit, and H[c i (t)] is the Hilbert transform, and its expression is: where c i (t) is the i-th component of the intrinsic mode function, and τ is the integration variable in the integration operation process; Then, obtain the envelope of the original vibration signal by calculating the absolute value of the corresponding analytic signal, and its expression is:

5. The method for bearing fault classification based on an improved convolutional neural network according to claim 2, characterized in that, In Step 1-2, the frequency transformation, its expression is: The envelope spectrum A(f) shows the frequency components in the envelope.

6. A method for bearing fault classification based on an improved convolutional neural network according to claim 1, characterized in that, In Step 2, the expression for extracting the frequency-domain features is: In the formula, f is the rotational frequency of the shaft, and n, d, D, θ are the geometric parameters of the bearing; where, n is the rotational speed of the bearing, d is the inner diameter of the bearing, D is the outer diameter of the bearing, and θ is the contact angle of the bearing.

7. A method for bearing fault classification based on an improved convolutional neural network according to claim 1, characterized in that, In Step 2, the methods of data augmentation include: Step 2.1: Move the envelope spectrum signal a certain distance in the up-down or left-right direction; p′(x′,y′) = p(x + m,y + n) In the formula, p(x,y) is a point in the two-dimensional space, m and n respectively represent the distances moved along the x-axis and y-axis, and p′(x′,y′) is the new point after movement; Step 2.2: Rotate the envelope spectrum signal a certain angle clockwise or counterclockwise around the origin: In the formula, θ represents the rotation angle; Step 2.3: Scale the envelope spectrum signal by multiplying it by a scaling factor k: When 0 < k < 1: The envelope spectrum signal is enlarged; When k > 1: x′ ki = x′ ki+1 = … = x′ ki+k-1 = x i y′ ki =y′ ki+1 =…=y′ ki+k-1 =ky i The envelope spectrum signal is shrunk.

8. A method for bearing fault classification based on an improved convolutional neural network according to claim 1, characterized in that In Step 3, the expression of the particle swarm optimization algorithm is: x ij (k + 1) = x ij (k) + ν ij (k + 1) where k represents the number of iterations, ν ij represents the velocity of the i-th particle in the j-th dimension, x ij represents the position of the i-th particle in the j-th dimension, p best is an i×j matrix, is the best position of the i-th particle in the j-th dimension, g best is a j-dimensional vector, is the global optimal position of all particles in the j-th dimension, ω is the inertia coefficient, c1 and c2 are learning factors, and r1 and r2 are random numbers generated from a uniform distribution within the range [0,1].

9. A method for bearing fault classification based on an improved convolutional neural network according to claim 1, characterized in that In Step 3, the convolutional neural network structure consists of two sets of convolutional layers and pooling layers, two fully connected layers, and a softmax classifier; The softmax layer as the classifier is the core component of the CNN classification task, and through probabilistic output, it realizes efficient multi-classification decision-making: The softmax output is: In the softmax layer, the process of probabilistic output includes: Convert the original scores output by the last layer of the neural network into a probability distribution, so that the probability value of each category is between 0 and 1, and the sum of the probabilities of all categories is 1: where, e zi for each score z i is exponentiated to amplify the effect of high scores, and K is the number of classes.

10. A method for bearing fault classification based on an improved convolutional neural network according to claim 1, characterized in that, In step 3, the random forest is a classifier that uses multiple decision trees to train and predict samples. Its classification process is as follows: First, randomly and with replacement select N samples from the original training set to construct a training subset for a decision tree. Then, for each decision tree, randomly select m candidate features from all features, and repeat to construct multiple decision trees to form a "forest". Finally, each decision tree votes on the class of the sample, and ultimately select the class with the most votes, and take the average of the classification results of all decision trees.

Images