CN117740381B - Bearing fault diagnosis method under low-speed heavy-load working condition - Google Patents

Abstract

The invention discloses a bearing fault diagnosis method under a low-speed heavy-load working condition, which adopts a single-component VMD (virtual machine model) to effectively avoid modal aliasing and end point effects, has better robustness, and has less noise interference on decomposition effects, thereby effectively separating out signal components containing main fault components in bearing vibration signals under the low-speed heavy-load working condition and forming a reconstruction signal; then, SAE is adopted to reconstruct the signal, the higher-order characteristics of the signal are effectively extracted through multi-layer coding and decoding, and the difference between different signal components is enlarged; then RF adopts a method based on Bagging random selection attribute, effectively reduces the correlation between trees, and simultaneously establishes a single non-pruning decision tree to achieve lower error, thereby constructing the mapping relation between the high-order characteristics and the fault class labels; and finally, combining the extracted high-order characteristic signals according to the mapping relation, thereby realizing the accurate diagnosis of bearing faults under the working condition of low speed and heavy load.

Description

Bearing fault diagnosis method under low-speed heavy-load working condition

Technical Field

The invention relates to a bearing fault diagnosis method, in particular to a bearing fault diagnosis method under a low-speed heavy-load working condition, and belongs to the technical field of bearing fault identification.

Background

In mechanical devices, rolling bearings are critical and common parts, which are extremely important for reliable and efficient operation of the rotating machine. The health of the rolling bearing directly affects the performance of the mechanical device. Under complex working conditions, the rolling bearing is easy to damage, and if the rolling bearing cannot be detected and maintained in time, the rolling bearing can possibly cause mechanical faults with larger area, so that the safe operation and the working efficiency of equipment are greatly influenced, the service life of the machine is shortened, and economic loss is caused. Therefore, research into fault diagnosis technology of rolling bearings is critical to ensure stability and performance of rotary machines.

Bearings are often used in closed mechanical devices, and analysis of bearing vibration signals is the most common method in bearing fault diagnosis. However, in the working condition of low speed and heavy load, bearing fault impact has certain energy, but the impact times in unit time are relatively less due to low-speed rotation, so that the generated fault characteristic signals are easily interfered by complex environmental noise, the fault characteristic signals are more difficult to extract compared with the conventional working condition, and finally the bearing fault condition cannot be accurately identified. Therefore, how to provide a new method can extract the fault characteristic signals of the bearing under the working condition of low speed and heavy load, so that the fault condition of the bearing can be accurately judged, and the method is one of the research directions in the industry.

Disclosure of Invention

Aiming at the problems in the prior art, the invention provides a bearing fault diagnosis method under the working condition of low speed and heavy load, which can extract the fault characteristic signals of the bearing under the working condition of low speed and heavy load, thereby accurately judging the bearing fault condition.

In order to achieve the above purpose, the technical scheme adopted by the invention is as follows: a bearing fault diagnosis method under the working condition of low speed and heavy load comprises the following specific steps:

step one, collecting mixed vibration signals under the working condition of low speed and heavy load, determining the center frequency of signal components and the number of signals based on the frequency spectrum distribution characteristics of the signals, and extracting the signal components one by adopting single component VMD (namely improved VMD, IVMD for short) (the parameter k value of the VMD is constantly set to be 1, and extracting one signal component at a time;

Step two, the internal information of the signal is extracted in the step one through SAE (namely a stacked self-encoder), high-order characteristics are fully extracted from each signal component, and the difference characteristics among different signal components are enhanced;

And thirdly, constructing a mapping relation between the high-order features and the fault class labels by adopting RF (i.e. random forest), and combining the high-order features of the signal components obtained in the second step according to the mapping relation so as to obtain the fault class labels corresponding to the current signals, thereby finally realizing the accurate diagnosis of the bearing faults under the low-speed heavy-load working condition.

Further, the first specific process of the step is as follows:

step ①, determining the decomposition times T;

Step ②, calculating coarse graining energy distribution spectrum of the signal spectrum;

Step ③, determining component initialization center frequency based on the energy distribution spectrum;

Step ④, performing a first single component VMD decomposition on the original signal based on the initialized center frequency to decompose a signal component, then removing the signal component from the original signal, and continuing to process the remaining signal as a new original signal;

step ⑤, repeat steps ① to ④ to perform next single component VMD decomposition, and repeat this until the determined decomposition number T is reached.

Further, in the second step, the SAE method is adopted to effectively extract the higher-order characteristics of the signal through multi-layer encoding and decoding, so as to enlarge the difference between different signal components, and provide a better basis for the subsequent fault type identification, and the specific steps are as follows:

Step a, coding a signal s= [ s ₁,s₂,...s_n ] through a weight W _x and a bias b _x to obtain a hidden layer vector h, wherein n is the length of a single sample;

h＝f(W_s·s+b_s) (1)

Wherein f is an activation function;

Step B, decoding the hidden layer through the weight W _x and the bias B _x to obtain an original signal;

s'＝f(W_h·h+b_h) (2)

step C, defining a reconstruction loss function;

single sample reconstruction error:

Dataset s= [ S ¹,s²,...s^m ] reconstruction error:

step D, optimizing a model by using a random gradient descent algorithm, and obtaining a required gradient by using a back propagation algorithm, and updating the weight W _x and the bias b _x;

Step E, repeating the steps A to D until the error requirement is met or the iteration times are reached;

and F, decoding and encoding the hidden layer vector h as a new signal to obtain a new hidden layer vector, and continuing until the set stacking layer number is reached, thereby completing the high-order feature extraction process of each signal component.

Further, in the third step, the RF adopts a method of randomly selecting attributes based on Bagging, so that the correlation between trees is effectively reduced, meanwhile, a single non-pruning decision tree established can reach lower error, the accuracy of RF classification is ensured, and the method comprises the following specific steps:

Step I, generating p training sets by using a Bagging method, and extracting n features from an original feature set by using a dropout method for each training set;

Step II, for each training set, generating a decision tree without pruning, and generating t decision trees according to the number p of the training sets;

Step III, classifying the samples, and determining the output class names through majority voting of trees in the forest;

wherein I is an oscillometric function, wherein c is the classification result of the tree h _i on the predicted class c, Number of nodes in the leaf; and constructing a mapping relation between the high-order features and the fault class labels through the process.

Compared with the prior art, the Improved VMD (IVMD) is adopted, so that modal aliasing and end-point effects can be effectively avoided, better robustness is realized, and the decomposition effect is less interfered by noise, so that signal components containing main fault components in the bearing vibration signal under the low-speed heavy-load working condition are effectively separated, and a reconstruction signal is formed; then SAE (namely a stacked self-encoder) is adopted for reconstruction signals, high-order features of the signals are effectively extracted through multi-layer encoding and decoding, the difference between different signal components is enlarged, and a better foundation is provided for subsequent fault type identification; then RF adopts a method based on Bagging random selection attribute, effectively reduces the correlation between trees, simultaneously establishes a single non-pruning decision tree to achieve lower error, ensures the accuracy of RF classification, and constructs the mapping relation between high-order characteristics and fault class labels; and finally, combining the extracted high-order characteristic signals according to the mapping relation, thereby realizing the accurate diagnosis of bearing faults under the working condition of low speed and heavy load.

Drawings

FIG. 1 is a graph showing the comparison of time domain signals and frequency spectrums of 5 bearing states under low-speed heavy-load conditions

FIG. 2 is a flow chart of a fault diagnosis method for a rolling bearing according to the present invention

FIG. 3 is a flow chart IVMD of the present invention

FIG. 4 is a graph showing the comparison of IVMD and VMD decomposition effects

FIG. 5 is a diagram of an SAE network architecture

FIG. 6 is a SAE procedure and result diagram

FIG. 7 is a visual representation of different types of features based on T-SNE

Fig. 8 is a voting example and classification result diagram of a test sample.

Detailed Description

The present invention will be further described below.

As shown in FIG. 1, the method is a comparison graph of time domain signals and frequency spectrums of 5 bearing states under the working condition of low speed and heavy load. From the time domain diagram, it can be seen that the 4 fault signals all have a certain impact component. But the fault signature is easily overwhelmed by background noise under low speed, heavy duty conditions. In the spectrogram, no obvious difference is shown, and the difference is mainly in the side frequency part of the low-frequency resonance frequency band. In order to solve the problem, the invention provides the following solutions:

As shown in fig. 2, the specific steps are as follows:

Step one, collecting mixed vibration signals under the working condition of low speed and heavy load, determining the central frequency of signal components and the number of signals based on the frequency spectrum distribution characteristics of the signals, adopting single-component VMDs to extract the signal components one by one, setting the k value of VMD parameters to be 1 constantly, and extracting one signal component at a time; the problem of inconsistent numbers of different signal components in batch processing is avoided, as shown in fig. 3, and the specific process is as follows:

step ①, determining the decomposition times T;

Step ②, calculating coarse graining energy distribution spectrum of the signal spectrum;

Step ③, determining component initialization center frequency based on the energy distribution spectrum;

Step ④, performing a first single component VMD decomposition on the original signal based on the initialized center frequency to decompose a signal component, then removing the signal component from the original signal, and continuing to process the remaining signal as a new original signal;

step ⑤, repeat steps ① to ④ to perform next single component VMD decomposition, and repeat this until the determined decomposition number T is reached.

As shown in fig. 4, to verify the effect of IVMD, the original signal is decomposed by the VMD at the same time, and both sideband constraint sizes are set to 1500, it can be seen that the two components extracted by the VMD are respectively distributed in the low frequency band and the high frequency band, and are not ideal signal components in the low frequency band. The low frequency signal has a higher amplitude in the frequency spectrum and the high frequency signal has a smaller amplitude in the frequency spectrum. IVMD can effectively optimize the representation of the signal spectrum distribution, avoid the problem of overhigh amplitude of individual frequency in the original signal, and finally realize the effective decomposition of the signal.

Step two, adopting an SAE method to effectively extract high-order characteristics of signals through multi-layer coding and decoding, expanding the difference between different signal components, providing a better foundation for the subsequent fault type identification, and specifically comprising the following steps:

Step a, coding a signal s= [ s ₁,s₂,...s_n ] through a weight W _x and a bias b _x to obtain a hidden layer vector h, wherein n is the length of a single sample;

h＝f(W_s·s+b_s) (1)

Wherein f is an activation function;

Step B, decoding the hidden layer through the weight W _x and the bias B _x to obtain an original signal;

s'＝f(W_h·h+b_h) (2)

step C, defining a reconstruction loss function;

single sample reconstruction error:

Dataset s= [ S ¹,s²,...s^m ] reconstruction error:

step D, optimizing a model by using a random gradient descent algorithm, and obtaining a required gradient by using a back propagation algorithm, and updating the weight W _x and the bias b _x;

Step E, repeating the steps A to D until the error requirement is met or the iteration times are reached;

and F, decoding and encoding the hidden layer vector h as a new signal to obtain a new hidden layer vector, and continuing until the set stacking layer number is reached, thereby completing the high-order feature extraction process of each signal component.

IVMD extract the signal components to form a reconstructed signal whose spectral distribution is mainly centered at the first 3000Hz. Taking the partial spectrum as an original characteristic, and extracting depth characteristics by utilizing SAE; using Sigmod as an activation function, the SAE network structure is shown in fig. 5; based on the SAE structure, carrying out normalization processing on the reconstructed signal spectrum, and then training SAE network parameters through the encoding and decoding processes; the feature extraction process and results are shown in fig. 6. From this figure, it can be seen that as the number of SAE network layers increases, the difference in characteristics of the different signals becomes increasingly apparent. By the time the third hidden layer vector is reached, there is already a significant difference in characteristics between the different signals. These differences provide good basis data for subsequent signal classification.

To verify the superiority of IVMD in combination with SAE, a comparison analysis of the resolved signal and feature extraction was performed using IVMD and TFF. The signal features are visualized with T-distributed random neighborhood embedding (T-SNE). The visualization results are shown in fig. 7. It can be seen that the method provided by the invention has better feature expression capability and can provide powerful support for subsequent pattern recognition.

Step three, adopting RF to construct a mapping relation between high-order features and fault class labels, wherein the RF adopts a Bagging random attribute selection method, so that the correlation between trees is effectively reduced, meanwhile, a single non-pruning decision tree can be built to achieve lower error, and the accuracy of RF classification is ensured, and the method comprises the following specific steps:

Step I, generating p training sets by using a Bagging method, and extracting n features from an original feature set by using a dropout method for each training set;

Step II, for each training set, generating a decision tree without pruning, and generating t decision trees according to the number p of the training sets;

Step III, classifying the samples, and determining the output class names through majority voting of trees in the forest;

wherein I is an oscillometric function, wherein c is the classification result of the tree h _i on the predicted class c, Number of nodes in the leaf; the mapping relation between the high-order features and the fault class labels is constructed through the process; and then, according to the mapping relation and combining the high-order characteristics of the signal components obtained in the step two, obtaining a fault class label corresponding to the current signal, and finally, accurately diagnosing the bearing fault under the low-speed heavy-load working condition.

The RF classification model introduces randomness, and effectively avoids the phenomenon of overfitting. In RF, the number of trees is 10 and the number of samples of predictors for splitting on each node is 2. The class 5 signals each contained 50 samples, 60% of which were used for training and 40% for testing. An example of voting and classification results for the test samples are shown in fig. 8.

The voting results for 5 signal test samples are given in fig. 8. The overall voting result has good consistency, which shows that the fault characteristics extracted based on IVMD and SAE have good difference characteristics, and the random forest constructed based on the independent decision tree has better robustness compared with a single decision tree. Through multiple tests, the comprehensive recognition rate can reach 97%. Therefore, the fault diagnosis algorithm provided by the invention can effectively monitor the health state of the bearing under the working condition of low speed and heavy load.

The foregoing is only a preferred embodiment of the invention, it being noted that: it will be apparent to those skilled in the art that various modifications and adaptations can be made without departing from the principles of the present invention, and such modifications and adaptations are intended to be comprehended within the scope of the invention.

Claims (2)

1. A bearing fault diagnosis method under the working condition of low speed and heavy load is characterized by comprising the following specific steps:

Step one, collecting mixed vibration signals under the working condition of low speed and heavy load, determining the center frequency of signal components and the number of signals based on the distribution characteristics of signal frequency spectrums, and extracting the signal components one by adopting single component VMD, wherein the specific process is as follows:

step ①, determining the decomposition times T;

Step ②, calculating coarse graining energy distribution spectrum of the signal spectrum;

Step ③, determining component initialization center frequency based on the energy distribution spectrum;

Step ④, performing a first single component VMD decomposition on the original signal based on the initialized center frequency to decompose a signal component, then removing the signal component from the original signal, and continuing to process the remaining signal as a new original signal;

Step ⑤, repeating steps ① to ④ to perform the next single-component VMD decomposition, and repeating the steps until the determined decomposition times T are reached;

Extracting internal information of the signal through the SAE mining step I, fully extracting high-order features from each signal component, and enhancing the difference characteristics among different signal components, wherein the method comprises the following specific steps of:

Step a, coding a signal s= [ s ₁,s₂,...s_n ] through a weight W _x and a bias b _x to obtain a hidden layer vector h, wherein n is the length of a single sample;

h＝f(W_s·s+b_s) (1)

Wherein f is an activation function;

Step B, decoding the hidden layer through the weight W _x and the bias B _x to obtain an original signal;

s'＝f(W_h·h+b_h) (2)

step C, defining a reconstruction loss function;

single sample reconstruction error:

Dataset s= [ S ¹,s²,...s^m ] reconstruction error:

step D, optimizing a model by using a random gradient descent algorithm, and obtaining a required gradient by using a back propagation algorithm, and updating the weight W _x and the bias b _x;

Step E, repeating the steps A to D until the error requirement is met or the iteration times are reached;

Step F, decoding and encoding the hidden layer vector h as a new signal to obtain a new hidden layer vector, and continuing until the set stacking layer number is reached, thereby completing the high-order feature extraction process of each signal component;

And thirdly, constructing a mapping relation between the high-order features and the fault class labels by adopting RF, and combining the high-order features of the signal components obtained in the second step according to the mapping relation so as to obtain the fault class labels corresponding to the current signals, thereby finally realizing the accurate diagnosis of the bearing faults under the low-speed heavy-load working condition.

2. The method for diagnosing bearing faults under the low-speed heavy-load working condition according to claim 1, wherein in the third step, RF adopts a method for randomly selecting attributes based on Bagging, so that the correlation between trees is effectively reduced, and meanwhile, a single non-pruning decision tree established can achieve lower error, thereby ensuring the accuracy of RF classification, and the method comprises the following specific steps:

Step I, generating p training sets by using a Bagging method, and extracting n features from an original feature set by using a dropout method for each training set;

Step II, for each training set, generating a decision tree without pruning, and generating t decision trees according to the number p of the training sets;

Step III, classifying the samples, and determining the output class names through majority voting of trees in the forest;

wherein I is an oscillometric function, wherein c is the classification result of the tree h _i on the predicted class c, Number of nodes in the leaf; and constructing a mapping relation between the high-order features and the fault class labels through the process.