Bearing fault diagnosis method based on multi-source resonance component feature fusion CNN

Abstract

The invention relates to the technical field of bearing fault diagnosis, in particular to a bearing fault diagnosis method based on multi-source resonance component feature fusion CNN, and the method comprises the following steps: obtaining the original vibration data of multiple sensors, carrying out the noise reduction of an original signal, carrying out the resonance sparse decomposition of the vibration signal after noise reduction, and obtaining a multi-source resonance component feature fusion CNN; the method comprises the following steps: acquiring a vibration signal of each sensor to obtain a high-resonance component and a low-resonance component in the vibration signal of each sensor, simultaneously performing symmetric point mode analysis on the high-resonance component and the low-resonance component of each sensor to obtain a multi-source resonance component feature fusion image, and finally constructing a bearing fault intelligent diagnosis model by combining a convolutional neural network and taking a fault feature fusion image as input. Compared with other bearing fault diagnosis methods, the bearing fault diagnosis method has the advantages that the accuracy is improved, a higher recognition effect on the rolling bearing fault under the variable load is achieved, and meanwhile the intelligent diagnosis model training process has the high convergence speed and the high robustness.

Description

Technical Field

[0001] The present invention relates to the technical field of bearing fault diagnosis, and in particular to a bearing fault diagnosis method based on multi-source resonance component feature fusion CNN. Background Art

[0002] Bearings are an important component widely used in modern rotating machinery. Their harsh working environment often leads to failures. In actual production, more than 40% of rotating machinery failures are related to bearing failures, which greatly reduces the stability and safety of equipment operation. In addition, due to the complex on-site noise, the impact components that characterize the fault information are easily submerged by the noise, affecting the recognition accuracy. Therefore, it is of great significance to achieve accurate and rapid bearing fault diagnosis in complex environments.

[0003] In response to the non-stationary bearing vibration signals under strong noise background, some researchers use signal feature extraction to distinguish faults, such as wavelet analysis, envelope analysis, empirical mode decomposition and its variants. Although these methods can extract fault information to a certain extent, they are easily restricted by human experience and subjective factors, resulting in low diagnostic efficiency.

[0004] In recent years, as mechanical diagnosis technology continues to develop in the direction of intelligence, the automatic identification of bearing faults has been achieved by combining machine learning models such as artificial neural networks (ANN), support vector machines (SVM), and decision trees (DT) with traditional signal processing technology.

[0005] Soualhi et al. used Hilbert-Huang transform to extract fault information from vibration signals and used SVM to classify bearing faults; Han et al. proposed an improved random forest (RF) algorithm to achieve bearing fault diagnosis under small samples. However, in actual operation, the motor speed and working conditions often change, and many signal feature extraction methods cannot separate and capture fault impact features under strong noise and multi-source interference on site. At the same time, the learning depth of shallow machine learning models is insufficient. When faced with massive industrial data today, the application scope and diagnostic effect of such methods will be restricted.

[0006] The emergence of deep learning solves the problem of limited feature extraction capabilities of signal processing methods and the learning depth in shallow machine learning.

[0007] Since the original vibration signal of rolling bearing is a one-dimensional time series, some scholars use one-dimensional convolutional neural network 1D-CNN to realize end-to-end rolling bearing fault diagnosis. However, 1D-CNN can only identify the vibration components related to the fault by training different convolution kernels. The unprocessed original bearing signal features are often not obvious, and the fault impact features are easily masked by gear harmonics, rotation frequency and other components, which will increase the recognition difficulty of the model. Using appropriate signal processing methods to fully extract fault features and visualize them as input of CNN, and finally achieve accurate classification, is also a current research hotspot. Han et al. used multi-wavelet transform (MWT) to analyze the vibration signal, extracted the multi-wavelet coefficient matrix to create feature maps as input of CNN, and effectively identified the composite faults of rolling bearings. Zhang et al. used short-time Fourier transform to convert the vibration signal into a time-frequency map as the input of CNN, realizing the fault diagnosis of rolling bearings.

[0008] In addition, it is difficult to obtain comprehensive fault characteristics of the equipment by analyzing the signal measured by a single sensor, which will affect the accuracy of fault identification. The signals captured by multiple sensors contain more information about the system operation.

[0009] In the past few years, the fusion strategy of multi-sensor data information based on CNN has also been widely used to comprehensively describe the status of equipment. The synchronous correlation of multi-sensor data helps to improve the model's diagnostic accuracy and anti-noise interference ability. Rational use of multi-sensor information and achieving accurate and rapid diagnosis require special attention. Summary of the invention

[0010] The purpose of the present invention is to provide a bearing fault diagnosis method based on multi-source resonance component feature fusion CNN. The three acceleration sensors of the bearing are used as signal sources. After the original data is subjected to wavelet threshold denoising processing, high and low resonance components are obtained by resonance sparse decomposition, which effectively realizes the separation of periodic harmonic components and transient impacts, and highlights the bearing fault characteristics. Then, the high and low resonance components from multiple sensors are simultaneously fused as a whole using a symmetrical point pattern to obtain a multi-source resonance component feature fusion image. Finally, the multi-source resonance component feature fusion image is studied through CNN, and the equipment status is comprehensively studied, which effectively improves the overall classification performance. Therefore, the information recorded by multiple sensors can be fully integrated, and the representation differences of different state characteristics are enhanced, thereby improving the recognition effect and recognition accuracy, so as to solve the existing problem that it is impossible to reasonably use multi-sensor information for accurate and rapid diagnosis of bearing faults, and it is difficult to obtain comprehensive fault characteristics of the equipment by analyzing the signal measured by a single sensor, which will affect the accuracy of fault identification.

[0011] The embodiment of the present invention is implemented by the following technical solution: a bearing fault diagnosis method based on multi-source resonance component feature fusion CNN, comprising the following steps:

[0012] Get raw vibration data from multiple sensors;

[0013] Performing noise reduction processing on the original vibration data of each sensor to eliminate the interference noise in the original vibration data and obtain a vibration signal after noise reduction;

[0014] Performing resonance sparse decomposition on the noise-reduced vibration signal to obtain a high resonance component and a low resonance component of each sensor;

[0015] The high resonance component and the low resonance component of each sensor are simultaneously analyzed in symmetrical point mode to obtain a multi-source resonance component feature fusion image;

[0016] A convolutional neural network fault diagnosis model is constructed, and the convolutional neural network fault diagnosis model is trained using the multi-source resonance component feature fusion image under different fault states as input, and bearing fault diagnosis is performed based on the trained convolutional neural network fault diagnosis model.

[0017] According to a preferred embodiment, the obtaining of raw vibration data of multiple sensors includes:

[0018] Vibration data of different positions of the target bearing are collected through multiple acceleration sensors.

[0019] According to a preferred embodiment, the denoising process is performed on the original vibration data of each sensor to eliminate the interference noise in the original vibration data and obtain the vibration signal after denoising, including:

[0020] The original vibration data of each sensor is subjected to wavelet threshold noise reduction processing to remove environmental noise in the original vibration data and obtain a vibration signal after noise reduction.

[0021] According to a preferred embodiment, the resonant sparse decomposition of the noise-reduced vibration signal to obtain the high resonance component and the low resonance component of each sensor includes:

[0022] Generate the wavelet basis function library S based on high quality factor, low quality factor, redundant factor and decomposition level 1 and S 2 ;

[0023] The vibration signal is filtered through two channels using wavelet transform with adjustable quality factor to obtain the initial coefficient matrix W of the corresponding subband. 1 and W 2 , generating the initial high resonance component S 1 W 1 and the initial low resonance component S 2 W 2 , the decomposition process expression is as follows:

[0024] X′=S 1 W 1 +S 2 W 2 +Res

[0025] In the above formula, X' represents the vibration signal, W 1 and W 2 is the corresponding coefficient matrix, and Res represents the residual component after decomposition.

[0026] According to a preferred embodiment, the resonant sparse decomposition of the vibration signal to obtain the high resonance component and the low resonance component of each sensor also includes:

[0027] Based on the initial high resonance component S 1 W 1 , initial low resonance component S 2 W 2 and establishing a dissipation function by presetting weight coefficients;

[0028] The split augmented Lagrangian shrinkage algorithm is used to iterate the dissipation function to obtain the optimal coefficient matrix when the dissipation function is minimized. and Complete signal reconstruction and obtain the high resonance component in the vibration signal and low resonance components

[0029] The expression of the dissipation function is as follows:

[0030]

[0031] In the above formula, X represents the original vibration data, λ 1 represents the regularization parameter of the high resonance component, λ 2 Represents the regularization parameter for the low-resonance component.

[0032] According to a preferred embodiment, performing symmetrical point pattern analysis on high resonance components and low resonance components of each sensor to obtain a multi-source resonance component feature fusion image includes:

[0033] The vibration signal is mapped in polar coordinates in the form of scattered points through symmetrical point pattern analysis;

[0034] In the time domain waveform, X i is the amplitude at time i, X i+l is the amplitude at time i+l, and the vibration signal is transformed into a point in the polar coordinate space s[r(i),θ(i),φ(i)] through symmetric point pattern analysis;

[0035] The calculation formula of r(i) is:

[0036]

[0037] In the above formula, X min Represents the minimum amplitude of the original time domain signal, X max Represents the maximum amplitude of the original time domain signal;

[0038] The calculation formula of θ(i) is:

[0039]

[0040] In the above formula, θ(i) represents the angle of rotation of the point in the counterclockwise direction about the mirror symmetry plane, θ represents the rotation angle of the specified mirror symmetry plane, θ=360m / n, m=1,2,…,n, n is the number of mirror symmetry planes, ζ represents the gain coefficient, ζ<0;

[0041] The calculation formula of φ(i) is:

[0042]

[0043] According to a preferred embodiment, the convolutional neural network model consists of three convolutional layers, three maximum pooling layers, a random dropout layer and a fully connected layer, wherein the convolutional layer operates as follows:

[0044]

[0045] In the above formula, N represents the number of convolution kernels in the l-1th layer. represents the input of the l-1th layer, represents the jth convolution kernel of the i-th feature map in layer l, is the bias of the convolutional layer, ReLU() represents the linear rectifier unit activation function;

[0046] The fully connected layer operates as follows:

[0047] Z L =σ(W L Y L-1 +b L )

[0048] In the above formula, Z L represents the neuron group of layer L, which is the last layer of the network, and W L represents the weight of the fully connected layer, Y L-1 represents the neuron group of L-1 layer, b L is the bias of the fully connected layer.

[0049] The technical solution of the bearing fault diagnosis method based on multi-source resonance component feature fusion CNN provided in the embodiment of the present invention has at least the following advantages and beneficial effects: (1) By performing wavelet threshold denoising processing on the original vibration data of multiple sensors, the influence of complex environment can be reduced and the large amount of environmental noise contained in the original signal can be reduced; (2) By performing RSSD resonance sparse decomposition, the complex original vibration signal is decomposed into a high resonance component of continuous oscillation and a low resonance component containing transient impact, thereby achieving the separation of continuous oscillation component and transient impact component, making the fault characteristics more prominent, and being able to effectively extract the characteristics of weak bearing faults; (3) By using the SDP analysis method to clearly map the complex time series in the form of scattered points in the polar coordinate diagram, the time domain signal after RSSD decomposition can be By improving the visualization capability through graphical means, the information fusion of multiple signals can be realized in the same polar coordinate system. A single RSSDP image can fully integrate the high and low resonance components of the three sensors within a limited range, which better reflects the distinguishability of different fault characteristics and highlights the differences between different faults. (4) By constructing a CNN model, the RSSDP fusion feature map is used as input to comprehensively learn the device status, effectively improving the overall classification performance, and can be used to complete complex multi-classification tasks. (5) Compared with the single-sensor solution and the multi-sensor solution that only uses SDP to generate SDP maps, this solution improves the accuracy. At the same time, the training process has a faster convergence speed and stronger robustness. It can classify rolling bearing faults under variable loads to a certain extent and has strong robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 A flow chart of a bearing fault diagnosis method based on multi-source resonance component feature fusion CNN provided in an embodiment of the present invention;

[0051] Figure 2 A schematic diagram of the final effect of RSSDP provided by an embodiment of the present invention;

[0052] Figure 3 A schematic diagram of the training results of the SDP graph fusion of high and low resonance components of a single sensor provided in an embodiment of the present invention;

[0053] Figure 4 A schematic diagram of the training results of the multi-sensor vibration signal SDP fusion graph provided in an embodiment of the present invention;

[0054] Figure 5 A schematic diagram of the RSSDP-CNN training results provided by an embodiment of the present invention;

[0055] Figure 6 A schematic diagram of the RSSDP-CNN confusion matrix analysis provided by an embodiment of the present invention;

[0056] Figure 7 A schematic diagram of cluster analysis of the fully connected layer of RSSDP-CNN provided in an embodiment of the present invention;

[0057] Figure 8 A schematic diagram of a variable operating condition data comparison test provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0058] In order to make the purpose, 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 in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings here can be arranged and designed in various different configurations.

[0059] like Figure 1 As shown, a flow chart of a bearing fault diagnosis method based on multi-source resonance component feature fusion CNN is shown. This technical solution can fully fuse the information recorded by multiple sensors, enhance the characterization difference of different state characteristics of rolling bearings, and realize rolling bearing state identification. The method may include the following steps:

[0060] Step S101: Acquire original vibration data of multiple sensors.

[0061] Specifically, in one implementation of this embodiment, radial vibration data of the load side of the target bearing, axial vibration data of the motor side of the target bearing, and radial vibration data of the output shaft of the motor side of the target bearing are collected by an acceleration sensor.

[0062] Affected by the complex environment, the original vibration data collected by the sensor contains a large amount of environmental noise. In order to ensure the decomposition effect, this embodiment further performs step S102 to reduce the noise of the original vibration data.

[0063] Step S102: performing noise reduction processing on the original vibration data of each sensor to eliminate the interference noise in the original vibration data and obtain a vibration signal after noise reduction.

[0064] Specifically, in one implementation of this embodiment, wavelet threshold denoising is performed on the original vibration data X of each sensor to remove the environmental noise in the original vibration data and obtain a vibration signal X'. By performing wavelet threshold denoising on the original vibration data of multiple sensors, this embodiment can reduce the degree of influence by complex environments and reduce a large amount of environmental noise in the original signal.

[0065] Considering that the rolling bearing is a rotating machine, when a fault occurs, the fault characteristics often exist in the form of periodic transient impacts; but as the fault severity deepens, a large number of harmonics will be generated. If the original vibration data is directly identified, it will be difficult to determine the fault form; based on this, this embodiment further executes step S103 to decompose the vibration signal.

[0066] Step S103: performing resonance sparse decomposition (RSSD) on the vibration signal to obtain a high resonance component and a low resonance component of each sensor.

[0067] Specifically, in one implementation of this embodiment, based on the high quality factor, the low quality factor, the redundancy factor and the number of decomposition levels, a wavelet basis function library S is generated. 1 and S 2 ; Further, the vibration signal is subjected to dual-channel filtering using the adjustable quality factor wavelet transform to obtain the initial coefficient matrix W of the corresponding subband 1 and W 2 , generating the initial high resonance component S 1 W 1 and the initial low resonance component S 2 W 2 .

[0068] The expression of the decomposition process is X′=S 1 W 1 +S 2 W 2 +Res, X′ represents the vibration signal, W 1 and W 2 is the corresponding coefficient matrix, and Res represents the residual component after decomposition.

[0069] Further, based on the initial high resonance component S 1 W 1 , initial low resonance component S 2 W 2 , and the preset weight coefficients are used to establish the dissipation function, and the split augmented Lagrangian shrinkage algorithm is used to iterate the dissipation function to obtain the optimal coefficient matrix when the dissipation function is minimized and Complete the final signal reconstruction and obtain the high resonance component in the vibration signal and low resonance components The dissipation function is X represents the original vibration data, λ 1 represents the regularization parameter of the high resonance component, λ 2 Represents the regularization parameter for the low-resonance component.

[0070] This embodiment performs RSSD resonance sparse decomposition on the vibration signal, so that the complex original vibration signal can be decomposed into a high-resonance component of continuous oscillation and a low-resonance component containing transient impact, thereby realizing the separation of the continuous oscillation component and the transient impact component, making the fault characteristics more prominent. This method can effectively extract the characteristics of weak bearing faults, which helps to improve the accuracy of subsequent classification.

[0071] Considering that the symmetrical point pattern (SDP) analysis can clearly map the complex time series in the form of scattered points in the polar coordinate diagram, the time domain signal after RSSD decomposition can be visualized in a graphical manner. Based on this, this embodiment further performs step S104 to fuse the decomposed features.

[0072] Step S104: perform symmetrical point pattern analysis on the high resonance component and the low resonance component of each sensor to obtain a multi-source resonance component feature fusion image.

[0073] Specifically, in one implementation of this embodiment, the vibration signal is mapped in polar coordinates in the form of scattered points through symmetrical point pattern analysis; in the time domain waveform, X i is the amplitude at time i, X i+l is the amplitude at time i+l, and the vibration signal is converted into a point in the polar coordinate space s[r(i),θ(i),φ(i)] through symmetric point pattern analysis.

[0074] The calculation formula of r(i) is X min Represents the minimum amplitude of the original time domain signal, X max Represents the maximum amplitude of the original time domain signal; the calculation formula of θ(i) is θ(i) represents the angle of rotation of the point in the counterclockwise direction about the mirror symmetry plane, θ represents the rotation angle of the specified mirror symmetry plane, θ=360m / n, m=1,2,…,n, n is the number of mirror symmetry planes, ζ represents the gain coefficient, ζ<0; the calculation formula of φ(i) is

[0075] It should be noted that after the one-dimensional signal is transformed by SDP, a "petal-shaped" scatter plot can be obtained, and the differences between different types of signals are mainly manifested in: (1) the curvature of the scatter petals; (2) the distribution and shape characteristics of the scatter points; (3) the geometric center position. Therefore, in the same polar coordinate system, the information fusion of multiple signals can be achieved.

[0076] This embodiment uses the SDP analysis method to clearly map the complex time series in the form of scattered points in the polar coordinate diagram, so that the time domain signal after RSSD decomposition can be visualized in a graphical way. In the same polar coordinate system, information fusion of multiple signals can be achieved. A single RSSDP image can fully integrate the high and low resonance components of the three sensors within a limited range, which better reflects the distinguishability of different fault characteristics and highlights the differences between different faults.

[0077] Considering that CNN is a feedforward neural network with a deep structure, it is well-known for its powerful representation learning and information extraction capabilities. Since its development, CNN has seen many variants, but the most classic CNN structure still includes a convolutional layer, a pooling layer, and a fully connected layer. Based on this, this embodiment further performs step S105 to perform fault classification through CNN.

[0078] Step S105: construct a convolutional neural network fault diagnosis model, use the multi-source resonance component feature fusion image as input to train the convolutional neural network fault diagnosis model, and perform bearing fault diagnosis according to the trained convolutional neural network fault diagnosis model.

[0079] Specifically, in one implementation of this embodiment, the convolutional neural network model consists of three convolutional layers (C1, C3, C5), three maximum pooling layers (P2, P4, P6), a random dropout layer (FC) and a fully connected layer, wherein overfitting can be suppressed by adding a random dropout layer before the fully connected layer.

[0080] The convolutional layer operation is N represents the number of convolution kernels in the l-1th layer, represents the input of the l-1th layer, represents the jth convolution kernel of the i-th feature map in layer l, is the bias of the convolutional layer, ReLU() represents the linear rectifier unit activation function; the fully connected layer operation is Z L =σ(W L Y L-1 +b L ), Z L represents the neuron group of layer L, which is the last layer of the network, and W L represents the weight of the fully connected layer, Y L-1 represents the neuron group of L-1 layer, b L is the bias of the fully connected layer.

[0081] Furthermore, in this embodiment, for the multi-classification problem, the classifier activation function of the fully connected output layer is Softmax, and supervised training and back propagation algorithm are used for training. The gradient descent method is used to minimize the loss function Loss through multiple iterations to complete the network training.

[0082] The filter size and number of each layer in this embodiment are shown in Table 1 below:

[0083] Table 1. CNN network structure

[0084]

[0085] This embodiment constructs a CNN model, takes the RSSDP fusion feature map as input, comprehensively learns the device status, effectively improves the overall classification performance, and can be used to complete complex multi-classification tasks.

[0086] Implementation Cases

[0087] The following uses the rolling bearing fault vibration dataset (CWRU dataset) from the Bearing Data Center of Case Western Reserve University in the United States as the first dataset (Dataset1) and the dataset collected from the wind turbine transmission system fault simulation test bench of the Power Engineering Laboratory of North China Electric Power University (Baoding) as the second dataset (Dataset2) to analyze the experimental signals of rolling bearing faults under different working conditions to verify the effectiveness of this method.

[0088] The faulty rolling bearing model used in the CWRU test platform is 6205–2RS JEM SKF. The relevant parameters are shown in Table 2 below:

[0089] Table 2.6205-2RS JEM SKF bearing parameters

[0090]

[0091] In the experiment, single-point fault defects were created on the inner ring, rolling element and outer ring of the bearing by electrospark machining. The defect diameters were 0.1778mm, 0.3556mm and 0.5334mm respectively. Each fault was collected under loads of 0hp, 1hp, 2hp and 3hp (corresponding speeds were 1797r / min, 1772r / min, 1750r / min and 1730r / min respectively).

[0092] This embodiment uses the drive end bearing data with a sampling frequency of 12KHz and the original vibration data obtained by three acceleration sensors installed at the drive end (DE), fan end (FE) and base (BA) as the data source.

[0093] In order to increase the number of learning samples and consider the continuity of the time series, this embodiment resamples the original vibration data with a sliding window. A total of 10 categories of data are used in each working condition, including 1 category of normal data and 9 categories of fault data (3 fault locations × 3 fault sizes). The single sample length is set to 1024, and each category of state has a total of 500 data. The data set is randomly divided into a training set and a test set at a ratio of 4:1. The normal state is encoded as 0, and the fault location is encoded from inside to outside and the size is encoded from small to large as 1-9 respectively. The detailed information of the data set is shown in Table 3:

[0094] Table 3. Dataset composition

[0095]

[0096]

[0097] The experimental bench of Dataset 2 is driven by an electric motor and uses a magnetic powder brake to simulate different loads. The faulty rolling bearing model used is SKF61800, which is arranged at the end of the planetary gear shaft. The three acceleration sensors are located in the radial direction of the load side of the faulty shaft of the planetary gearbox, the axial direction of the motor side of the faulty shaft, and the radial direction of the output shaft on the motor side. The fault signal is collected by NI USB4432 data acquisition card with a sampling frequency of 10240Hz. In order to make the experimental device closer to the actual working conditions of the wind turbine gearbox, the speed of the drive motor is randomly controlled at 0-1500rpm to simulate the fluctuation of the actual wind turbine speed. The load controlled by the magnetic brake used is 15Nm, accompanied by certain noise and slight misalignment. The faulty rolling bearings include outer ring faults, inner ring faults, rolling element faults and composite faults (inner ring, outer ring and rolling element faults). The fault damage point is processed by electrospark machining. The width of the damage point is 0.3mm and the depth is 0.2mm. There are a total of 5 state modes for the faulty parts and normal bearings. The single sample length is set to 1024, with a total of 500 data for each state. The data set is randomly divided into training set and test set at a ratio of 4:1. The 5 state modes are sequentially encoded as 0-4.

[0098] Since it is difficult to obtain comprehensive fault characteristics of the equipment by analyzing the signal measured by a single sensor, which will affect the accuracy of fault identification, the embodiment proposes an SDP information fusion method of high and low resonance components of multiple sensors, namely RSSDP.

[0099] In order to make full use of the information fusion capability of SDP, the mirror symmetry planes are set to 30°, 90°, 150°, 210°, 270°, and 330°, and the parameter is 20°. The high and low resonance components of the three position sensors at the BA, DE, and FE ends are taken as inputs to form 6 petals and obtain the RSS DP image.

[0100] Taking the three types of faults with a fault size of 0.1778 mm, i.e. the inner ring, rolling element, and outer ring at six o'clock, as an example, three samples of each type of fault are randomly selected for demonstration. The final effect of RSSDP is as follows: Figure 2 shown.

[0101] Depend on Figure 2 It can be seen that a single RSSDP image fully integrates the high and low resonance components of the three sensors within a limited range, which further reflects the distinguishability of different fault features. However, manually identifying image features often requires a lot of pre-learning and memory, and it is difficult to identify small differences between images, and it cannot meet the needs of intelligent recognition; therefore, this embodiment proposes a CNN-based bearing vibration signal RSSDP information fusion image recognition method, and realizes rapid diagnosis of rolling bearing faults through the CNN model.

[0102] It should be noted that all experimental results of this embodiment are achieved on a PC computer on a Windows 10 64-bit system. The RSSDP conversion process is implemented by Matlab2020b, and the CNN model is built by the Tensorflow2.3 framework based on the Python language. All training processes set the learning rate to 0.0001, the batch number to 32, and the iteration to 20 rounds.

[0103] As a comparative example, a single sensor is used to detect the rolling bearing fault characteristics. The specific implementation process is as follows:

[0104] Based on Dataset 1, the SDP fusion map (double petals) of the high and low resonance components of a single sensor is used as the input of CNN to explore the impact of a single sensor. Experiments are conducted on the three position sensors at the BA end, DE end, and FE end under four working conditions of 0-3hp, for a total of 12 groups of experiments; the final accuracy and loss value of the training process are as follows: Figure 3 As shown, the overall accuracy and loss function of the test set are shown in Table 3:

[0105] Table 3. Results of the single sensor high and low component fusion SDP graph test set

[0106]

[0107] Table 3 shows that the final accuracy of the test set is 83.67%-90.42% on the BA side, 86.29%-92.24% on the DE side, and 73.49%-85.28% on the FE side.

[0108] Depend on Figure 3 It can be seen that after 20 rounds of training, the data of the three sensors have reached 94.77%-99.08% on the BA side, 92.85%-98.18% on the DE side, and 86.73%-96.9% on the FE side. The training curves of the BA and DE data are relatively stable, and have converged at 17 epochs, while the FE side has a large fluctuation. It can be seen from the curve that at the 20th epoch, the curve still shows a slight upward trend.

[0109] Based on the results of the training set and the test set, it can be inferred that this method has a certain feature extraction capability and can provide effective learning information for CNN. However, the accuracy and loss function of the test set are somewhat different from those of the training set. The recognition accuracy of the test set on the DE side is the highest, and the recognition accuracy of the test set on the FE side is the worst, which is prone to misdiagnosis.

[0110] It can be inferred that the DE end is closer to the faulty part and contains more vibration information, while the FE end is farther away and carries less vibration information; a certain degree of overfitting occurred in the training process. In order to improve the recognition effect without increasing the training data or using a more complex network, we should consider fusing the information of multiple sensors to make the data distinguishable as a whole.

[0111] As a comparative example, multi-sensor SDP fusion is used to detect rolling bearing fault characteristics. The specific implementation process is as follows:

[0112] The original data of the sensors at the BA, DE, and FE ends that have not been processed by wavelet threshold denoising and RSSD are used only with the SDP algorithm to generate the SDP graph (containing only three petals) as the input of the CNN, and the training process accuracy and loss value under the four working conditions of 0-3hp are obtained, as shown in the figure. Figure 4 shown.

[0113] from Figure 4 It can be seen that after SDP image fusion is performed directly on the raw data of the three sensors that have not been decomposed by RSSD, after 20 epochs, the final accuracy rates under the four working conditions are: 0hp-96.38%, 1hp-97.58%, 2hp-95.75%, 3hp-96.60%, respectively. The overall training effect is relatively stable, and the curve has converged around the 15th epoch.

[0114] Table 4 shows the overall classification performance of the method on the test set. The accuracy of the test set is lower than that of the training set, especially the final accuracy of the test set under 3hp only reaches 92.94%.

[0115] Table 4. Test set results of multi-sensor SDP fusion graph

[0116]

[0117] The experimental results show that although directly fusing multi-sensor data with SDP can enable CNN to achieve higher classification accuracy, the original signals of the three sensors at the BA, DE, and FE ends are not further extracted when fusing SDP, which will lead to unclear differences in fault characteristics, thus limiting the further improvement of CNN recognition accuracy.

[0118] As a comparative example, RSSDP is integrated to detect the rolling bearing fault characteristics. The specific implementation process is as follows:

[0119] The high and low resonance components of the three single sensors are fused in polar coordinates using the SDP method. Figure 2 The RSSDP graph (six petals) shown in the figure is used as the input of CNN. It is still trained under four working conditions of 0-3hp to view its training results and test results. At the same time, in order to evaluate the classification effect of this method, the confusion matrix and T-SNE cluster analysis are used to visualize the classification results of the data on the test set. Figure 5 The comparison results of accuracy and loss value under four working conditions of 0-3hp are shown. The results show that the method has good adaptability to the data of four different working conditions. After 20 training cycles, the classification accuracy of the final training set has reached more than 99%.

[0120] Compared with the comparison of single sensor and multi-sensor SDP fusion, the method of integrating RSSDP improves the accuracy. At the same time, it can be seen that the training process has a faster convergence speed and stronger robustness.

[0121] The test set performance under four working conditions is shown in Table 5:

[0122] Table 5. RSSDP-CNN test set results

[0123]

[0124] After 20 epochs, the final classification accuracy reached: 0hp-98.25%, 1hp-98.39%, 2hp-98.49%, 3hp-99.04%, all higher than the classification accuracy in Table 2 and Table 3. The average recognition rate under four different loads was 98.54%, which means that the proposed RSSDP-CNN method effectively suppresses the risk of overfitting. At the same time, it can be inferred that the SDP information fusion of high and low resonance components of multiple position sensors can ensure the benefits of data complementarity, which far exceeds the adverse effects of data redundancy.

[0125] In order to evaluate the model effect more clearly and intuitively, the confusion matrix of all classification results of the test set of 1000 images is plotted in the form of a heat map, as shown in the figure below: Figure 6 At the same time, the T-SNE method is introduced to reduce the dimension of the fully connected output layer data after the test set data is classified, generate clusters, and further visualize the results, as shown in Figure 7 shown.

[0126] comprehensive Figure 6 and Figure 7 It can be seen that the misclassification labels of the confusion matrix correspond to the stacking of color clusters in the T-SNE graph. Under the four working conditions, the probability of 10 states being misclassified is very small. Although a certain degree of misdiagnosis can still be observed, the clustering results show that the feature boundaries between different categories are obvious and the feature distribution between different categories is compact.

[0127] In general, the proposed method can fully integrate the vibration signal features and improve the overall classification performance.

[0128] Actual working condition experiment: In the actual operation of the equipment, the working conditions are often constantly changing, and the collected data is not limited to a single working condition. Since the same state characteristics of changing working conditions are somewhat different, if the fault diagnosis system has poor generalization ability, this will lead to difficulties in diagnosis when the working conditions change.

[0129] In order to explore the classification effect of the model under actual working conditions, the following Dataset2 variable working condition data is used to perform RSSDP-CNN classification similar to RSSDP fusion. It should be noted that due to the large dynamic and accidental errors of the data under variable working conditions, 10 experiments were conducted, and the final mean and standard deviation of the test set after 20 epochs were used as evaluation indicators. The final classification accuracy is shown in Table 6:

[0130] Table 6. Results of variable working condition test set

[0131]

[0132] The above results show that after 10 tests, the average accuracy of the test set reached 92.18%, and the standard deviation was 0.76%. Since the test bench contains multiple components such as gears, the frequency components are more complex, and the test bench simulates the actual operating conditions, the data stability is poor, and the difficulty of identifying faulty bearing information is correspondingly increased; but in general, the proposed method can still classify rolling bearing faults under variable loads to a certain extent, and has strong robustness.

[0133] Comparative study: In order to compare the effects of other intelligent bearing fault diagnosis methods, several common representative methods are selected to compare with this method on Dataset2:

[0134] (1) EMD-SVM method, using EMD method for signal processing, and then using SVM method for state recognition; (2) DBN method, extracting time domain features from multiple sensor vibration data and fusing them, and then inputting them into DBN network to complete classification; (3) 1D-CNN method, building a tightly structured 1D-CNN model, using original data as model input; (4) STFT-CNN method, using short-time Fourier transform to convert vibration signals into two-dimensional time-frequency graphs and inputting them into convolutional neural network; (5) RSSD-CNN method, decomposing the original vibration signal into high and low resonance components through RSSD, and inputting them into a dual-stream CNN for fusion learning to obtain classification results; (6) The RSSDP-CNN method proposed in this method, denoising the vibration data of multiple sensors, and obtaining the high and low resonance components after RSSD decomposition, using SDP method to obtain the RSSDP graph of fused information, and inputting it into CNN for recognition.

[0135] The comparative experiment was conducted on dataset2. The experimental results are as follows Figure 8 As shown, taking the average result of ten accuracy rates as an indicator, the comparative experimental results are shown in Table 7:

[0136] Table 7. Average results of variable operating condition data comparison test

[0137]

[0138] Among them, EMD-SVM is a traditional machine learning diagnosis method, but the unreasonable signal processing method and insufficient learning ability of the shallow model are the reasons for the poor accuracy; DBN uses information from multiple sensors, but feature selection is a dimensionality reduction process, which will inevitably reduce the learnable information of the model and lead to poor accuracy; 1D-CNN has the lowest diagnostic complexity, but it does not mine the original data information, which also brings difficulties to model learning; since the decomposition window of STFT is fixed, it cannot adaptively decompose the signal for actual non-steady-state signals, so STFT-CNN also has limitations; RSSD-CNN is similar to the method in this paper, but the network structure is complex and it is impossible to effectively fuse multi-sensor information in advance; in contrast, the RSSDP-CNN of this method effectively extracts and fuses the multi-sensor information of bearing faults under variable working conditions, and realizes the bearing fault diagnosis task with higher accuracy.

[0139] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A bearing fault diagnosis method based on multi-source resonance component feature fusion CNN, characterized in that: The following steps are involved: Get raw vibration data from multiple sensors; Performing noise reduction processing on the original vibration data of each sensor to eliminate the interference noise in the original vibration data and obtain a vibration signal after noise reduction; Performing resonance sparse decomposition on the noise-reduced vibration signal to obtain a high resonance component and a low resonance component of each sensor; The high resonance component and the low resonance component of each sensor are simultaneously analyzed in symmetrical point mode to obtain a multi-source resonance component feature fusion image; A convolutional neural network fault diagnosis model is constructed, and the convolutional neural network fault diagnosis model is trained using the multi-source resonance component feature fusion image under different fault states as input, and bearing fault diagnosis is performed based on the trained convolutional neural network fault diagnosis model.

2. The bearing fault diagnosis method based on multi-source resonance component feature fusion CNN as claimed in claim 1 is characterized in that: The obtaining of raw vibration data of multiple sensors includes: Vibration data of different positions of the target bearing are collected through multiple acceleration sensors.

3. The bearing fault diagnosis method based on multi-source resonance component feature fusion CNN as claimed in claim 2 is characterized in that: The denoising process is performed on the original vibration data of each sensor to eliminate the interference noise in the original vibration data and obtain the vibration signal after denoising, including: The original vibration data of each sensor is subjected to wavelet threshold noise reduction processing to remove environmental noise in the original vibration data and obtain a vibration signal after noise reduction.

4. The bearing fault diagnosis method based on multi-source resonance component feature fusion CNN as claimed in claim 3 is characterized in that: The resonant sparse decomposition of the noise-reduced vibration signal to obtain high resonance components and low resonance components of each sensor includes: Generate wavelet basis function libraries S1 and S2 based on high quality factor, low quality factor, redundant factor and decomposition level; The vibration signal is subjected to dual-channel filtering using wavelet transform with adjustable quality factor to obtain the initial coefficient matrices W1 and W2 of the corresponding subbands, and the initial high resonance component S1W1 and the initial low resonance component S2W2 are generated. The decomposition process expression is as follows: X′=S1W1+S2W2+Res In the above formula, X′ represents the vibration signal after denoising, W1 and W2 are the corresponding coefficient matrices, and Res represents the residual component after decomposition.

5. The bearing fault diagnosis method based on multi-source resonance component feature fusion CNN as claimed in claim 4 is characterized in that: The resonant sparse decomposition of the vibration signal to obtain a high resonance component and a low resonance component of each sensor also includes: Establishing a dissipation function based on the initial high resonance component S1W1, the initial low resonance component S2W2 and a preset weight coefficient; The split augmented Lagrangian shrinkage algorithm is used to iterate the dissipation function to obtain the optimal coefficient matrix when the dissipation function is minimized. and Complete signal reconstruction and obtain the high resonance component in the vibration signal and low resonance components The expression of the dissipation function is as follows: In the above formula, X represents the original vibration data, λ1 represents the regularization parameter of the high resonance component, and λ2 represents the regularization parameter of the low resonance component.

6. The bearing fault diagnosis method based on multi-source resonance component feature fusion CNN as claimed in claim 5 is characterized in that: The simultaneous symmetrical point pattern analysis of the high resonance component and the low resonance component of each sensor to obtain a multi-source resonance component feature fusion image includes: The vibration signal is mapped in polar coordinates in the form of scattered points through symmetrical point pattern analysis; In the time domain waveform, X i is the amplitude at time i, X i+l is the amplitude at time i+l, and the vibration signal is transformed into a point in the polar coordinate space s[r(i), θ(i), φ(i)] through symmetric point pattern analysis; The calculation formula of r(i) is: In the above formula, X min Represents the minimum amplitude of the original time domain signal, X max Represents the maximum amplitude of the original time domain signal; The calculation formula of θ(i) is: In the above formula, θ(i) represents the angle of rotation of the point in the counterclockwise direction about the mirror symmetry plane, θ represents the rotation angle of the specified mirror symmetry plane, θ=360m / n, m=1, 2, …, n, n is the number of mirror symmetry planes, ζ represents the gain coefficient, ζ<0; The calculation formula of φ(i) is:

7. The bearing fault diagnosis method based on multi-source resonance component feature fusion CNN according to any one of claims 1 to 6 is characterized in that: The convolutional neural network model consists of three convolutional layers, three maximum pooling layers, a random dropout layer and a fully connected layer, wherein the convolutional layer operates as follows: In the above formula, N represents the number of convolution kernels in the l-1th layer. represents the input of the l-1th layer, represents the jth convolution kernel of the i-th feature map in layer l, is the bias of the convolutional layer, ReLU() represents the linear rectifier unit activation function; The fully connected layer operates as follows: WITH L =σ(W L Y L - 1 +b L ) In the above formula, Z L represents the neuron group of layer L, where layer L is the last layer of the network, and W L represents the weight of the fully connected layer, Y L-1 represents the neuron group of L-1 layer, b L is the bias of the fully connected layer.

Images