Multi-source signal processing method fusing intrinsic mode function and point symmetry mode

Abstract

The invention discloses a multi-source signal processing method fusing an intrinsic mode function and a point symmetry mode, and the method comprises the steps: obtaining the multi-source signal data of a rolling bearing, and generating a fused multi-source signal SDP image according to the multi-source signal data through employing a point symmetry mode-based fusion intrinsic mode function multi-source signal processing method; carrying out classification prediction on the SDP image of the multi-source signal by adopting a MobileViT-SCConv classification model; the method comprises the following steps: randomly dividing a multi-source data fusion image into a training set, a verification set and a test set, updating model parameters of the training set by using a MobileViT-SCConv classification model, carrying out training according to result iteration on the verification set, and then evaluating a classification prediction effect by using the test set to realize fault diagnosis of the rolling bearing. According to the invention, the fusion intrinsic mode function is added, so that the phenomenon that global features and local features are difficult to consider at the same time during multi-source signal processing is improved.

Description

Technical Field

[0001] The invention relates to the technical field of fault diagnosis of rolling bearings, which are key components in rotating mechanical devices, and in particular discloses a multi-source signal processing method integrating intrinsic mode functions and point symmetric modes. Background Art

[0002] Rolling bearings are key components widely used in mechanical equipment, and their operating status directly affects the performance and reliability of the entire mechanical system. However, due to the complex working environment and long-term operation, rolling bearings often have various faults, such as spalling, wear and fatigue cracks. These faults will not only reduce the efficiency of the equipment, but may also cause serious safety hazards.

[0003] Traditional methods for fault diagnosis of rotating machinery mainly reveal the changes of amplitude and frequency in fault signals over time by analyzing the time or frequency characteristics of monitoring signals. Traditional analysis methods include wavelet transform, variational mode decomposition, empirical mode decomposition, etc. Although these traditional methods can achieve fault feature extraction to a certain extent, they still face some problems in practical problems. For example, under working conditions such as variable speed and variable load, the monitoring signals at industrial sites often show non-stationarity and are easily interfered by noise, which makes feature extraction difficult. In contrast, the visual information in the equipment status signal can reflect the global characteristics of the fault state, thereby suppressing the influence of working condition changes. Therefore, using visual information to extract fault features is an appropriate method to more effectively identify and diagnose mechanical faults. Most researchers use time-frequency images based on Cartesian coordinates, which do not have an intuitive mapping relationship with the actual fault features.

[0004] The Symmetrized Dot Pattern (SDP) method can directly transform the original signal into a mirror-symmetrical snowflake image, which accurately reflects the signal changes through the visual difference of the graphics. However, the rolling bearing fault diagnosis method based on the SDP method mainly relies on a single data source, such as vibration signals, temperature signals, etc. A single data source is often difficult to fully reflect the complex fault mode of rolling bearings, resulting in limited accuracy and robustness of fault diagnosis. In addition, in order to achieve accurate fault diagnosis, it is necessary to extract and classify the processed signal image. Classifiers based on deep learning methods have been applied in bearing fault diagnosis. Although deep learning-based methods have shown excellent feature extraction and classification capabilities in bearing fault diagnosis, they usually face challenges in processing multi-source data. These deep learning methods often rely on a single type of data source, such as vibration signals, and cannot fully utilize the comprehensive information from different sensors or different signal types (such as torque and current).

[0005] In summary, how to consider the complementarity between different sensors under complex working conditions and unstable operating environment and extract effective fault vibration characteristics through multi-source data is a major challenge faced by rolling bearing fault diagnosis. Summary of the invention

[0006] The present invention provides a multi-source signal processing method integrating intrinsic mode functions and point symmetric modes, aiming to solve at least one of the above-mentioned defects in the prior art.

[0007] The present invention relates to a multi-source signal processing method integrating intrinsic mode functions and point symmetric modes, comprising the following steps:

[0008] Acquire multi-source signal data of rolling bearings, and generate fused multi-source signal SDP images based on the multi-source signal data by adopting a fused intrinsic mode function multi-source signal processing method based on point symmetry mode; the source signal data includes vibration, torque, current, force signals collected by multiple sensors and data acquisition description files;

[0009] The MobileViT-SCConv classification model is used to classify and predict multi-source signal SDP images;

[0010] The multi-source data fusion images are randomly divided into training set, validation set and test set. The MobileViT-SCConv classification model is used to update the model parameters of the training set. The training is iteratively performed according to the results on the validation set. Then the test set is used to evaluate the effect of classification prediction and realize rolling bearing fault diagnosis.

[0011] Further, the steps of acquiring multi-source signal data of a rolling bearing and generating a fused multi-source signal SDP image by using a fused intrinsic mode function multi-source signal processing method based on a point symmetric mode according to the multi-source signal data include:

[0012] The original signals collected by each sensor are normalized, and the normalized multi-source signal data are used to generate a fused multi-source signal SDP image through the SDP method;

[0013] Perform empirical mode decomposition on the multi-source signal SDP image and decompose it into a finite number of eigenmode functions;

[0014] The first intrinsic mode function of the intrinsic mode function is selected, the first intrinsic mode function is normalized first, and then the first intrinsic mode function containing torque and force signals is printed on the inner circle of the SDP image of multi-source data fusion respectively; the first intrinsic mode function containing vibration, first current and second current signals is printed on the outer circle of the polar coordinate image.

[0015] Furthermore, the original signal collected by each sensor is normalized, and in the step of generating a fused multi-source signal SDP image by the SDP method using the normalized multi-source signal data, the original signal is normalized by the following formula:

[0016]

[0017] in, and X k are the kth sample in the normalized dataset and the original dataset respectively, max(X) and min(X) are the maximum and minimum values ​​of the training data respectively.

[0018] Furthermore, in the step of normalizing the original signals collected by each sensor and generating a fused multi-source signal SDP image by using the SDP method, the normalized multi-source signal data X tq , X fc , X vb , X cr1 , X cr2 The fused multi-source signal SDP image is generated by the SDP method, where X tq , X fc , X vb , X cr1 , X cr2 They refer to torque, force, vibration, first current and second current sensor data respectively:

[0019]

[0020]

[0021] Where R(k) is X k The polar diameter in polar coordinates, X k is the kth sample in the original data set, θ s is the initial rotation angle, i is the time lag coefficient used for plotting, X k+i is the k+ith sample in the original data set, γ is the magnification factor, N s is the sum of the total number of sensors; ψ + (k) is the positive angle of the point in polar coordinates, ψ - (k) is the negative angle of the point in polar coordinates, n s The sensor category.

[0022] Furthermore, in the step of performing empirical mode decomposition on the multi-source signal SDP image and decomposing it into a finite number of intrinsic mode functions, the intrinsic mode function is:

[0023]

[0024] Where X(k) is the intrinsic mode function, n is the number of IMF components, and IMF n (k) is the nth eigenmode function and r(k) is the residual component.

[0025] Further, the first intrinsic mode function of the intrinsic mode function is selected, the first intrinsic mode function is first normalized, and then the first intrinsic mode function containing torque and force signals is printed on the inner circle of the SDP image of multi-source data fusion respectively; in the step of printing the first intrinsic mode function containing vibration, first current and second current signals on the outer circle of the polar coordinate image, the first intrinsic mode function includes X′ tq , X′ fc , X′ vb , X′ cr1 and X′ cr2 , where X′ tq , X′ fc , X′ vb , X′ cr1 and X′ cr2 Refers to X tq , X fc , X vb , X cr1 , X cr2 The first eigenmode function after empirical mode decomposition.

[0026] Furthermore, the steps of using the MobileViT-SCConv classification model to classify and predict the multi-source signal SDP image include:

[0027] The multi-source signal SDP image is input into the 3x3 convolutional layer for preliminary feature extraction;

[0028] Use the inverted residual block of MobileNetV2 to perform feature conversion and downsampling on the preliminary features: first, expand the channels of the preliminary features through 1x1 convolution, then use 3x3 depth-separable convolution to extract spatial features, and finally restore the number of feature map channels to the original size through 1x1 convolution;

[0029] By inputting feature X T ∈R H×W×C Apply a 3x3 convolution kernel, where X T ∈R H×W×C Represents higher-level features that are gradually extracted through multi-layer convolution and nonlinear activation functions to achieve local perception and feature extraction to obtain X L ∈R H×W×d , where X L ∈R H×W×d Represents the local structural information features through multi-layer convolution, where H and W are the height and width of the input image, respectively, and C and d are the number of channels of input and output, respectively;

[0030] Flatten the feature map and divide it into several non-overlapping image blocks X U ∈R P×N×d , then, the pixels p∈{1,...,P} at the same position in each image block are input into the Transformer encoder for encoding to obtain the output X G ∈R P×N×d ; where X G ∈R P×N×d Represents the global structural information features passed through the Transformer encoder, P = w × h, P is the size of the flattened image block; N = H × W / P, N is the number of image blocks; w is the width of the image block P, h is the height of the image block P;

[0031] X G ∈R P×N×d By superposition we get X F ∈R H×W×d , where X F ∈R H×W×d Represents X G ∈R P×N×d The structural information features obtained by superposition are X F ∈R H×W×d Reintegrate into the shape of the initial feature X through a 1x1 convolution operation and combine with the input feature X through a cat operation T ∈R H×W×C Combined, and then use a 3x3 convolution layer for feature fusion to get the output feature Y∈R H×W×C , where Y∈R H×W×C Represents local and global structural information features through multi-layer convolution and Transformer encoder;

[0032] The output feature Y∈R H×W×C The spatial refinement feature is obtained through the SRU operation in the SCConv module, and then the channel refinement feature is obtained by the CRU operation. The feature map is converted into a feature vector of fixed length through the global average pooling layer; finally, the feature vector is sent to the fully connected layer for classification.

[0033] Furthermore, the feature map is flattened and divided into several non-overlapping image blocks X U ∈R P×N×d , then, the pixels p∈{1,...,P} at the same position in each image block are input into the Transformer encoder for encoding to obtain the output X G ∈R P×N×d In the step, the output X G ∈R P×N×d Calculated by the following formula:

[0034] X G (p) = Transformer(X U (p)),1≤p≤P

[0035] Among them, X G (p) is the output information feature of the pth position after being processed by the Transformer encoder, X U (p) refers to the input feature vector at the pth position in the flattened image patch, Transformer (X U (p)) indicates that X U (p) encoding operation, p is the pixel at the same position in each image block, and P is the size of the flattened image block.

[0036] The beneficial effects achieved by the present invention are:

[0037] The present invention provides a multi-source signal processing method for fusing intrinsic mode functions and point symmetric patterns, by acquiring multi-source signal data of rolling bearings, and according to the multi-source signal data, using a multi-source signal processing method for fusing intrinsic mode functions based on point symmetric patterns to generate a fused multi-source signal SDP image; using a MobileViT-SCConv classification model to classify and predict the multi-source signal SDP image; randomly dividing the multi-source data fusion image into a training set, a validation set and a test set, using the MobileViT-SCConv classification model to update the model parameters of the training set, iteratively training according to the results on the validation set, and then using the test set to evaluate the effect of classification prediction, thereby realizing the fault diagnosis of rolling bearings. The multi-source signal processing method for fusing intrinsic mode functions and point symmetric patterns provided by the present invention improves the technical problems that a single data source cannot fully reflect complex fault modes, thereby causing missed diagnosis, misdiagnosis and the like, and effectively reduces the influence of a single data source on fault diagnosis. Adding a fused intrinsic mode function improves the phenomenon that it is difficult for multi-source signal processing to simultaneously take into account global features and local features. By constructing a classification model based on the attention mechanism to extract the fault features of the fused image, reducing information loss through channel and spatial reconstruction units, and dynamically adjusting the attention weight, the accuracy of fault identification is improved, and the classification accuracy of fault diagnosis is further improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 It is a flow chart of a multi-source signal processing method integrating intrinsic mode functions and point symmetric modes according to the present invention;

[0039] Figure 2 This is a comparison diagram of feature visualization results in the multi-source signal processing method of the present invention that integrates the intrinsic mode function and the point symmetric mode;

[0040] Figure 3A comparison diagram of confusion matrix results in a multi-source signal processing method that integrates intrinsic mode functions and point symmetric patterns according to the present invention;

[0041] Figure 4 This is a comparison chart of prediction results in the multi-source signal processing method of the present invention that integrates the intrinsic mode function and the point symmetric mode. DETAILED DESCRIPTION

[0042] In order to better understand the above technical solution, the above technical solution will be described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0043] like Figures 1 to 4 As shown, the first embodiment of the present invention proposes a multi-source signal processing method integrating an intrinsic mode function and a point symmetric mode, comprising the following steps:

[0044] Step S100, acquiring multi-source signal data of a rolling bearing, and generating a fused multi-source signal SDP image based on the multi-source signal data by adopting a multi-source signal processing method of a fused intrinsic mode function based on a point symmetric mode; the source signal data includes vibration, torque, current, force signals collected by multiple sensors and a data acquisition description file.

[0045] Acquire multi-source signal data of rolling bearings, including vibration, torque, current, force signals and data acquisition description files collected by multiple sensors, and use the point-symmetric pattern-based fusion intrinsic mode function multi-source signal (Fusion of Intrinsic Mode Functions Multisource Signal Symmetrized Dot Pattern, IMFMSSDP) processing method proposed in this embodiment to generate a fused multi-source signal SDP (Symmetrized Dot Pattern, symmetrical dot pattern) image.

[0046] Step S200: Use the MobileViT-SCConv classification model to classify and predict the multi-source signal SDP image.

[0047] The components of the empirical mode decomposition of the multi-source signal are added to the SDP image to synthesize the SDP-EMD (Symmetrized Dot Pattern-Empirical Mode Decomposition) image, so that the global and local fault feature information of the multi-source signal can be integrated. The MobileViT-SCConv (Mobile Vision Transformer Self-Calibrated Convolution) classification model proposed in this embodiment is used to classify the multi-source signal SDP image obtained in step S100.

[0048] Step S300: randomly divide the multi-source data fusion image into a training set, a validation set, and a test set, use the MobileViT-SCConv classification model to update the model parameters of the training set, iterate the training according to the results on the validation set, and then use the test set to evaluate the effect of classification prediction to achieve rolling bearing fault diagnosis.

[0049] The multi-source data fusion image generated in step S200 is randomly divided into a training set, a validation set, and a test set. MobileViT-SCConv is used to update the model parameters of the training set. The training is iterated according to the results on the validation set. Then, the test set is used to evaluate the effect of classification prediction, thereby achieving fault diagnosis.

[0050] The multi-source data fusion images are used as training samples to classify the bearing fault categories through the improved MobileViT-SCConv classification model, thereby realizing accurate fault diagnosis of rolling bearings efficiently and accurately.

[0051] Further, see Figures 1 to 4 In the multi-source signal processing method for fusing intrinsic mode functions and point symmetric modes proposed in this embodiment, step S100 includes:

[0052] Step S110: normalize the original signal collected by each sensor, and generate a fused multi-source signal SDP image using the SDP method for the normalized multi-source signal data.

[0053] In the step of normalizing the original signal collected by each sensor and generating a fused multi-source signal SDP image by the SDP method, the original signal is normalized by the following formula:

[0054]

[0055] In formula (1), and X k are the kth sample in the normalized dataset and the original dataset respectively, max(X) and min(X) are the maximum and minimum values ​​of the training data respectively.

[0056] The normalized signal data X tq , X fc , X vb , X cr1 , X cr2 The fused multi-source signal SDP image is generated by the SDP method, namely:

[0057]

[0058]

[0059] In formulas (2) and (3), X tq , X fc , X vb , X cr1 , X cr2 Corresponding to the data collected by the torque sensor, force sensor, vibration sensor, first current sensor and second current sensor, X max and X min are the maximum and minimum values ​​of the sample time domain signal respectively, and R(k) is k The polar diameter in polar coordinates, θ s is the initial rotation angle, the time lag coefficient i = 4, the amplification factor γ = 35°, and the total number of sensors N s =5,n s It refers to the nth s Sensors, when n s =1, 2, 3, 4, 5, corresponding to vibration, torque, first current, second current and force sensor, X k It refers to the nth s The sample data of the sensor is the kth sample in the original data set.

[0060] Step S120 , performing empirical mode decomposition on the multi-source signal SDP image to decompose it into a finite number of eigenmode functions.

[0061] Perform Empirical Mode Decomposition (EMD) on the collected original signal to decompose the complex signal into a finite number of Intrinsic Mode Functions (IMFs). The Intrinsic Mode Functions are:

[0062]

[0063] In formula (4), X(k) is the intrinsic mode function, n is the number of IMF components, and IMF n (k) is the nth eigenmode function and r(k) is the residual component.

[0064] Step S130, select the first intrinsic mode function of the intrinsic mode function, normalize the first intrinsic mode function first, and then print the first intrinsic mode function containing torque and force signals on the inner circle of the SDP image of multi-source data fusion respectively; print the first intrinsic mode function containing vibration, first current and second current signals on the outer circle of the polar coordinate image.

[0065] Select the first intrinsic mode function after EMD decomposition of each multi-source data, that is, X′ tq , X′fc , X′ vb , X′ cr1 , X′ cr2 ,In order to better display the multi-source data SDP image of the fused intrinsic mode function, the first intrinsic mode function after decomposition is first normalized, and then the first intrinsic mode function containing torque and force signals is printed on the inner circle of the SDP image of multi-source data fusion, and the first intrinsic mode function containing vibration, first current and second current signals is printed on the outer circle of the polar coordinate image.

[0066] Preferably, see Figures 1 to 4 In the multi-source signal processing method for fusing intrinsic mode functions and point symmetric modes proposed in this embodiment, step S200 includes:

[0067] Step S210: Input the multi-source signal SDP image into the 3x3 convolution layer for preliminary feature extraction.

[0068] The multi-source signal SDP image generated in step S100 is input into a convolutional layer with a step size of 3x3 for processing to extract preliminary features.

[0069] Step S220, use the inverted residual block of MobileNetV2 to perform feature conversion and downsampling on the preliminary features: first, expand the channels of the preliminary features through 1x1 convolution, then use 3x3 depth-separable convolution to extract spatial features, and finally restore the number of feature map channels to the original size through 1x1 convolution.

[0070] The inverted residual block of MobileNetV2 is used to perform feature conversion and downsampling on the preliminary features extracted in step S100. The preliminary features are first expanded through 1x1 convolution, and then spatial features are extracted using 3x3 depth-separable convolution. Finally, the number of feature map channels is restored to the original size through 1x1 convolution.

[0071] Step S230: By inputting feature X T ∈R H×W×C Apply a 3x3 convolution kernel, where X T ∈R H×W×C Represents higher-level features that are gradually extracted through multi-layer convolution and nonlinear activation functions to achieve local perception and feature extraction to obtain X L ∈R H×W×d , where X L ∈R H×W×d Represents the local structural information features through multi-layer convolution, where H and W are the height and width of the input image, respectively, and C and d are the number of channels of input and output, respectively.

[0072] By inputting feature X T ∈RH×W×C Apply 3x3 convolution kernel to achieve local perception and feature extraction to obtain X L ∈R H ×W×d In this embodiment, the height of the input image H=256, the width of the input image W=256, the number of input channels C=3, and the number of output channels d=16.

[0073] Step S240: Flatten the feature map and divide it into a number of non-overlapping image blocks X U ∈R P×N×d , then, the pixels p∈{1,...,P} at the same position in each image block are input into the Transformer encoder for encoding to obtain the output X G ∈R P×N×d ; where X G ∈R P×N×d Represents the global structural information features passed through the Transformer encoder, P = w × h, P is the size of the flattened image block; N = H × W / P, N is the number of image blocks; w is the width of image block P, and h is the height of image block P.

[0074] X U is a real number matrix or tensor with dimensions of P×N×d; in this embodiment, the size of the flattened image block is P=4, w=h=2, N=H×W / P, and N is the number of image blocks. G is a real matrix or tensor of dimension P×N×d.

[0075] Output X G ∈R P×N×d Calculated by the following formula:

[0076] X G (p) = Transformer(X U (p)),1≤p≤P (5)

[0077] In formula (5), X G (p) is the output information feature of the pth position after being processed by the Transformer encoder. Transformer(X U (p)) indicates that X U (p) encoding operation, p is the pixel at the same position in each image block, and P is the size of the flattened image block.

[0078] Step S250: X G ∈R P×N×d By superposition we get X F ∈R H×W×d , where X F ∈R H×W×d Represents XG ∈R P×N×d By superimposing the structural information features obtained, X F ∈R H×W×d Reintegrate into the shape of the initial feature X through a 1x1 convolution operation and combine with the input feature X through a cat operation T ∈R H×W×C Combined, and then use a 3x3 convolution layer for feature fusion to get the output feature Y∈R H×W×C , where Y∈R H×W×C Represents local and global structural information features through multi-layer convolution and Transformer encoder.

[0079] The cat operation, full name catenate or concatenate, is a command used in Linux system to connect, display or process text files. It can connect the contents of multiple text files and display them on the terminal, or write the results into a new file.

[0080] Step S260: Output feature Y∈R H×W×C The spatial refinement feature is obtained through the SRU operation in the SCConv module, and then the channel refinement feature is obtained by the CRU operation. The feature map is converted into a feature vector of fixed length through the global average pooling layer; finally, the feature vector is sent to the fully connected layer for classification.

[0081] The SCConv (Spatial and Channel reconstruction Convolution) module can obtain more discriminative features through simple latent-space representation enhancement.

[0082] CRU (Cross Refinement Unit) is an operation that optimizes and extracts features through a series of operations. The operation of CRU mainly involves the optimization and extraction of features. Through specific algorithms and calculation processes, the input features are finely processed to obtain more accurate and useful feature representations. This operation is widely used in machine learning and deep learning, especially when processing complex data such as images and texts. CRU can help the model better understand and extract key information from the data, thereby improving the performance and accuracy of the model.

[0083] In order to prove the effectiveness of the method of the present invention, the vibration, torque, force and two current sensor data of the Paderborn data set in Germany are used to compare it with different classification methods for further description. The Paderborn rolling bearing data set in Germany has three system states, namely inner ring fault, outer ring fault and rolling element fault. The vibration and two current sensor data sets are sliced ​​according to 2048 data points, and the torque and force sensor data sets are sliced ​​according to 128 data points, and then divided into training sets and test sets. The rolling bearing data set is classified by the method of the present invention. For the data set, the original signal is first processed by the IMFMSDP (Fusion of Intrinsic Mode Functions Multisource SignalSymmetrized Dot Pattern, based on the fusion of point symmetric patterns of intrinsic mode functions multisource signals) method to generate image data of fused multisource signals, and then the MobileViT-SCConv algorithm is used to extract features for classification and diagnosis. The classification accuracy of the method of the present invention is compared with that of AlexNet, DenseNet, GoogleNet, ResNet18, CNN, MobileViT, and MobileNetV2. The method of the present invention can achieve a classification accuracy of 99.43%, which is 1.71%, 0.88%, 1.62%, 0.96%, 1.49%, 1.80%, and 1.84% higher than the above methods, respectively. This shows that the method of the present invention has good prediction accuracy and can efficiently diagnose faults. The feature visualization results are compared. Figure 2 As shown, the confusion matrix results are compared to Figure 3 As shown, the prediction results are compared with Figure 4 shown.

[0084] The multi-source signal processing method for fusing intrinsic mode functions and point symmetric patterns provided in this embodiment, compared with the prior art, obtains multi-source signal data of rolling bearings, and according to the multi-source signal data, uses a multi-source signal processing method for fusing intrinsic mode functions based on point symmetric patterns to generate a fused multi-source signal SDP image; uses the MobileViT-SCConv classification model to classify and predict the multi-source signal SDP image; randomly divides the multi-source data fusion image into a training set, a validation set, and a test set, uses the MobileViT-SCConv classification model to update the model parameters of the training set, iterates the training according to the results on the validation set, and then uses the test set to evaluate the effect of classification prediction, thereby realizing the fault diagnosis of rolling bearings. The multi-source signal processing method for fusing intrinsic mode functions and point symmetric patterns provided in this embodiment improves the problem that a single data source cannot fully reflect complex fault modes, thereby causing problems such as missed diagnosis and misdiagnosis, and effectively reduces the impact of a single data source on fault diagnosis. The addition of fused intrinsic mode functions improves the phenomenon that it is difficult for multi-source signal processing to take into account both global and local features at the same time. By constructing a classification model based on the attention mechanism to extract the fault features of the fused image, reducing information loss through channel and spatial reconstruction units, and dynamically adjusting the attention weight, the accuracy of fault identification is improved, and the classification accuracy of fault diagnosis is further improved.

[0085] Although preferred embodiments of the present invention have been described, additional changes and modifications may be made to these embodiments by those skilled in the art once the basic inventive concepts are known. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention. Obviously, those skilled in the art may make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A multi-source signal processing method integrating intrinsic mode functions and point symmetric modes, characterized in that: The following steps are involved: Acquire multi-source signal data of a rolling bearing, and generate a fused multi-source signal SDP image based on the multi-source signal data by using a fused intrinsic mode function multi-source signal processing method based on a point symmetric mode; the source signal data includes vibration, torque, current, force signals collected by multiple sensors and a data acquisition description file; Using the MobileViT-SCConv classification model to classify and predict the multi-source signal SDP image; The multi-source data fusion image is randomly divided into a training set, a validation set and a test set. The MobileViT-SCConv classification model is used to update the model parameters of the training set. The training is iteratively performed according to the results on the validation set. Then, the test set is used to evaluate the effect of classification prediction to achieve rolling bearing fault diagnosis.

2. The multi-source signal processing method for fusing intrinsic mode functions and point symmetric modes according to claim 1, characterized in that: The step of acquiring multi-source signal data of a rolling bearing and generating a fused multi-source signal SDP image by using a fused intrinsic mode function multi-source signal processing method based on a point symmetric mode according to the multi-source signal data comprises: The original signals collected by each sensor are normalized, and the normalized multi-source signal data are used to generate a fused multi-source signal SDP image through the SDP method; The multi-source signal SDP image is subjected to empirical mode decomposition and decomposed into a finite number of eigenmode functions: The first intrinsic mode function of the intrinsic mode function is selected, the first intrinsic mode function is normalized first, and then the first intrinsic mode function containing torque and force signals is printed on the inner circle of the SDP image of multi-source data fusion respectively; the first intrinsic mode function containing vibration, first current and second current signals is printed on the outer circle of the polar coordinate image.

3. The multi-source signal processing method for fusing intrinsic mode functions and point symmetric modes as claimed in claim 2, characterized in that: In the step of normalizing the original signal collected by each sensor and generating a fused multi-source signal SDP image by the SDP method using the normalized multi-source signal data, the original signal is normalized by the following formula: in, and X k are the kth sample in the normalized dataset and the original dataset respectively, max(X) and min(X) are the maximum and minimum values ​​of the training data respectively.

4. The multi-source signal processing method for fusing intrinsic mode functions and point symmetric patterns as claimed in claim 3, characterized in that: In the step of normalizing the original signals collected by each sensor and generating a fused multi-source signal SDP image using the SDP method, the normalized multi-source signal data X tq , X fc , X vb , X cr1 , X cr2 , the fused multi-source signal SDP image is generated by the SDP method, where X tq , X fc , X vb , X cr1 , X cr2 They refer to torque, force, vibration, first current and second current sensor data respectively: Where R(k) is X k The polar diameter in polar coordinates, X k is the kth sample in the original data set, θ s is the initial rotation angle, i is the time lag coefficient used for plotting, X k+i is the k+ith sample in the original data set, γ is the magnification factor, N s is the sum of the total number of sensors; ψ + (k) is the positive angle of the point in polar coordinates, ψ - (k) is the negative angle of the point in polar coordinates, n s The sensor category.

5. The multi-source signal processing method for fusing intrinsic mode functions and point symmetric modes as claimed in claim 4, characterized in that: In the step of performing empirical mode decomposition on the multi-source signal SDP image to decompose it into a finite number of intrinsic mode functions, the intrinsic mode function is: Where X(k) is the intrinsic mode function, n is the number of IMF components, and IMF n (k) is the nth eigenmode function and r(k) is the residual component.

6. The multi-source signal processing method for fusing intrinsic mode functions and point symmetric modes as claimed in claim 5, characterized in that: In the step of selecting the first intrinsic mode function of the intrinsic mode function, first normalizing the first intrinsic mode function, and then printing the first intrinsic mode function containing torque and force signals on the inner circle of the SDP image of multi-source data fusion respectively; in the step of printing the first intrinsic mode function containing vibration, first current and second current signals on the outer circle of the polar coordinate image, the first intrinsic mode function includes X′ tq , X′ fc , X′ vb , X′ cr1 and X cr2 , where X′ tq , X′ fc , X′ vb , X′ cr1 and X′ cr2 Refers to X tq , X fc , X vb , X cr1 , X cr2 The first eigenmode function after empirical mode decomposition.

7. The multi-source signal processing method for fusing intrinsic mode functions and point symmetric modes according to claim 1, characterized in that: The step of using the MobileViT-SCConv classification model to classify and predict the multi-source signal SDP image includes: Input the multi-source signal SDP image into a 3x3 convolutional layer for preliminary feature extraction; The preliminary features are converted and downsampled using the inverted residual block of MobileNetV2: first, the preliminary features are channel-expanded by 1x1 convolution, then spatial features are extracted by 3x3 depth-separable convolution, and finally, the number of feature map channels is restored to the original size by 1x1 convolution; By inputting feature X T ∈R H×W×C Apply 3x3 convolution kernel to achieve local perception and feature extraction to obtain X L ∈R H×W×d , where X T ∈R H×W×C Represents higher-level features X that are gradually extracted through multiple layers of convolution and nonlinear activation functions L ∈R H×W×d Represents the local structural information features through multi-layer convolution, H and W are the height and width of the input image, C and d are the number of input and output channels, respectively; Flatten the feature map and divide it into several non-overlapping image blocks X U ∈R P×N×d , then, the pixels p∈{1,...,P} at the same position in each image block are input into the Transformer encoder for encoding to obtain the output X G ∈R P×N×d ; Among them, X G ∈R P×N×d Represents the global structural information feature passed through the Transformer encoder, R is a set of real numbers, P = w × h, P is the size of the flattened image block; N = H × W / P, N is the number of image blocks; w is the width of the image block P, h is the height of the image block P; X G ∈R P×N×d By superposition we get X F ∈R H×W×d , where X F ∈R H×W×d Represents X G ∈R P×N×d By superimposing the structural information features obtained, X F ∈R H×W×d Reintegrate into the shape of the initial feature X through a 1x1 convolution operation and combine with the input feature X through a cat operation T ∈R H×W×C Combined, and then use a 3x3 convolution layer for feature fusion to get the output feature Y∈R H×W×C , where Y∈R H×W×C Represents local and global structural information features through multi-layer convolution and Transformer encoder; The output feature Y∈R H×W×C The spatial refinement feature is obtained through the SRU operation in the SCConv module, and then the channel refinement feature is obtained by the CRU operation. The feature map is converted into a feature vector of fixed length through the global average pooling layer; finally, the feature vector is sent to the fully connected layer for classification.

8. The multi-source signal processing method of fusing intrinsic mode functions and point symmetric modes as claimed in claim 7, characterized in that: The feature map is flattened and divided into several non-overlapping image blocks X U ∈R P ×NX d , then, the pixels p∈{1,...,P} at the same position in each image block are input into the Transformer encoder for encoding to obtain the output X G ∈R P×N×d In the step, the output X G ∈R P×N×d Calculated by the following formula: X G (p)=Transformer(X U (p)),1≤p≤P Among them, X G (p) is the output information feature of the pth position after being processed by the Transformer encoder, X U (p) refers to the input feature vector at the pth position in the flattened image patch, Transformer (X U (p)) indicates that X U (p) encoding operation, p is the pixel at the same position in each image block, and P is the size of the flattened image block.

Images