Ball screw fault diagnosis method based on one-dimensional RegNet multi-scale feature fusion

By using the one-dimensional RegNet multi-scale feature fusion method, combined with the path aggregation network and the gray wolf optimization algorithm, the problems of large number of parameters and low recognition rate of ball screw fault diagnosis model are solved, and efficient and accurate fault diagnosis is achieved.

CN118913679AActive Publication Date: 2024-11-08NANJING GONGDA CNC TECH

Patent Information

Application Number
CN202410676445.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-05-29
Publication Date
2024-11-08
Estimated Expiration
2044-05-29

Smart Images

  • Figure CN118913679A_ABST
    Figure CN118913679A_ABST
Patent Text Reader

Abstract

The invention discloses a ball screw fault diagnosis method based on one-dimensional RegNet multi-scale feature fusion, and the method is characterized in that the method comprises the following steps: S1, collecting the one-dimensional current signal data of a ball screw under different fault conditions through an experiment, carrying out the preprocessing of the collected one-dimensional current signal data, and building a ball screw data set; s2, extracting features in four stages based on a one-dimensional RegNet network structure, and adding an efficient channel attention module in a block of each stage; s3, designing a multi-scale fusion module according to the path aggregation network, inserting a convolution attention mechanism behind four feature branches of different scales, and fusing the features of different scales; s4, searching an optimal parameter value by using a grey wolf optimization algorithm to optimize model parameters and improve model performance; and S5, performing fault diagnosis on the ball screw by using the trained one-dimensional RegNet-based multi-scale feature fusion diagnosis model, and outputting a ball screw fault diagnosis result. The embodiment of the invention aims to provide the ball screw fault diagnosis method based on one-dimensional RegNet multi-scale feature fusion, so that the problems of large parameter quantity and low fault feature recognition rate of the conventional diagnosis model can be solved, the diagnosis precision is improved, and the model is easier to converge.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of ball screw fault diagnosis of a key component of rotating mechanical equipment, and in particular to a ball screw fault diagnosis method based on one-dimensional RegNet multi-scale feature fusion. Background Art

[0002] With the ever-changing manufacturing technology, ball screw pairs are gradually developing towards the trend of green manufacturing. Ball screw pairs are key transmission components of processing equipment and are widely used in various mechanical equipment. The accuracy, sensitivity and stability of their transmission directly affect the positioning accuracy and processing accuracy of CNC machine tools. In the feed transmission system of CNC machine tools, ball screw pairs convert rotary motion into high-precision and high-efficiency linear motion. However, due to long-term high-speed operation, load impact and processing dust and other factors, wear and degradation will inevitably occur, which will lead to various faults, seriously affecting the processing accuracy and workpiece quality of machine tools, and even evolving into the failure of the entire machine tool, bringing safety hazards to the manufacturing process. Therefore, it is of great significance to conduct fault diagnosis research on ball screw pairs during their full life cycle.

[0003] Previous fault diagnosis methods have used basic networks that pursue deeper models. In order to capture more complex features, a large amount of data and computing resources are required. The RegNet selected here uses the neural architecture search technology to design the space, and provides a scalable and efficient deep learning model through structural design and parameter optimization. The core idea of ​​RegNet is to control the complexity and computational efficiency of the network through a method called "structural parameterization", which allows it to maintain performance while having fewer parameters compared to some deep and large-parameter models, which means faster training and inference, and easier model convergence.

[0004] In the field of deep learning, the size of the receptive field is directly proportional to the depth of the network, that is, the deeper the network model, the larger the receptive field of the pixels on the network. However, the characteristics of features at different levels are not only the receptive field, but also have their own different focuses. The path aggregation network adds an adaptive feature pooling layer, which helps the model integrate the information data of all feature layers into each recommended area to avoid arbitrary allocation. It also proposes a classification prediction structure that integrates the full convolutional network and the fully connected layer to improve the accuracy of classification while maintaining high resolution. It uses two paths, bottom-up and top-down, to increase the representation ability of the backbone network, which can shorten the information path and effectively improve the fusion effect of the feature pyramid. Summary of the invention

[0005] The purpose of the embodiment of the present invention is to provide a ball screw fault diagnosis method based on one-dimensional RegNet multi-scale feature fusion, which can solve the problems of large number of parameters and poor fault feature recognition rate of previous diagnosis models, improve the diagnosis accuracy, and make the model easier to converge.

[0006] In order to achieve the above object, the present invention provides the following technical solution, a ball screw fault diagnosis method based on one-dimensional RegNet multi-scale feature fusion, comprising the following steps:

[0007] Step S1, collecting one-dimensional current signal data of the ball screw under different fault conditions through experiments, preprocessing the collected one-dimensional current signal data, and establishing a ball screw data set;

[0008] Step S2, based on the one-dimensional RegNet network structure, the feature extraction is divided into four stages, and an efficient channel attention module is added to the block of each stage;

[0009] Step S3, a multi-scale fusion module is designed based on the path aggregation network, and a convolutional attention mechanism is inserted after the feature branches of four different scales to fuse the features of different scales;

[0010] Step S4, using the Gray Wolf Optimization Algorithm to search for the best parameter values ​​to optimize the model parameters and improve the model performance;

[0011] Step S5, use the trained one-dimensional RegNet-based multi-scale feature fusion diagnosis model to perform fault diagnosis on the ball screw, and output the ball screw fault diagnosis result.

[0012] In step S1, the preprocessing specifically uses a sliding window repeated sampling function with a fixed sliding window length of 784 and an overlap size of 392 to divide the data samples on the ball screw current signal, first replace the outliers on the ball screw data set, and then perform normalization processing.

[0013] In step S2, the one-dimensional RegNet network consists of a three-layer structure, wherein:

[0014] Input layer: Based on the original 1X1 standard convolution, dilated convolutions with dilation rates of 3 and 5 are added to obtain a larger receptive field while maintaining the original receptive field.

[0015] Backbone layer: It consists of four stages. The resolution r of these stages is gradually reduced. Each stage is composed of several blocks. The default step size of the first block in each stage is 2. The main branch of the structure consists of a 1×1 convolution, a 3×3 group convolution, and finally a 1×1 convolution. When the step size is 1, the shortcut branch is not processed. When the step size is 2, the shortcut branch only passes through a 1×1 convolution.

[0016] Output layer: It consists of a global average pooling layer and a fully connected layer for predicting classification.

[0017] In step S2, the attention mechanism inserts an efficient channel attention module after the shortcut branch of the block structure, allowing the model to pay more attention to the key information directly transmitted, effectively improving the feature expression ability, where each channel feature y n The weight w n Through the interaction with p parameters in the neighborhood, the process can be expressed as:

[0018]

[0019] In step S3, the multi-scale feature fusion module performs path aggregation network feature fusion on the features output by the four stages through a convolutional attention mechanism and an upsampling operation to form a comprehensive feature containing information at different levels.

[0020] In step S4, the gray wolf optimization algorithm has the ability to identify the potential optimal solution position, and the gray wolf algorithm is used to optimize the learning rate, the discard rate and the regularization parameter, wherein the gray wolf position update formula is:

[0021]

[0022] Where: t is the current iteration number; E and A are coefficient vectors; Xp(t) represents the position vector of the prey; X(t) represents the position vector of the gray wolf.

[0023] In step S5, after the multi-scale feature fusion, the global average pooling layer is used on the output feature and then the Softmax classifier is used to realize ball screw fault classification. The gradient is back-propagated according to the cross entropy loss function to update the parameters of the ball screw fault diagnosis model again to obtain the optimized ball screw fault diagnosis model. The ball screw data to be diagnosed is input into the optimized and trained ball screw classification model, and the fault diagnosis result of the ball screw is output.

[0024] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0025] The present invention provides a fault diagnosis method for a ball screw pair. Ball screw current signals under four working conditions are collected through an experimental platform, important features are extracted through a one-dimensional Regnet basic network, the extracted multi-scale features are fused through a path aggregation module, and a gradient optimization model is back-propagated through a cross entropy loss function to improve the stability and accuracy of the diagnosis.

[0026] The present invention uses convolutional attention to filter out the most relevant features from the feature branches obtained at each stage in the model, which helps to improve the efficiency and accuracy of the model. While the model is simple and easy to understand, it can handle high computational load, has better performance, and is more generalizable than other models. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 It is the overall flow chart of the ball screw fault diagnosis method based on one-dimensional RegNet multi-scale feature fusion of the present invention;

[0028] Figure 2 It is a schematic diagram of the multi-scale feature fusion module of the present invention. DETAILED DESCRIPTION

[0029] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0030] Example 1

[0031] like Figure 1 As shown, a ball screw fault diagnosis method based on one-dimensional RegNet multi-scale feature fusion includes the following steps:

[0032] Step S1, collecting one-dimensional current signal data of the ball screw under different fault conditions through experiments, preprocessing the collected one-dimensional current signal data, and establishing a ball screw data set;

[0033] Step S1.1, first increase the number of samples of the collected current signal through the sliding window segmentation enhancement technology, and perform linear normalization processing on all samples to divide them into training set and test set.

[0034] Step S2, based on the one-dimensional RegNet network structure, the feature extraction is divided into four stages, and an efficient channel attention module is added to the block of each stage;

[0035] Step S2.1, the input layer replaces the original 3*3 ordinary convolution layer with a 3*3 hole convolution layer, with hole rates of 1, 3, and 5 respectively. Then, the calculation results of the three convolution layers are spliced ​​in the feature dimension to obtain the output. The output layer consists of a global average pooling layer and a fully connected layer for prediction and classification. The backbone layer consists of four stages, each of which contains a different number of block structures. The number of blocks in the four stages is 1, 1, 4, and 7 respectively. In each stage, except for the first block, which uses a convolution with a step size of 2, the remaining blocks are convolutions with a step size of 1, where:

[0036] The main branch is a 1x1 convolution, a 3x3 group convolution, and then a 1x1 convolution. When the step size is 1, no processing is done on the shortcut branch. When the step size is 2, downsampling is performed through a 1x1 convolution.

[0037] Step S2.2, the efficient channel attention module is added after the shortcut branch in each block, which helps to identify more critical features and suppress the effects of other useless features. While reducing the computational complexity of the model, it can avoid the impact of channel dimension reduction on the accuracy of the prediction network. Step S3, then a multi-scale fusion module is designed based on the path aggregation network. The convolutional attention mechanism is inserted after the feature branches of different scales in the four stages to fuse the features of different scales through the path aggregation network;

[0038] Step S3.1, a multi-scale fusion module is designed using the path aggregation network. The specific structure is as follows: Figure 2 As shown in the figure, the features extracted by the backbone network are integrated with the top-down information and the bottom-up information to further improve the feature extraction capability and ensure the integrity of the features. A bottom-up path aggregation module is added on the basis of the top-down feature pyramid network to transmit the information of the low-level features to the high-level features. At the same time, the number of convolutional layers that the information flow from the high-level features to the low-level features is reduced in the top-down path, and a green line is added in the path to flow the bottom-level features to the N5 layer.

[0039] Step S3.2, where a convolutional attention mechanism is added before each branch is fused, which includes channel attention and spatial attention, effectively saving parameters and computing power. The output of channel attention is:

[0040] M c (F)=σ(MLP(AvgPool(F))+MLP(MaxPool(F))),

[0041] Where F is the input feature map, AvgPool and MaxPool represent global average pooling and maximum pooling operations respectively, MLP represents multi-layer perceptron, and σ represents Sigmoid activation function;

[0042] The output of spatial attention M s (F):

[0043] M s (F) = σ(f 7×7 ([AvgPool(F);MaxPool(F)])),

[0044] Among them, f 7×7Represents a 7×7 convolution operation, and [AvgPool(F); MaxPool(F)] means concatenating the average pooling and maximum pooling results along the channel axis.

[0045] Step S4, using the Gray Wolf Optimization Algorithm to optimize the learning rate, dropout rate and regularization parameters to improve model performance;

[0046] In the gray wolf algorithm, the optimization process is to calculate the location of the prey through the α wolf, β wolf, and δ wolf in each generation of the population, and continuously shorten the distance between themselves and the prey to achieve the purpose of capturing the prey. The position of the wolf pack is constantly adjusted according to the current updated position of the three wolves until the best prey position is found. The position of the prey corresponds to the optimal fitness value. The movement formula of the three wolves is as follows:

[0047]

[0048] Where D δ , D β , D δ They are the direction vectors of α wolf, β wolf and δ wolf respectively; X1, X2 and X3 are the direction vectors used to determine the direction of movement of the next generation;

[0049] In this model, the fitness function is selected as follows:

[0050]

[0051] Where test_right and train_right are the numbers of correct diagnoses in the test set and training set, respectively; test_total and train_total are the total numbers in the test set and training set, respectively.

[0052] Step S5, use the trained one-dimensional RegNet-based multi-scale feature fusion diagnosis model to perform fault diagnosis on the ball screw, and output the ball screw fault diagnosis result.

[0053] According to the classification loss function, the gradient is back-propagated to update the parameters of the ball screw fault diagnosis model again, and the optimized diagnosis model is saved after the training test is completed.

[0054] Example 2

[0055] A ball screw fault diagnosis method based on one-dimensional RegNet multi-scale feature fusion, comprising the following steps:

[0056] Step S1, collecting one-dimensional current signal data of the ball screw under different fault conditions through experiments, preprocessing the collected one-dimensional current signal data, and establishing a ball screw data set;

[0057] Step S2, based on the one-dimensional RegNet network structure, the feature extraction is divided into four stages, and an efficient channel attention module is added to the block of each stage;

[0058] Step S3, then a multi-scale fusion module is designed based on the path aggregation network, and a convolutional attention mechanism is inserted after the feature branches of four different scales to fuse the features of different scales;

[0059] Step S4, using the Gray Wolf Optimization Algorithm to search for the best parameter values ​​to optimize the model parameters and improve the model performance;

[0060] Step S5, use the trained one-dimensional RegNet-based multi-scale feature fusion diagnosis model to perform fault diagnosis on the ball screw, and output the ball screw fault diagnosis result.

[0061] According to the ball screw fault diagnosis method based on one-dimensional RegNet multi-scale feature fusion of the present invention, in step S1, the ball screw current data is measured on-site through the constructed experimental platform, and four operating current data of normal, screw misalignment fault, bearing plate bolt looseness and screw wear are obtained respectively.

[0062] According to the ball screw fault diagnosis method based on one-dimensional RegNet multi-scale feature fusion of the present invention, when the ball screw data set is preprocessed in step S1, the ball screw current data is first subjected to sliding window repeated sampling enhancement data to obtain a ball screw four-classification data set.

[0063] Furthermore, the window size of the sliding window repeated sampling function is fixed at 784. The function intercepts a fixed-size window from the original data as a subsequence each time, and then at the next sampling, the starting position moves backward by a distance of 392. If the current window has exceeded the range of the original data, the starting position is reset to 0.

[0064] Then, the outliers in the ball screw data set are first replaced by the median, and then linear normalization is performed. The principle of linear normalization is:

[0066]

[0067] Wherein, x′ represents the value of a single data, min is the minimum value of the column where the data is located, and max is the maximum value of the column where the data is located. According to the ball screw fault diagnosis method based on one-dimensional RegNet multi-scale feature fusion of the present invention, in step S2, the one-dimensional RegNet basic network consists of a three-layer structure, namely, an input layer, a backbone layer and an output layer, wherein:

[0068] The input layer replaces the original 3*3 ordinary convolution layer with a 3*3 hole convolution layer, with hole rates of 1, 3, and 5 respectively. Then, the calculation results of the three convolution layers are spliced ​​in the feature dimension to obtain the output;

[0069] The output layer consists of a global average pooling layer and a fully connected layer for predicting classification;

[0070] The backbone layer consists of four stages, each of which contains a different number of block structures. The four stages use 1, 1, 4, and 7 blocks respectively. In each stage, except for the first block, which uses a convolution with a step size of 2, the rest of the blocks are convolutions with a step size of 1, where:

[0071] The main branch is a 1x1 convolution, a 3x3 group convolution, and then a 1x1 convolution. When the step size is 1, no processing is done on the shortcut branch. When the step size is 2, downsampling is performed through a 1x1 convolution.

[0072] Group convolution: The input feature map is divided into 8 groups by channel, and then conventional convolution is performed on each group. After grouping, the number of channels of each group of input feature maps is So the number of channels of each convolution kernel is also reduced to The amount of computation and parameters of grouped convolution is reduced. The number of parameters of the entire grouped convolution is:

[0073]

[0074] Among them, C out is the number of output channels, C in is the number of input channels, g is the number of groups, K1×K2 is the size of the convolution kernel, and bias is the bias. According to the ball screw fault diagnosis method based on one-dimensional RegNet multi-scale feature fusion of the present invention, in step S2, the efficient channel attention module is added after the shortcut branch in each block, which helps to identify more critical features and suppress the effects of other useless features. While reducing the computational complexity of the model, it can avoid the impact of channel dimension reduction on the accuracy of the prediction network. The adaptive one-dimensional convolution mainly adjusts the size of the input channel C through the parameter K:

[0075]

[0076] in, represents the nearest odd number, and γ and b represent the linear function coefficients.

[0077] According to the ball screw fault diagnosis method based on one-dimensional RegNet multi-scale feature fusion of the present invention, in step S3, a multi-scale fusion module is designed using a path aggregation network, and the features extracted by the backbone network are fused from top to bottom with bottom to top, so as to further improve the feature extraction capability and ensure the integrity of the features. A convolutional attention mechanism is added before each branch is fused to help the network focus more on the key features of the input.

[0078] According to the ball screw fault diagnosis method based on one-dimensional RegNet multi-scale feature fusion of the present invention, in step S4 gray wolf algorithm, the optimization process is to calculate the location of the prey by the α wolf, β wolf, and δ wolf in each generation of population, and continuously shorten the distance between themselves and the prey to achieve the purpose of capturing prey. The position of the wolf pack is continuously adjusted according to the currently updated position of the three wolves until the best prey position is found. Step S5, after multi-scale feature fusion, the global average pooling layer is used on the output feature, and finally the Softmax classifier is used to realize the ball screw fault classification, and the gradient back propagation is used according to the classification loss function to update the parameters of the ball screw fault diagnosis model again, and the optimized diagnosis model is saved after the training test is completed.

[0079] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A ball screw fault diagnosis method based on one-dimensional RegNet multi-scale feature fusion, characterized in that: The following steps are involved: Step S1, collecting one-dimensional current signal data of the ball screw under different fault conditions through experiments, preprocessing the collected one-dimensional current signal data, and establishing a ball screw data set; Step S2, based on the one-dimensional RegNet network structure, the feature extraction is divided into four stages, and an efficient channel attention module is added to the block of each stage; Step S3, a multi-scale fusion module is designed based on the path aggregation network, and a convolutional attention mechanism is inserted after the feature branches of four different scales to fuse the features of different scales; Step S4, using the Gray Wolf Optimization Algorithm to search for the best parameter values ​​to optimize the model parameters and improve the model performance; Step S5, use the trained one-dimensional RegNet-based multi-scale feature fusion diagnosis model to perform fault diagnosis on the ball screw, and output the ball screw fault diagnosis result.

2. The ball screw fault diagnosis method based on one-dimensional RegNet multi-scale feature fusion according to claim 1 is characterized in that: In step S1, the preprocessing specifically uses a sliding window repetitive sampling function with a fixed sliding window length of 784 and an overlap size of 392 to divide the data samples on the ball screw current signal, replace the outliers of the ball screw data set and then perform normalization.

3. The ball screw fault diagnosis method based on one-dimensional RegNet multi-scale feature fusion according to claim 1 is characterized in that: In step S2, the one-dimensional RegNet network consists of a three-layer structure, wherein: Input layer: Based on the original 1X1 standard convolution, dilated convolutions with dilation rates of 3 and 5 are added to obtain a larger receptive field while maintaining the original receptive field. Backbone layer: It consists of four stages. The resolution r of these stages is gradually reduced. Each stage is composed of several blocks. The default step size of the first block in each stage is 2. The main branch of the structure consists of a 1×1 convolution, a 3×3 group convolution, and finally a 1×1 convolution. When the step size is 1, the shortcut branch is not processed. When the step size is 2, the shortcut branch only passes through a 1×1 convolution. Output layer: It consists of a global average pooling layer and a fully connected layer for predicting classification.

4. The ball screw fault diagnosis method based on one-dimensional RegNet multi-scale feature fusion according to claim 1 is characterized in that: In step S2, the attention mechanism inserts an efficient channel attention module after the shortcut branch of the block structure, allowing the model to pay more attention to the key information directly transmitted, effectively improving the feature expression ability, where each channel feature y n The weight w n Through the interaction with p parameters in the neighborhood, the process can be expressed as:

5. The ball screw fault diagnosis method based on one-dimensional RegNet multi-scale feature fusion according to claim 1 is characterized in that: In step S3, the multi-scale feature fusion module fuses the features output by the four stages through a convolutional attention mechanism and an upsampling path aggregation network to form a comprehensive feature containing information at different levels.

6. The ball screw fault diagnosis method based on one-dimensional RegNet multi-scale feature fusion according to claim 1 is characterized in that: In step S4, the gray wolf optimization algorithm has the ability to identify the potential optimal solution position, and the gray wolf algorithm is used to optimize the learning rate, the discard rate and the regularization parameter, wherein the gray wolf position update formula is: Where: t is the current iteration number; E and A are coefficient vectors; Xp(t) represents the position vector of the prey; X(t) represents the position vector of the gray wolf.

7. The ball screw fault diagnosis method based on one-dimensional RegNet multi-scale feature fusion according to claim 1 is characterized in that: In step S5, after the multi-scale feature fusion, the global average pooling layer is used on the output feature and then the Softmax classifier is used to realize ball screw fault classification. The gradient is back-propagated according to the cross entropy loss function to update the parameters of the ball screw fault diagnosis model again to obtain the optimized ball screw fault diagnosis model. The ball screw data to be diagnosed is input into the optimized and trained ball screw classification model, and the fault diagnosis result of the ball screw is output.

Citation Information

Patent Citations

  • Rolling bearing fault diagnosis method and device, electronic equipment and storage medium

    CN114840938A

  • Bearing fault diagnosis method based on 1dCNN-GWO-SVM model

    CN115392286A

  • Ball screw fault diagnosis method based on channel parallel fusion multi-attention mechanism

    CN117218508A

  • Rotating machine fault diagnosis method based on multi-sensor fusion and ECA-CNN

    CN117312948A

  • Switchable Powertrain

    US20210070280A1

Cited By

  • Wind power bearing fault diagnosis method based on dual-channel feature fusion

    CN120541716A

  • Adaptive chaotic disturbance optimization OLTC fault diagnosis method based on multi-feature fusion

    CN121278666A