CN113378725B - Multi-scale-channel attention network-based tool fault diagnosis method, equipment and storage medium - Google Patents

Abstract

The invention relates to a tool fault diagnosis method, equipment and storage medium based on a multi-scale-channel attention network, which comprises the following steps: (1) data acquisition; (2) data preprocessing; (3) constructing a multi-scale-channel attention network model; (4) training; (5) testing. According to the invention, the multi-scale-channel attention network is adopted to carry out tool wear fault diagnosis, and a channel attention mechanism is introduced into residual connection of the multi-scale network, so that different importance degrees of vibration signals of a machine tool spindle in three directions on tool wear state classification tasks are mapped to a characteristic learning process, and data are better fused.

Description

A tool fault diagnosis method, device and storage medium based on multi-scale-channel attention network

Technical Field

The present invention relates to the field of intelligent manufacturing product quality control, and in particular to a tool fault diagnosis method, device and storage medium based on a multi-scale-channel attention network.

Background Art

As the "mother machine of industry", CNC machine tools are widely used in the production and processing process. As the cutting tool of CNC machine tools, the real-time health status of the tool directly affects the processing efficiency and product quality of the machine tool. The tool is in direct contact and interacts with the workpiece. In the process of high-speed cutting of the tool, inevitable wear and damage will occur. Accurate monitoring of the tool wear status helps to avoid product quality problems caused by tool failure.

At present, manual inspection methods are still used in actual industrial sites. This inspection method is time-consuming and labor-intensive, and there are errors caused by the operators themselves. As the manufacturing industry transforms from automation to intelligence, the deep integration of artificial intelligence technology and manufacturing has become the key. We hope to introduce deep learning into the fault diagnosis method to replace the traditional manual direct contact inspection method, so as to achieve accurate fault diagnosis of cutting tools.

The general intelligent detection process is to collect monitoring signals (such as cutting force signals, vibration signals, acoustic emission signals, spindle current signals, etc.) and analyze the mapping relationship between these signals and tool wear status. At present, a large number of domestic and foreign scholars have carried out such research. Hu et al. proposed a multiscale network (MSNet), which contains a three-branch structure, each branch has a different convolution layer depth, so that the features of different levels of one-dimensional vibration signals can be extracted and merged through full connection. Li Peng et al. used a one-dimensional convolutional network to perform preliminary feature extraction on the vibration and acoustic emission signals of the machine tool spindle and worktable during processing, and then input the signal features into the long short-term memory network for analysis, and finally obtained the evaluation results of the tool wear status, with a classification accuracy of 93.8%. Hsieh et al. used fast Fourier transform to convert the vibration time domain signal into frequency domain signal, and extracted the frequency domain features through the mean scattering criterion, and input the features into a single-layer convolutional neural network to classify the tool wear status. According to the experimental results, it was found that the combination of Z-direction vibration signal and X-direction or Y-direction vibration signal can obtain better classification results. In order to optimize the gradient propagation of the model, the article adds residual connections between convolutional layers with the same feature scale. However, the above method has the following problems: (1) Due to the validity and singleness of the data set, only single-channel vibration signals can be converted into single-channel images. The single channel contains insufficient features, which may affect the accuracy of tool fault classification; (2) Due to the fact that feature extraction and feature analysis are completed step by step, the self-learning ability of the model is weak, and the accuracy of the final tool fault diagnosis is low; (3) Due to the simple model structure, no effective data fusion method is used, resulting in poor fusion effect of vibration signal features in different directions, and the inability to effectively extract data features.

Summary of the invention

In view of the above problems, the present invention proposes a tool fault diagnosis method based on a multi-scale-channel attention network.

The present invention uses a tool wear test platform to collect machine tool spindle vibration signals and tool wear values that conform to actual industrial production sites. The vibration signals include three directions: X, Y, and Z. Based on the idea of feature fusion, the vibration signals in the three directions of X, Y, and Z are spliced to form a three-channel multi-channel feature map.

The present invention uses a convolutional neural network to extract the features of the feature map and perform classification tasks. The convolutional neural network has excellent adaptive feature learning capabilities and can well mine the potential features of the data.

In the present invention, in order to map the different importance of the vibration signals in three directions of the machine tool spindle to the tool wear state classification task into the feature learning process, so that the data can be better integrated, the multi-scale network is improved and channel attention is introduced into the multi-scale network.

The invention also provides a computer device and a storage medium.

The technical solution of the present invention is:

A tool fault diagnosis method based on a multi-scale-channel attention network comprises the following steps:

(1) Data collection:

The vibration signals of the three axes X, Y, and Z of the machine tool spindle and the tool wear value after each tool pass are collected respectively. The spindle coordinate system of the three axes X, Y, and Z of the machine tool spindle is established according to the right-handed Cartesian rectangular coordinate system;

(2) Data preprocessing:

The tool wear stages are classified according to the tool wear values, and the vibration signals are classified using the classification results as labels;

The vibration signals of the machine tool spindle X, Y, and Z axes are segmented into slices of length n ² in time order, and then the slices of the machine tool spindle X, Y, and Z axes are spliced to construct a three-channel input feature map of n×n×3, where n is the height or width of the feature map when the vibration signal slice is converted into a feature map;

Select training set, validation set and test set from the data after data preprocessing;

(3) Constructing a multi-scale-channel attention network model:

The multi-scale-channel attention network model includes an input layer, three branches, and a fully connected layer;

The three branches include a first branch, a second branch, and a third branch. The first branch includes 5 convolutional layers; the second branch includes 2 convolutional layers; and the third branch includes 1 convolutional layer.

Residual connections are set between the second convolution layer of the first branch and the first convolution layer of the second branch, between the fourth convolution layer of the first branch and the second convolution layer of the second branch, and between the first convolution layer of the second branch and the first convolution layer of the third branch. The residual summation results are respectively input into the third convolution layer of the first branch of the convolution layer, the fourth convolution layer of the second branch, and the second convolution layer of the second branch;

The output feature map of the first convolutional layer of the second branch, the output feature map of the second convolutional layer of the second branch, and the output feature map of the first convolutional layer of the third branch are input into the channel attention module before the residual connection.

(4) Training: The training set is input into the multi-scale-channel attention network model for training, and the training set accuracy and loss function of each training cycle are recorded during training;

(5) Testing: The test set is input into the trained multi-scale-channel attention network model, and the tool wear stage corresponding to the test set data is output.

Preferably according to the present invention, 80% of the data after data preprocessing is used as a training set, and 20% is used as a test set, 80% of the training set is used for training, and 20% of the training set is used as a validation set.

Further preferably, the three branches include the first branch, the second branch, and the third branch, and their mathematical expressions are shown in formula (I), formula (II), and formula (III), respectively:

In formula (I), formula (II) and formula (III), i ₁ =1, 2, ... 5, i ₂ =1, 2, i ₃ =1;

是指第一分支的第i₁个卷积层的输出特征图，

是指第一分支的第i₁个卷积层，

是指第一分支的第i₁-1个卷积层的输出特征图；

refers to the output feature map of the _i1th convolutional layer of the first branch,

refers to the _i1th convolutional layer of the first branch,

It refers to the output feature map of the i ₁ -1th convolutional layer of the first branch;

是指第二分支的第i₂个卷积层的输出特征图，

是指第二分支的第i₂个卷积层，

是指第二分支的第i₂-1个卷积层的输出特征图；

refers to the output feature map of the _i2th convolutional layer of the second branch,

refers to the i ₂ th convolutional layer of the second branch,

It refers to the output feature map of the i ₂ -1th convolutional layer of the second branch;

是指第三分支的第i₃个卷积层的输出特征图，

是指第三分支的第i₃个卷积层，

是指第三分支的第i₃-1个卷积层的输出特征图。

refers to the output feature map of the i _3rd convolutional layer of the third branch,

refers to the _3rd convolutional layer of the third branch,

It refers to the output feature map of the i ₃ -1 th convolutional layer of the third branch.

Further preferably, the convolution kernel of each convolution layer is 3×3, and a ReLU nonlinear activation function is added after each convolution layer.

Further preferably, the output feature map scales of the five convolutional layers of the first branch are 64×64×3, 32×32×3, 16×16×3, 8×8×3, and 4×4×3 respectively; the output feature map scales of the two convolutional layers of the second branch are 32×32×3 and 8×8×3 respectively; and the output feature map scale of one convolutional layer of the third branch is 32×32×3.

Further preferably, the channel attention module uses global average pooling operation and global maximum pooling operation to obtain the global receptive field, and then inputs them into a two-layer neural network, the number of neurons in the first layer is 1, the activation function is ReLU, and the number of neurons in the second layer is 3. The two-layer neural network is shared, and then they are added to generate channel weights, and finally the channel weights are normalized to between (0,1) through the normalization layer.

Further preferably, the output feature map of the first convolutional layer of the second branch, the output feature map of the second convolutional layer of the second branch, and the output feature map of the first convolutional layer of the third branch are input into the channel attention module before the residual connection, and the channel weight is obtained after the pooling operation. Then, the channel weight is multiplied by the original output feature map, and the channel weight is mapped to the output feature map of the shallow branch convolutional layer. The feature map with the channel weight is residually connected with the output feature map of the deep branch convolutional layer including the second convolutional layer of the first branch, the fourth convolutional layer of the first branch, and the first convolutional layer of the second branch, and is input into the next convolutional layer of the deep branch convolutional layer.

Preferably, according to the present invention, in step (1), a vibration sensor of model KS903 is used to collect vibration signals of the three axes X, Y, and Z of the machine tool spindle, and the sampling frequency is 10240 Hz; and a 19JC digital universal tool microscope is used to collect tool wear values.

According to the preferred embodiment of the present invention, in step (2), the tool wear stage is classified according to the tool wear value, and the vibration signal is classified using the classification result as a label; specifically, it refers to:

Draw a curve graph of the maximum wear value of the secondary flank of the tool during its entire life cycle, and a time domain graph of the vibration signal of the X-axis of the machine tool spindle during its entire life cycle; the abscissa of the curve graph of the maximum wear value is the number of tool passes, and the ordinate is the maximum wear value; the abscissa of the time domain graph of the vibration signal is time, and the ordinate is the vibration signal amplitude;

The vibration signal is divided into three categories according to the change of the slope of the maximum wear value curve. When the slope of the maximum wear value curve is not greater than 0.01, the corresponding vibration signal is in the rapid initial wear stage. When the slope of the maximum wear value curve is between 0.01 and 0.3, the corresponding vibration signal is in the steady-state wear stage. When the slope of the maximum wear value curve is not less than 0.3, the corresponding vibration signal is in the rapid wear stage.

Preferably, according to the present invention, in step (4), a cross entropy function L is used as a loss function, as shown in formula (IV):

表示预测值，N为样本数量。In formula (IV), _yi represents the true value,

represents the predicted value, and N is the number of samples.

A computer device comprises a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the steps of a tool fault diagnosis method based on a multi-scale-channel attention network are implemented.

A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of a tool fault diagnosis method based on a multi-scale-channel attention network.

The beneficial effects of the present invention are:

1. The present invention uses a convolutional neural network to diagnose tool faults based on the mapping relationship between the machine tool spindle vibration signal and the tool wear status.

2. The present invention is based on the idea of feature fusion. The vibration signals in three directions are used as the three channels of the input feature map. The vibration signals of the three axes of the machine tool spindle are spliced and then input into the convolutional neural network.

3. In order to more effectively fuse the vibration signals in three directions, the present invention introduces the channel attention mechanism into the multi-scale convolutional neural network, and utilizes the interdependence between channels to map the different importance of the vibration signals in three directions to the tool state classification into the feature learning process.

BRIEF DESCRIPTION OF THE DRAWINGS

Figure 1(a) is a graph showing the maximum wear value of the secondary flank during the tool life cycle;

Figure 1(b) is a time domain diagram of the full life cycle vibration signal of the machine tool spindle in the X direction;

Figure 2 is a schematic diagram of the structure of a multi-scale-channel attention network model;

Figure 3(a) is a graph showing the accuracy of the multi-scale-channel attention network model training process;

Figure 3(b) is a graph of the loss function during the training process of the multi-scale-channel attention network model;

Figure 4 is a schematic diagram of a confusion matrix;

DETAILED DESCRIPTION

The present invention will be further defined below in conjunction with the accompanying drawings and embodiments, but is not limited thereto.

Example 1

A tool fault diagnosis method based on a multi-scale-channel attention network comprises the following steps:

(1) Data collection:

In order to collect real data that conforms to actual industrial production scenarios, the present invention designs a tool wear test platform. Using the tool wear test platform, the vibration signals of the three axes of the machine tool spindle X, Y, and Z and the tool wear value after each tool pass are collected respectively, wherein the spindle coordinate system of the three axes of the machine tool spindle X, Y, and Z is established according to the right-handed Cartesian rectangular coordinate system;

(2) Data preprocessing:

The slope of the tool wear value curve reflects the degree of tool wear at the current stage. The present invention classifies the tool wear stages according to the tool wear value, and uses the classification results as labels to classify the vibration signals;

The vibration signals of the X, Y, and Z axes of the machine tool spindle are segmented into slices of length n ² in time order, and then the slices of the X, Y, and Z axes of the machine tool spindle are spliced to construct a three-channel input feature map of n×n×3, where n is the height or width of the feature map when the vibration signal slice is converted into a feature map; this number is an artificially defined size;

80% of the preprocessed data is used as the training set, 20% is used as the test set, 80% of the training set is used for training, and 20% of the training set is used as the validation set.

(3) Constructing a multi-scale-channel attention network model:

The multi-scale network consists of three branches. The depth of the convolution layer of different branches is different. The deeper network branches can extract more local information, and the shallower network branches can extract more local information. Finally, the features of different levels are fused in the fully connected layer. The multi-scale network sets residual connections between the convolution layers of different branches with the same output feature map scale, which helps the network gradient transmission and improves the network performance.

Before the convolution layer of the shallower network branch is connected with the residual connection, the output feature map of the convolution layer is input into the channel attention module. Through the pooling operation, the correlation of the vibration signals in the three directions of X, Y, and Z is captured to obtain the channel attention weight. Then the channel weight is multiplied by the output feature map pixel by pixel, and the channel weight should be set on the feature map. The feature map with the channel attention weight is connected with the output feature map of the convolution layer of the deeper network branch with the residual connection, so as to map the different importance of the three directions to the feature learning process.

As shown in Figure 2, the multi-scale-channel attention network model includes an input layer, three branches, and a fully connected layer;

The three branches include a first branch, a second branch, and a third branch. The first branch includes 5 convolutional layers; the second branch includes 2 convolutional layers; and the third branch includes 1 convolutional layer.

To achieve identity mapping between convolutional layers of different branches with the same output feature map scale, the feature map of the shallow convolutional network branch is residually connected with the feature map of the deep convolutional network branch, and then input into the next convolutional layer of the deep convolutional network branch. Residual connections are set between the second convolutional layer of the first branch and the first convolutional layer of the second branch, the fourth convolutional layer of the first branch and the second convolutional layer of the second branch, and the first convolutional layer of the second branch and the first convolutional layer of the third branch. The residual summation results are input into the third convolutional layer of the first branch of the convolutional layer, the fourth convolutional layer of the second branch, and the second convolutional layer of the second branch respectively; the residual connection can help the features in the network to be identically mapped in the forward process. When the output of the shallow network has reached the optimal level, the features are directly transferred to the deep network; in the reverse process, it helps to conduct gradients, so that deeper models can be successfully trained and the performance of the network can be improved.

The channel attention module is input before the residual connection is made to the output feature map of the shallow branch convolution layer, and the channel attention module is input before the residual connection is made to the output feature map of the first convolution layer of the second branch, the output feature map of the second convolution layer of the second branch, and the output feature map of the first convolution layer of the third branch.

The mathematical expressions of the three branches including the first branch, the second branch and the third branch are shown in formula (I), formula (II) and formula (III) respectively:

In formula (I), formula (II) and formula (III), i ₁ =1, 2, ... 5, i ₂ =1, 2, i ₃ =1;

是指第一分支的第i₁个卷积层的输出特征图，

是指第一分支的第i₁个卷积层，

是指第一分支的第i₁-1个卷积层的输出特征图；

refers to the output feature map of the _i1th convolutional layer of the first branch,

refers to the _i1th convolutional layer of the first branch,

It refers to the output feature map of the i ₁ -1th convolutional layer of the first branch;

是指第二分支的第i₂个卷积层的输出特征图，

是指第二分支的第i₂个卷积层，

是指第二分支的第i₂-1个卷积层的输出特征图；

refers to the output feature map of the _i2th convolutional layer of the second branch,

refers to the i ₂ th convolutional layer of the second branch,

It refers to the output feature map of the i ₂ -1th convolutional layer of the second branch;

是指第三分支的第i₃个卷积层的输出特征图，

是指第三分支的第i₃个卷积层，

是指第三分支的第i₃-1个卷积层的输出特征图。

refers to the output feature map of the i _3rd convolutional layer of the third branch,

refers to the _3rd convolutional layer of the third branch,

It refers to the output feature map of the i ₃ -1 th convolutional layer of the third branch.

The convolution kernel of each convolution layer is 3×3, and a ReLU nonlinear activation function is added after each convolution layer.

The output feature map scales of the five convolutional layers of the first branch are 64×64×3, 32×32×3, 16×16×3, 8×8×3, and 4×4×3 respectively; the output feature map scales of the two convolutional layers of the second branch are 32×32×3 and 8×8×3 respectively; the output feature map scale of the one convolutional layer of the third branch is 32×32×3.

The channel attention module uses global average pooling and global maximum pooling operations to obtain the global receptive field, and then inputs them into a two-layer neural network. The number of neurons in the first layer is 1, the activation function is ReLU, and the number of neurons in the second layer is 3. The two-layer neural network is shared, and then they are added to generate channel weights. Finally, the channel weights are normalized to between (0,1) through the normalization layer.

Before the output feature map of the shallow branch convolution layer is connected with the residual connection, the channel attention module is input. The mutual dependence between channels is used to capture the correlation of the vibration signals in the three directions of X, Y, and Z, so that the different importance of the three directions can be mapped to the feature extraction process through channel attention learning. The output feature map of the first convolution layer of the second branch, the output feature map of the second convolution layer of the second branch, and the output feature map of the first convolution layer of the third branch are input with the channel attention module before the residual connection. The channel weight is obtained through pooling operation, and then the channel weight is multiplied with the original output feature map. The channel weight is mapped to the output feature map of the shallow branch convolution layer. The feature map with the channel weight is residually connected with the output feature map of the deep branch convolution layer, including the second convolution layer of the first branch, the fourth convolution layer of the first branch, and the first convolution layer of the second branch, and input into the next convolution layer of the deep branch convolution layer.

和

为例：

表示第一分支的第三个卷积层的输入，A表示通道注意力模块，

表示第一分支的第二个卷积层的输出特征图，

表示第二分支的第一个卷积层的输出特征图。通过高效通道注意力模块，提升影响因素较强的通道特征权重，而抑制影响因素较弱的通道特征权重。Output feature map with convolutional layer

and

For example:

represents the input of the third convolutional layer of the first branch, A represents the channel attention module,

represents the output feature map of the second convolutional layer of the first branch,

The output feature map of the first convolutional layer of the second branch. Through the efficient channel attention module, the weights of channel features with stronger influencing factors are increased, while the weights of channel features with weaker influencing factors are suppressed.

(4) Training: The training set is input into the multi-scale-channel attention network model for training, and the training set accuracy and loss function of each training cycle are recorded during training;

(5) Testing: Input the test set into the trained multi-scale-channel attention network model and output the tool wear stage corresponding to the test set data. Record the accuracy of each category in the test set and draw the confusion matrix. Use the four indicators of accuracy, precision, recall, and F1 score to illustrate the performance of the network.

Example 2

The tool fault diagnosis method based on a multi-scale-channel attention network according to embodiment 1 is different in that:

In step (1), a vibration sensor of model KS903 is used to collect vibration signals of the three axes X, Y, and Z of the machine tool spindle, and the sampling frequency is 10240Hz; a 19JC digital universal tool microscope is used to collect tool wear values. Data collection is carried out on a tool wear test platform, and the workpiece cutting process is completed on a CNC machine tool vertical machining center (VDF-850). The tool is a three-edge end mill with a diameter of 10mm, and the cutting workpiece is a 45# steel cylindrical workpiece. The spindle speed of the CNC machine tool is 2000r/min, and the feed speed is 764mm/min. The milling depth and width are 2mm and 5mm respectively. In order to accelerate tool wear, a dry cutting method without cutting fluid is adopted. The total processing length of the milling cutter is about 110 meters, so that the tool reaches a serious degree of wear and there is a small area of milling chips. 35-47 sets of test data are collected for each tool, and each set of milling process takes 4 minutes and 17 seconds. At the same time, after each group of cutting is completed, a 19JC digital universal tool microscope is used to collect tool wear values, and the maximum wear width VBmax of the secondary flank is used as the classification label of the tool wear stage.

In step (2), the tool wear stage is classified according to the tool wear value, and the vibration signal is classified using the classification result as a label; specifically, it means:

Draw a curve graph of the maximum wear value of the secondary flank of the tool during its entire life cycle, and a time domain graph of the vibration signal of the X-axis of the machine tool spindle during its entire life cycle; the abscissa of the curve graph of the maximum wear value is the number of tool passes, and the ordinate is the maximum wear value; the abscissa of the time domain graph of the vibration signal is time, and the ordinate is the vibration signal amplitude;

The vibration signal is divided into three categories according to the change of the slope of the maximum wear value curve. When the slope of the maximum wear value curve is not greater than 0.01, the corresponding vibration signal is in the rapid initial wear stage. When the slope of the maximum wear value curve is between 0.01 and 0.3, the corresponding vibration signal is in the steady-state wear stage. When the slope of the maximum wear value curve is not less than 0.3, the corresponding vibration signal is in the rapid wear stage.

According to the change of the slope of the maximum wear value curve, the vibration signal is divided into three categories, corresponding to three different tool wear stages: rapid initial wear stage, steady-state wear stage and rapid wear stage. The maximum wear value of the tool in the whole life cycle is shown in Figure 1(a). As can be seen from Figure 1(a), the tool wears faster in the 1-5 pass stage, and the slope of the wear curve in this stage is larger. In the 6-41 pass stage, the wear value increases evenly until it reaches the limit value. This stage is the effective working time of the tool. In the 42-47 pass stage, the tool wear value rises rapidly and causes the tool to fail. The slope of the wear curve in this stage increases rapidly. Although the initial wear stage is not very obvious in Figure 1(a), in the vibration signal time domain diagram of Figure 1(b), the amplitude of the vibration signal in the initial stage is large. The reason for this phenomenon is that the surface of the new tool is rough and uneven, the contact stress is large, and the new tool has surface defects caused by decarburization and oxidation layer. According to the above analysis, the vibration signal is divided into three stages.

将切片构建为64×64×3的输入特征图输入多尺度-通道注意力网络模型。Before the vibration signal is input into the multi-scale-channel attention network model, the vibration signal is segmented into slices of length 64 ² according to the time order of the signal, and then the vibration signal slices in the three directions of X, Y, and Z are spliced as the three channels of the input feature map. Assuming there are three signal slices in the same period of the X, Y, and Z directions,

The slices are constructed as 64×64×3 input feature maps and input into the multi-scale-channel attention network model.

In step (4), the number of training iterations is set to 100, and the cross entropy function L is used as the loss function, as shown in formula (IV):

表示预测值，N为样本数量。In formula (IV), _yi represents the true value,

represents the predicted value, and N is the number of samples.

During the training process, the accuracy and loss function values of the training set and validation set are recorded for each training cycle. After the training process is completed, the optimal validation model is saved for testing, and the training set accuracy curve and the training process loss function curve are plotted. Figure 3(a) is the accuracy curve of the multi-scale-channel attention network model training process; Figure 3(b) is the loss function curve of the multi-scale-channel attention network model training process;

Record the accuracy of each category of the test set and draw a confusion matrix. Figure 4 is a schematic diagram of the confusion matrix. Table 1 is a table of four evaluation indicators of the multi-scale-channel attention network model. The four indicators of accuracy, precision, recall, and F1 score illustrate the performance of the multi-scale-channel attention network model.

Table 1

As can be seen from Table 1, the four indicators of accuracy, precision, recall and F1 score are all high, indicating that the multi-scale-channel attention network model has better performance.

Example 3

A computer device includes a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the steps of the tool fault diagnosis method based on a multi-scale-channel attention network described in Example 1 or 2 are implemented.

Example 4

A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the tool fault diagnosis method based on a multi-scale-channel attention network described in Example 1 or 2.

Claims (9)

1. The tool fault diagnosis method based on the multiscale-channel attention network is characterized by comprising the following steps of:

(1) And (3) data acquisition:

respectively collecting vibration signals of three axes of a machine tool spindle X, Y, Z and cutter abrasion values after each feeding, wherein a spindle coordinate system of the machine tool spindle X, Y, Z is established according to a right-hand Cartesian rectangular coordinate system;

(2) Data preprocessing:

classifying the cutter abrasion stage according to the cutter abrasion value, and classifying the vibration signals by taking the classification result as a label;

the vibration signals of the three axes of the main shaft X, Y, Z of the machine tool are segmented into the length n according to the time sequence ² Then splicing the sections of the three axes of the main shaft X, Y, Z of the machine tool to construct an n multiplied by 3 three-channel input feature map, wherein n is the height or width of the feature map when the vibration signal section is converted into the feature map;

selecting a training set, a verification set and a test set from the data after the data preprocessing;

(3) Constructing a multi-scale-channel attention network model:

the multi-scale-channel attention network model comprises an input layer, three branches and a full-connection layer;

the three branches comprise a first branch, a second branch and a third branch, and the first branch comprises 5 convolution layers; the second branch comprises 2 convolution layers; the third branch comprises 1 convolution layer;

residual error connection is arranged between the second convolution layer of the first branch and the first convolution layer of the second branch, between the fourth convolution layer of the first branch and the second convolution layer of the second branch and between the first convolution layer of the second branch and the first convolution layer of the third branch, and residual error addition results are respectively input into the third convolution layer of the first branch, the fifth convolution layer of the first branch and the second convolution layer of the second branch;

the input channel attention module is used for carrying out residual connection on the output characteristic diagram of the first convolution layer of the second branch, the output characteristic diagram of the second convolution layer of the second branch and the output characteristic diagram of the first convolution layer of the third branch;

(4) Training: inputting the training set into a multi-scale-channel attention network model for training, and recording the accuracy and the loss function of the training set in each training period while training;

(5) And (3) testing: inputting the test set into a trained multi-scale-channel attention network model, and outputting a cutter abrasion stage corresponding to the test set data;

the mathematical expressions of the three branches including the first branch, the second branch and the third branch are respectively shown as the formula (I), the formula (II) and the formula (III):

in the formula (I), the formula (II) and the formula (III), i ₁ ＝1,2,…5，i ₂ ＝1,2，i ₃ ＝1；

Refers to the ith branch of the first branch ₁ Output feature map of the convolutional layers, +.>

Refers to the ith branch of the first branch ₁ The number of the convolution layers is one,

refers to the ith branch of the first branch ₁ -an output profile of 1 convolutional layer;

refers to the ith of the second branch ₂ Output feature map of the convolutional layers, +.>

Refers to the ith of the second branch ₂ The number of the convolution layers is one,

refers to the ith of the second branch ₂ -an output profile of 1 convolutional layer;

refers to the ith branch of the third branch ₃ Output feature map of the convolutional layers, +.>

Refers to the ith branch of the third branch ₃ Convolutional layers->

Refers to the ith branch of the third branch ₃ -an output profile of 1 convolutional layer;

the channel attention module acquires global receptive fields by adopting global average pooling operation and global maximum pooling operation respectively, inputs the global receptive fields into a two-layer neural network respectively, wherein the number of neurons of the first layer is 1, the activation function is ReLU, the number of neurons of the second layer is 3, the two-layer neural networks are shared, then the two-layer neural networks are added to generate channel weights, and finally the channel weights are normalized to be between (0 and 1) through a normalization layer;

and (3) inputting a channel attention module before residual connection to obtain channel weights by pooling operation, multiplying the channel weights by the original output characteristic diagram, mapping the channel weights onto the output characteristic diagram of the shallow branch convolution layer, and carrying out residual connection on the characteristic diagram with the channel weights and the deep branch convolution layer comprising the second convolution layer of the first branch, the fourth convolution layer of the first branch and the output characteristic diagram of the first convolution layer of the second branch.

2. The multi-scale channel attention network based tool fault diagnosis method according to claim 1, wherein the convolution kernel of each convolution layer is 3 x 3, and a ReLU nonlinear activation function is added after each convolution layer.

3. The multi-scale-channel attention network based tool fault diagnosis method according to claim 1, wherein the output feature map scale of the 5 convolution layers of the first branch is 64×64×3, 32×32×3, 16×16×3,8×8×3,4×4×3, respectively; the output feature map scales of the 2 convolution layers of the second branch are respectively 32×32×3,8×8×3; the output feature map scale of 1 convolution layer of the third branch is 32×32×3.

4. The method for diagnosing a tool failure based on a multi-scale and channel attention network according to claim 1, wherein in the step (1), vibration signals of three axes of a main shaft X, Y, Z of a machine tool are collected by using a vibration sensor with a model KS903, and the sampling frequency is 10240Hz; and acquiring the cutter abrasion value by using a 19JC digital universal tool microscope.

5. The multi-scale channel attention network based tool fault diagnosis method according to claim 1, wherein 80% of the data after the data preprocessing is used as a training set, 20% is used as a test set, 80% of the training set is used as training, and 20% of the training set is used as a verification set.

6. The multi-scale-channel attention network based tool fault diagnosis method according to claim 1, wherein in the step (2), tool wear stage classification is performed according to tool wear values, and vibration signals are classified by using classification results as labels; the method specifically comprises the following steps:

drawing a graph of the maximum abrasion value of the tool surface of the auxiliary tool in the whole life cycle and a time domain graph of a vibration signal in the whole life cycle of the X axis of a main shaft of a machine tool; wherein, the abscissa of the graph of the maximum abrasion value is the number of times of feeding, and the ordinate is the maximum abrasion value; the abscissa of the time domain graph of the vibration signal is time, and the ordinate is vibration signal amplitude;

vibration signals are classified into three types according to the slope change of the maximum abrasion value curve, when the slope of the maximum abrasion value curve is not more than 0.01, the corresponding vibration signals are in a rapid initial abrasion stage, when the slope of the maximum abrasion value curve is between 0.01 and 0.3, the corresponding vibration signals are in a steady abrasion stage, and when the slope of the maximum abrasion value curve is not less than 0.3, the corresponding vibration signals are in a rapid abrasion stage.

7. The multi-scale-channel attention network based tool fault diagnosis method according to any one of claims 1 to 6, wherein in step (4), a cross entropy function L is used as a loss function, as shown in formula (iv):

in the formula (IV), y _i The true value is represented by a value that is true,

representing the predicted value, N is the number of samples.

8. A computer device comprising a memory and a processor, the memory storing a computer program, characterized in that the processor, when executing the computer program, implements the steps of the multi-scale-channel attention network based tool fault diagnosis method of any of claims 1-7.

9. A computer readable storage medium having stored thereon a computer program, characterized in that the computer program when executed by a processor implements the steps of the multi-scale-channel attention network based tool fault diagnosis method of any of claims 1-7.

Images