CN111325112B - Cutter wear state monitoring method based on depth gate control circulation unit neural network - Google Patents

Abstract

The invention discloses a tool wear state monitoring method based on a depth-gated cyclic unit neural network, which includes: using a sensor to collect vibration signals generated during tool processing in real time, and inputting them into a one-dimensional convolutional neural network after wavelet threshold denoising Carry out local feature extraction of time-series signals at a single time step, and then input the improved deep gated recurrent unit neural network CABGRUs for time-series feature extraction of time-series signals, introduce the Attention mechanism to calculate network weights and allocate them reasonably, and finally, the different weights The signal feature information is put into the Softmax classifier to classify the tool wear state, avoiding the complexity and limitations caused by manual feature extraction; at the same time, it effectively solves the problem of single convolutional neural network ignoring the front and rear correlation of time series signals, by introducing the Attention mechanism Improve the accuracy of the model. Therefore, the present invention has the characteristics of improving the real-time and accuracy of tool wear state monitoring.

Description

Tool wear state monitoring method based on deep gated recurrent unit neural network

Technical Field

The present invention belongs to the field of manufacturing process monitoring, and in particular relates to a tool wear state monitoring method based on a deep gated recurrent unit neural network.

Background Art

In the process of mechanical processing, cutting is the most important processing method for part forming. The wear state of the tool will directly affect the processing accuracy, surface quality and production efficiency of the parts. Therefore, tool condition monitoring (TCM) technology is of great significance for ensuring processing quality and realizing continuous automated processing. The tool condition monitoring method currently mainly adopts the indirect measurement method, which can collect signals in real time through sensors during the tool cutting process. After data processing and feature extraction, the machine learning (ML) model is used to monitor the tool wear.

In the prior art, Zhang Ansi et al. proposed a model for predicting the remaining life of equipment based on transfer learning and long short-term memory network (LSTM). The model was trained on different but related equipment remaining life prediction data sets in advance, and then the network structure and training parameters were fine-tuned for the target data set. Experimental results show that the transfer learning method can improve the prediction accuracy of the model with only a small number of samples. Jin Qi et al. proposed a planetary gearbox fault diagnosis method based on deep learning diversity feature extraction and information fusion. The multi-objective optimization algorithm was used to optimize multiple stack denoising autoencoders (SDAE), and the diversity of fault features was extracted at the same time. Then, the multi-response linear regression model was used to integrate the diversity of fault features to achieve information fusion. The experimental results show that the method of diversity feature extraction and information fusion can effectively improve the accuracy and stability of fault diagnosis, and has strong generalization ability. Zhang Cunji et al. proposed to transform the vibration signal of the tool during the machining process into an energy spectrum through wavelet packet transform (WPT), and then input it into the convolutional neural network to automatically extract features and accurately classify them. The experimental results show that the accuracy of the deep convolutional neural network is better than that of the traditional neural network model. Cao Dali et al. proposed to construct a deep neural network DenseNet using dense connections to adaptively extract the hidden subtle features in the tool processing signal from the original time series signal. The experimental results show that deepening the network layers helps to mine the high-dimensional features hidden in the processing signal and improve the accuracy of the tool wear monitoring model. The above methods all use deep learning to adaptively extract features, but the convolutional neural network they use relies too much on high-dimensional feature extraction. Too many convolution layers are prone to gradient diffusion, too few convolution layers cannot grasp the overall situation, and do not take into account the important feature of the correlation between the time series signal samples generated during tool processing.

Summary of the invention

The purpose of the present invention is to overcome the above-mentioned shortcomings and propose a tool wear state monitoring method based on a deep gated recurrent unit neural network that can improve the real-time and accuracy of tool wear state monitoring.

A tool wear state monitoring method based on a deep gated recurrent unit neural network of the present invention comprises the following steps:

Step 1: Use an acceleration sensor to collect the three-axis vibration signal generated by the tool during milling, use the wavelet threshold denoising method to denoise the collected original signal, and cut the vibration signal generated by each tool feed into multiple short sequence time series signals with a length of 2000. The specific steps are: intercept 100,000 consecutive points in each sampling signal, and divide the intercepted points into 50 samples with a sample number of 2000. These 50 samples correspond to the same wear status label.

Step 2: Extraction of local features of single time step timing signal: A one-dimensional convolutional neural network is used to process the short sequence timing signal generated in the above tool processing process. The convolutional neural network part includes two layers of convolutional layers (CONV) and one layer of pooling layer (POOL). The convolutional layer performs neighborhood filtering on the timing signal of each dimension through a one-dimensional convolution operation to generate a feature map. Each feature map can be regarded as a convolution operation of different filters on the timing signal of the current time step; the calculation formula of the one-dimensional convolutional layer operation is as follows:

表示第l层的第j个特征映射，f表示激活函数，M表示输入特征映射的数量，

表示第l-1层的第i个特征映射，

表示可训练的卷积核，

表示偏置参数；所述激活函数采用Relu激活函数；in:

represents the jth feature map of the lth layer, f represents the activation function, M represents the number of input feature maps,

represents the i-th feature map of the l-1th layer,

represents a trainable convolution kernel,

represents the bias parameter; the activation function adopts the Relu activation function;

The pooling layer uses maximum pooling to take the maximum value of the feature points in the neighborhood. The formula is as follows:

表示第l层的第i个特征矢量中的第t个神经元的值，t∈[(j-1)w+1,jw]；w为池化区域的宽度；P_i ^l+1(j)表示第l+1层神经元对应的值。in:

represents the value of the tth neuron in the i-th feature vector of the l-th layer, t∈[(j-1)w+1,jw]; w is the width of the pooling area; Pi _l ⁺¹ (j) represents the value corresponding to the neuron in the l+1-th layer.

Step 3: The vibration signals generated during tool processing have a time series relationship. In order to explore the time series variation rules of relatively long intervals in the time series, an improved gated recurrent unit is used to extract the time series features of the time series signals and learn the dependency relationship between the time series features of the time series signals.

The improved gated recurrent unit constructs two bidirectional BiGRU networks and superimposes them to form a CABGRUs network. At the same time, an Attention mechanism is introduced into the CABGRUs network and an Attention layer is added, so that the model can not only extract time series signal features from both the forward and reverse directions, but also obtain the ability to selectively learn key information in signal features.

表示在时间步t时的候选隐藏状态，h_t表示在时间步t时的隐藏状态，x_t表示在时间步t时的输入向量。更新门z_t用于控制当前状态更新了多少状态信息，z_t越趋近于1，表示当前状态对先前时刻的信息利用的越多。重置门r_t用于控制从先前状态中移除哪些状态信息，r_t越趋近于0，表示从先前时刻的输出状态所占的比例越小。公式如下：Each bidirectional BiGRU network in the improved deep gated recurrent unit neural network CABGRUs contains 256 neurons, and the forward and reverse BiGRU networks are composed of 128 neurons. Each BiGRU neuron includes an update gate and a reset gate, which are represented by z _t and r _t respectively, where,

represents the candidate hidden state at time step t, h _t represents the hidden state at time step t, and x _t represents the input vector at time step t. The update gate z _t is used to control how much state information is updated in the current state. The closer z _t is to 1, the more the current state uses the information of the previous moment. The reset gate r _t is used to control which state information is removed from the previous state. The closer r _t is to 0, the smaller the proportion of the output state from the previous moment. The formula is as follows:

r _t =σ(x _t W _xr +h _t-1 W _hr +b _r )

z _t =σ(x _t W _xz +h _t-1 W _hz +b _z )

Wherein: _Wxr and _Whr represent the weight vector of the reset gate, _Wxz and _Whz represent the weight vector of the update gate, _Wxh and _Whh represent the weight vector of the candidate hidden state, _br , _bz , _bh represent bias vectors, □ represents the Hadamard product, that is, the dot product of the matrix, σ(·) represents the Sigmod function, and the tanh function represents the hyperbolic tangent activation function.

后向BiGRU网络输出隐藏状态

在时间步t时CABGRUs网络输出隐藏状态P_t。The high-dimensional features of the input time series signal are output through the forward BiGRU network to generate hidden states

The backward BiGRU network outputs the hidden state

At time step t, the CABGRUs network outputs the hidden state _Pt .

The formula is as follows:

Step 4: Introduce the Attention mechanism to calculate the importance distribution of the features of the time series signal at consecutive time steps. The introduced Attention mechanism performs weighted summation on the output vectors of each time step of the BiGRU layer by assigning different initialization probability weights, and finally calculates the value through the Sigmod function. The calculation formula of the Attention mechanism is as follows:

u _t = tanh(W _s P _t + b _s )

ν＝∑α _t P _t

Among them, _Pt represents the output feature vector of the BiGRU layer at time step t, _ut represents the hidden layer representation of _Pt obtained through the neural network layer, _us represents the randomly initialized context vector, _αt represents the importance weight obtained by normalizing _ut through the Softmax function, ν represents the feature vector of the final text information, that is, _us is randomly generated during the training process, and finally the output value ν of the Attention layer is mapped through the Softmax function to obtain the real-time classification result of the tool wear status.

Step 5: Training of the network model: Dropout technology is introduced to prevent the model from overfitting during the training process; the activation function of the network model uses Softmax, and the loss function uses Categorical_crossentropy. The time series signal features obtained in the above steps are used for wear classification to obtain the classification results.

The input data is a time series signal, and the feature extraction and expression of the time series signal are realized through convolutional layers (C1 and C2), Dropout layers, pooling layers (P1), Flatten layers, BiGRU (B1 and B2) layers, Attention layers A1 and fully connected layers (F1 and F2); wherein, the time series signal of size (2000,3) is input into the deep learning neural network as input data, the convolutional layer C1 uses a 3×1 convolution kernel to convolve the time series signal, the convolution kernel step size is 1, and a (20,98,128) feature map is generated, which is input from the convolutional layer C1 to the convolutional layer C2, the convolutional layer C2 uses a 3×1 convolution kernel to convolve the time series signal, the convolution kernel step size is 1, and a (20,96,128) feature map is generated, which is input from the convolutional layer C2 to the pooling layer P1 through Dropout, D The ropout is 0.5, and the maximum value pooling method is used to generate a (20, 48, 128) feature map, which is input from the pooling layer P1 to the Flatten layer L1 to generate (20, 6144) features, which is input from the Flatten layer L1 to the BiGRU layer B1 to generate (20, 256) features, which is input from the BiGRU layer B1 to the BiGRU layer B2 to generate (20, 256) features, which is input from the BiGRU layer B2 to the Attention layer A1 via Dropout, and the Dropout is 0.5 to generate a 256×1 feature map, which is input from the Attention layer A1 to the fully connected layer F1 to output a 128×1 feature map, which is input from the fully connected layer F1 to the fully connected layer F2, and finally outputs three types of tool wear status values to confirm the wear status of the tool at the current moment.

Compared with the prior art, the present invention has obvious beneficial effects. It can be seen from the above scheme that the existing gated recurrent unit (GRU) neural network is improved, and two bidirectional BiGRU networks are constructed to form a CABGRUs network. At the same time, the Attention mechanism is introduced into the CABGRUs network, and the Attention layer is added, so that the model can not only extract the characteristics of the time series signal from the forward and reverse directions at the same time, but also obtain the ability to selectively learn the key information in the signal characteristics. Among them, in the CABGRUs network, each bidirectional GRU network contains 256 neurons, and the forward and reverse GRU networks are composed of 128 neurons, and each GRU neuron includes an update gate and a reset gate. The real-time monitoring model of the tool wear state can better learn the dependency relationship between the time series characteristics of the time series signals, and improve the accuracy of the model classification. In addition, the Attention mechanism is introduced, and the Attention mechanism is a brain signal processing mechanism similar to human vision. By assigning different initialization probability weights and weighted summation with the output vectors of each time step of the BiGRU layer, and then bringing them into the Sigmod function, the final calculation is performed and the numerical value is obtained. It is possible to selectively filter out some key information from a large number of signal features and focus on it. The focusing process is reflected in the calculation of weight coefficients. Different weights are assigned to different key information. The weight is increased to strengthen the proportion of key information and reduce the loss of key information in long-sequence time series signals.

In summary, the present invention uses sensors to collect vibration signals generated during tool processing in real time, and after wavelet threshold denoising, inputs them into a one-dimensional convolutional neural network to extract local features of single time-step time series signals, and then inputs them into an improved deep gated recurrent unit neural network CABGRUs to extract time series features of time series signals, introduces an Attention mechanism to calculate network weights and distributes them reasonably, and finally, puts signal feature information of different weights into a Softmax classifier to classify tool wear status, avoiding the complexity and limitations caused by manual feature extraction; at the same time, it effectively solves the problem of single convolutional neural networks ignoring the before and after association of time series signals, and avoids the gradient diffusion and gradient explosion problems of recurrent neural networks, and improves the accuracy of the model by introducing an Attention mechanism. Therefore, the present invention has the characteristics of improving the real-time and accuracy of tool wear status monitoring.

The beneficial effects of the present invention are further illustrated below through specific implementation modes.

BRIEF DESCRIPTION OF THE DRAWINGS

Fig. 1 is a flow chart of the method of the present invention;

FIG2 is a schematic diagram of the CABGRUs network structure of the present invention;

FIG3 is a diagram of CNN model training and verification in an embodiment;

FIG4 is a diagram of BiGRU model training and verification in an embodiment;

FIG5 is a diagram of CBLSTMs model training and verification in an embodiment;

FIG6 is a diagram of CABGRUs model training and verification in an embodiment.

DETAILED DESCRIPTION

In combination with the accompanying drawings and preferred embodiments, the specific implementation method, features and efficacy of a tool wear state monitoring method based on a deep gated recurrent unit neural network proposed in accordance with the present invention are described in detail as follows.

Referring to FIG1 , a tool wear state monitoring method based on a deep gated recurrent unit neural network of the present invention comprises the following steps:

Step 1: Use the acceleration sensor to collect the vibration signal generated by the CNC machining equipment in the process of machining the workpiece in real time. The signal after wavelet threshold denoising is used as the input signal of the real-time monitoring model of the tool wear status, including α _x , α _y , and α _z vibration signals. The original vibration signal is continuously sampled and cropped into a tensor (2000,3) consisting of 2000 sampling points in the x, y, and z directions, and is brought into the model as the input data of the model.

Step 2: Extract local features of a single time-step time series signal: Input the time series signal of size (2000, 3) as input data into a one-dimensional convolutional neural network (CNN) for neighborhood filtering, and use a sliding window for calculation to finally obtain the high-dimensional features of the single time-step time series signal;

A one-dimensional convolutional neural network is used to directly process the time series signals generated during tool processing. The convolutional neural network part includes two convolutional layers (CONV) and one pooling layer (POOL). The convolutional layer performs neighborhood filtering on the time series signal of each dimension through a one-dimensional convolution operation to generate a feature map. Each feature map can be regarded as a convolution operation of different filters on the time series signal of the current time step. When the input time series signal is _xt , the filter is _wt , and the feature map _yt of the convolutional layer can be expressed as:

In the convolutional layer, each neuron in the lth layer is only connected to a neuron in a local window of the l-1th layer, forming a local connection network. The calculation formula of the one-dimensional convolutional layer is as follows:

表示第l层的第j个特征映射，f表示激活函数，M表示输入特征映射的数量，

表示第l-1层的第i个特征映射，

表示可训练的卷积核，

表示偏置参数。考虑到收敛速度和过拟合问题，本发明的非线性激活函数选用收敛速度较快的修正线性单元(Rectified Linear,Relu)，用于提高网络的稀疏性，减少参数的相互依存关系，缓解过拟合现象的发生。Relu激活函数的公式为：in:

represents the jth feature map of the lth layer, f represents the activation function, M represents the number of input feature maps,

represents the i-th feature map of the l-1th layer,

represents a trainable convolution kernel,

Represents the bias parameter. Considering the convergence speed and overfitting problems, the nonlinear activation function of the present invention uses the rectified linear unit (Relu) with a faster convergence speed to improve the sparsity of the network, reduce the interdependence of parameters, and alleviate the occurrence of overfitting. The formula of the Relu activation function is:

表示卷及操作的输出值；

表示

的激活值。in:

Represents the output value of the volume and operation;

express

The activation value of .

The convolutional layer is followed by a pooling layer to obtain the local maximum or local mean, namely, maximum pooling (MaxPooling) and mean pooling (Mean Pooling). The pooling layer has a function similar to feature selection, which can ensure that the feature has anti-deformation ability while achieving the purpose of reducing feature dimensions, speeding up network training, reducing the number of parameters, and improving feature robustness. The present invention uses maximum pooling to take the maximum value of the feature points in the neighborhood, and the formula is as follows:

表示第l层的第i个特征矢量中的第t个神经元的值，t∈[(j-1)w+1,jw]；w为池化区域的宽度；P_i ^l+1(j)表示第l+1层神经元对应的值。in:

represents the value of the tth neuron in the i-th feature vector of the l-th layer, t∈[(j-1)w+1,jw]; w is the width of the pooling area; Pi _l ⁺¹ (j) represents the value corresponding to the neuron in the l+1-th layer.

By extracting the features of the original data through a one-dimensional convolutional neural network, the three-dimensional features of the time series signal can be better expressed as high-dimensional features, which facilitates the subsequent network to extract time series features.

Step 3: Extract features from the time series of the time series signal: Use an improved gated recurrent unit to process the high-dimensional features generated by the continuous time step time series signal, and gradually synthesize the vector feature representation of the input signal;

The improved gated recurrent unit constructs two bidirectional gated recurrent unit BiGRU networks and superimposes them to form a CABGRUs network. At the same time, an Attention mechanism is introduced into the CABGRUs network and an Attention layer is added, so that the model can not only extract time series signal features from both forward and reverse directions, but also selectively learn key information in signal features.

The original signals generated during tool processing have a time series relationship. The RNN network can encode the time series of the time series signal and explore the time series variation rules of relatively long intervals in the time series. In order to enable the real-time monitoring model of tool wear status to better learn the dependency relationship between the time series features of the time series signals, the accuracy of model classification is improved. The present invention improves the existing gated recurrent unit (GRU) by constructing two bidirectional gated recurrent unit BiGRU networks to form a CABGRUs network. At the same time, the Attention mechanism is introduced into the CABGRUs network, and the Attention layer is added, so that the model can not only extract the time series signal features from both the forward and reverse directions, but also obtain the ability to selectively learn key information in the signal features.

Each bidirectional BiGRU network in the deep gated recurrent unit neural network CABGRUs constructed by the present invention contains 256 neurons, and the forward and reverse BiGRU networks are composed of 128 neurons. Each BiGRU neuron includes an update gate and a reset gate, which are represented by z _t and r _t respectively.

表示在时间步t时的候选隐藏状态，h_t表示在时间步t时的隐藏状态，x_t表示在时间步t时的输入向量。更新门z_t用于控制当前状态更新了多少状态信息，z_t越趋近于1，表示当前状态对先前时刻的信息利用的越多。重置门r_t用于控制从先前状态中移除哪些状态信息，r_t越趋近于0，表示从先前时刻的输出状态所占的比例越小。公式如下：in,

represents the candidate hidden state at time step t, h _t represents the hidden state at time step t, and x _t represents the input vector at time step t. The update gate z _t is used to control how much state information is updated in the current state. The closer z _t is to 1, the more the current state uses the information of the previous moment. The reset gate r _t is used to control which state information is removed from the previous state. The closer r _t is to 0, the smaller the proportion of the output state from the previous moment. The formula is as follows:

r _t =σ(x _t W _xr +h _t-1 W _hr +b _r )

z _t =σ(x _t W _xz +h _t-1 W _hz +b _z )

Wherein: _Wxr and _Whr represent the weight vector of the reset gate, _Wxz and _Whz represent the weight vector of the update gate, _Wxh and _Whh represent the weight vector of the candidate hidden state, _br , _bz , _bh represent bias vectors, □ represents the Hadamard product, that is, the dot product of the matrix, σ(·) represents the Sigmod function, and the tanh function represents the hyperbolic tangent activation function.

后向BiGRU网络输出隐藏状态

在时间步t时CABGRUs网络输出隐藏状态P_t。The high-dimensional features of the input time series signal are output through the forward BiGRU network to generate hidden states

The backward BiGRU network outputs the hidden state

At time step t, the CABGRUs network outputs the hidden state _Pt .

The formula is as follows:

Step 4: Introduce the Attention mechanism: Use the Attention mechanism to calculate the importance distribution of the time series signal features of continuous time steps, and generate a time series signal feature model containing the attention probability distribution; that is, the introduced Attention mechanism assigns different initialization probability weights to the output vectors of each time step of the BiGRU layer, and finally calculates the value through the Sigmod function.

The Attention mechanism introduced in the present invention performs weighted summation by assigning different initialization probability weights to the output vectors of each time step of the BiGRU layer, and finally obtains the value through Sigmod function calculation. It can selectively filter out some key information from a large number of signal features and focus on them. The focusing process is reflected in the calculation of weight coefficients. Different weights are assigned to different key information, and the proportion of key information is strengthened by increasing the weight, thereby reducing the loss of key information in long sequence time series signals. The calculation formula of the Attention mechanism is as follows:

u _t = tanh(W _s P _t + b _s )

ν＝∑α _t P _t

Among them, _Pt represents the output feature vector of the BiGRU layer at time step t, _ut represents the hidden layer representation of _Pt obtained through the neural network layer, _us represents the randomly initialized context vector, _αt represents the importance weight obtained by normalizing _ut through the Softmax function, and ν represents the feature vector of the final text information. _us is randomly generated during the training process, and finally the output value ν of the Attention layer is mapped through the Softmax function to obtain the real-time classification result of the tool wear state.

Step 5: Training of the network model: Dropout technology is introduced to prevent the model from overfitting during the training process; the activation function of the network model uses Softmax, and the loss function uses Categorical_crossentropy. The time series signal features obtained in the above steps are used for wear classification to obtain the classification results.

The present invention introduces the Dropout technology into the real-time monitoring model of tool wear status to prevent the model from overfitting during the training process. The activation function of the network model adopts Softmax, and the loss function adopts Categorical_crossentropy to classify the wear of the obtained time series signal features.

The formula is as follows:

y is a vector with the dimension of the number of categories. The value of each dimension is between [0,1], and the sum of all dimensions is 1. This value represents the probability that the tool wear state belongs to a certain category. M is the number of possible categories. During the model training process, the entire model is trained through Categorical_crossentropy Loss. The cross entropy error calculation formula is:

…

表示刀具磨损状态真实类别标签向量中的第i个值，y_im表示Softmax分类器的输出向量y的第i个值。对于所获得的交叉熵误差，最后取其平均作为模型的损失函数。在训练模型的时候采用Adam方法来最小化目标函数，Adam本质上是带有动量项的RMSprop，它利用梯度的一阶矩估计和二阶矩估计动态调整每个参数的学习率。Adam的优点主要在于经过偏置校正后，每一次迭代学习率都有个确定范围，使得参数变化比较平稳。Among them: m represents the number of categories, n represents the number of samples,

represents the i-th value in the true category label vector of the tool wear state, and _yim represents the i-th value of the output vector y of the Softmax classifier. For the obtained cross entropy error, the average is finally taken as the loss function of the model. When training the model, the Adam method is used to minimize the objective function. Adam is essentially RMSprop with a momentum term. It uses the first-order moment estimate and the second-order moment estimate of the gradient to dynamically adjust the learning rate of each parameter. The main advantage of Adam is that after bias correction, the learning rate of each iteration has a certain range, which makes the parameter change relatively stable.

As shown in Figure 2, the CABGRUs network structure diagram. The input data of the CABGRUs model-based neural network includes time series signals, and the feature extraction and expression of time series signals are realized through convolutional layers (C1 and C2), Dropout layers, pooling layers (P1), Flatten layers, BiGRU layers (B1 and B2), Attention layers A1 (attention mechanism) and fully connected layers (F1 and F2).

The time series signal of size (2000,3) is input as input data to the deep learning neural network. The convolution layer C1 uses a 3×1 convolution kernel to convolve the time series signal. The convolution kernel step size is 1 to generate a (20,98,128) feature map, which is input from the convolution layer C1 to the convolution layer C2. The convolution layer C2 uses a 3×1 convolution kernel to convolve the time series signal. The convolution kernel step size is 1 to generate a (20,96,128) feature map, which is input from the convolution layer C2 to the pooling layer P1 through Dropout. The Dropout is 0.5 and the maximum value pooling is used to generate a (20,48,128) feature map, which is input from the pooling layer P1 to the Flatt en layer L1, generates (20,6144) features, which are input to the BiGRU layer B1 by the Flatten layer L1, and generate (20,256) features. The BiGRU layer B1 is input to the BiGRU layer B2, and the (20,256) features are generated. The BiGRU layer B2 is input to the Attention layer A1 through Dropout, and the Dropout is 0.5, generating a 256×1 feature map. The Attention layer A1 is input to the fully connected layer F1, and a 128×1 feature map is output. The fully connected layer F1 is input to the fully connected layer F2, and finally three types of tool wear status values are output to confirm the wear status of the tool at the current moment.

The examples are as follows:

1 Experimental design

(1) Condition monitoring

The experiment of the present invention uses a high-precision CNC vertical milling machine (model: VM600) for milling workpieces. No coolant is added during the milling process. The milling workpiece is mold steel (S136). The milling tool uses an ultra-fine particle tungsten steel carbide four-edge milling cutter, and the blade surface is covered with a TiAIN layer. Table 1 shows the cutting parameters of the milling experiment.

Table 1 Cutting parameters of milling experiment

In the experiment, three acceleration sensors (model: INV9822) were magnetically attached to the machine tool fixture in the x, y, and z directions to collect the original vibration signals generated during the tool processing process in real time; a high-precision digital acquisition instrument (model: INV3018CT) from the Beijing Oriental Vibration and Noise Research Institute was used to process the real-time signals and transmit them to the computer. The sampling frequency of the signal was 20KHz, and each pass of the tool milled 200mm along the x direction, which was recorded as a milling stroke. Each tool milled 330 strokes. After each milling stroke, a pre-calibrated high-precision digital microscope was used to measure the back face wear value of each milling edge of the milling cutter.

(2) Data Analysis

The deep learning hardware platform of the experiment uses a high-performance server: Intel Xeon E5-2650 processor, main frequency 2.3GHz, 256GB memory, GPU uses NVIDIA GeForce TITAN X graphics processor. The software platform uses Ubuntu16.04.4 operating system, and the deep learning framework uses Keras as the front end and TensorFlow as the back end for data analysis.

The experiment uses 4 milling cutters (C1, C2, C3, C4) to complete the milling operation, milling 1320 times, and obtain 1320 original signal samples. The data of 3 milling cutters (C1, C2, C3) are used for model training and verification, and the data of 1 milling cutter (C4) is used for model testing. 80% of the 990 samples are randomly selected as training sets and 20% as verification sets. A sufficient number of samples are required in the deep learning training process to improve the learning quality of the neural network. Data expansion can increase experimental data on the basis of the original order of magnitude of data and improve robustness. The data sample of the original processing signal is a long sequence periodic time series signal. According to the signal sampling principle, the present invention continuously samples and cuts each sample into multiple short sequence time series signals of equal length, which are used as the input of the model after data normalization.

Each sample contains a three-dimensional signal and the wear values of the four back tool faces. To prevent the mutual interference of different blade wear values, the maximum value among the four blades is selected as the wear value of the tool for this milling. The wear state of the tool is divided into: initial wear, normal wear, and rapid wear. The present invention defines the wear state of the tool according to the actual wear curve of each milling cutter, which is used to determine the degree of wear of the tool, divides the degree of wear of the tool into three types of label data, and converts the label data in one-hot encoding form to facilitate the final classification of the tool wear state.

2 Deep Learning Comparative Experimental Results

Table 2 Model specific training parameters

In the experiment, the original signal generated by the milling process is sampled and cropped and then input into the CABGRUs neural network model. The model adaptively extracts the high-dimensional features implicit in the time series signal, calculates the error distance between the actual output value of the model and the true value, and uses the Adam algorithm to reduce the loss and continuously update the network weights, so that the actual output value of the model is closer to the true value. The present invention uses CNN, BiGRU, CBLSTMs deep learning neural networks to compare with the CABGRUs model proposed in the present invention. The same training parameters are set for the four models during the training process. Table 2 is a table of specific training parameters for the model.

After deep learning neural network training and verification, different loss function values and accuracy rates are obtained. Figures 3, 4, 5, and 6 are the loss function values of the training set and verification set output by the CNN, BiGRU, CBLSTMs, and CABGRUs models, as well as the accuracy of the verification set. The x-axis represents the number of iterations of the milling cutter data set, and the double y-axes represent the loss function value and the model verification accuracy rate. It can be concluded from the figure that the loss function value of the network model training set decreases with the increase of the number of iterations and eventually stabilizes. The loss function value of the verification set fluctuates periodically. The oscillation amplitude of the loss function of the CNN and BiGRU network models is large. The CBLSTMs and CABGRUs network models are relatively stable. The overall trend of the loss function continues to decrease and eventually converges. There is no gradient explosion or diffusion phenomenon, and the network converges quickly. The accuracy of the validation set of the CNN and BiGRU network models is 89.75% and 88.02% respectively, and the prediction accuracy is low. This shows that although a single deep learning network can predict the tool wear state, it cannot capture the deeper features hidden in the tool vibration signal due to the limitation of the network model capacity. The CABGRUs proposed in this paper is better than the CNN and BiGRU network models. This is because the network structure is relatively deep, which is conducive to mining deeper features. First, the CNN network can effectively extract the local features hidden in the time series signal, and at the same time compress the length of the time series signal features, which is convenient for the subsequent network to learn the dependency relationship between the time series features of the time series signals, and improve the prediction ability of the model. Compared with the deep CBLSTMs network model, the CABGRUs network model proposed in this paper has achieved higher prediction accuracy. CBLSTMs constructs a two-layer BiLSTM network and uses a bidirectional LSTM network to access past and future information, that is, it can extract time series signal features from both forward and reverse directions, and mine more abundant information features. After 22 iterations, the accuracy of the validation set is basically stable at more than 96%, and the accuracy is 96.75% after 50 iterations. CABGRUs improves the internal structure of neurons on the basis of CBLSTMs and introduces the Attention mechanism to selectively filter out some key information from a large amount of information and focus on it, reducing the loss of key information features of long sequence texts. After 20 iterations, the accuracy of the validation set is basically stable at more than 96%, and after 50 iterations, the accuracy is 98.02%, and the loss function value reaches 0.0595, indicating high network stability. Table 3 shows the loss function and accuracy of the model validation.

Table 3 Loss function and accuracy of model verification

The data of the milling cutter (C4) was selected for the test set of the network model. The total number of test samples was 330, including 23 initial wear samples, 232 normal wear samples, and 75 acute wear samples. The above samples were randomly introduced into the trained CABGRUs network model. From the test results, it can be seen that the CABGRUs network model proposed in the present invention has a strong generalization ability. Although the test time is not as good as some comparative models, the algorithm has found a good balance between time and accuracy. Table 4 shows the test time of a single sample and the accuracy of the model test set.

Table 4 Test time of a single sample and accuracy of the model test set

3. Comparison between Deep Learning and Machine Learning

Table 5 Accuracy of machine learning and deep learning predictions

To further verify the feasibility of the algorithm proposed in the present invention, the experimental data collected by the present invention are used in the tool wear state monitoring model of BP neural network (BPNN), support vector machine (SVM), hidden Markov model (HMM), and fuzzy neural network (FNN). The original signal collected by the acceleration sensor is subjected to wavelet threshold denoising. After feature extraction and feature screening, features with less noise interference and a high degree of correlation with tool wear are obtained. Feature extraction includes time domain features, frequency domain features, and time-frequency domain features. The Pearson correlation coefficient method is used to reflect the degree of correlation between features and wear, and features with a correlation coefficient greater than 0.9 are selected as extraction objects to achieve feature dimensionality reduction, and the extracted features are used as inputs to the machine learning model. The experimental results are compared with those of the CABGRUs network model proposed in the present invention. It can be concluded from the table that the accuracy of traditional machines varies greatly. This is because the instability of manually extracted features and the construction of the model will affect the prediction results. The prediction accuracy of deep learning is significantly higher than that of machine learning BPNN, SVM, and HMM. However, the prediction accuracy of machine learning FNN reaches 94.24%, because FNN uses neural networks to learn the rules of fuzzy systems, automatically designs and adjusts the design parameters of fuzzy systems according to the input and output learning samples, and realizes the self-learning and self-adaptive functions of fuzzy systems. Compared with other algorithm models, the method in this paper has a significant improvement in performance. The test sample speed of the CABGRUs model can reach 8ms, which meets the requirements of real-time monitoring of tool wear status in actual industrial production. Table 5 shows the prediction accuracy of machine learning and deep learning.

In summary, the present invention applies the machine learning method of CNN and RNN fusion to the task of real-time monitoring of tool wear status, and modifies the network parameters and structure in view of the characteristics of high-frequency vibration signal noise and sample redundancy, so as to monitor the degree of tool wear in real time. In the preprocessing stage, the time series signal collected by the acceleration sensor is denoised by wavelet threshold, and the lengthy signal generated by each tool feed is divided into multiple training samples by combining data expansion. Experimental data is added on the basis of the original order of magnitude data to filter out noise and improve the robustness of the algorithm; it is proposed to define the wear state of the tool based on the actual wear curve to determine the degree of tool wear, thereby improving the accuracy of data label classification; a one-dimensional convolutional neural network is used for local feature extraction to mine rich high-dimensional features from the denoised signal, better characterize the tool wear status information hidden in the original signal, and shorten the network model training time; the idea of the Attention mechanism is innovatively introduced into the improved CABGRUs network model, which effectively improves the recognition accuracy and generalization performance of the network model for real-time monitoring. The feasibility of this method is verified by using a real-time monitoring system for tool wear status in this paper. The experiment built a signal acquisition unit and a host computer analysis unit, and used a deep learning framework to predict the tool wear status in real time. The experimental results show that the prediction accuracy of the CABGRUs network model we proposed reached 97.58%, which is better than the traditional machine learning algorithm. At the same time, it can adapt to the hardware system of most production environments and meet industrial requirements in terms of recognition accuracy and recognition speed.

The above description is only a preferred embodiment of the present invention and does not limit the present invention in any form. Any simple modification, equivalent change and modification made to the above embodiment according to the technical essence of the present invention without departing from the technical solution of the present invention still falls within the scope of the technical solution of the present invention.

Claims (3)

1. A cutter wear state monitoring method based on a depth gating cyclic unit neural network comprises the following steps:

the method comprises the following steps: collecting vibration signals generated by a cutter by using a sensor, carrying out denoising treatment on the collected original vibration signals by using a wavelet threshold denoising method, and cutting the vibration signals generated by cutter feeding each time into short sequence time sequence signals with the length of 2000 points;

step two: local feature extraction of single time step time sequence signals: processing the short sequence time sequence signal by adopting a one-dimensional convolutional neural network, wherein the convolutional neural network part comprises 2 convolutional layers CONV and 1 pooling layer POOL, the convolutional layers perform neighborhood filtering on the time sequence signal of each dimension in a one-dimensional convolutional operation mode to generate feature mapping, each feature map can be regarded as convolution operation of different filters on the current time step time sequence signal, namely, the signal self-adaptive feature extraction is performed through the one-dimensional convolutional neural network, the input parameters of a subsequent network are reduced, the calculation speed is improved, meanwhile, the feature map is reduced to a certain extent in the vector dimension, the features of a vibration signal are highlighted, and the subsequent neural network can perform time sequence feature extraction conveniently;

step three: time series signal time series feature extraction: in order to dig the time sequence change rule of relatively longer intervals in a time sequence, extracting time sequence signal time sequence characteristics by adopting an improved depth gate control cycle unit neural network CABGRUS and learning the dependency relationship of the time sequence characteristics among the time sequence signals;

the improved depth gating circulating unit neural network CABGRUs is formed by constructing two depth bidirectional gating circulating unit BiGRU networks which are overlapped together, and meanwhile, an Attention mechanism is introduced into the CABGRUs network to increase an Attention layer, so that the model not only obtains the capability of simultaneously extracting time sequence signal characteristics from the forward direction and the reverse direction, but also obtains the capability of selectively learning key information in the signal characteristics;

each bidirectional BiGRU network in the improved depth-gated cyclic unit neural networks CABGRUs comprises 256 neurons, each of the forward and reverse BiGRU networks consists of 128 neurons, each of the BiGRU neurons comprises an update gate and a reset gate, and the two neurons are used respectively

And &>

Is expressed in that>

Indicates at a time step->

Candidate hidden states in time, based on a time delay>

Indicating at a time step>

Hidden state in time>

Indicates at a time step->

The timed input vector updates the gate>

For controlling how much state information is updated for the current state, based on the status information>

The closer to 1, the more information is utilized indicating that the current state has for the previous time, the reset gate->

For controlling which status information is removed from a previous status, based on the status information determined by the previous status>

The closer to 0, the smaller the proportion of the output state from the previous time, the formula is as follows:

wherein:

and &>

A weight vector representing a reset gate>

And &>

Indicating the right to update the doorA override vector>

And

a weight vector representing a candidate hidden state>

Represents a bias vector, <' > based on the status of the bias>

Representing a Hadamard product, i.e. a dot multiplication of a matrix, is->

Represents a Sigmod function, and the tanh function represents a hyperbolic tangent activation function;

step four: an Attention mechanism is introduced to calculate importance distribution numerical values of continuous time step time sequence signal features, the introduced Attention mechanism carries out weighted summation by distributing different initialization probability weights and each time step output vector of a deep bidirectional gating circulating unit BiGRU layer, and finally numerical values are obtained through calculation of a Sigmod function, so that the proportion of key information is enhanced in a weight increasing mode, and the loss of the key information of long sequence time sequence signals is reduced;

step five: training of a network model: dropout technology is introduced to prevent the model from being over-fitted during the training process; and (3) performing wear classification on the time sequence signal characteristics obtained in the steps by adopting Softmax as an activation function and adopting category _ cross control as a loss function of the network model to obtain a classification result, and confirming the wear state of the tool at the current moment.

2. The tool wear state monitoring method based on the deep gated cyclic unit neural network as claimed in claim 1, wherein: the method is characterized in that a vibration signal generated by cutter feeding each time is cut into a short sequence time sequence signal with the length of 2000, and the method specifically comprises the following steps: and intercepting continuous 100000 points in each sampling signal, and dividing the intercepted points into 50 samples by taking 2000 as the number of the samples, wherein the 50 samples correspond to the same wear state label.

3. The tool wear state monitoring method based on the deep gated cyclic unit neural network as claimed in claim 1, wherein: the equation for the Attention mechanism in step four is as follows:

wherein,

represents an output feature vector at a time step for the BiGRU layer, based on the comparison of the output feature vector at the time step>

Represents->

Hidden layer representation by neural network layer>

Represents a randomly initialized context vector, <' > or>

Represents->

Importance weights normalized by the Softmax function>

Feature vectors representing the final text information, i.e. </or >>

Randomly generated in the training process, and finally the output value of the Attention layer is subjected to the method of Softmax>

And mapping to obtain a real-time classification result of the wear state of the cutter. />

Images