Tool wear prediction method based on time convolution network and auxiliary learning

Abstract

A tool wear prediction method based on time convolution network and auxiliary learning, in the numerical control machining process, the current signal of the spindle motor is collected through the current sensor arranged on the numerical control machine tool, the tool wear after each machining process is measured, the main machining parameters used in the machining process are recorded, and the training sample is generated by intercepting the current signal of the spindle motor according to the time corresponding to the stable feed stage, which is used for training the neural network, in the online test stage, the current signal of the spindle motor in the feed stage of the tool to be predicted is collected and selected as the test sample, and the real-time prediction value of the current tool wear is obtained after inputting the trained neural network.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of mechanical processing, and particularly relates to a tool wear amount prediction method based on a time convolution network and auxiliary learning. BACKGROUND

[0002] The existing tool wear monitoring method is based on direct measurement of offline tool wear amount, needs to take the form of stoppage measurement, needs certain time cost and cannot realize real-time wear monitoring, and cannot meet the production demand of modern high efficiency and high reliability. Therefore, the indirect monitoring method based on a sensing signal is widely researched and applied. The indirect monitoring method generally collects sensing signals in a processing process, such as a spindle current, vibration, acoustic emission and the like, extracts a series of features related to tool wear conditions, and then adopts a regression model to predict the current tool wear amount, so as to realize online tool wear amount prediction. The sensing signals in the machine tool processing process include a large amount of information, and how to obtain the features highly related to the tool wear conditions is the key to determining the prediction accuracy of the wear amount. The existing technology does not have special time sequence processing properties when modeling the time sequence, so some time sequence characteristics included in the sensing signals cannot be well utilized and the prediction accuracy of the wear amount when the tool is in a variable working condition cannot be guaranteed. SUMMARY

[0003] The present application proposes a tool wear amount prediction method based on a time convolution network and auxiliary learning to solve the problems that the existing tool wear amount prediction method is difficult to adapt to variable working conditions and has low prediction accuracy, realizes real-time prediction of the variable working condition adaptability tool wear amount according to the original sensing signal, and effectively improves the prediction accuracy and universality.

[0004] The present application is realized by the following technical scheme:

[0005] The present application relates to a tool wear amount prediction method based on a time convolution network and auxiliary learning. In a numerical control processing process, a current sensor arranged on a numerical control machine tool collects a current signal of a spindle motor, measures the wear amount of a tool after each processing process, records main processing parameters used in the processing process, and generates a training sample by intercepting the current signal of the spindle motor according to the time corresponding to the stable tool walking stage to train a neural network. In an online test stage, the current signal of the spindle motor in the tool walking stage of a tool to be predicted is collected and selected as a test sample, and the current wear amount of the tool is obtained after the test sample is input into the trained neural network.

[0006] The training sample is to divide the stable tool walking stage signal in the collected spindle current signal into a plurality of signal samples with the same length.

[0007] The neural network, in particular a neural network based on a time convolution network and auxiliary learning, is composed of a plurality of sequentially connected time convolution residual blocks, and a main task module and an auxiliary task module connected in parallel, wherein: the time convolution residual block outputs a multi-channel sequence according to the input current signal; the main task module is configured to obtain an h-dimensional vector y composed of the last element in the multi-channel sequence output by the last time convolution residual module T , and obtain a predicted tool wear value through a regression layer, in particular a wear real-time prediction value , wherein: W o and b o are a weight matrix and a bias matrix of the regression layer, respectively; the auxiliary task module is configured to obtain an h-dimensional vector y composed of the last element in the multi-channel sequence output by the last time convolution residual module T , and obtain a predicted machining parameter through a fully connected layer and then through a regression layer, in particular: , wherein: W z and b z are a weight matrix and a bias matrix of the fully connected layer, respectively, is a nonlinear activation function, and z is an output of the fully connected layer; W c and b c are a weight matrix and a bias matrix of the regression layer, respectively, is an output of the regression layer, that is, a predicted machining parameter.

[0008] The training is specifically performed by establishing a loss function based on a network output obtained by inputting a training sample into the neural network, a measured value of the tool wear, and a recorded value of the machining parameter, iteratively optimizing the loss function value, and stopping when a preset iteration step number is reached.

[0009] The measured value of the tool wear refers to a wear value VB of a tool rear face.

[0010] The loss function adopts, but is not limited to, a mean square error function, in particular: , wherein: M is a sample number, and are a main task module output and an auxiliary task module output of the jth sample, o j is a true wear value of the jth sample, c j is a vector composed of a recorded value of the machining parameter of the jth sample, and a is a weight parameter of the auxiliary task, and ||·||2 is a 2-norm of a vector.

[0011] The machining parameter record includes: the feed amount, the cutting depth, and the spindle speed recorded in the numerical control machining process.

[0012] The time convolution residual block comprises two time convolution modules connected in sequence and identical.

[0013] The time convolution module comprises, in sequence: a dilated causal convolution layer, a weight normalization layer, a ReLU layer, and a dropout layer. wherein [f(0), f(1), …, f(k-1)] is a convolution kernel, k is the size of the convolution kernel, and d is the dilation factor; the weight normalization layer normalizes the weight vector of the convolution kernel in the dilated causal convolution layer, that is, wherein w is the weight vector, v and g are a parameter vector and a parameter scalar respectively; the ReLU layer uses a ReLU function on the output sequence of the dilated causal convolution layer, and the ReLU function is f(x) = max{0, x}; and the dropout layer randomly sets part of the channel data to 0 in each training process to improve the generalization performance.

[0014] The dilated causal convolution layer comprises h channels performing dilated causal convolution operations, and thus the output sequence also comprises h channels.

[0015] The weight vector can be replaced by new parameters v and g.

[0016] The dimension of the predicted machining parameter is the same as that of the machining parameter record.

[0017] The training sample has the same length as the sample to be predicted.

[0018] The present application relates to a system for implementing the above method, comprising: a signal acquisition unit, a tool wear amount prediction model training unit, a tool wear amount real-time prediction unit, wherein: the signal acquisition unit is connected with the current sensor arranged on the spindle motor and acquires the spindle current signal in the stable cutting stage, the tool wear amount prediction model training unit acquires the spindle current signal collected by the signal acquisition unit and intercepts a plurality of training samples of the same length, combines the measured value of the tool wear amount and the machining parameter, trains the neural network based on the time convolution network and the auxiliary learning in the unit, arranges the trained neural network in the tool wear amount real-time prediction unit, the tool wear amount real-time prediction unit acquires the real-time cutting signal collected by the signal acquisition unit, intercepts the same length of the predicted sample as the training sample and inputs the neural network arranged in the unit, and obtains the tool wear amount real-time prediction value.

[0019] Technical effects

[0020] The present application solves the problem that the prior art does not consider the time sequence characteristics of the sensor signal or the adaptability of the method under variable machining conditions, resulting in low tool wear amount prediction accuracy, improves the prediction accuracy of the tool wear amount under variable conditions, and has better variable condition tool wear amount prediction accuracy compared with the prior art. BRIEF DESCRIPTION OF DRAWINGS

[0021] Figure 1 The method flowchart of the present application is shown in the figure.

[0022] Figure 2 The structure diagram of the neural network of the embodiment is shown in the figure. DETAILED DESCRIPTION

[0023] The present embodiment relates to a tool wear amount prediction method based on time convolution network and auxiliary learning, which uses NASA's milling experiment data to verify the tool wear amount real-time prediction method proposed by the present application. The experiment uses two different materials of the tool, namely cast iron and stainless steel, and carries out milling experiment under variable conditions on the numerical control machine tool. In each experiment, a new tool is used for repeated milling processing, and the spindle motor current, spindle vibration, spindle acoustic emission and other signals in the processing process are collected. After each processing, the tool wear amount VB is measured under the microscope. The spindle speed during milling is 826 revolutions per minute, and the cutting depth and feed rate are selected differently in different experiments. The present embodiment selects the first 12 experiments for verification, and the specific parameter settings of each experiment are shown in Table 1.

[0024] Table 1 Specific parameter settings of each experiment

[0025]

[0026]

[0027] As Figure 1 shown, the specific steps of the embodiment include:

[0028] Step 1: Perform numerical control machining, collect the current signal of the spindle motor from the current sensor arranged on the numerical control machine tool, and record the main machining parameters used in the machining process;

[0029] Step 2: Measure the tool wear after each machining process under a microscope;

[0030] Step 3: Cut the stable cutting stage signal from the collected spindle current signal, and divide it into multiple data samples of the same length N for training;

[0031] Step 4: Construct a neural network based on time convolution network and auxiliary learning, including multiple time convolution residual blocks connected in sequence, a main task module, and an auxiliary task module, and input the data samples for training into the network;

[0032] Step 5: After inputting the training samples into the neural network based on time convolution network and auxiliary learning, the network output is obtained, and a loss function is established with the real measured value of tool wear and the recorded value of machining parameters. The loss function value is iteratively optimized, and the training is stopped after reaching the preset iteration step number to obtain the trained network;

[0033] Step 6: Collect the spindle current signal of the tool to be predicted during the cutting stage, select a certain length of data sample and input it into the trained network, and the output of the main task module is the real-time prediction value of the current tool wear.

[0034] Preferably, the main machining parameters in step 1 generally include feed rate, cutting depth, spindle speed, etc.

[0035] In this embodiment, the cutting depth and feed rate are selected as the main machining parameters for subsequent consideration because the spindle speed of each group of experiments remains unchanged.

[0036] In this embodiment, the wear measurement in step 2 selects the wear value VB of the tool flank face.

[0037] In this embodiment, step 3 is implemented as follows: the spindle current signal collected during each machining process can clearly reflect that the signal amplitude of the tool increases significantly when it contacts the workpiece and decreases significantly when it leaves the workpiece. Remove the signals in these two stages, and select the signals in the stable amplitude stage during each machining. In this embodiment, the 3500-5500 data points of each machining process are selected and divided into data samples with a length of N=100.

[0038] Preferably, the neural network based on temporal convolutional network and auxiliary learning in step 4 includes multiple sequentially connected temporal convolutional residual blocks, a main task module, and an auxiliary task module. The main task module and the auxiliary task module share the output of the last temporal convolutional residual module as the module input.

[0039] like Figure 2 As shown, the network includes five sequentially connected temporal convolutional residual blocks. Each temporal convolutional residual block consists of two identical temporal convolutional modules connected in sequence. The output of the temporal convolutional residual block is the sum of the input and the output of the second temporal convolutional module.

[0040] The temporal convolution module comprises, in sequence: a dilated causal convolutional layer, a weight normalization layer, a ReLU layer, and a dropout layer. Specifically, the dilated causal convolutional layer performs a dilated causal convolution operation on the input sequence x to obtain the output sequence, where the nth element of the output sequence... Where: [f(0), f(1), ..., f(k-1)] is a convolution kernel, k is the kernel size, and d is the dilation factor; the weight normalization layer normalizes the weight vector of the convolution kernel in the dilated causal convolution layer to improve training speed, specifically as follows: Where: w is the weight vector, v and g are the parameter vector and parameter scalar, respectively, and the weight vector can be replaced by new parameters v and g; the ReLU layer uses the ReLU function to act on the output sequence of the dilated causal convolutional layer, and the expression of the ReLU function is f(x) = max{0, x}; the dropout layer randomly sets a certain proportion of channel data to 0 in each round of training to improve generalization.

[0041] In this embodiment, the dilated causal convolutional layer includes h channels that perform dilated causal convolution operations, and therefore the output sequence also includes h channels. In this embodiment, the number of channels h of the dilated causal convolutional layer is 32, the kernel size k is 7, and the dilation factor d increases in multiples of 2 with the order of the temporal convolutional residual blocks, that is, d=1 in the first temporal convolutional residual block, d=2 in the second temporal convolutional residual block, d=4 in the third temporal convolutional residual block, and so on.

[0042] In this embodiment, the proportion of certain channel data set to 0 during each round of training in the dropout layer is 0.1.

[0043] The main task module uses the multi-channel sequence output by the last temporal convolutional residual module as input, and selects the last element of the sequence to form an h-dimensional vector y. T The predicted tool wear is obtained through a regression layer. Among them: W o and bo are the weight matrix and bias matrix of the regression layer, respectively.

[0044] The auxiliary task module also takes the multi-channel sequence output by the last time convolution residual module as input, selects the last element of the sequence to form a h-dimensional vector y T , first passes through a fully connected layer where W f and b f are the weight matrix and bias matrix of the fully connected layer, respectively, represents a nonlinear activation function. Then, a regression layer is used to obtain the predicted machining parameters, and the output dimension is set to be the same as the number of main machining parameters recorded in step 1 where W c and b c are the weight matrix and bias matrix of the regression layer, respectively.

[0045] In the embodiment, the output dimension of the fully connected layer is set to 8, and the output dimension of the regression layer is consistent with the number of selected main machining parameters, which is 2. The nonlinear activation function of the fully connected layer is ReLU function.

[0046] Preferably, the network output includes the output of the main task module and the output of the auxiliary task module

[0047] Preferably, the loss function is generally selected as mean square error function where M is the number of samples, and are the main task module output and auxiliary task module output of the jth sample, o j is the true wear amount of the jth sample, c j is the vector composed of the machining parameter record value of the jth sample, and α is the weight parameter of the auxiliary task. ||·||2 is the 2-norm of the vector.

[0048] In this embodiment, the weight parameter of the auxiliary task is set to 0.5, the Adam optimization algorithm is used to iteratively optimize the loss function value, the learning rate is 0.001, and the training is stopped after 20 iterations.

[0049] Preferably, step 6 selects the spindle current signal of the tool to be predicted according to the same sample length N = 100 in step 3.

[0050] In the embodiments, the 2nd and 10th groups of experiments with different processing parameters are selected as test for real-time prediction of the wear amount, and the other 10 groups of experimental data are used for training of the model. The spindle current signals collected in the test experiments are also intercepted as test samples with a length of 100, which are input into the trained network to obtain the predicted value of the wear amount of the tool. The root-mean-squared-error (RMSE) and the mean-absolute-error (MAE) of the predicted results and the true values are used to measure the accuracy of the predicted results, and Table 2 shows the root-mean-squared-error and the mean-absolute-error of the predicted results of the test samples of the 2nd and 10th groups of experiments obtained by using the method provided in the embodiments and the convolutional neural network and the long short-term memory neural network.

[0051] Table 2 Root-mean-squared-error and mean-absolute-error of the predicted results obtained by using the method and the convolutional neural network and the long short-term memory neural network

[0052]

[0053] The results in Table 2 show that the error of the predicted wear amount of the milling cutter by using the method provided in the embodiments is smaller compared with the true value, and compared with the two prior art methods, the method has lower prediction error, i.e. higher prediction accuracy of the tool wear amount in the two groups of experiments.

[0054] In summary, the method provided in the embodiments uses the time convolution network to perform time series processing on the spindle current signals and then predicts the tool wear amount, and at the same time, the adaptability of the method under varying working conditions is improved by constructing an auxiliary learning task based on identification of the working conditions, and by constructing the neural network based on the time convolution network and the auxiliary learning, the time series features are extracted by using multiple time convolution residual blocks connected in sequence, and an auxiliary task module is included for identifying the working conditions.

[0055] The above specific embodiments can be adjusted in different ways by those skilled in the art without departing from the principles and purposes of the embodiments, and the protection scope of the embodiments is subject to the claims and is not limited by the above specific embodiments, and each implementation scheme within the scope is subject to the constraints of the embodiments.

Claims

1. A tool wear amount prediction method based on a time convolution network and auxiliary learning, characterized by, In the numerical control machining process, the current signal of the main shaft motor is collected through the current sensor arranged on the numerical control machine tool, the wear amount of the tool after each machining process is measured, the main machining parameters used in the machining process are recorded, and the training sample is generated by intercepting the current signal of the main shaft motor corresponding to the stable tool running stage time for training the neural network. In the online test stage, the current signal of the main shaft motor in the tool running stage of the tool to be predicted is collected and selected as the test sample, which is input into the trained neural network to obtain the real-time prediction value of the current wear amount of the tool. The training sample is to divide the stable tool running stage signal in the collected main shaft current signal into a plurality of signal samples with the same length. The neural network, in particular, a neural network based on a time convolution network and auxiliary learning, is composed of a plurality of time convolution residual blocks connected in sequence, and a main task module and an auxiliary task module connected in parallel, wherein the time convolution residual block outputs a multi-channel sequence according to an input current signal; the main task module is composed of the last element in the multi-channel sequence output by the last time convolution residual module dimensional vector The predicted tool wear amount is obtained through a regression layer, in particular, a wear amount real-time prediction value , wherein and are a weight matrix and a bias matrix of the regression layer respectively; the auxiliary task module is composed of the last element in the multi-channel sequence output by the last time convolution residual module dimensional vector The predicted machining parameter is obtained through a fully connected layer and then through a regression layer, in particular, , , wherein and are a weight matrix and a bias matrix of the fully connected layer respectively, is a nonlinear activation function, and z is an output of the fully connected layer; and are a weight matrix and a bias matrix of the regression layer respectively, is an output of the regression layer, that is, a predicted machining parameter.

2. The tool wear amount prediction method based on a time convolution network and auxiliary learning according to claim 1, characterized in that, The training is specifically to establish a loss function by inputting the network output obtained by inputting the training sample into the neural network and the measured value of the tool wear amount, i.e. the measured wear value VB of the tool flank face and the recorded value of the machining parameter, iteratively optimize the loss function value, and stop after reaching the preset iteration step number.

3. The tool wear amount prediction method based on a time convolution network and auxiliary learning according to claim 2, characterized by, The loss function mentioned uses the mean squared error function, specifically: ,in: Sample size and The first The main task module output and the auxiliary task module output for each sample No. The actual wear amount of each sample No. A vector composed of the processing parameter records of each sample. Weight parameters for auxiliary tasks, The 2-norm of a vector.

4. The tool wear amount prediction method based on a time convolution network and auxiliary learning according to claim 1, characterized in that, The machining parameter record includes the feed amount, cutting depth and main shaft speed recorded in the numerical control machining process, i.e. the cutting feed amount, cutting depth and main shaft speed set before the machining starts, which can be obtained from the numerical control code of the machining process. Only the main machining parameters, i.e. the parameters that change significantly in multiple machining processes, need to be recorded to simplify the calculation amount.

5. The tool wear amount prediction method based on a time convolution network and auxiliary learning according to claim 1, characterized in that, The time convolution residual block includes two time convolution modules connected in sequence and completely the same, and the output of the time convolution residual block is equal to the sum of the input of the first time convolution module and the output of the second time convolution module.

6. The tool wear amount prediction method based on a time convolution network and auxiliary learning according to claim 1, characterized by, The time convolution module sequentially comprises a dilated causal convolution layer, a weight normalization layer, a ReLU layer and a dropout layer, wherein the dilated causal convolution layer performs a dilated causal convolution operation on an input sequence x to obtain an output sequence, the nth element of the output sequence is wherein is a convolution kernel, is a convolution kernel size, is a dilation factor; the weight normalization layer normalizes a weight vector of the convolution kernel in the dilated causal convolution layer, that is is a weight vector, and are a parameter vector and a parameter scalar respectively; the ReLU layer uses a ReLU function on the output sequence of the dilated causal convolution layer, the ReLU function is ; the dropout layer randomly sets part of the channel data to 0 in each training process to improve the generalization performance.

7. The tool wear amount prediction method based on a time convolution network and auxiliary learning according to claim 1, characterized in that, The dimension of the predicted machining parameter is the same as that of the machining parameter record, and the length of the training sample is the same as that of the test sample.

8. A system for implementing the tool wear amount prediction method based on a time convolution network and auxiliary learning according to any one of claims 1 to 7, characterized by, It comprises: a signal acquisition unit, a tool wear amount prediction model training unit and a tool wear amount real-time prediction unit, wherein the signal acquisition unit is connected with the current sensor arranged on the main shaft motor and collects the main shaft current signal in the stable tool running stage, the tool wear amount prediction model training unit obtains the main shaft current signal collected by the signal acquisition unit and intercepts it into a plurality of training samples with the same length, trains the neural network based on the time convolution network and auxiliary learning in the unit combined with the measured value of the tool wear amount and the machining parameter, arranges the trained neural network in the tool wear amount real-time prediction unit, the tool wear amount real-time prediction unit obtains the real-time tool running signal collected by the signal acquisition unit, intercepts it into a test sample with the same length as the training sample, and inputs it into the neural network arranged in the unit to obtain the real-time prediction value of the tool wear amount.

Images