1. A method for predicting the wear state of a cutter in a process is characterized by comprising the following steps:

firstly, obtaining a cutter abrasion curve and cutter abrasion data through a basic cutter abrasion test;

dividing the tool wear data obtained in the step one into a training set T1 and a testing set T2 through moving a sliding window;

thirdly, constructing a dense residual error neural network, and training the dense residual error neural network by using the training set T1 in the second step;

301, constructing a residual error neural network, wherein the residual error neural network is formed by a batch processing standardized layer and a nonlinear function layer which skip two to three layers; when the input to the neural network is x, the desired output is h (x); the residual neural network directly transmits the input x of the model to the layer behind the network, and the output y of the network is the superposition of the mapping and the input, as shown in the formula:

y＝F(x,{W_i})+x (1)

where y denotes the output vector of the network, x denotes the input of the network, and F (x, { W) }_i}) represents the residual map that needs to be learned, W_iRepresenting network needsA weight to learn;

by means of recursive derivation, the network output when the layer depth is m can be obtained as follows:

wherein, y_mRepresents the output at layer depth m, x_iRepresents a feature with a layer depth of i, W_iA weight representing a layer depth of i;

302 based on the residual error neural network in step 301, combining the regular information flow and the output of the dense layer of the residual error module, and directly connecting the dense layer and the input to complete the construction of the dense residual error neural network;

303, training the weights in the dense residual error neural network by using the training set in the step two, and training the model by using a Huber loss function as an optimization function, wherein the Huber loss function and corresponding derivatives are shown as a formula (3) and a formula (4):

wherein p is_w(xⁱ) Representing the predicted values, q, obtained by a dense residual neural networkⁱRepresenting the actual tool wear value, δ representing the cutoff over-parameter;

304 training parameters in the dense residual error neural network by using a Huber loss function as a target and through a chain derivation method of back propagation, wherein the expression is as follows:

wherein ε represents the Huber loss function,representing the derivation of a partial derivative;

fourthly, the prediction of the tool wear state in the process is realized through the trained dense residual error neural network in the third step, and the tool wear prediction error of a training set T1 is obtained;

fifthly, training a moving average integration autoregressive model by using the tool wear prediction error of the training set T1 in the fourth step, and correcting the tool wear prediction error;

501, training the moving average integration autoregressive model by using the tool wear prediction error data in the step IV;

502 converts a non-stationary time series into a stationary time series, for time series x (t), t 1.., N, a moving average integrated autoregressive model as:

wherein φ (B) represents an autoregressive model, φ₁,...,φ_pDenotes an autoregressive coefficient, p denotes an autoregressive order, B denotes a delay operator, theta (B) denotes a moving average function, theta₁,...,θ_qRepresents a moving average coefficient, q represents a moving average index;

503, based on the step 502, optimizing the autoregressive order p and the moving average index q, and determining the values of p and q by using the akage pool information criterion, wherein the expression of the akage pool information criterion is shown as the following formula:

AIC＝2k-2ln(L) (7)

wherein AIC represents the Chichi information criterion, k represents the number of parameters, and L represents a likelihood function;

504, predicting the tool wear prediction error of m steps in the future by using the trained moving average integrated autoregressive model on the premise of giving out the initial tool prediction error of the test set;

and sixthly, accurately predicting the tool wear trend by using the prediction error of the tool wear in the process obtained by prediction in the fifth step and the prediction result of the tool wear state in the process in the fourth step, namely, predicting the tool wear state in the future process by adding the prediction error of the tool wear in the future m steps in the step fifth step and the prediction value of the tool wear in the future m steps in the step fourth step.

2. The method for predicting the wear state of the in-process tool according to claim 1, wherein the step (ii) comprises:

201, dividing cutter wear data in the first step into sub-data groups through a mobile sliding window operation, wherein the sub-data groups are divided into a training set T1 and a testing set T2;

202 divide the tool wear curve VB ═ {0.04,0.05,0.07,0.08,0.09,0.10,0.11,0.12,0.13,.., 0.3} into sub-data sets by moving the sliding window, which operates as shown in equation:

3. the method for predicting the wear state of the in-process tool according to claim 1, wherein the step (iv) comprises:

401 establishing a mapping relationship between tool wear data and tool wear states in a future process through a trained dense residual error neural network, as follows:

(VB_t+1,VB_t+2,...,VB_t+m)＝H(VB_t-n+1,VB_t-n+1,...,VB_t) (9)

where H represents the well-trained dense residual neural network, (VB)_t+1,VB_t+2,...,VB_t+m) Representing the tool wear Value (VB) of the future m steps predicted by the model_t-n+1,VB_t-n+1,...,VB_t) Historical tool wear data representing the last n steps;

402, predicting the cutter wear value of the training set T1 by using the trained dense residual neural network, and subtracting the cutter wear value predicted by the model from the actual cutter wear value to obtain the cutter wear prediction error of the training set T1.

Background

In the cutting process, the surface quality of a workpiece is reduced due to tool abrasion, particularly in the severe abrasion stage of a tool, the tool abrasion value is greatly changed in a short time, so that the dimensional precision of a machined part cannot meet the use requirement, and in addition, compared with the monitoring of the current tool abrasion state, the prediction of the multi-step tool abrasion state is more meaningful for data-driven intelligent manufacturing. Therefore, in order to obtain parts satisfying dimensional accuracy, it is necessary to predict changes in the tool wear value over a future period of time. The prediction of the future multi-step tool wear value belongs to a time series prediction problem, and with the development of artificial intelligence technology, a deep learning model and machine learning make the prediction of the multi-step tool wear value possible. At present, most researches on tool wear prediction are still on the basis of multi-sensor signal fusion, a current tool wear value is monitored through a machine learning or deep learning model, few researches relate to prediction of future multi-step tool wear values, and in addition, few researches for correcting tool wear prediction errors so as to further improve the tool wear prediction precision. Therefore, the current tool wear prediction research has limitations, the obtained tool wear monitoring model can only realize the prediction of the current tool wear value, and the condition that the tool wear value cannot be changed too much in the severe wear stage of the tool is not considered, so that the dimensional accuracy of the processed part can meet the use requirement.

Disclosure of Invention

The technical problem to be solved by the invention is to provide a method for predicting the wear state of a multi-step tool by taking tool wear data as a basis, considering the nonlinear dependence relationship between the tool wear trend and the existing tool wear data, realizing the prediction of the wear state of the multi-step tool through a dense residual error neural network, correcting the tool wear prediction error by utilizing a moving average integration autoregressive model, and accurately realizing the prediction of the wear trend of the multi-step tool.

In order to solve the technical problems, the invention adopts the following technical scheme that the invention comprises the following steps:

firstly, obtaining a cutter abrasion curve and cutter abrasion data through a basic cutter abrasion test;

dividing the tool wear data obtained in the step one into a training set T1 and a testing set T2 through moving a sliding window;

thirdly, constructing a dense residual error neural network, and training the dense residual error neural network by using the training set T1 in the second step;

301, constructing a residual error neural network, wherein the residual error neural network is formed by a batch processing standardized layer and a nonlinear function layer which skip two to three layers; when the input to the neural network is x, the desired output is h (x); the residual neural network directly transmits the input x of the model to the layer behind the network, and the output y of the network is the superposition of the mapping and the input, as shown in the formula:

y＝F(x,{W_i})+x (1)

where y denotes the output vector of the network, x denotes the input of the network, and F (x, { W) }_i}) represents the residual map that needs to be learned, W_iA weight indicating that the network needs to learn;

by means of recursive derivation, the network output when the layer depth is m can be obtained as follows:

wherein, y_mRepresents the output at layer depth m, x_iRepresents a feature with a layer depth of i, W_iA weight representing a layer depth of i;

302 based on the residual error neural network in step 301, combining the regular information flow and the output of the dense layer of the residual error module, and directly connecting the dense layer and the input to complete the construction of the dense residual error neural network;

303, training the weights in the dense residual error neural network by using the training set in the step two, and training the model by using a Huber loss function as an optimization function, wherein the Huber loss function and corresponding derivatives are shown as a formula (3) and a formula (4):

wherein p is_w(xⁱ) Representing the predicted values, q, obtained by a dense residual neural networkⁱRepresenting the actual tool wear value, δ representing the cutoff over-parameter;

304 training parameters in the dense residual error neural network by using a Huber loss function as a target and through a chain derivation method of back propagation, wherein the expression is as follows:

wherein ε represents the Huber loss function,representing the derivation of a partial derivative;

fourthly, the prediction of the tool wear state in the process is realized through the trained dense residual error neural network in the third step, and the tool wear prediction error of a training set T1 is obtained;

fifthly, training a moving average integration autoregressive model by using the tool wear prediction error of the training set T1 in the fourth step, and correcting the tool wear prediction error;

501, training the moving average integration autoregressive model by using the tool wear prediction error data in the step IV;

502 converts a non-stationary time series into a stationary time series, for time series x (t), t 1.., N, a moving average integrated autoregressive model as:

wherein φ (B) represents an autoregressive model, φ₁,...,φ_pDenotes an autoregressive coefficient, p denotes an autoregressive order, B denotes a delay operator, theta (B) denotes a moving average function, theta₁,...,θ_qRepresents a moving average coefficient, q represents a moving average index;

503, based on the step 502, optimizing the autoregressive order p and the moving average index q, and determining the values of p and q by using the akage pool information criterion, wherein the expression of the akage pool information criterion is shown as the following formula:

AIC＝2k-2ln(L) (7)

wherein AIC represents the Chichi information criterion, k represents the number of parameters, and L represents a likelihood function;

504, predicting the tool wear prediction error of m steps in the future by using the trained moving average integrated autoregressive model on the premise of giving out the initial tool prediction error of the test set;

and sixthly, accurately predicting the tool wear trend by using the prediction error of the tool wear in the process obtained by prediction in the fifth step and the prediction result of the tool wear state in the process in the fourth step, namely, predicting the tool wear state in the future process by adding the prediction error of the tool wear in the future m steps in the step fifth step and the prediction value of the tool wear in the future m steps in the step fourth step.

The step II comprises the following steps: 201, dividing cutter wear data in the first step into sub-data groups through a mobile sliding window operation, wherein the sub-data groups are divided into a training set T1 and a testing set T2;

202 divide the tool wear curve VB ═ {0.04,0.05,0.07,0.08,0.09,0.10,0.11,0.12,0.13,.., 0.3} into sub-data sets by moving the sliding window, which operates as shown in equation:

the fourth step comprises the following steps: 401 establishing a mapping relationship between tool wear data and tool wear states in a future process through a trained dense residual error neural network, as follows:

(VB_t+1,VB_t+2,...,VB_t+m)＝H(VB_t-n+1,VB_t-n+1,...,VB_t) (9)

where H represents the well-trained dense residual neural network, (VB)_t+1,VB_t+2,...,VB_t+m) Representing the tool wear Value (VB) of the future m steps predicted by the model_t-n+1,VB_t-n+1,...,VB_t) Historical tool wear data representing the last n steps;

402, predicting the cutter wear value of the training set T1 by using the trained dense residual neural network, and subtracting the cutter wear value predicted by the model from the actual cutter wear value to obtain the cutter wear prediction error of the training set T1.

The invention has the following positive effects: according to the method, cutter wear data are obtained through a cutter wear test, a group of cutter wear data are divided into a plurality of groups of cutter wear data through a data processing method of a movable sliding window, the cutter wear data are divided into a training set T1 and a testing set T2, a dense residual error neural network is established by combining a dense layer and an input layer on the basis of a typical residual error neural network, a Huber loss function is taken as an optimization target, parameters in the dense residual error neural network are trained by using the training set, the trained dense residual error neural network is taken as the basis, recent historical cutter wear data are taken as model input, and the cutter wear state in the future process is predicted; the method takes the prediction error of the time sequence prediction problem into consideration, trains the moving average integration autoregressive model by utilizing the tool wear prediction error of a training set, obtains the tool wear prediction error in the future process by predicting the trained moving average integration autoregressive model on the premise of giving the initial tool wear prediction error, and accumulates the predicted tool wear prediction error and the predicted tool wear value in the future process to correct the tool wear prediction precision, thereby further improving the prediction accuracy of the tool wear state in the process; the method can judge whether the tool wear value changes greatly in a period of time in the future so as to ensure that the dimensional accuracy of the processed part meets the use requirement, and makes a decision on whether the tool is replaced according to the predicted tool wear state in the future process, so that the production efficiency is improved and the cost is reduced on the premise of ensuring the processing quality, and the method has important significance for the current data-driven intelligent manufacturing.

Drawings

FIG. 1 is a general flow chart of future multi-step tool wear prediction

FIG. 2 is a process diagram of tool wear data processing

FIG. 3 is a diagram of a constructed dense residual error neural network

FIG. 4 is a graph of predicted tool wear without error correction

FIG. 5 is a graph showing the prediction result of the tool wear prediction error

FIG. 6 is a graph showing the result of prediction of tool wear in consideration of error correction

Detailed Description

The invention is described in detail below with reference to the drawings and the specific cutting examples. As shown in fig. 1, in order to realize the overall flow chart of multi-step tool wear prediction, the main steps are as follows:

firstly, obtaining a cutter abrasion curve and cutter abrasion data through a basic cutter abrasion test;

dividing the tool wear data obtained in the step one into a training set T1 and a testing set T2 through moving a sliding window;

201, preprocessing tool wear data, as shown in fig. 2, dividing the tool wear data obtained by the experiment into subgroup data through a moving sliding window operation, and dividing the subgroup data into a training set T1 and a testing set T2;

202, the sliding window operation can expand a data set, ensure the relevance among data and realize the pretreatment of tool wear data;

203, for a set of tool wear curves VB, {0.04,0.05,0.07,0.08,0.09,0.10,0.11,0.12,0.13,. and 0.3}, the tool wear curves can be divided into multiple sets of data through a moving sliding window operation, and the data are used for training and verifying a subsequent multi-step tool wear prediction model, and the specific operation is as shown in formula 1:

204, after the data in 203 is processed, the data can be divided into a training set T1 and a test set T2, wherein the training set T1 is used for training the time series model, and the test set T2 is used for verifying the prediction effect of the time series model;

205, for the dense residual error neural network, the historical tool wear data of the previous 2 steps is used as the input of a model, and the tool wear state of the future 5 steps is predicted; and (5) predicting the tool wear prediction error of the future 5 steps by using the moving average integrated autoregressive model.

Thirdly, building a dense residual error neural network:

301, in order to realize the prediction of the wear state of the cutter in the process, a time sequence prediction model needs to be built;

compared with the classical machine learning technology (decision tree, multilayer perceptron, convolutional neural network, long-time memory network and the like), the 302 dense residual neural network can accurately realize the association mapping between the historical cutter wear data and the future cutter wear state, the parameters involved in the model are less, and the model training time can be greatly shortened, wherein the structure of the network is shown in FIG. 3, wherein W represents a hidden layer unit, and ReLU represents an activation function;

303, building a dense residual error neural network through a Tensorflow framework. Dense residual neural networks were developed based on the residual neural networks. A typical residual neural network may be formed by skipping two to three batch normalization layers and nonlinear function layers;

the 304 residual neural network is realized as follows: assuming that the input of the neural network is x, the expected output is h (x), the traditional network structure fits h (x) through the superposition of nonlinear mapping, each layer of the network involves a complex gradient derivation problem, and the training difficulty of the model is large. The residual error neural network directly transmits the input x of the model to a layer behind the network in a shortcut connection mode, and the output y of the network is the superposition of mapping and input, as shown in formula 2:

y＝F(x,{W_i})+x (2)

where y denotes the output vector of the network, x denotes the input of the network, and F (x, { W) }_i}) represents the residual map that needs to be learned, W_iRepresenting the weights that the network needs to learn.

By means of recursive derivation, the network output when the layer depth is m can be obtained as follows:

wherein, y_mRepresents the output at layer depth m, x_iRepresents a feature with a layer depth of i, W_iA weight representing a layer depth of i; the 304 dense residual error neural network is a further extension of the residual error neural network, combines the regular information flow and the output of the dense layer of the previous residual error module, and can realize the construction of the dense residual error neural network by directly connecting the subsequent dense layer and the input; 305 then training the weights in the dense residual neural network by using a training set T1 in the step II, and training the model by using a Huber loss function as an optimization function in the training process, wherein the Huber loss function and corresponding derivatives are shown as formulas 4 and 5:

wherein p is_w(xⁱ) Representing the predicted values, q, obtained by a dense residual neural networkⁱRepresenting the actual cutter wear value, delta representing a cutoff over-parameter, and generally taking the value of 2;

306, training parameters in the dense residual error neural network is realized by taking a Huber loss function as a target through a chain derivation method of back propagation, and the expression is shown as the following formula:

wherein ε represents the Huber loss function,representing derivation；

307, it can be seen from the back propagation derivation formula that the dense residual error neural network has no problem of gradient explosion or gradient disappearance in the training process, and the jump connection mode ensures that the model has a faster convergence speed.

Prediction of future multi-step tool wear status:

401, a mapping relation between historical tool wear data and future multi-step tool wear states can be established through a trained dense residual error neural network, as shown in the following formula:

(VB_t+1,VB_t+2,...,VB_t+m)＝H(VB_t-n+1,VB_t-n+1,...,VB_t) (7)

where H represents the well-trained dense residual neural network, (VB)_t+1,VB_t+2,...,VB_t+m) Representing the tool wear Value (VB) of the future m steps predicted by the model_t-n+1,VB_t-n+1,...,VB_t) Historical tool wear data representing the last n steps;

402, predicting to obtain a tool wear value of the training set T1 by using the trained dense residual neural network, and subtracting the tool wear value predicted by the model from the actual tool wear value to obtain a tool wear error of the training set T1.

Establishing a cutter wear prediction error correction model:

501, in order to further improve the prediction accuracy of the future multi-step tool wear value, training a moving average integration autoregressive model by using tool wear error data in 402;

the 502 moving average integrated autoregressive model belongs to a time series prediction model, and the basic theory of the moving average integrated autoregressive model is to convert a non-stationary time series into a stationary time series, and for the time series x (t), t is 1.

Wherein φ (B) represents an autoregressive model, φ₁,...,φ_pDenotes an autoregressive coefficient, p denotes an autoregressive order, B denotes a delay operator, theta (B) denotes a moving average function, theta₁,...,θ_qRepresents a moving average coefficient, q represents a moving average index;

503, it can be known from 502 that in order to ensure that the moving average integrated autoregressive model has a better prediction effect, the autoregressive order p and the moving average index q need to be optimized, and the values of p and q are generally determined by using the akachi pool information criterion, whose expression is shown as the following formula:

AIC＝2k-2ln(L) (9)

wherein AIC represents the akabane information criterion, k represents the number of parameters, and L represents the likelihood function.

504, the trained moving average integrated autoregressive model is utilized, and on the premise of giving out an initial tool prediction error, the tool wear prediction error of m steps in the future can be predicted.

And (3) final realization of future multi-step tool wear state prediction:

601, adding the predicted future m-step tool wear error in the step 401 to the predicted future m-step tool wear value in the step 504 to finally obtain a multi-step tool wear predicted value.

DETAILED DESCRIPTION OF EMBODIMENT (S) OF INVENTION

The part is mainly based on the turning test, workpieces of different models are cut under different cutting parameters and different types of tools, the rear tool face of the tool is shot after each test is finished, and the wear width of the rear tool face of the turning tool is measured to serve as a wear standard. Wherein the cutting parameter combinations are shown in table 1:

table 1: combination of cutting parameters for different tool and material types

The material adopted in the test is 1Cr18Ni9Ti and 38CrSi, the cutting test is carried out by adopting a blade with the model number of VBMT160408-HMP and DNMG150408HQ, the test platform is TC-HAWK150 CNC, and the wear width of the back face of the cutting tool is measured by utilizing a Keyence VK-100 shape measurement laser microscope.

A training set and a test set are obtained by moving a sliding window according to the tool wear data, the tool wear data obtained when 1Cr18Ni9Ti material is turned is used as the training set, and the tool wear data obtained when 38CrSi material is turned is used as the test set.

Firstly, training a dense residual error neural network by using training set data, verifying the quality of a model by using tool wear data of a test set after the training is finished, wherein the quality of the model is verified by using tool wear data obtained when a 38CrSi material is turned at a cutting speed of 160m/min, the tool wear states of five future steps are predicted by using historical tool wear of the previous two steps as the input of the model, and the tool wear prediction results of the five future steps are shown in fig. 4. The actual value of the tool wear is consistent with the predicted value of the tool wear in trend, the prediction accuracy of the model is high, in order to quantitatively evaluate the performance of the model, the average error calculated according to the prediction result is 1.8212 mu m, the mean square error is 3.4908 mu m, and the model can be used for predicting the wear state of the future multi-step tool.

Observing fig. 4, it can be known that: the prediction errors are large in the initial tool wear stage and the severe tool wear stage, and in order to further improve the prediction accuracy of the future multi-step tool wear state, the prediction errors of the parts need to be corrected. The tool wear data of the training set is used as the input of the dense residual error neural network, the multi-step tool wear state prediction result corresponding to the training set can be obtained, the tool wear prediction error data set can be obtained by subtracting the tool wear value obtained through prediction from the actual tool wear value of the training set, the moving average integration autoregressive model is trained on the basis of the tool wear prediction error data set, on the premise that the initial five-step tool wear prediction error of the test set is given, the correction of the future multi-step tool wear prediction result can be achieved, and the actual tool wear prediction error and the tool wear prediction error obtained through prediction are shown in figure 5.

The tool wear prediction error of the five-step tool in the future can be corrected by adding the tool wear prediction error of the five-step tool in the future obtained by the prediction of the moving average integrated autoregressive model and the tool wear value of the five-step tool in the future obtained by the prediction of the dense residual error neural network, the prediction result after the tool wear prediction error is corrected is considered to be shown in FIG. 6, the average error obtained by calculation according to the prediction result is 1.0758 micrometers, the mean square error is 3.4908 micrometers, the tool wear state prediction accuracy of the five steps in the future is improved by 40.92% by taking the average error as an evaluation standard, the robustness of the time series prediction model is greatly improved, and the time series model built on the basis of the dense residual error neural network and the moving average integrated autoregressive model can be used for predicting the wear state of the multi-step tool in the future.

The embodiments described above are only preferred embodiments of the invention and are not exhaustive of the possible implementations of the invention. Any obvious modifications to the above would be obvious to those of ordinary skill in the art, but would not bring the invention so modified beyond the spirit and scope of the present invention.