A tool wear monitoring method based on IWOA-GRU

Abstract

The application discloses an IWOA-GRU-based tool wear monitoring method, belongs to the field of metal cutting monitoring, and aims at solving the problem of low tool wear state monitoring precision caused by the incapability of capturing long-term dependence relationship in the prior art. The method comprises the following steps: step one, collecting tool machining process data to obtain an original data set; step two, processing the original data set and then dividing the original data set into a training set and a test set; step three, inputting the training set into a GRU model for training; step four, adopting an IWOA whale swarm optimization algorithm to optimize the weight matrix of the GRU model; and step five, using the trained IWOA-GRU model to monitor tool wear in the tool machining process, and determining the tool wear state according to the tool wear value Y output by the model t .

Description

TECHNICAL FIELD

[0001] The application relates to a tool wear monitoring method based on IWOA-GRU and belongs to the field of metal cutting monitoring. BACKGROUND

[0002] With the development of manufacturing technology in the modern aerospace field, the performance of aircraft materials is gradually improved, and difficult-to-machine materials with comprehensive properties such as high temperature resistance and corrosion resistance are mostly used, and the part structure is complex and the machining precision is high. In the cutting process of the part, due to the interaction between the tool and the workpiece, the cutting edge of the tool is inevitably worn, which affects the machining precision, so it is very important to monitor the tool wear state. The tool will be worn quickly and severely during the cutting process in the traditional machining mode, thereby seriously reducing the material performance and carrying capacity. Among these difficult-to-machine materials, most of the tools are forced to be scrapped before reaching the service life due to tool failure causing workpiece damage and machine tool failure, which leads to about 20% to 25% of the tool life not being effectively used, increasing the processing cost. Therefore, it is a great challenge to manufacture and process high-precision aviation industry parts that meet the current scientific and technological level requirements, and the tool wear state affects the machining precision, efficiency and economic benefit, so it is of great significance to monitor the tool wear state in the machining process of aviation structure parts, which can greatly improve the cutting machining quality and production efficiency and reduce the cost. Generally, the replacement of tools is determined according to the subjective judgment of operators, and if the tool is replaced too early, resources will be wasted and the processing cost will be increased, and if the tool is not replaced in time, the workpiece machining quality will be reduced and the production efficiency will be reduced. Research shows that if the tool wear can be accurately monitored, the precision of the machined parts can be improved, the total processing cost can be saved by 10% to 40%, and the cutting speed can be improved by 10% to 50%. When processing long time series, the traditional RNN is prone to gradient disappearance problem, that is, in the process of back propagation, the gradient gradually becomes smaller, which causes the model to be unable to effectively learn the long-term dependence relationship between the tool wear value and the sensor data length. SUMMARY

[0003] In view of the problem that the tool wear state monitoring precision is low due to the inability of the prior art to capture long-term dependence relationship, the application provides a tool wear monitoring method based on IWOA-GRU.

[0004] The tool wear monitoring method based on IWOA-GRU comprises the following steps:

[0005] Step one, collect tool machining process data to obtain an original data set, the original data set is Y t =F(X t ), wherein Y t is a tool wear value, Xt sensor data for online monitoring, wherein a vibration signal, an acoustic emission signal, a main shaft current signal;

[0006] Step two, processing the original data set in step one, and then dividing it into a training set and a test set;

[0007] Step three, inputting the training set into the IWOA-GRU model, wherein the training set is trained using the GRU model, the weight matrix of the GRU model is optimized using the IWOA whale swarm algorithm, and the individuals of the initial population of the IWOA whale swarm algorithm come from the training set; Step four, the trained IWOA-GRU model is used to monitor tool wear in the tool machining process, and the tool wear value Y t determining the tool wear state.

[0008] Preferably, the process of processing the original data set in step two is:

[0009] Step two one, a preprocessing step, the time domain signal x i is filtered, smoothed, and empty values, outliers, and repeated values are removed; the time domain signal x i refers to the sensor data collected during each cutting process, i = 1, 2,..., N, and N is the length of the intercepted sensor data;

[0010] Step two two, a step of extracting three-domain feature values, including time domain feature values, frequency domain features, and time-frequency domain feature values;

[0011] The time domain feature values include mean, standard deviation, root mean square, peak value, pulse factor, kurtosis, shape coefficient, and skewness, which are obtained by processing the time domain signal x i ;

[0012] The frequency domain feature values include maximum frequency, minimum frequency, frequency average, center of gravity frequency, frequency deviation, and mean square frequency, which are obtained by performing Fourier transform on the time domain signal x i ;

[0013] The time-frequency domain feature values are obtained using wavelet packet transform, which decomposes to generate multi-band data, and then splices the multi-band data at each time into time-frequency domain feature values;

[0014] Step three, the three-domain features are screened by principal component analysis, and the top ten three-domain features are retained. Then, the effective data corresponding to the top ten three-domain features in the original data set are screened out, and the screened data set is divided into a training set and a test set.

[0015] Preferably, the time-domain feature values are obtained according to the following formula:

[0016] Mean value Standard deviation Root mean square

[0017] Peak value X pv = max |x i |, pulse factor Kurtosis

[0018] Shape coefficient Skewness

[0019] Preferably, the frequency-domain feature values are obtained by performing Fourier transform on the time-domain signal x i to obtain the frequency f i and the corresponding amplitude p i , and the frequency-domain feature values are obtained according to the following formula:

[0020] Maximum frequency f max = max (f1, f2,..., f N ), minimum frequency f min = min (f1, f2,..., f N ),

[0021] Frequency average Center of gravity frequency

[0022] Frequency deviation Mean square frequency

[0023] Preferably, the process of optimizing the weight matrix of the GRU model in step three by using the IWOA whale swarm algorithm is as follows:

[0024] Extract m data in the training set as whale individuals in the population, and set the maximum number of iterations;

[0025] The IWOA whale swarm algorithm obtains the optimal whale individual position in each iteration, and updates the positions of other whale individuals;

[0026] According to the whale individual position , the update gate weight matrix W zand reset gate weight matrix W r :

[0027]

[0028] wherein, is the optimal whale individual position of the tth iteration, W rand is the random position of the whale individual, W z (t) and W r (t) are the update gate weight matrix and reset gate weight matrix of the tth iteration, W z (t+1) and W r (t+1) are the predicted update gate weight matrix and reset gate weight matrix of the (t+1)th iteration, which are fed back to the GRU model for updating;

[0029] H is a scaling factor,

[0030] H = 2 e -λ·t

[0031] wherein, lambda is a constant.

[0032] Preferably, the judgment principle of the tool wear state in step four is:

[0033] the tool wear value Y t =[0, 0.07), indicating that the tool is in the primary wear state;

[0034] the tool wear value Y t =[0.07, 0.12), indicating that the tool is in the normal wear state;

[0035] the tool wear value Y t =[0.12, 0.3), indicating that the tool is in the sharp wear state.

[0036] The present application has the beneficial effects: the present project intends to study a method for intelligently monitoring the tool wear state, to establish an IWOA-GRU deep learning model according to the cutting data set after feature extraction and screening, to continuously optimize and train the model, to ensure the adaptability and robustness of the model, to improve the model training efficiency and monitoring accuracy, and to realize the accurate monitoring of the tool wear state, so as to prolong the service life of the tool, and to reduce the cost and improve the production efficiency.

[0037] The whale population algorithm is used to optimize the parameters of the GRU. Compared with the currently popular prediction models such as SVM, Transformer and BP, the IWOA-GRU is more flexible and has stronger adaptability in time series prediction. The IWOA optimizes the parameters of the GRU, and the model training efficiency and prediction accuracy are higher than those of other models, and the adaptability and robustness are also improved. BRIEF DESCRIPTION OF DRAWINGS

[0038] Figure 1 is a flow chart of the IWOA-GRU based tool wear monitoring method described in the present application;

[0039] Figure 2 is a flow chart of the joint training of the IWOA-GRU model;

[0040] Figure 3 is a schematic diagram of the gated recurrent unit GRU;

[0041] Figure 4 is a schematic diagram of the update gate of the GRU;

[0042] Figure 5 is a schematic diagram of the reset gate of the GRU;

[0043] Figure 6 is a schematic diagram of the hidden output state of the GRU;

[0044] Figure 7 is a schematic diagram of the candidate hidden output state of the GRU. DETAILED DESCRIPTION

[0045] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of the present application.

[0046] It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict.

[0047] The present application will be further described below with reference to the drawings and specific embodiments, but not as a limitation of the present application.

[0048] DETAILED DESCRIPTION Figures 1 to 7 The present embodiment describes an IWOA-GRU based tool wear monitoring method, which comprises the following steps:

[0049] Step 1, collect tool machining process data to obtain an original data set, the original data set is Y t = F(X t ), wherein Y t is the tool wear value, X t is the sensor data of online monitoring, wherein is a vibration signal, is an acoustic emission signal, The spindle current signal;

[0050] Step two, the original data set in step one is processed, and then divided into a training set and a test set;

[0051] Step three, the training set is input into the IWOA-GRU model, wherein the training set is trained by using the GRU model, and the weight matrix of the GRU model is optimized by using the IWOA whale swarm algorithm, and the individuals of the initial population of the IWOA whale swarm algorithm come from the training set;

[0052] Step four, the trained IWOA-GRU model is used for monitoring tool wear in tool machining process, and the tool wear value Y t The tool wear state is judged.

[0053] The original data set in step one is constructed by collecting a large amount of data in the tool machining process, wherein the data of each sensor is collected online, and the tool wear value is collected offline.

[0054] The data in the original data set cannot be directly used for model training, and the original data set needs to be processed to extract effective data for model training, so as to make the model training accurate and the output result accurate.

[0055] The process of processing the original data set in step two is:

[0056] Step two one, preprocessing step, the time domain signal x i is filtered, smoothed, and the null value, abnormal value and repeated value are removed; the time domain signal x i refers to the sensor data collected in each cutting process, i=1,2,...,N, and N is the length of the intercepted sensor data;

[0057] Step two two, the step of extracting three domain characteristic values, the three domain characteristic values include time domain characteristic values, frequency domain characteristics and time-frequency domain characteristic values;

[0058] The time domain analysis reflects the change of the cutting process parameter signal waveform on the time axis, and the characteristics retained by the time domain signal are more comprehensive, the time domain characteristic values include mean, standard deviation, root mean square, peak value, pulse factor, kurtosis, shape coefficient and skewness, and the time domain characteristic values are obtained by i processing.

[0059] The time domain characteristic values are obtained according to the following formula:

[0060] Mean Standard deviation Root mean square

[0061] Peak X pv = max |x i |, pulse factor kurtosis

[0062] shape coefficient skewness

[0063] The time domain analysis is remarkable in stationary periodic signals, and can more intuitively reflect the change trend of tool wear, but the signals collected in actual machining are mostly non-stationary signals. Therefore, the analysis of time domain characteristics cannot fully represent the characteristics of the signal. With the increase of tool wear, the frequency of the signal will often change, the time domain signal is transformed by fast Fourier transform (FFT) to obtain the relationship between frequency and amplitude, and then the frequency domain characteristics of the tool with wear are analyzed. The frequency domain characteristic values include maximum frequency, minimum frequency, average frequency, center of gravity frequency, frequency deviation and mean square frequency, and the time domain signal x i is Fourier transformed to obtain;

[0064] The frequency domain characteristic values are obtained by Fourier transforming the time domain signal x i , obtaining the frequency f i and the corresponding amplitude p i , and the frequency domain characteristic values are obtained according to the following formula:

[0065] The maximum frequency f max = max (f1, f2,..., f N ), the minimum frequency f min = min (f1, f2,..., f N ),

[0066] The average frequency The center of gravity frequency

[0067] The frequency deviation The mean square frequency

[0068] The common characteristics of the time domain or the frequency domain describe the change of the tool wear state in the whole cutting process, but cannot observe the frequency information of the non-stationary signal changing with time, and cannot analyze the change of the cutting process parameter signal locally, that is, the time frequency resolution is not high. With the help of time-frequency domain analysis, this feature extraction method can determine the state information of tool wear. In order to obtain more comprehensive information related to tool wear, the local and global of the signal need to be analyzed at the same time, and the wavelet packet transform is used for time-frequency domain analysis in the embodiment, the wavelet packet transform is decomposed to generate multi-band data, and then the multi-band data at each time is spliced into time-frequency domain characteristic values;

[0069] The time domain, frequency domain and time-frequency domain features of the cutting process parameter signal are extracted, and the extracted features change with the different tool wear states, but these features are not all highly related to the tool wear state, and all of them as inputs of the model will cause information redundancy, which will reduce the model operation and even cause overfitting and performance decline. Therefore, in order to effectively monitor the tool state, it is necessary to extract the effective information related to the tool wear, and the principal component analysis method is used for screening in the embodiment.

[0070] Step two, the principal component analysis method is used to reduce the dimension of the three-domain features and select the top ten three-domain features with contribution rate, and then the effective data corresponding to the top ten three-domain features with contribution rate in the original data set are screened out, and the screened data set is divided into a training set and a test set.

[0071] Step three, inputting the training set into the GRU model for training;

[0072] The gated recurrent unit (GRU) is a deep learning model specially used for processing time series data, as shown in Figure 3 . The GRU contains 1 memory unit 2 gates: update gate z t and reset gate r t , x t represents the input data at the current t time, h t-1 represents the historical t-1 time hidden state output. σ and tanh represent the sigmoid function and hyperbolic tangent function respectively.

[0073] Referring to Figure 4 , the update gate z t : decides how much data set historical information of the last training and the current time training data set information are used to update the current hidden state. The update gate at the t time is:

[0074] z t =σ(W z ·[h t-1 ,x t ])

[0075] Wherein, z t is the gating update signal, the size of which determines the memory degree of the candidate hidden state, h t-1 is the historical hidden state, x t represents the new input data at t time, and W z is the weight matrix.

[0076] Referring to Figure 5 , the reset gate r t: Mainly determines how much trained vibration, acoustic emission, current signal and dataset information is retained, how much past vibration, acoustic emission, current signal and dataset information needs to be forgotten, and the reset gate at the t time is:

[0077] r t = sigma (W r ·[h t-1 ,x t ])

[0078] Wherein, r t is the reset signal, the greater the reset signal indicates that the amount of historical information to be remembered is more, W r is the weight matrix.

[0079] Referring to Figure 6 , the implicit output state h t : Under the action of the update gate and the reset gate, the implicit output state h t of the current time can be updated as:

[0080]

[0081] Wherein, h t represents the final implicit output state of the current time dataset, including memory and forgetting two processes, the greater the element value in z t , the greater the Hadamard product of z t and h t-1 , which means that the information in the h t-1 of the previous time vibration, acoustic emission, current signal and dataset is retained to a higher degree. On the contrary, the smaller the element value in z t , the more new implicit state is derived from

[0082] Referring to Figure 7 , the candidate implicit state

[0083]

[0084] The input new vibration, acoustic emission, current signal and data and the vibration, acoustic emission, current signal and dataset feature information of the historical implicit state, this operation is related to the reset signal r t obtained by the reset gate, and the candidate implicit state is obtained through the tanh function.

[0085] GRU is good at processing sequence data, but it is easy to fall into local optimal solution and has low convergence precision, and the optimization algorithm of IWOA is used to enhance the parameter tuning ability of GRU. The improved whale optimization algorithm optimizes the time prediction of the gated recurrent unit, the model has good generalization ability, and the core idea is to optimize the parameters of GRU through the whale optimization algorithm.

[0086] The process of using IWOA whale optimization algorithm to optimize the weight matrix of GRU model in step three is:

[0087] Extract m data in the training set as whale individuals in the population, and set the maximum number of iterations;

[0088] The IWOA whale optimization algorithm optimizes the optimal whale individual position in each iteration and update the position of other whale individuals;

[0089] According to the position of the whale individual Get the update gate weight matrix W z and the reset gate weight matrix W r :

[0090]

[0091] wherein, is the optimal whale individual position of the tth iteration, W rand is the random position of the whale individual, W z (t) and W r (t) are the update gate weight matrix and the reset gate weight matrix of the tth iteration, the initial value W z (0) and W r (0) are random values, W z (t+1) and W r (t+1) are the predicted update gate weight matrix and reset gate weight matrix of the t+1th iteration, which are fed back to the GRU model for updating;

[0092] H is a scaling factor for controlling the search step;

[0093] H = 2·e -λ·t

[0094] wherein λ is a constant.

[0095] In the IWOA optimization algorithm, the maximum number of iterations is set to update the optimal position parameters of the population individuals in each iteration, and the optimal solution is the individual position closest to the prey position in the population individuals formed by the data set, that is, the most consistent vibration, acoustic emission and current signal and data set. Then change the position of other individuals, and then IWOA improves the traditional WOA model in the late local development by using the adaptive weight strategy and random differential variation strategy, which is prone to local optimization and premature convergence. The individual position is updated again, which has better optimization effect, and then it is mapped back to the GRU model to optimize the weight matrix and other parameters of the GRU model, replacing the original parameters. The specific process of IWOA is shown in Figure 2 .

[0096] The judgment principle of the tool wear state in step four is:

[0097] Tool wear value Y t = [0, 0.07), representing that the tool is in the primary wear state;

[0098] Tool wear value Y t = [0.07, 0.12), representing that the tool is in the normal wear state;

[0099] Tool wear value Y t = [0.12, 0.3), representing that the tool is in the sharp wear state.

[0100] Whether to replace the tool is decided according to the tool wear state.

[0101] While the application has been described with reference to particular embodiments, it will be understood that the examples are merely illustrative of the principles and applications of the application. It will be understood that numerous modifications can be made to the illustrative embodiments, and that other arrangements can be devised without departing from the spirit and scope of the present application as defined by the appended claims. It will be understood that the features of the various embodiments can be combined with each other, in different ways than as described herein. It will be understood that features described with reference to one embodiment can be used in other embodiments.

Claims

1. A tool wear monitoring method based on IWOA-GRU, characterized in that, The method comprises the following steps: Step one, collect tool machining process data to obtain original data set, the original data set is Y t = F(X t ), wherein Y t is the tool wear value, X t is the online monitoring sensor data, wherein is a vibration signal, is an acoustic emission signal, is a spindle current signal; Step two, the original data set in step one is processed, and then divided into a training set and a test set; Step three, the training set is input into an IWOA-GRU model, wherein the training set is trained by using a GRU model, a weight matrix of the GRU model is optimized by using an IWOA whale swarm algorithm, and individuals of an initial population of the IWOA whale swarm algorithm come from the training set; The process of optimizing the weight matrix of the GRU model by using the IWOA whale swarm algorithm is as follows: m data in the training set are extracted as whale individuals of a population, and a maximum iteration number is set; The IWOA whale population algorithm obtains the optimal whale individual position in each iteration and updates the positions of other whale individuals According to the position of the individual whale Obtaining an update gate weight matrix W for optimizing the GRU model z and a reset gate weight matrix W r : wherein, is the optimal whale individual position of the tth iteration, W rand is a random position of a whale individual, W z (t) and W r (t) are the update gate weight matrix and the reset gate weight matrix of the tth iteration, W z (t+1) and W r (t+1) are the predicted update gate weight matrix and the reset gate weight matrix of the t+1th iteration, for feedback to the GRU model for updating; H is a scaling factor, H = 2 e -λ·t wherein, λ is a constant; Step four, the trained IWOA-GRU model is used to monitor the tool wear in the tool machining process, and the tool wear value Y t judges the tool wear state.

2. The IWOA-GRU-based tool wear monitoring method according to claim 1, wherein, The process of processing the original data set in step two is as follows: Step two, preprocessing step, the time domain signal x collected for each cutting process i Filtering, smoothing, and removing null, outliers and repeated values; the time domain signal x i Refers to the sensor data collected during each cutting process, i = 1, 2,..., N, N is the length of the intercepted sensor data; Step two two, a three-domain feature value extraction step is performed, and the three-domain feature values include time-domain feature values, frequency-domain features and time-frequency domain feature values; The time-domain characteristic values include mean value, standard deviation, root mean square, peak value, pulse factor, kurtosis, shape coefficient and skewness, and the time-domain characteristic values are obtained by taking the time-domain signal x i processing obtained; The frequency domain characteristic values include maximum frequency, minimum frequency, frequency average, barycentric frequency, frequency deviation and mean square frequency, which are obtained by performing Fourier transform on the time domain signal x i ​ The time-frequency domain feature values are obtained by using wavelet packet transformation, wavelet packet transformation is decomposed to generate multi-frequency band data, and then the multi-frequency band data at each time point are spliced into time-frequency domain feature values; Step two three, principal component analysis is used to reduce and screen the three-domain features, the top ten three-domain features in contribution rate are retained, and then effective data in the original data set corresponding to the top ten three-domain features in contribution rate are screened out, and the screened data set is divided into the training set and the test set.

3. The IWOA-GRU-based tool wear monitoring method according to claim 2, wherein, The time-domain feature values are obtained according to the following formula: average value standard deviation root mean square peak x pv = max |x i |, pulse factor kurtosis Shape factor Skewness 4. The IWOA-GRU-based tool wear monitoring method according to claim 2, wherein, Obtaining frequency-domain characteristic value: Fourier transform is performed on time-domain signal x i to obtain frequency f i and corresponding amplitude p i , and the frequency-domain characteristic value is obtained according to the following formula: Maximum frequency f max = max(f1, f2,..., f N ), minimum frequency f min = min(f1, f2,..., f N ), frequency average value gravity frequency Frequency deviation Mean square frequency 5. The IWOA-GRU-based tool wear monitoring method according to claim 1, wherein, The judgment principle of the tool wear state in step four is as follows: Tool wear value Y t = [0, 0.07), characterizing the tool in a primary wear state; Tool wear value Y t = [0.07, 0.12), indicating that the tool is in a normal wear state; Tool wear value Y t = [0.12, 0.3), characterizing the tool in a state of sharp wear.

Images