A tool wear monitoring method based on discrete wavelet decomposition

Abstract

The application discloses a tool wear monitoring method based on discrete wavelet decomposition, and belongs to the technical field of tool monitoring.The method is as follows: a tool wear data set is constructed; a periodic two-dimensional reconstruction is performed on a time series signal; and a tool wear monitoring model is obtained.The application converts a one-dimensional time series signal into a two-dimensional tensor feature according to a period, extracts multi-scale feature information, fits a tool wear value by capturing deep features of a periodic signal, establishes a tool wear monitoring model for historical periodic data generated in a tool machining process, can adaptively extract optimal features for monitoring, can dynamically adjust parameters according to signal characteristics of different process parameters in tool machining, and thus the generalization ability and the prediction precision of the model are improved, and accurate monitoring of the tool wear value is realized.

Description

TECHNICAL FIELD

[0001] The application relates to a tool wear monitoring method based on discrete wavelet decomposition and belongs to the technical field of tool monitoring. BACKGROUND

[0002] A data-driven model is a scheme for learning the mapping relationship between signals and tool states from a large amount of processing data. In the context of industrial big data, with a large amount of data mined in the processing process, the data-driven model shows better prediction results than the physical-driven model and is easier to fit the nonlinear function relationship in the complex processing process.

[0003] Since the data-driven model can learn the corresponding relationship from a large amount of process information, researchers have begun to consider the data-driven mode of multiple source sensors. The emergence of deep learning has rapidly developed the field of tool wear monitoring, and a large number of RNN network models have been used to solve the tool wear monitoring problem. However, the length of the processed information is limited, and the process signal contains sparse information and a large amount of noise. The training process is prone to learn the sparsity of the features, and the RNN is limited in processing the length of the signal. Moreover, the mapping of the tool state evolution and the multi-source sensing information is not clear. SUMMARY

[0004] To solve the problems in the background art, the application provides a tool wear monitoring method based on discrete wavelet decomposition.

[0005] To achieve the above object, the application adopts the following technical scheme: a tool wear monitoring method based on discrete wavelet decomposition, the method comprising the following steps:

[0006] S1: constructing a tool wear data set;

[0007] S2: periodically reconstructing a time series signal;

[0008] S3: obtaining a tool wear monitoring model.

[0009] Compared with the prior art, the application has the following beneficial effects:

[0010] 1. The application converts one-dimensional time series signals into two-dimensional tensor features according to a period, extracts multi-scale feature information through Inception, fits tool wear values by capturing deep features of the periodic signal, establishes a DWT-Timesnet tool wear monitoring model for historical periodic data generated in the tool processing process, can adaptively extract optimal features for monitoring, can dynamically adjust parameters according to the signal characteristics of different process parameters in the tool processing, thereby improving the generalization ability and prediction accuracy of the model, and realizing accurate monitoring of the tool wear value.

[0011] 2、The application replaces the global frequency analysis of FFT in the original timesnet model with discrete wavelet transform, which can analyze signals of different time scales, and can capture the local time-frequency characteristics of time series and has good noise resistance by changing the scale and translation parameters of the wavelet. BRIEF DESCRIPTION OF DRAWINGS

[0012] Figure 1 is a flowchart of the application;

[0013] Figure 2 is a Timesnet network workflow diagram based on discrete wavelet decomposition;

[0014] Figure 3 is a discrete wavelet transform period extraction and two-dimensional tensor conversion schematic diagram;

[0015] Figure 4 is a Timesnet network structure schematic diagram;

[0016] Figure 5 is a monitoring result diagram of the application example 1 on the test set. DETAILED DESCRIPTION

[0017] The technical solutions in the application will be described clearly and completely in the embodiments of the application combined with the accompanying drawings, obviously, the described embodiments are only a part of the embodiments of the application, not all the embodiments, based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor belong to the protection scope of the application.

[0018] A tool wear monitoring method based on discrete wavelet decomposition, the method comprises the following steps:

[0019] S1: constructing a tool wear data set;

[0020] S101: acquiring time series signals in the machining process for representing the tool wear state through a sensor, obtaining multi-working condition historical data of force sensor information, the time series signals selecting force signals, vibration signals and power signals in the machining process;

[0021] S102: measuring the offline wear value after each machining to obtain the actual tool wear value W true of each machining;

[0022] S103: Because the timing signal has regular periodic characteristics, by observing the change of the periodic characteristics of the timing signal with tool wear, the tool wear process is divided into three stages of initial wear, mid-wear and late-wear; through frequency domain analysis, it is found that the frequency spectrum energy of the timing signal gradually moves to the high frequency band with the tool wear process. It shows that the frequency domain characteristics of the timing signal have strong correlation with tool wear. Therefore, the TimesNet network is used to convert the one-dimensional timing signal of the timing signal into two-dimensional features for multi-convolution kernel extraction, and the two-dimensional change in time is obtained;

[0023] S104: The timing signal is sliced along the time dimension, and the actual tool wear value W true is linearly interpolated to expand the wear data set label to form a tool wear data set;

[0024] S2: The timing signal is periodically reconstructed in two dimensions;

[0025] S201: The input timing signal is decomposed by discrete wavelet to obtain low-frequency timing signal and high-frequency timing signal, the process is as follows:

[0026]

[0027] In formula (1):

[0028] x[n] represents a discrete one-dimensional timing signal;

[0029] h[k] represents the coefficient of the high-pass filter;

[0030] g[k-2n] represents the coefficient of the low-pass filter;

[0031] * represents convolution operation;

[0032] S202: The timing signal is decomposed into approximation coefficients cA and detail coefficients cD;

[0033] S203: Calculate the amplitude list through the approximation coefficient cA;

[0034] S204: Determine the threshold T through the mean and standard deviation of the amplitude list, the process is as follows:

[0035]

[0036] In formula (2):

[0037] N represents the number of approximation coefficients cA obtained by discrete wavelet decomposition;

[0038] cA i represents the i-th in the approximation coefficient cA;

[0039] S205: Select the low-frequency signal corresponding to the approximate coefficient cA whose amplitude coefficient is greater than the threshold T, and obtain its corresponding frequency {f1, f2,..., f k} and the corresponding period length {p1, p2,..., p k};

[0040] S206: Periodic two-dimensional reconstruction is performed on the time sequence signal.

[0041] S3: Obtain a tool wear monitoring model.

[0042] S301: Input the reconstructed two-dimensional feature into the Timesnet network;

[0043] S302: Fold the original one-dimensional time sequence signal by selecting the period, and perform zero extension on the original one-dimensional time sequence signal through the Padding layer to obtain the two-dimensional tensor feature of the time sequence signal, the process is as follows:

[0044]

[0045] In formula (3):

[0046] Padding(·) represents zero extension of the original one-dimensional time sequence signal to make its period length aligned;

[0047] X 1D represents the original one-dimensional time sequence signal;

[0048] Reshape(·) represents two-dimensional tensor feature conversion, and the size is p i ,f i ;

[0049] represents the two-dimensional tensor feature;

[0050] S303: Adopt timesblock to capture the information of the time sequence signal under different periods through different scale convolution kernels in the classic Inception model;

[0051] S304: Convert the extracted time sequence feature to one-dimensional space and perform adaptive fusion;

[0052] S30401: Convert the extracted time sequence feature to one-dimensional space through two-dimensional tensor feature,

[0053] S30402: Perform weighted summation on the intensity of the corresponding frequency of the one-dimensional space representation to realize adaptive fusion, the process is as follows:

[0054] ​

[0055]

[0056]

[0057] In formula (4) :

[0058] Inception(·) represents the operation of the Inception module;

[0059] In formula (5) :

[0060] Trunc(·) represents that after the original two-dimensional tensor feature is reshaped to a one-dimensional feature, the original one-dimensional signal is truncated according to the length of the original one-dimensional signal, so as to keep the length consistent with the original input one-dimensional signal;

[0061] S305: After the adaptive fusion feature is activated by GELU, the time and channel dimensions are fused, and mapped to the tool wear value through a fully connected layer to obtain the network wear prediction value W pre ;

[0062] W pre = Linear(GELU(X 1D )) (7)

[0063] In formula (7) :

[0064] Linear(·) represents a fully connected layer;

[0065] S306: Calculate the error between the network wear prediction value W pre and the actual tool wear value W true , and optimize the tool wear monitoring model parameters;

[0066] S307: Obtain the tool wear monitoring model.

[0067] In the online monitoring process, the time sequence signal collected under unknown working conditions is segmented and input into the trained tool wear monitoring model one by one to obtain the predicted wear value W' pre , if the predicted wear value reaches the tool life critical value, stop and replace the tool.

[0068] Embodiment 1:

[0069] The numerical control machine tool adopts a three-axis high-speed milling machine tool VMC855, the adopted milling cutter is an end milling cutter, the cutter bar diameter is 20mm, and the blade is PCD APMT1135. The workpiece adopts aluminum-based silicon carbide with a SiC volume fraction content of 65%, and the volume is 100mmx150mmx80mm.

[0070] The workpiece surface is transversely cut by a tool along multiple paths. Force sensors, vibration sensors and power sensors are used to collect processing signals, and the tool microscope measures the tool flank wear width of the blade as the tool wear during this processing.

[0071] The force sensor uses a Swiss Kistler intelligent force handle to collect three orthogonal direction force signals at a sampling frequency of 10 kHz, the vibration sensor uses a Donghua test IEPE piezoelectric acceleration sensor, the power signal is the spindle driving motor current and voltage signal, the tool microscope uses a MV-HS2000GM industrial camera and a double telecentric microscope to measure the tool flank wear width, and the tool failure standard is 300μm of tool flank wear after the tool flank wear width experiment. The monitoring algorithm is on the NVIDIARTX308012GB GPU platform on the Python 3.8-Anaconda 1.10.0-pytorch platform.

[0072] S1: Obtain the time sequence signal x composed of force signals, vibration signals and power signals, and segment it to obtain signal segments x j , and the actual tool wear value W measured by the offline method true After interpolation, the tool wear data set is formed.

[0073] S2: Input the data set into the Timesnet network model based on discrete wavelet decomposition for training and optimization.

[0074] In the online monitoring process, the force signals, vibration signals and power signals collected under the same working condition are segmented in real time and sent into the trained network model to monitor the wear value in real time, so as to judge whether the tool needs to be replaced.

[0075] S3: The Timesnet network model based on discrete wavelet decomposition is constructed, and the aluminum-based silicon carbide high-speed machining experiment is carried out, the spindle speed is 2000rpm, the feed speed is 600mm / min, and the machining depth is 0.1mm. The data collected during actual processing is segmented and the wear value is expanded to obtain the tool wear data set. The first 80% of the sample data set is used as the training set, and the last 20% is used as the test set. The training set is used to train and optimize the Timesnet network model based on discrete wavelet decomposition. The model uses Adam optimizer, the initial learning rate is 0.001, and the sample size is 10. The test set is used to test the trained model, and the final result is as follows Figure 5As shown, the measured MAE is 3.57 μm, the MAPE is 1.42%, and the MSE is 29.79. Among them, MAE is the average absolute error, MAPE is the average absolute percentage error, and MSE is the mean square error. The closer these values are to 0, the higher the model accuracy. Experiments show that the proposed model can accurately and effectively predict the current tool wear value.

[0076] It is obvious to those skilled in the art that the present application is not limited to the details of the above exemplary embodiments, and can be implemented in other forms without departing from the spirit or essential characteristics of the present application. Therefore, the embodiments should be regarded as exemplary and non-limiting, and the scope of the present application is defined by the appended claims rather than the above description, and all changes falling within the meaning and scope of the equivalent conditions of the claims are intended to be included in the present application. Any reference signs in the claims should not be regarded as limiting the claims involved.

[0077] In addition, it should be understood that although the present specification is described in terms of embodiments, not every embodiment contains only one independent technical solution, and the description of the specification is only for the sake of clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that those skilled in the art can understand.

Claims

1. A tool wear monitoring method based on discrete wavelet decomposition, characterized in that: The method includes the following steps: S1: Construct a tool wear dataset; S1 includes the following steps: S101: Acquires timing signals and obtains multi-condition historical data of force sensor information; S102: Perform offline wear value measurement to obtain the actual tool wear value for each machining operation. ; S103: The tool wear process is divided into three stages: early wear, middle wear, and late wear; the one-dimensional time series signal is converted into two-dimensional features for multi-convolution kernel extraction; S104: Slice the timing signal along the time dimension and apply it to the actual tool wear value. Perform linear interpolation to construct a tool wear dataset; S2: Perform periodic two-dimensional reconstruction of the time-series signal; S2 includes the following steps: S201: The input time-series signal is decomposed by discrete wavelet decomposition to obtain low-frequency and high-frequency time-series signals. The process is as follows: （1） In formula (1): Represents a discrete one-dimensional time-series signal; These represent the coefficients of the high-pass filter; These represent the coefficients of the low-pass filter; This represents the convolution operation; S202: Decompose the time-series signal into approximation coefficients cA and detail coefficients cD; S203: A list of amplitudes calculated using the approximation coefficient cA; S204: Determine the threshold using the mean and standard deviation of the amplitude list. The process is as follows: （2） In formula (2): This represents the number of approximation coefficients cA obtained from discrete wavelet decomposition; This represents the i-th approximation coefficient in cA; S205: Select an approximation coefficient cA corresponding to a low-frequency signal whose amplitude coefficient is greater than the threshold. The low-frequency signal can be obtained with the corresponding frequency as and the corresponding period length ; S206: Perform periodic two-dimensional reconstruction of the time-series signal; S3: Obtain the tool wear monitoring model; S3 includes the following steps: S301: Input the reconstructed two-dimensional features into the TimesNet network; S302: The original one-dimensional time series signal is folded by selecting a period, and zero-expanded by a padding layer to obtain the two-dimensional tensor features of the time series signal. The process is as follows: （3） In formula (3): This represents zero-spreading of the original one-dimensional time-series signal; Represents the original one-dimensional time-series signal; This represents a two-dimensional tensor feature transformation, and its size is... ; Representing the characteristics of a two-dimensional tensor; S303: Captures information about time-series signals under different periods using two-dimensional tensor features; S304: The extracted temporal features are transformed into a one-dimensional space and then adaptively fused. S304 includes the following steps: S30401: The extracted temporal features are transformed into a one-dimensional space using two-dimensional tensor features. S30402: Representation of one-dimensional space With the intensity of its corresponding frequency The weighted summation is performed to achieve adaptive fusion, and the process is as follows: （4） （5） （6） In equation (4): This indicates the operations performed by the Inception module; In formula (5): This means that after reshaping the original two-dimensional tensor features into one-dimensional features, the features are truncated according to the length of the original one-dimensional signal. S305: After GELU activation of the adaptive fusion features, the temporal and channel dimensions are fused, and then mapped onto the tool wear value through a fully connected layer to obtain the network's wear prediction value. ; （7） In equation (7): Indicates a fully connected layer; S306: Calculate the network wear prediction value Compared with actual tool wear value After adjusting the parameters of the tool wear monitoring model to account for the error; S307: Obtain the tool wear monitoring model.

Images