A method of cutting tool wear monitoring

Abstract

The present application relates to a kind of cutting tool wear monitoring methods, comprising: acquisition tool cutting process in multiple directions signal, the signal includes cutting force, vibration and sound signal;Acquisition tool wear after cutting is completed, signal and wear are spliced, obtain state matrix x and wear target value vector c;Using sliding window, data interception is carried out to state matrix x, obtain M data section;The time domain characteristics of the data section are calculated to obtain;Interpolation is carried out to target value vector c, and the wear of each data section corresponding is obtained;Using, training neural network and constructing tool wear monitoring model.The present application only needs to input current cutting data and the prior distribution parameter of wear, can carry out tool wear monitoring, and using codec structure improves the explainability of model.

Description

Technical Field

[0001] This invention belongs to the field of metal cutting technology, and specifically relates to a method for monitoring the wear of cutting tools. Background Technology

[0002] In metal cutting processes, tool wear refers to the gradual degradation of cutting tools such as drills, end grinders, and inserts as they interact with the workpiece during machining. Tool wear monitoring is a crucial aspect of modern machining processes, playing a key role in ensuring production efficiency, product quality, and cost-effectiveness. By monitoring tool wear, tool conditions can be dynamically and in real-time assessed, tool life extended, and machining parameters and tool change schedules optimized.

[0003] Currently, tool wear monitoring is mainly divided into direct and indirect methods. The direct method involves using optical instruments to acquire images of tool wear for monitoring. While this method offers high monitoring accuracy, it requires interrupting the machining process for image acquisition, and the machining environment can cause significant interference and noise, making continuous monitoring impossible. The indirect method involves analyzing signals from various sensors to assess tool wear. These sensors include cutting force, vibration, sound, and motor power. The indirect method allows for continuous, real-time signal acquisition during machining without interrupting the process.

[0004] Current indirect methods mainly include three stages: feature extraction, feature selection, and model construction. For example, Chinese invention patent CN116638374A discloses a tool wear state recognition method based on feature selection and sensor combination, including the following steps: combining three types of sensors that generate cutting force, vibration, and acoustic emission signals to obtain seven sensor combinations; collecting multi-sensor signals during the machining process; extracting signals from the intermediate stable milling stage and processing outliers and environmental interference noise signals to obtain a feature subset ①; removing features with weak correlation to tool wear and analyzing the correlation between features to obtain a feature subset ③ with redundant features removed; dividing the data into training and testing sets, and inputting them into a forest to construct a tool wear state recognition model to identify the tool wear state. However, this method has the following drawbacks: in actual operation, the number of tool wear value samples is limited, and the tool wear value increases with usage time. The above method ignores the relationship between tool wear values ​​at different time points, resulting in an inaccurate trained model. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a cutting tool wear monitoring method, aiming to solve the problem of low accuracy in tool wear monitoring.

[0006] To address the above problems, this invention provides a method for monitoring cutting tool wear, which includes the following steps:

[0007] Step S1: Acquire signals from multiple directions during the cutting process. The signals include cutting force, vibration, and sound signals; the directions are based on the x-axis, y-axis, and z-axis directions of the machining coordinate system.

[0008] Step S2: Collect the tool wear after cutting. , will signal and wear The states are concatenated to obtain the state matrix x and the wear target value vector c.

[0009] Step S3: Using a sliding window, extract data from the state matrix x to obtain M data segments. ; Calculate the data segment The time-domain features are obtained Interpolate the target value vector c to obtain the wear amount corresponding to each data segment. ;

[0010] Step S4: Use , Train a neural network and build a tool wear monitoring model.

[0011] Furthermore, the aforementioned Data for seven channels ,in, For cutting force signal, It is a vibration signal. It is a sound signal.

[0012] Furthermore, the aforementioned The initial value is 0.

[0013] Further, the calculation of the data segment The time-domain features are obtained Specifically: Calculate the data statistics for each channel vertically to obtain the maximum value. mean Covariance ,in This is the index value of the channel.

[0014] Furthermore, the use , The specific calculation formula for training a neural network is as follows: ,in This is the predicted tool wear monitoring value at time t. For a neural network, Q is the model encoding length. Wear amount The prior distribution parameters.

[0015] Furthermore, the amount of wear The prior distribution parameters are as follows: in, Wear amount The mean, Wear amount The covariance.

[0016] Furthermore, the tool wear monitoring model includes: gated residual network GRN, feature selection network VSN, static oblique variance vector encoder SCE, cyclic encoder-decoder LSTMEnDe, and temporal fusion decoder TFD.

[0017] Furthermore, wear amount The prior distribution parameters s are processed by the feature selection network VSN to obtain the selected features. Wear amount The prior distribution parameters s are processed by a static orthometric vector encoder (SCE) to obtain the output encoded feature vector. , , and .

[0018] Furthermore, the input features of the LSTMEnDe loop codec and characteristics and Output encoded temporal features .

[0019] Furthermore, the temporal fusion decoder (TFD) is used to process the encoded temporal features. Combination Data augmentation is performed; a temporal self-attention multi-head mechanism is adopted to calculate attention and weights, and then feature calculation and fusion are performed based on the weights; the fused features are subjected to gating calculation, skip-level connection and normalization processing, and element-level feedforward calculation is performed through a gated residual network GRN, and finally the tool wear monitoring value is output.

[0020] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0021] (1) The method proposed in this patent does not require previous wear data. It only requires inputting the current cutting data and the prior distribution parameters of wear to monitor tool wear.

[0022] (2) The method proposed in this patent is to improve the interpretability of the model by using an encoder-decoder structure. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0024] Figure 1 This is a block diagram of the cutting tool wear monitoring method of the present invention;

[0025] Figure 2 This is a flowchart of the feature extraction process of the present invention;

[0026] Figure 3 This is a flowchart of the feature selection process of the present invention. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. Obviously, the specific embodiments described herein are merely illustrative of the invention and represent only a portion, not all, of the embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without inventive effort are within the scope of protection of this invention.

[0028] Example 1

[0029] like Figures 1-3 As shown, this is a diagram illustrating a cutting tool wear monitoring method according to an embodiment of the present invention. The method includes:

[0030] This invention provides a method for monitoring the wear of cutting tools, the method comprising the following steps:

[0031] Step S1: Acquire signals from multiple directions during the cutting process. The signals include cutting force, vibration, and sound signals; the directions are based on the x-axis, y-axis, and z-axis directions of the machining coordinate system.

[0032] Step S2: Collect the tool wear after cutting. , will signal and wear The states are concatenated to obtain the state matrix x and the wear target value vector c.

[0033] Step S3: Using a sliding window, extract data from the state matrix x to obtain M data segments. ; Calculate the data segment The time-domain features are obtained Interpolate the target value vector c to obtain the wear amount corresponding to each data segment. ;

[0034] To ensure the accuracy of the model, feature extraction and feature selection are crucial stages, specifically as follows: Figures 2-3 As shown, during the cutting process, sensors are used to collect cutting force signals in three directions. Vibration signal sound signal This results in a 7-channel signal. , The corresponding N sets of data collected, i.e. The tool performs multiple cuts, and the wear of the tool cutting edge is measured after each cut. During the corresponding K cutting processes, a total of K sets of data were collected. ,Right now .

[0035] Treating the entire lifecycle of the tool as a complete time series, and collecting K sets of data. and wear By concatenating the data along the time dimension, we obtain the state matrix x and the wear target value vector c.

[0036] Using a sliding window, data is extracted from the state matrix x according to the interval wid, resulting in a total of M data segments. Interpolating the target value vector c yields the wear amount corresponding to each data segment. Because a new tool is used at the start of cutting, the initial value for the wear interpolation is 0. Therefore, all processed data segments are obtained. and wear The correspondence between them, that is Each data segment after processing This contains data from 7 channels (wid). The statistical value of each channel is calculated vertically, and the maximum value is obtained. mean Covariance ,in This is the index value of the channel. After the above feature selection, the data segment is transformed into a single 3*7 channel data segment. Corresponding wear amount This forms a new correspondence, namely .

[0037] Step S4: Use , Train a neural network and build a tool wear monitoring model.

[0038] This step is the model building process, including: a gated residual network (GRN), a feature selection network (VSN), a static oblique variance vector encoder (SCE), a recurrent encoder-decoder (LSTMEnDe), and a temporal fusion decoder (TFD). The gated residual network (GRN) is the basic unit constituting the entire network, controlling data flow and the nonlinearity of the model by selectively allowing and blocking data. The feature selection network (VSN), based on the gated residual network (GRN), dynamically selects the most relevant features at each time step. The static oblique variance vector encoder (SCE), also based on the gated residual network (GRN), handles time-independent features. The recurrent encoder-decoder (LSTMEnDe) calculates and represents the changes in data over local time periods. The model implementation process includes key steps such as generating contextual information, generating recurrent encoder-decoder features, and generating temporal fusion features, as detailed below:

[0039] Generate context information, wear level The prior distribution parameters s are processed by the feature selection network VSN to obtain the selected features. Wear amount The prior distribution parameters s are processed by a static orthometric vector encoder (SCE) to obtain the output encoded feature vector. , , and .

[0040] Generate loop encoder / decoder features, LSTMEnDe input features and characteristics and Output encoded temporal features Specifically: ① The sensor acquires the raw data signal; ② Feature extraction and feature selection are performed to obtain the processed data signal. ③ According to the set model coding length Input data segment sequence: ; ④ before Each data segment is processed sequentially through its corresponding Feature Selection Network (VSN), combined with Output features: ⑤ The data segments that need to be monitored at present The features are processed by the corresponding Feature Selection Network (VSN) and output as features. ⑥ will and As the first data segment feature The initial memory state of the LSTM unit, inputting the features of the first data segment. The calculation is performed using LSTM units, and the output unit outputs the data. And memory data; ⑦ The memory data output in the previous step is used as the feature of the second data segment. The initial memory state of the LSTM unit is then input, followed by the features of the second data segment. The calculation is performed using LSTM units, and the output unit outputs the data. and memorize the data; ⑧ Calculate sequentially until the output is reached. Features of each data segment Output data of the LSTM unit ; ⑨ Output data of each LSTM unit Through gated computation, hierarchical data fusion, and normalization, the processed data is output. The process can be expressed by the following formula: ;⑩ This refers to the generated loop encoding / decoding features.

[0041] Generate temporal fusion features, and the temporal fusion decoder (TFD) is used to process the encoded temporal features. Combination Data augmentation is performed; a temporal self-attention multi-head mechanism is adopted to calculate attention and weights, and then feature calculation and fusion are performed based on the weights; the fused features are subjected to gating calculation, skip-level connections and normalization, and element-level feedforward calculation is performed through a gated residual network (GRN) to finally output tool wear monitoring values. Specifically: ① Cyclic encoding and decoding features By using the gated residual network (GRN) corresponding to each sequence, combined with contextual information... The process involves: 1) data augmentation; 2) employing a temporal self-attention multi-head mechanism to perform attention and weight calculations on features of different time lengths; 3) feature calculation and fusion based on the weights; 4) data processing through gating computation, skip-level connections, and normalization; 5) element-level feedforward computation through a gated residual network (GRN); 6) data processing through gating computation, skip-level connections, and normalization; and 7) regression output of the current tool wear monitoring value from a fully connected layer. In other words, the above process can be summarized as: model construction and training of neural networks. That is, input wear amount Prior distribution parameters and Output the predicted wear amount at time t. Where Q is the encoding length during model prediction, and the wear amount is... Prior distribution parameters Including model code length Q, wear amount mean ,variance and index number. ,in The method proposed in this patent does not require prior wear data; it only requires input of the current cutting data and prior distribution parameters of the wear amount to perform tool wear monitoring. Furthermore, it employs an encoder-decoder structure to improve the interpretability of the model.

[0042] The above embodiments merely illustrate preferred embodiments of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A method for monitoring the wear of cutting tools, characterized in that, Includes the following steps: Step S1: Acquire signals x from multiple directions during the cutting process. N The signals include cutting force, vibration, and sound signals; the directions are the x-axis, y-axis, and z-axis directions based on the machining coordinate system; the x N For seven channels of data x N ={f x,n f y,n f z,n v x,n v y，n v z,n a e,n }, n∈[1,N], where, f x,n f y,n f z,n For cutting force signal, v x,n v y,n v z,n For vibration signals, a e,n It is a sound signal; Step S2: Collect the tool wear c after cutting is completed. K , will signal x N and wear amount c K The states are concatenated to obtain the state matrix x and the wear target value vector c. Step S3: Using a sliding window, extract data from the state matrix x to obtain M data segments Sec. m ; Calculate the data segment Sec m The time-domain features are obtained from Sec′ m Interpolate the target value vector c to obtain the wear amount Csec corresponding to each data segment. m The Csec m The initial value is 0; The calculation of the data segment Sec m The time-domain features are obtained from Sec′ m Specifically: Calculate the data statistics for each channel vertically to obtain the maximum value MAX. m，u Mean m，u Covariance VAR m，u , where u∈[1,7] is the index value of the channel; Step S4: Use Sec′ m Csec m Train a neural network and build a tool wear monitoring model; the method uses Sec′ m Csec m The specific calculation formula for training a neural network is as follows: y t =f(s,Sec′) (t-Q-1)：(t-1) ,Sec′ t ), where y t Let f(.) be the tool wear monitoring value at time t, Q be the model encoding length, and s be the wear amount Csec. m The prior distribution parameters; The specific calculation formula is as follows: y t = f(s, Sec′ (t-Q-1)：(t-1) , Sec′ t ), where s = {Q, MEAN Csec , VAR Csec , u}; Among them, y t Let f(.) be the predicted wear amount at time t, f(.) be the training neural network, Q be the model encoding length, and s be the wear amount Csec. m The prior distribution parameters of MEAN Csec Wear amount Csec m Mean, VAR Csec Wear amount Csec m The variance and u are index numbers; the wear amount Csc m The prior distribution parameters are: s={Q,MEAN Csec VAR Csec ,u}, where MEAN Csec Wear amount Csec m The mean of VAR Csec Wear amount Csc m covariance; The tool wear monitoring model includes: gated residual network GRN, feature selection network VSN, static orthometric vector encoder SCE, cyclic encoder-decoder LSTMEnDe, and temporal fusion decoder TFD; Wear amount Csec m The prior distribution parameters s are processed by the feature selection network VSN to obtain the selected features. Wear amount Csec m The prior distribution parameters s are processed by a static orthometric vector encoder (SCE) to obtain the output encoded feature vector c. e c c c h and c s ; LSTMEnDe Input Features and feature c c and c h Output encoded temporal features Temporal Fusion Decoder (TFD) is used to process encoded temporal features. Combined with c e Data augmentation is performed; a temporal self-attention multi-head mechanism is adopted to calculate attention and weights, and then feature calculation and fusion are performed based on the weights; the fused features are subjected to gating calculation, skip-level connection and normalization processing, and element-level feedforward calculation is performed through a gated residual network GRN, and finally the tool wear monitoring value is output.

Images