Cutter face abrasion loss prediction method and system and digital twin system

Abstract

The invention relates to the field of machining, in particular to the field of tool face abrasion prediction, and particularly discloses a tool face abrasion loss prediction method and system and a digital twin system. Inputting the preprocessed real-time machine tool spindle current data into the trained blunt state classification model to obtain a blunt state classification result; if the blunt state classification result is a non-blunt state, inputting the preprocessed real-time machine tool spindle current data into a trained first tool face abrasion loss prediction model to obtain a tool face abrasion loss prediction value; and if the blunt state classification result is a blunt state, inputting the preprocessed real-time machine tool spindle current data into the trained second tool face abrasion loss prediction model to obtain a tool face abrasion loss prediction value. According to the method, the tool wear states are classified, wear loss prediction models are trained for different wear states, wear state information is fully utilized, and higher prediction precision is achieved.

Description

Technical Field

[0001] The present invention relates to the field of machining, in particular to the field of prediction of tool flank wear, and specifically to a method and system for predicting tool flank wear amount and a digital twin system. Background Art

[0002] As the core tool in the machining process, the wear condition of the tool affects the machining efficiency and quality. The wear of the tool leads to a decline in cutting performance, which may in turn cause a series of problems such as a reduction in machining accuracy and deterioration of surface quality. Therefore, accurately and timely grasping the tool wear state is of great significance for ensuring the smooth progress of production.

[0003] The traditional detection method for tool wear amount mainly relies on manual regular inspection. This inspection method is not only time-consuming and laborious, but also affected by human factors, making it difficult to accurately and timely reflect the actual wear state of the tool. Especially in the high-speed and high-precision machining process, the tool wears quickly, and the manual inspection method can no longer meet the production requirements.

[0004] With the development of intelligent manufacturing technology, advanced technologies such as sensor technology, data analysis technology, and machine learning algorithms have been widely applied, providing new solutions for the real-time monitoring and prediction of tool wear state. By collecting the real-time operation data of the machine tool through sensors and using data analysis technology and machine learning algorithms to process and analyze these data, a prediction model relationship between the tool wear amount and the machine tool operation data can be established to achieve the prediction of tool wear condition. However, the traditional tool life prediction model directly uses the full-state data set for training, resulting in the insufficient utilization of tool wear state information and affecting the accuracy of the tool prediction model. Summary of the Invention

[0005] To solve the above problems, the present invention provides a method and system for predicting tool flank wear amount and a digital twin system, which classify the tool wear state and train the wear amount prediction models for different wear states respectively, making full use of the wear state information to achieve higher prediction accuracy.

[0006] In a first aspect, the technical solution of the present invention provides a method for predicting tool flank wear amount, including the following steps: Training a blunt state classification model and a tool flank wear amount prediction model based on historical machine tool spindle current data; the tool flank wear amount prediction model includes a first tool flank wear amount prediction model and a second tool flank wear amount prediction model; Collecting real-time machine tool spindle current data and performing preprocessing operations; Inputting the preprocessed real-time machine tool spindle current data into the trained blunt state classification model to obtain a blunt state classification result; If the classification result of the blunt state is the non-blunt state, input the preprocessed real-time machine tool spindle current data into the trained first flank wear amount prediction model to obtain the predicted value of the flank wear amount; If the classification result of the blunt state is the blunt state, input the preprocessed real-time machine tool spindle current data into the trained second flank wear amount prediction model to obtain the predicted value of the flank wear amount.

[0007] In an optional implementation manner, training the blunt state classification model based on historical machine tool spindle current data specifically includes: Configure blunt state labels for each historical machine tool spindle current data; Preprocess the historical machine tool spindle current data; Construct a first initial data set from the preprocessed historical machine tool spindle current data and the corresponding blunt state labels; Use the first initial data set to train the SSA-RF algorithm to obtain the blunt state classification model.

[0008] In an optional implementation manner, using the first initial data set to train the SSA-RF algorithm to obtain the blunt state classification model specifically includes: Initialize the parameters of the SSA algorithm. The initialized parameters include population size, maximum number of iterations, proportion of discoverers, proportion of guardians, and warning threshold; Use the initialized parameters to iterate the SSA algorithm to output the initial parameters of the RF algorithm; Input the initial parameters of the RF algorithm into the RF model, and use the first initial data set to train the RF model to obtain the trained blunt classification model.

[0009] In an optional implementation manner, using the first initial data set to train the RF model to obtain the trained blunt classification model specifically includes: Divide the first initial data set into N mutually exclusive subsets; Based on the N mutually exclusive subsets, use the N-fold cross-validation method to train the RF model to obtain the trained blunt classification model.

[0010] In an optional implementation manner, training the flank wear amount prediction model based on historical machine tool spindle current data specifically includes: Obtain the flank wear amount corresponding to each historical machine tool spindle current data; Preprocess the historical machine tool spindle current data; Construct a second initial data set from the preprocessed historical machine tool spindle current data and the corresponding flank wear amount; Use the second initial data set to train the flank wear amount prediction model.

[0011] In an alternative embodiment, the flank wear amount prediction model is trained using a second initial data set, which specifically includes: The second initial data set is divided into an unblunted state data set and a blunted state data set according to the blunting state corresponding to the flank wear amount; The Bootstrap method is used to expand the unblunted state data set and the blunted state data set; The expanded unblunted state data set is used to train the first flank wear amount prediction model; The expanded blunted state data set is used to train the second flank wear amount prediction model; Among them, the first flank wear amount prediction model and the second flank wear amount prediction model have the same structure, including an input layer, a CNN layer, an attention mechanism layer, a BiLSTM layer, a fully connected layer, and an output layer; the CNN layer includes a convolutional layer, a ReLU layer, and a global pooling layer; the attention mechanism layer uses a sigmoid activation function; the BiLSTM layer adopts a single-layer BiLSTM layer.

[0012] In an alternative embodiment, the historical machine tool spindle current data is preprocessed, which specifically includes: Wavelet threshold denoising processing is performed on the historical machine tool spindle current data; Feature extraction is performed on the denoised historical machine tool spindle current data, including time-domain features and frequency-domain features; The extracted features are normalized; PCA dimensionality reduction processing is performed on the normalized features.

[0013] In a second aspect, the technical solution of the present invention provides a tool flank wear amount prediction system, including: A model training unit: training a blunting state classification model and a flank wear amount prediction model based on historical machine tool spindle current data; the flank wear amount prediction model includes a first flank wear amount prediction model and a second flank wear amount prediction model; A current acquisition and processing unit: acquiring real-time machine tool spindle current data and performing preprocessing operations; A data classification unit: inputting the preprocessed real-time machine tool spindle current data into the trained blunting state classification model to obtain a blunting state classification result; A first wear amount prediction unit: if the blunting state classification result is the unblunted state, inputting the preprocessed real-time machine tool spindle current data into the trained first flank wear amount prediction model to obtain a flank wear amount prediction value; Second wear amount prediction unit: If the wear state classification result is the worn state, input the preprocessed real-time machine tool spindle current data into the trained second flank wear amount prediction model to obtain the predicted flank wear amount value.

[0014] Thirdly, the technical solution of the present invention provides a tool flank wear amount prediction digital twin system, including: Data communication and processing module: used for data acquisition, data transmission, data processing and data storage; the acquired data includes machine tool spindle current data, and the data processing executes the method described in any one of the above; Visualization module: used for displaying the real-time machining state of the machine tool and moving and scaling the virtual model of the machine tool; Data display module: used for displaying the real-time monitored machine tool spindle current data, machine tool spindle power, tool wear state, and historical data; Data management module: used for data modification, deletion and update processing.

[0015] In an optional embodiment, the visualization interface of the visualization module is built based on the Libigl framework, and the virtual model of the machine tool displayed on the visualization interface is modeled using SOLIDWORKS software; Modeling is performed in SOLIDWORKS software and a STEP format file is output. The STEP format file includes a machine tool base device, a workbench X-direction moving device, a saddle Y-direction moving device, and a spindle Z-direction moving device; The STEP file exported from SOLIDWORKS is imported into Magics, and a STL file with color information is output by Magics and input into Libigl.

[0016] A tool flank wear amount prediction method, system and digital twin system provided by the present invention have the following beneficial effects compared with the prior art: A wear state classification model and two flank wear amount prediction models are established. First, the wear state of the tool is classified by the wear state classification model into an unworn state and a worn state. For the unworn state, the first flank wear amount prediction model is used for wear amount prediction, and for the worn state, the second flank wear amount prediction model is used for wear amount prediction. By separately establishing flank wear amount prediction models for the non-worn state and the worn state, the present invention realizes the full utilization of wear state information and the accurate prediction of tool wear amount, improves the accuracy of prediction, and can reflect the wear state of the tool in real time, providing strong support for timely tool replacement and ensuring machining quality. Description of the Drawings

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0018] Figure 1 It is a schematic flowchart of a method for predicting the flank wear amount of a cutting tool provided by an embodiment of the present invention.

[0019] Figure 2 It is a schematic diagram of the overall framework of the classification prediction method in an embodiment of the present invention.

[0020] Figure 3 It is a schematic block diagram of the structure of a system for predicting the flank wear amount of a cutting tool provided by an embodiment of the present invention.

[0021] Figure 4 It is a schematic diagram of the framework structure of a digital twin system for predicting the flank wear amount of a cutting tool provided by an embodiment of the present invention.

[0022] Figure 5 It is a schematic diagram of the module function division and its content of a digital twin system for predicting the flank wear amount of a cutting tool provided by an embodiment of the present invention. Detailed implementation manners

[0023] In order to enable those skilled in the art of this technology to better understand the solution of the present invention, the following will further elaborate on the present invention in combination with the drawings and specific implementation manners. Obviously, the described embodiments are only some embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0024] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the technical field to which the present invention belongs. The terms used in the specification of the present invention herein are only for the purpose of describing specific embodiments, and are not intended to limit the present invention.

[0025] The following explains the key terms that appear in the present invention.

[0026] SSA algorithm: Sparrow Search Algorithm, a sparrow search algorithm.

[0027] RF algorithm: Random Forests, a random forest algorithm, which is an ensemble algorithm composed of decision trees.

[0028] PCA: Principal Component Analysis technology, also known as principal component analysis technology, aims to use the idea of dimensionality reduction to transform multiple indicators into a few comprehensive indicators.

[0029] BiLSTM layer: Bidirectional Long Short-Term Memory Network.

[0030] Figure 1 It is a schematic flowchart of a method for predicting the tool flank wear amount provided by an embodiment of the present invention. Among them, Figure 1 The execution subject can be a tool flank wear amount prediction system. The tool flank wear amount prediction method provided by the embodiment of the present invention is executed by a computer device. Correspondingly, the tool flank wear amount prediction system runs in the computer device. According to different requirements, the order of the steps in this flowchart can be changed, and some can be omitted.

[0031] Such as Figure 1 shown, the method includes the following steps.

[0032] S1. Train the blunt state classification model and the flank wear amount prediction model based on historical machine tool spindle current data; the flank wear amount prediction model includes a first flank wear amount prediction model and a second flank wear amount prediction model.

[0033] S2. Collect real-time machine tool spindle current data and perform preprocessing operations.

[0034] S3. Input the preprocessed real-time machine tool spindle current data into the trained blunt state classification model to obtain a blunt state classification result.

[0035] S4. If the blunt state classification result is the non-blunt state, input the preprocessed real-time machine tool spindle current data into the trained first flank wear amount prediction model to obtain a flank wear amount prediction value.

[0036] S5. If the blunt state classification result is the blunt state, input the preprocessed real-time machine tool spindle current data into the trained second flank wear amount prediction model to obtain a flank wear amount prediction value.

[0037] The tool flank wear prediction method of this embodiment establishes a blunt state classification model and two flank wear prediction models. First, the wear state of the tool is classified by the blunt state classification model into an unworn state and a worn state. For the unworn state, the first flank wear prediction model is used to predict the wear amount, and for the worn state, the second flank wear prediction model is used to predict the wear amount. It should be noted that during the initial wear stage, the tool wear value increases rapidly and the duration is short. The initial wear stage and the normal wear stage are uniformly classified as the unblunted state, while the rapid wear stage is classified as the blunt state. There are differences in the wear characteristics of tools in different states. If a single model is directly used for tool wear prediction, problems such as large differences in prediction accuracy between different states will occur, reducing the overall prediction accuracy. Therefore, this embodiment adopts a two-stage tool state monitoring and life prediction method of classifying first and then predicting. By separately establishing flank wear prediction models for the unblunted state and the blunted state, this embodiment realizes the full utilization of wear state information and the accurate prediction of tool wear amount, improves the accuracy of prediction, and can also reflect the wear state of the tool in real time, providing strong support for timely tool replacement and ensuring machining quality.

[0038] In this embodiment, the spindle current data of the machine tool is used as the monitoring object to predict the flank wear amount, and the monitoring principle is as follows. During the tool machining process, as the wear degree increases, the friction force between the tool flank and the workpiece increases, which will lead to an increase in cutting force and cutting power. The additional cutting force caused by tool wear includes the radial force caused by the flank wear of the tool and the friction force between the tool and the workpiece caused by wear : (1-1) (1-2) In the formula, H represents the Brinell hardness of the workpiece material, VB represents the flank wear amount of the tool, μ represents the sliding friction coefficient between the tool and the workpiece, s represents the flank wear land length of the tool.

[0039] Then, without considering the changes in cutting parameters and cutting conditions, the increase in cutting power caused by tool wear is: (1-3) In the formula, represents the linear velocity between the contact points of the tool and the workpiece during machining.

[0040] It can be seen from formula (1-3) that when the external conditions are constant, the machining power of the machine tool spindle increases with the flank wear value of the tool VBincreases linearly with the increase of . Since it takes a long time to actually collect the power signal, there are certain difficulties. Under the condition that the spindle motor voltage is constant, the spindle current data can reflect the change of the spindle power, that is, the change of the spindle power during the machining process can be reflected by monitoring the spindle motor current signal.

[0041] The tool wear process can be divided into the initial wear stage, the normal wear stage, and the rapid wear stage. In the initial wear stage, the tool wear value increases rapidly and the duration is short. The initial wear stage and the normal wear stage can be unified and classified as the unblunted state, while the rapid wear stage is classified as the blunted state. There are differences in the tool wear characteristics of different states. If a single model is directly used for tool wear prediction, there will be problems such as a large difference in prediction accuracy between different states, reducing the overall prediction accuracy. Therefore, this embodiment adopts a two-stage tool flank wear prediction method of classification first and then prediction. Specifically, this embodiment trains the blunted state classification model and the flank wear amount prediction model based on the historical machine tool spindle current data. The flank wear amount prediction model includes two prediction models, the first flank wear amount prediction model and the second flank wear amount prediction model, which are used to predict the tools in the unblunted state and the blunted state respectively. Figure 2 is the overall framework schematic diagram of the classification prediction method in this embodiment. The blunted state classification model adopts the SSA-RF model, and the flank wear amount prediction model adopts the CNN-Attention-BiLSTM model. The first step is the classification stage, where the tool signal features are used as the input of the SSA-RF model to predict the current wear state of the tool. In the second step, for the unblunted state and the blunted state of the tool, there are corresponding CNN-Attention-BiLSTM prediction models respectively. The classified tool signal features are used as the input, and the predicted tool wear value is used as the output. This embodiment avoids using the tool full life cycle feature data for the training of a single model, makes full use of the information of the tool wear state category to improve the accuracy of the tool wear prediction value, and improves the utilization rate of the monitoring information and the reliability of the prediction result.

[0042] The training of the blunted state classification model based on the historical machine tool spindle current data in this embodiment specifically includes the following steps.

[0043] S1.11, Configure blunted state labels for each historical machine tool spindle current data.

[0044] In this step, classification labels are configured for each original data to train the blunted state classification model. An optional implementation manner is to obtain the flank wear amount corresponding to each historical machine tool spindle current data, and set the tool blunting standard. For example, the tool blunting standard is selected as , historical machine tool spindle current data less than this standard is configured as the non-blunted state, and historical machine tool spindle current data greater than or equal to this standard is configured as the blunted state.

[0045] S1.12, preprocess the historical machine tool spindle current data.

[0046] S1.12.1, perform wavelet threshold denoising on the historical machine tool spindle current data.

[0047] In a complex production site environment, the acquisition of sensor data is inevitably contaminated with many noises and interferences, such as voltage pulses generated during machine tool startup and braking, power grid voltage fluctuations, etc. To avoid distortion of the current signal and more accurately achieve tool condition monitoring, it is necessary to preprocess the current signal to reduce the influence of noise and interference.

[0048] At the same time, in the acquired current signal, in addition to random interference, there will also be periodic interference containing components with power frequency and double power frequency. To eliminate random interference and periodic interference in the current signal, it is necessary to preprocess the acquired sensor signal to obtain higher-quality data.

[0049] The basic idea of wavelet threshold denoising is that after the signal is wavelet-transformed, the wavelet coefficients generated by the signal contain important information of the signal. After the signal is wavelet-decomposed, the wavelet coefficients of the signal are larger, the wavelet coefficients of the noise are smaller, and the wavelet coefficients of the noise are smaller than those of the signal. By selecting a suitable threshold, the wavelet coefficients greater than the threshold are considered to be generated by the signal and should be retained, while those less than the threshold are considered to be generated by the noise and are set to zero to achieve the purpose of denoising.

[0050] The specific steps of the wavelet threshold denoising process are as follows: 1) Select a wavelet basis type and the decomposition level N of the wavelet, and perform N-layer wavelet decomposition on the noisy signal to obtain the wavelet coefficients of each layer; 2) Quantify the high-frequency wavelet coefficients from the first layer to the Nth layer with a suitable threshold, and use their correlation to remove noise; 3) Reconstruct the processed signal to obtain the denoised signal.

[0051] In this embodiment, fixed threshold denoising is adopted to make the denoising more effective.

[0052] S1.12.2, extract features from the historical machine tool spindle current data after denoising processing, including time-domain features and frequency-domain features.

[0053] The methods for extracting data features mainly include time-domain analysis and frequency-domain analysis. Time-domain feature analysis takes a short time to calculate and can intuitively and accurately reflect data features; frequency-domain feature analysis can dig deeper information from data and obtain characteristics such as signal frequency structure and phase. The data signals collected by the sensor system are discrete signals and need to be processed using the Discrete Fourier Transform (DFT).

[0054] In a specific embodiment, the extracted time-domain features include maximum value, minimum value, peak-to-peak value, rectified average value, variance, root mean square value, peak factor, waveform factor, skewness, kurtosis, impulse factor, and margin factor. The extracted frequency-domain features include centroid frequency, mean square frequency, frequency variance, and frequency band energy.

[0055] S1.12.3, normalize the extracted features.

[0056] Before dimensionality reduction of the features, the extracted feature data needs to be normalized. In this embodiment, the MATLAB built-in function mapminmax is used for data normalization.

[0057] S1.12.4, perform PCA dimensionality reduction on the normalized features.

[0058] PCA maps high-dimensional data to a low-dimensional space through linear mapping and is one of the classic methods for feature dimensionality reduction. PCA calculates the centroid of the data and establishes a new coordinate system so that the projection of each data in this coordinate system is the most dispersed, that is, the direction with the largest variance is the principal component.

[0059] S1.13, construct the first initial dataset from the preprocessed historical spindle current data of the machine tool and the corresponding wear state labels.

[0060] S1.14, use the first initial dataset to train the SSA-RF algorithm to obtain a wear state classification model.

[0061] S1.14.1, initialize the parameters of the SSA algorithm. The initialized parameters include population size, maximum number of iterations, proportion of discoverers, proportion of guardians, and guard threshold.

[0062] S1.14.2, use the initialized parameters to iterate the SSA algorithm to output the initial parameters of the RF algorithm.

[0063] S1.14.3, input the initial parameters of the RF algorithm into the RF model and use the first initial dataset to train the RF model to obtain a trained wear classification model.

[0064] S1.14.3.1, divide the first initial data set into N mutually exclusive subsets.

[0065] S1.14.3.2, based on the N mutually exclusive subsets, use the N-fold cross-validation method to train the RF model to obtain the trained blunt classification model.

[0066] In a specific embodiment, according to the tool wear change trend in the NASA data set, the tool blunt standard is selected as , then 225 groups of data in the non-blunt state and 192 groups of blunt data can be obtained, and the data distribution is relatively uniform and reasonable. Randomly shuffle the data set and set the test set to account for 30% of the total data set. However, due to the small total amount of the data set, the number of training sample sets and test sample sets generated after segmentation will be even smaller, which will have a greater impact on subsequent modeling. Therefore, the 5-fold cross-validation method is used to solve the problem of dividing the small sample data set. After completing the data set division and using the MATLAB function mapminmax for data normalization processing, use the SSA optimization algorithm to optimize the initial parameters of the random forest. Set the population size to 30, the maximum number of iterations to 40, the proportion of discoverers to 0.8, the proportion of vigilants to 0.2, and the vigilance threshold to 0.8. When the SSA optimization algorithm iterates 15 times, its fitness value has already tended to be stable, and it has a fast convergence speed. According to the results obtained by SSA optimization, input the initial parameters into the RF model and train on the training set to obtain a better classification effect.

[0067] In this embodiment, the first flank wear prediction model and the second flank wear prediction model have the same structure, both adopting the CNN-Attention-BiLSTM model. The CNN can re-extract features, but it cannot discover the long-term dependence problem in the data, while the BiLSTM can well learn time series data. At the same time, the Attention mechanism can capture features more relevant to the current task, reduce the attention to non-critical features, and thus improve the prediction accuracy of the model. The CNN-Attention-BiLSTM model includes an input layer, a CNN layer, an attention mechanism layer, a BiLSTM layer, a fully connected layer, and an output layer. The input layer obtains tool wear data; the convolutional layer in the CNN layer re-extracts the input data features, selects important feature data, and the pooling layer reduces the dimension of the data output by the convolutional layer; the Attention layer selectively weights different feature information and selects some key information; the BiLSTM layer passes the feature data with attention from the Attention layer through the forward and backward LSTM units respectively to obtain the output data; the fully connected layer is used to map the output of the BiLSTM layer to the desired output dimension, realizing feature combination and dimension reduction; the output layer performs output calculation on the data output by the BiLSTM layer to obtain the final tool wear prediction value.

[0068] In this embodiment, the flank wear prediction model is trained based on historical machine tool spindle current data, which specifically includes the following steps.

[0069] S1.21, Obtain the flank wear amount corresponding to each historical machine tool spindle current data.

[0070] The purpose of this step is to obtain the flank wear amount corresponding to each original data to construct a data set for training the flank wear prediction model.

[0071] S1.22, Preprocess the historical machine tool spindle current data.

[0072] The preprocessing in this step is the same as that in training the blunt state classification model, which will not be elaborated here.

[0073] S1.23, Construct a second initial data set from the preprocessed historical machine tool spindle current data and the corresponding flank wear amount.

[0074] S1.24, Use the second initial data set to train the flank wear prediction model.

[0075] S1.24.1, Divide the second initial data set into an unblunted state data set and a blunted state data set according to the blunt state corresponding to the flank wear amount.

[0076] It should be noted that the tool wear standard is set to divide the wear state, and then the second initial data set is divided into an un-worn state data set and a worn state data set.

[0077] S1.24.2. Use the Bootstrap method to expand the un-worn state data set and the worn state data set.

[0078] It should be noted that if the amount of data in the data set itself is small, the data set obtained after classification will be even smaller, which will affect the model training. Therefore, before inputting the data into the model for training, first use the Bootstrap method to expand the data set and divide it into a training set and a test set.

[0079] S1.24.3. Use the expanded un-worn state data set to train the first flank wear amount prediction model.

[0080] S1.24.4. Use the expanded worn state data set to train the second flank wear amount prediction model.

[0081] In a specific embodiment, a sequence input layer and a sequence folding layer are set to receive and process sequence data with different numbers of features, facilitating subsequent model iteration. In the convolutional layer, in order to overcome the problem of gradient dispersion, the ReLU activation function is selected. The ReLU function can set negative values to zero and keep positive values unchanged. For the two convolutional layers used, convolutional kernels of size (1,1) are adopted, the convolutional stride is 1, and the numbers of convolutional kernels are 32 and 64 respectively. The model introduces a global pooling layer to reduce the number of parameters and computational complexity, which can prevent overfitting and accelerate model convergence. A single-layer BiLSTM layer is used, which has 6 hidden units. The SE (Squeeze-and-Excitation) attention mechanism layer is used to adaptively learn the important features in the input sequence. In the SE attention mechanism, the sigmoid activation function is used to generate weights. The number of iterations is set to 700, and the initial learning rate is 0.01. The training set data is used to train the model. Through experiments, when the model iterates 400 times, it has already achieved good results. After training the CNN-Attention-BiLSTM model using the data set, select as the model performance evaluation index. Through experiments, the tool flank wear amount prediction method of this embodiment can achieve a good fitting effect.

[0082] In the above text, an embodiment of a tool flank wear amount prediction method is described in detail. Based on the tool flank wear amount prediction method described in the above embodiment, an embodiment of the present invention also provides a tool flank wear amount prediction system corresponding to this method.

[0083] Figure 3The present invention provides a schematic block diagram of a tool face wear prediction system structure. The tool face wear prediction system can be divided into multiple functional units according to the functions it performs. The unit referred to in the present invention refers to a series of computer program segments that can be executed by at least one processor and can complete fixed functions, which are stored in a memory.

[0084] Model training unit: Based on the historical machine tool spindle current data, the blunting state classification model and the tool surface wear prediction model are trained; the tool surface wear prediction model includes the first tool surface wear prediction model and the second tool surface wear prediction model.

[0085] Current acquisition and processing unit: collects real-time machine tool spindle current data and performs pre-processing operations.

[0086] Data classification unit: input the preprocessed real-time machine tool spindle current data into the trained blunt state classification model to obtain the blunt state classification result.

[0087] The first wear amount prediction unit: if the blunt state classification result is not blunt, the pre-processed real-time machine tool spindle current data is input into the trained first tool face wear amount prediction model to obtain the tool face wear amount prediction value.

[0088] Second wear prediction unit: If the blunt state classification result is blunted, the pre-processed real-time machine tool spindle current data is input into the trained second tool face wear prediction model to obtain the tool face wear prediction value.

[0089] The tool face wear prediction system of the present embodiment is used to implement the aforementioned tool face wear prediction method, so the specific implementation method of the system can be seen in the embodiment section of the tool face wear prediction method in the previous text, so its specific implementation method can refer to the description of the corresponding embodiments of each part, and will not be introduced in detail here.

[0090] In addition, since the tool face wear prediction system of the present embodiment is used to implement the aforementioned tool face wear prediction method, its function corresponds to that of the aforementioned method and will not be described in detail here.

[0091] This embodiment provides a digital twin system for predicting tool surface wear based on the prediction model provided in the above embodiment. The digital twin system uses information technology and sensor technology and twin data to achieve real-time interaction, thereby achieving the effect of comprehensive control of tool wear status during the processing process. Figure 4 It is a schematic diagram of the framework structure of the digital twin system, which is based on the five-dimensional digital twin model and includes five parts: physical entity, virtual entity, service system, twin data and connection. The various parts interact with each other through twin data.

[0092] (1)Physical entity Physical entities mainly include entities in the physical space such as machine tools, cutting tools, workpieces, sensors, and data acquisition devices. They mainly undertake production and processing tasks, and at the same time, they also need to collect and store data for the rest of the parts.

[0093] (2)Virtual model The virtual model is the core of the digital twin system. It can describe the physical entity from dimensions such as geometric, physical, behavior, and rule models. Its shape, size, material, motion coupling relationship, and instruction response are all consistent with the physical entity, and it can realize the mapping of the entity tool state based on the tool state monitoring and life prediction model.

[0094] (3)Twin data Twin data is the basis for the operation of the digital twin system, including the collection of real-time data and the processing of historical data. The data content includes sensor data, processing parameters, simulation prediction data, etc. After combining various data, it is transmitted to the service system for real-time decision-making.

[0095] (4)Service system The main function of the service system is to realize data visualization analysis and equipment monitoring. It can monitor the real-time wear state and life of the cutting tool, and then realize intelligent decision-making.

[0096] (5)Connection Connect the physical entity, virtual entity, service system, and twin data. The real-time, dynamic, and interactive connection is the artery of the digital twin system.

[0097] The digital twin system of this embodiment collects data in real time through sensors. After preprocessing, feature extraction, and dimensionality reduction of the data, the data is input into the two-stage tool state monitoring and life prediction model to display the current wear state and predicted life of the tool in real time, realizing the twin mapping of tool wear. At the same time, the storage and display of historical data are also required.

[0098] Based on the above functional expectations, the digital twin system can be divided into a data communication and processing module, a visualization module, a data display module, and a data management module. Figure 5 It is a schematic diagram of the module function division and its content of the digital twin system of this embodiment.

[0099] Data communication and processing module: used for data acquisition, data transmission, data processing, and data storage; the collected data includes machine tool spindle current data, and the data processing executes the method described in any one of the above.

[0100] Visualization module: used for displaying the real-time machining state of the machine tool and moving and scaling the virtual model of the machine tool.

[0101] Data display module: used to display real-time monitored spindle current data of the machine tool, spindle power of the machine tool, tool wear status, and historical data.

[0102] Data management module: used to perform data modification, deletion, and update processing.

[0103] In an optional implementation, in the data communication and processing module, data acquisition is based on the FANUC 0i system and an external spindle current sensor. Through the open Ethernet interface CD38A of the FANUC system, a wired connection is made using a network cable to ensure stable and reliable data transmission. At the same time, the IP address of the PC needs to be configured so that the PC and the numerical control system are under the same local area network and gateway.

[0104] The FOCAS1 / 2 (FANUC Open CNC API Specifications 1 or 2) functions are used to read the machine tool data. FOCAS1 is mainly applied to series such as 0i and 16i / 18i, while FOCAS2 is for series such as 30i / 31i.

[0105] To implement the digital twin system, appropriate 3D modeling tools and virtual scene development tools are selected. In this embodiment, SOLIDWORKS software is used as the modeling tool software, and Libigl is used to build the virtual scene.

[0106] Using 3D software modeling is the first step in constructing the digital space. First, the overall dimensions of the machine tool, the area of the workbench, the types and specification dimensions of the tools used need to be measured and recorded on-site. Subsequently, modeling is carried out in SOLIDWORKS software according to the actual situation of the machine tool and a STEP format file is output. In order for the subsequent machine tool model to reproduce the actual machining operations of the machine tool, the output STEP format file needs to be divided into four major parts, including the machine tool base device, the X-direction moving device of the workbench, the Y-direction moving device of the saddle, and the Z-direction moving device of the spindle.

[0107] Import the STEP file exported from SOLIDWORKS into Magics. Output the STL file with color information from Magics and input it to Libigl. The readSTL() method built into Libigl can be used to read it, and the vertex information and face information of the 3D model are saved in the form of Eigen::Matrix. Call the minCoeff() and maxCoeff() methods to calculate the coordinate range of the model, and calibrate the positions of the four parts of the machine tool base, saddle, workbench, and spindle. Call viewer.data().set_colors() to set the model colors respectively, and call the viewer.data().set_mesh() and viewer.launch() methods to render the 3D model using OpenGL functions. Rewrite the viewer.callback_pre_draw() function, write the moving algorithm for the machine tool module, and offset the model coordinates. Set the animation frame rate through the viewer.core().animation_max_fps method, and call the viewer.core().is_animating() method to realize the real-time display of the positions of each module in the machine tool model.

[0108] In an alternative embodiment, the processing of data and the construction of the tool state monitoring and life prediction model are both realized by programming based on MATLAB software, while the digital twin visualization interface used is built based on the Libigl framework and written in C++ language. Since the two languages are not compatible, joint programming needs to be realized in a calling manner. That is, after the main program.m file is written and compiled into a DLL file after being packaged, it can be called by the C++ program.

[0109] The above-disclosed are only the preferred embodiments of the present invention, but the present invention is not limited thereto. Any non-creative changes that can be thought of by those skilled in the art, as well as several improvements and refinements made without departing from the principle of the present invention, should fall within the protection scope of the present invention.

Claims

1. A method for predicting tool surface wear, characterized in that: The following steps are involved: Based on the historical machine tool spindle current data, the blunting state classification model and the tool face wear prediction model are trained; the tool face wear prediction model includes the first tool face wear prediction model and the second tool face wear prediction model; Collect real-time machine tool spindle current data and perform preprocessing operations; The preprocessed real-time machine tool spindle current data is input into the trained blunt state classification model to obtain the blunt state classification result; If the blunt state classification result is not blunt, the pre-processed real-time machine tool spindle current data is input into the trained first tool face wear prediction model to obtain the tool face wear prediction value; If the blunt state classification result is blunted, the preprocessed real-time machine tool spindle current data is input into the trained second tool face wear prediction model to obtain the tool face wear prediction value.

2. The tool surface wear prediction method according to claim 1, characterized in that: The blunting state classification model is trained based on historical machine tool spindle current data, including: Configure blunting status tags for each historical machine tool spindle current data; Preprocess the historical machine tool spindle current data; Constructing a first initial data set using the preprocessed historical machine tool spindle current data and the corresponding blunting state labels; The first initial data set is used to train the SSA-RF algorithm to obtain a dulling state classification model.

3. The tool surface wear prediction method according to claim 2, characterized in that: The SSA-RF algorithm is trained using the first initial data set to obtain a blunting state classification model, specifically including: Initialize the parameters of the SSA algorithm, including population size, maximum number of iterations, proportion of discoverers, proportion of alerters, and alert threshold; Use the initialization parameters to iterate the SSA algorithm and output the initial parameters of the RF algorithm; The initial parameters of the RF algorithm are input into the RF model, and the RF model is trained using the first initial data set to obtain a trained sharpness classification model.

4. The method for predicting tool surface wear according to claim 3, characterized in that: The RF model is trained using the first initial data set to obtain a trained blunt classification model, specifically including: Divide the first initial data set into N mutually exclusive subsets; The RF model is trained based on N mutually exclusive subsets using the N-fold cross validation method to obtain the trained blunt classification model.

5. The method for predicting tool surface wear according to claim 2, characterized in that: The tool wear prediction model is trained based on historical machine tool spindle current data, including: Obtain the tool surface wear corresponding to each historical machine tool spindle current data; Preprocess the historical machine tool spindle current data; The preprocessed historical machine tool spindle current data and the corresponding tool surface wear amount are used to construct a second initial data set; The tool surface wear prediction model is trained using the second initial data set.

6. The method for predicting tool surface wear according to claim 5, characterized in that: The tool wear prediction model is trained using the second initial data set, specifically including: Dividing the second initial data set into a non-blunt state data set and a blunt state data set according to the blunt state corresponding to the tool surface wear amount; The Bootstrap method is used to expand the unblunted state data set and the blunted state data set; The first tool face wear prediction model is trained using the expanded un-blunted state data set; The expanded blunt state data set is used to train the second tool face wear prediction model; Among them, the first blade wear prediction model and the second blade wear prediction model have the same structure, including input layer, CNN layer, attention mechanism layer, BiLSTM layer, fully connected layer and output layer; the CNN layer includes convolution layer, ReLU layer and global pooling layer; the attention mechanism layer uses the sigmoid activation function; the BiLSTM layer uses a single-layer BiLSTM layer.

7. The method for predicting tool surface wear according to any one of claims 2 to 6, characterized in that: Preprocess the historical machine tool spindle current data, including: Perform wavelet threshold denoising on historical machine tool spindle current data; Feature extraction is performed on the historical machine tool spindle current data after denoising, including time domain features and frequency domain features; Normalize the extracted features; The normalized features are subjected to PCA dimensionality reduction.

8. A tool surface wear prediction system, characterized in that: include: Model training unit: training the blunting state classification model and the tool surface wear prediction model based on the historical machine tool spindle current data; the tool surface wear prediction model includes the first tool surface wear prediction model and the second tool surface wear prediction model; Current acquisition and processing unit: collects real-time machine tool spindle current data and performs preprocessing operations; Data classification unit: input the preprocessed real-time machine tool spindle current data into the trained blunt state classification model to obtain the blunt state classification result; The first wear amount prediction unit: if the blunt state classification result is not blunt state, the pre-processed real-time machine tool spindle current data is input into the trained first tool face wear amount prediction model to obtain the tool face wear amount prediction value; Second wear prediction unit: If the blunt state classification result is blunted, the pre-processed real-time machine tool spindle current data is input into the trained second tool face wear prediction model to obtain the tool face wear prediction value.

9. A digital twin system for predicting tool surface wear, characterized in that: include: Data communication and processing module: used for data acquisition, data transmission, data processing and data storage; the acquired data includes the spindle current data of the machine tool, and the data processing executes the method described in any one of claims 1 to 7; Visualization module: used to display the real-time processing status of the machine tool and move and scale the virtual model of the machine tool; Data display module: used to display the real-time monitoring of machine tool spindle current data, machine tool spindle power, tool wear status, and historical data; Data management module: used for data modification, deletion and update processing.

10. The tool surface wear prediction digital twin system according to claim 9, characterized in that: The visualization interface of the visualization module is built based on the Libigl framework, and the virtual model of the machine tool displayed in the visualization interface is modeled using SOLIDWORKS software; Modeling is performed in SOLIDWORKS software and a STEP format file is output. The STEP format file includes a machine tool base device, a worktable X-direction moving device, a saddle Y-direction moving device, and a spindle Z-direction moving device; Import the STEP file exported from SOLIDWORKS into Magics, and then output the STL file with color information from Magics and import it into Libigl.

Images