Method and device for predicting remaining service life of cutter

Abstract

The invention provides a method and a device for predicting the remaining service life of a cutter. The method comprises the following steps of: making a plurality of labels for wear values in a cutter wear experiment data set; vibration signal data and cutting force signal data of the tool are collected and preprocessed; randomly dividing the plurality of labels and the preprocessed vibration signal data and cutting force signal data into a training set, a verification set and a test set; extracting multi-domain features of the training set, the verification set and the test set; inputting the multi-domain features into a first training model for training, and outputting nonlinear features; inputting the multi-domain features into a second training model for training, and outputting linear features; fusing the nonlinear feature and the linear feature to obtain a fused feature; and inputting the fusion feature into a third training model, and outputting a predicted value of the remaining service life of the cutter. The method improves the prediction precision of the residual service life of the cutter.

Description

Technical Field

[0001] The present invention belongs to the technical field of high-end equipment, and specifically relates to a method and device for predicting the remaining service life of a tool. Background Art

[0002] At present, the remaining service life prediction methods for CNC machine tool tools are mainly divided into three categories: manual prediction methods, prediction methods based on physical models, and prediction methods based on data-driven.

[0003] Among them, the manual prediction method requires CNC machine tool operators to subjectively judge the remaining service life of the tool based on the sound of the processing site, the quality of the processed parts and other information. This method requires the operator to have rich processing experience, is highly subjective, cannot be widely learned by most people, and is prone to wrong predictions due to different operator experience.

[0004] The prediction method based on the physical model needs to find out the characteristics of tool failure according to the specific conditions of the CNC machining process and establish a mathematical model for predicting the remaining tool life. In this method, the wear of CNC tool during machining is affected by the material of the machined parts, tool material, machining temperature, feed rate, spindle speed and other conditions. The wear mechanism is very complex and it is difficult to accurately establish a mathematical model for tool wear. Therefore, the prediction method based on the physical model cannot be widely promoted and used.

[0005] Compared with the prediction method based on physical models, the prediction method based on data-driven uses the machine learning method of artificial intelligence, mainly based on the accumulation of a large amount of data in the production process, and uses the existing processing data for feature extraction, model prediction, etc., to achieve real-time prediction and operation and maintenance of tool wear, which can be widely used. However, in the related technology, most of the data-driven tool remaining service life prediction methods rely on knowledge reserves when extracting features, and require manual modification of relevant parameters for learning, which will increase the workload of researchers, and the model prediction accuracy will vary with the different knowledge reserves of different researchers. Too much interference from human factors leads to poor model prediction accuracy. Summary of the invention

[0006] In view of the shortcomings of the prior art, the present invention discloses a method for predicting the remaining service life of a tool. The method combines a denoising auto encoder (Denoising Auto Encoder, DAE) and a kernel principal component analysis (Kernel Principal Component Analysis, KPCA) to improve the data-driven prediction method of the related technology, so as to improve the prediction accuracy of the remaining service life of the tool.

[0007] In order to achieve the above object, the present invention provides a method for predicting the remaining service life of a tool, comprising:

[0008] Several labels are made for the wear values ​​in the tool wear experimental data set, among which: the average wear value of the cutting edge of the tool flank is selected as the prediction label of the tool's remaining service life;

[0009] Collect the vibration signal data and cutting force signal data of the tool and perform pre-processing;

[0010] The labels and the preprocessed vibration signal data and cutting force signal data are randomly divided into a training set, a validation set and a test set;

[0011] Extracting multi-domain features of the training set, validation set, and test set;

[0012] Inputting the multi-domain features into a first training model for training, and outputting nonlinear features;

[0013] The multi-domain features are input into a second training model for training, and linear features are output;

[0014] The nonlinear feature is fused with the linear feature to obtain a fused feature as a sensitive feature representation related to tool degradation information in the sample;

[0015] The fused feature is input into a third training model to output a prediction value of the remaining service life of the tool; wherein the third training model comprises a stacked bidirectional gated recurrent unit network and a fully connected network, and an input end of the fully connected network is connected to an output end of the stacked bidirectional gated recurrent unit network;

[0016] The stacked bidirectional gated recurrent unit network is composed of two stacked bidirectional gated recurrent units.

[0017] Each bidirectional gated recurrent unit includes a forward GRU network and a backward GRU network. The forward GRU network receives the sequence information of the current time series calculated by the fused features, and the backward GRU network receives the reverse sequence information of the same time series calculated by the fused features. The bidirectional gated recurrent unit outputs the weighted value of the sequence information and the reverse sequence information to obtain the predicted value of the remaining service life of the tool.

[0018] In one embodiment of the present invention, the first training model adopts an improved denoising autoencoder, comprising a noise module, an encoder module, and a decoder module; the input end of the encoder module is connected to the input end of the noise module and the output end of the decoder module, and the input end of the decoder module is connected to the output end of the encoder module; wherein:

[0019] The noise module is used to dynamically adjust the type and intensity of added noise to generate noisy multi-domain features according to the training progress or the continuity of the input multi-domain features;

[0020] The encoder module and the decoder module adopt a fully connected network with a symmetric number of neural units; wherein:

[0021] Convolutional layers and recurrent layers are introduced into the encoder module and the decoder module to capture local features and temporal dependencies of data respectively;

[0022] Increasing the number of layers of the encoder module and the decoder module, and adding a Dropout layer after each layer of the encoder module and the decoder module to randomly discard the output of a portion of neurons;

[0023] A residual connection or a skip connection is added between the encoder module and the decoder module.

[0024] In one embodiment of the present invention, dynamically adjusting the type and intensity of added noise to generate a noisy multi-domain feature according to the training progress or the continuity of the input multi-domain feature includes:

[0025] In the early stage of training the first training model, injecting noise of greater intensity and multiple types;

[0026] In the later stage of training of the first training model, gradually reducing the noise intensity;

[0027] If the multi-domain feature is identified as a continuously changing feature, adding Gaussian noise is selected;

[0028] If the multi-domain feature is identified as a discretized feature, salt and pepper noise and / or Poisson noise is added.

[0029] In one embodiment of the present invention, inputting the multi-domain features into the first training model for training includes:

[0030] The multi-domain features extracted from the training set and the validation set are input into the first training model, wherein:

[0031] After the multi-domain feature passes through the noise module, the type and intensity of the added noise are dynamically adjusted according to the training progress or the continuity of the input multi-domain feature to generate a noisy multi-domain feature;

[0032] The encoder module compresses the noisy multi-domain features into low-dimensional hidden layer representations through nonlinear transformation and outputs the noisy features;

[0033] The noise-containing feature is further input into the decoder module to output a reconstructed feature, and the reconstructed feature is fed back to the input end of the encoder module;

[0034] Calculating an error value between the reconstructed feature and the multi-domain feature to obtain a reconstruction error, and adjusting a network parameter of the first training model according to the reconstruction error;

[0035] When it is determined that the training of the first training model is completed,

[0036] The multi-domain features extracted from the training set, the validation set, and the test set are input into the trained first training model, and the encoder module outputs the nonlinear features.

[0037] In one embodiment of the present invention, binary cross entropy or the like is used as a loss function, and the back propagation algorithm is combined with the AdamW algorithm to adjust the parameters of the encoder module and the decoder module so as to minimize the reconstruction error.

[0038] In one embodiment of the present invention, determining whether the first training model has been trained includes:

[0039] Compare the loss of this round of the validation set with the loss of the previous round, and save the model parameters if the loss of this round is less than the loss of the previous round;

[0040] It is determined whether the current number of network iterations meets the preset number value. If the preset number value is reached, the training of the first training model is completed, and the trained first training model is saved.

[0041] In one embodiment of the present invention, the second training model adopts a support vector regression model based on kernel principal component analysis, maps the multi-domain features to a high-dimensional space by introducing a kernel function, performs linear analysis in the high-dimensional space, and captures the linear features, including:

[0042] Define the parameter grid for kernel principal component analysis;

[0043] Using parameter grid search for cross validation, find the best dimension parameters by training the second training model on different dimensions and evaluating the performance;

[0044] Using the optimal dimension parameters, the second training model is retrained.

[0045] In one embodiment of the present invention, before the fusion feature is input into the third training model, the method further comprises:

[0046] The fused features are adjusted to be suitable for input data of the third training model.

[0047] In one embodiment of the present invention, a sliding window method is used to adjust the dimension of the fusion feature to be suitable for the third training model.

[0048] Another aspect of the present invention also discloses a tool remaining service life prediction device, adopting the above-mentioned tool remaining service life prediction method, the device comprises:

[0049] The data cleaning and labeling module is used to create several labels for the wear values ​​in the tool wear experiment data set, including: selecting the average wear value of the cutting edge of the tool back face as the prediction label of the tool's remaining service life; and collecting and pre-processing the tool's vibration signal data and cutting force signal data;

[0050] A data set partitioning module randomly divides the labels and the preprocessed vibration signal data and cutting force signal data into a training set, a validation set and a test set;

[0051] A feature extraction module, used to extract multi-domain features of the training set, validation set and test set;

[0052] A first training module, used for inputting the multi-domain features into a first training model for training, and outputting nonlinear features;

[0053] A second training module is used to input the multi-domain features into a second training model for training and output linear features;

[0054] A feature fusion module is used to fuse the nonlinear feature with the linear feature to obtain a fused feature as a sensitive feature representation related to tool degradation information in the sample;

[0055] A third training module is used to input the fusion feature into a third training model to output a prediction value of the remaining service life of the tool;

[0056] Wherein: the third training model includes a stacked bidirectional gated recurrent unit network and a fully connected network, and the input end of the fully connected network is connected to the output end of the stacked bidirectional gated recurrent unit network; wherein the stacked bidirectional gated recurrent unit network is composed of two stacked bidirectional gated recurrent units; each bidirectional gated recurrent unit includes a forward GRU network and a backward GRU network, the forward GRU network receives the sequence information of the current time series calculated by the fused features, and the backward GRU network receives the reverse sequence information of the same time series calculated by the fused features, and the bidirectional gated recurrent unit outputs the weighted value of the sequence information and the reverse sequence information to obtain the remaining service life prediction value of the tool.

[0057] On the other hand, the present invention also provides a readable storage medium, on which a program or instruction is stored. When the program or instruction is executed by a processor, the steps of the above-mentioned tool remaining service life prediction method are implemented and the same technical effect can be achieved.

[0058] It can be seen from the above scheme that the advantages of the present invention are:

[0059] The method for predicting the remaining service life of a tool provided by the present invention makes a number of labels for the wear values ​​in the tool wear experimental data set; and randomly divides the number of labels and the pre-processed vibration signal data and cutting force signal data into a training set, a validation set and a test set, and extracts the multi-domain features of the training set, the validation set and the test set. The multi-domain features are input into a first training model for training to extract nonlinear features; and the multi-domain features are input into a second training model for training to extract linear features. By extracting nonlinear features and linear features respectively, the diversity of feature representation is enhanced, the risk of overfitting caused by a single feature extraction method is reduced, various feature information in the data is captured more comprehensively, and the diversity, robustness and information content of the features are improved. In addition, the nonlinear features are further fused with the linear features, and the fused features are input into a third training model to output the prediction value of the remaining service life of the tool. The multi-scale context information of the tool processing data is obtained through the third training model, which adapts to the complex changes of tool wear in the numerical control processing process, and solves the gradient problem caused by the rough feature perception and the deepening of the network. The method reduces the dependence on the learning process of artificially modifying related parameters and improves the prediction accuracy of the tool life. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 The overall process diagram of the method for predicting the remaining service life of a tool provided by an embodiment of the present invention is shown;

[0061] Figure 2 1 is a schematic diagram showing the structure of the first training model;

[0062] Figure 3 A detailed flow chart of step S5 is shown;

[0063] Figure 4 A schematic diagram of the stacked bidirectional gated recurrent unit network structure is shown;

[0064] Figure 5 The following is a schematic diagram of a tool remaining service life prediction device.

[0065] 300-Tool Remaining Life Prediction Device

[0066] 301-Data cleaning and annotation module

[0067] 302-Dataset Partitioning Module

[0068] 303-Feature Extraction Module

[0069] 304-First training module

[0070] 305-Second training module

[0071] 306-Feature Fusion Module

[0072] 307-Third Training Module

[0073] S1-S8, S51-S54: steps. DETAILED DESCRIPTION

[0074] In order to make the above features and effects of the present invention more clearly understood, embodiments are given below and described in detail with reference to the accompanying drawings.

[0075] See also Figure 1-Figure 5 As shown, Figure 1 The overall process diagram of the method for predicting the remaining service life of a tool provided by an embodiment of the present invention is shown; Figure 2 1 is a schematic diagram showing the structure of the first training model; Figure 3 A detailed schematic diagram of the process of step S5 is shown; Figure 4 A schematic diagram of the structure of the stacked bidirectional gated recurrent unit network (SBiGRU) is shown; Figure 5 The schematic diagram of the remaining tool life prediction device is shown.

[0076] See also Figure 1 As shown in, it shows the overall flow chart of the remaining service life prediction method of the tool provided in this embodiment. The remaining service life prediction method of the tool includes at least the following steps:

[0077] Step S1, selecting wear values ​​in a tool wear experimental data set to create a number of labels, wherein: selecting an average wear value of a cutting edge on a tool flank as a prediction label for a tool's remaining service life.

[0078] In one embodiment, the tool wear experimental data set includes experimental data of the entire life cycle, and the wear values ​​in the tool wear experimental data set are selected to make several labels. In addition, according to the national standard GB / T16460-2016 "End Mill Life Test", when the average wear width VB of the cutting edge of the tool back face is selected to be 0.3mm, it is considered that the tool wear is very serious and is in a failure state. Therefore, the average wear width is selected as the average wear value indicator, and the several labels are screened, and the average wear value of the cutting edge of the tool back face is selected as the prediction label of the remaining service life of the tool.

[0079] Step S2: collecting the vibration signal data and cutting force signal data of the tool and performing preprocessing.

[0080] In this embodiment, in addition to the above tool wear experimental data set, vibration signal data and cutting force signal data of the tool are further collected in real time to construct a data set together with the tool wear experimental data set.

[0081] In one embodiment, the original data is standardized and transformed by detecting whether there are missing values, abnormal values ​​or repeated values ​​in the vibration signal data and cutting force signal data of the tool, and performing corresponding preprocessing such as filling missing values, deleting abnormal values ​​or repeated values ​​to meet model assumptions or improve model performance.

[0082] Step S3: Then, the plurality of labels and the preprocessed vibration signal data and cutting force signal data are randomly divided into a training set, a validation set and a test set.

[0083] Step S4: extracting multi-domain features of the training set, validation set and test set.

[0084] In one embodiment, the extracted multi-domain features include time domain, frequency domain, and time-frequency domain analysis of the signal, totaling 31 features, including 11 time domain features, 12 frequency domain features, and 8 time-frequency domain features obtained based on wavelet packet decomposition.

[0085] Step S5: input the multi-domain features into the first training model for training, and output nonlinear features.

[0086] In view of the problems in the related technologies, when predicting the remaining service life of the tool, there are problems such as the inability to automatically capture the nonlinear features in the data, poor dimensionality reduction and denoising capabilities, high computational complexity, and the ability to extract time-series dependencies. In this regard, in this embodiment, starting from one aspect, an improved denoising autoencoder model is used as the first training model, which can automatically learn the feature representation of the data without manually designing features; it can better capture the nonlinear features in the data, realize the dimensionality reduction and denoising of the data, thereby extracting the most important information in the data and improving the accuracy and generalization ability of the feature representation.

[0087] Specifically, in this embodiment, see Figure 2 , Figure 3 As shown in Figure 2 : is a schematic diagram showing the structure of the first training model. Figure 3 A detailed flow chart of step S5 is shown.

[0088] The first training model adopts an improved denoising autoencoder model, which specifically includes a noise module, an encoder module, and a decoder module, wherein the input end of the encoder module is connected to the input end of the noise module and the output end of the decoder module, and the input end of the decoder module is connected to the output end of the encoder module.

[0089] For the noise module, the noise module is used to dynamically adjust the type and intensity of the added noise according to the training progress or the continuity of the input multi-domain features to generate a multi-domain feature containing noise. Specifically, in order to simulate a more realistic data noise environment in the use of the tool of the CNC machine tool, in one embodiment, in addition to Gaussian noise, other types of noise, such as salt and pepper noise, Poisson noise, etc., can also be considered to be introduced to simulate a more realistic data noise environment. Specifically, if the multi-domain feature is identified as a continuously changing feature, Gaussian noise is selected to be added; if the multi-domain feature is identified as a discretized feature, salt and pepper noise and / or Poisson noise are selected to be added. In addition, in practice, too much noise may cause the model to be unable to learn effective feature representations; while too little noise may not be able to improve the robustness of the model. Therefore, in this embodiment, the type and intensity of the noise are dynamically adjusted during the training process. In the early stage of the training of the first training model, a large intensity and multiple types of noise are injected to enhance the robustness of the model; in the later stage of the training of the first training model, the noise intensity is gradually reduced so that the first training model pays more attention to the original multi-domain features in the later stage of the training.

[0090] For the encoder module and the decoder module, in this embodiment, the encoder module and the decoder module adopt a fully connected network with symmetric number of neural units; wherein: the encoder module is responsible for processing the input noisy multi-domain feature sequence, compressing the noisy multi-domain features into a low-dimensional hidden layer representation through nonlinear transformation, and this internal representation captures the key information of the noisy multi-domain feature sequence, especially the time dependency information. The decoder module gradually generates the final reconstructed features based on the noisy features output by the encoder (the noisy features are the low-dimensional hidden layer representations into which the noisy multi-domain features are compressed by the encoder module) and the generated partial target sequence. During the generation process, the decoder module will consider the time dependency information in the input noisy feature sequence to ensure the coherence and accuracy of the output reconstructed features.

[0091] In addition, in order to better capture the temporal dependency of input data in the encoder module and the decoder module, in this embodiment, a recurrent layer is introduced in both the encoder module and the decoder module. In the recurrent layer, each neuron is connected to the output of the previous moment and the input of the current moment. This connection mode enables the network to save information from the previous moment and use this information in the calculation of the current moment, and can process the timing information in the input multi-domain feature sequence and encode it into the internal representation, thereby realizing the capture of the temporal dependency.

[0092] In addition, in one embodiment, convolutional layers are further introduced in the encoder module and the decoder module to capture local features of the input data. Specifically, in the encoder module, the input noisy multi-domain features are first processed through a series of convolutional layers, which gradually extract low-level and high-level features of the input data. The decoder module receives the noisy features generated after encoding as input, and converts them back to the original data space or the target space through a series of deconvolutional layers and convolutional layers, and generates the final reconstructed features.

[0093] In one embodiment, the number of layers of the encoder module and the decoder module is increased to extract a deeper feature representation of the input multi-domain features. In addition, a Dropout layer is further added after each layer of the encoder module and the decoder module to randomly discard the output of a portion of neurons to prevent overfitting.

[0094] In one embodiment, a residual connection or a skip connection is added between the encoder module and the decoder module, which can help the gradient propagate better and alleviate the gradient vanishing problem in the deep network.

[0095] Specific reference Figure 3 As shown in , in one embodiment, step S5 specifically includes the following steps:

[0096] First, the multi-domain features extracted from the training set and the validation set are input into the first training model. First, in step S51, the multi-domain features are passed through the noise module, and the type and intensity of the added noise are dynamically adjusted according to the training progress or the continuity of the input multi-domain features to generate noisy multi-domain features. Step S52, the encoder module compresses the noisy multi-domain features into a low-dimensional hidden layer representation through nonlinear transformation, and outputs the noise-containing features. Then, step S53, the noise-containing features are further input into the decoder module to output the reconstructed features, and the reconstructed features are fed back to the input end of the encoder module, and together with the multi-domain features to which the Gaussian noise is added, they are used as the input of the encoder module again. Step S54, the error value between the reconstructed features and the original multi-domain features is calculated to obtain the reconstruction error, and the network parameters of the first training model are adjusted according to the reconstruction error.

[0097] In one embodiment, in step S54, the error value between the reconstructed feature and the multi-domain feature is calculated using a mean square error loss function. In addition, in one embodiment, the parameters of the encoder module and the decoder module are adjusted using a back propagation algorithm in combination with an AdamW algorithm so that the reconstruction error is minimized. The back propagation algorithm is responsible for calculating the gradient, i.e., the partial derivative of the loss function to the model parameters, while the AdamW algorithm uses these gradients to update the model parameters to achieve optimization. This combination enables the neural network to be trained efficiently and improves the accuracy of the training results.

[0098] Finally, in step S55, when it is determined that the training of the first training model is completed, the multi-domain features extracted from the training set, the validation set, and the test set are input into the trained first training model, and the encoder module outputs the nonlinear features.

[0099] Specifically, in step S54, as to how to determine whether the first training model has been trained, in one embodiment, by comparing the loss of the current round of the validation set with the loss of the previous round, the model parameters are saved when the loss of the current round is less than the loss of the previous round, otherwise the model training continues. In addition, by judging whether the current number of network iterations meets the preset number value, if the preset number value is reached, the training of the first training model is completed, and the trained first training model is saved, otherwise, if it is not reached, it returns to step S51 to continue the iteration. Then, when it is determined that the training of the first training model is completed, the multi-domain features extracted from the training set, the validation set, and the test set are input into the trained first training model, and the encoder module outputs the nonlinear features.

[0100] Step S6: input the multi-domain features into the second training model for training, and output linear features.

[0101] In view of the problems in the related art in predicting the remaining service life of the tool, such as the inability to automatically capture the nonlinear features in the data, poor dimensionality reduction and denoising capabilities, high computational complexity, and the ability to extract time series dependencies. In this regard, starting from another aspect, in this embodiment, the second training model adopts a support vector regression model based on kernel principal component analysis, and maps the multi-domain features to a high-dimensional space by introducing a kernel function, performs linear analysis in the high-dimensional space, and captures the linear features in the data, which can more effectively extract important information from the data, reduce the dimension of the data, and reduce the dimension of the data while maintaining the important information of the data, thereby reducing the computational complexity.

[0102] Specifically, a parameter grid of kernel principal component analysis is defined, including the dimension parameters to be adjusted. Then, a parameter grid search is used for cross-validation, and the optimal dimension parameters are found by training the second training model on different dimensions and evaluating the performance. Finally, the second training model is retrained using the optimal dimension parameters, and the performance of the retrained model is evaluated.

[0103] In this embodiment, the first training model can extract nonlinear features in the data, and the second training module can extract linear features. By connecting these two methods, their advantages can be comprehensively utilized to enhance the diversity of feature representation, reduce the risk of overfitting caused by a single feature extraction method, more comprehensively capture various feature information in the data, and improve the diversity, robustness and information content of the features.

[0104] Step S7: fuse the nonlinear feature with the linear feature to obtain a fused feature as a sensitive feature representation related to tool degradation information in the sample.

[0105] In this embodiment, after the nonlinear features of the multi-domain features are extracted through step S5 and the linear features of the multi-domain features are extracted through step S6, the nonlinear features are further fused with the linear features to obtain fused features, and the fused features are used as sensitive feature representations related to tool degradation information in the sample, that is, the features after fusing different noise reduction methods.

[0106] Step S8: input the fused features into the third training model, and output the predicted value of the remaining service life of the tool.

[0107] In this embodiment, after the fusion feature is obtained in step S7, the fusion feature is input into the third training model for training, and the prediction value of the remaining service life of the tool is output.

[0108] Specifically, the third training model includes a stacked bidirectional gated recurrent unit network and a fully connected network, and the input end of the fully connected network is connected to the output end of the stacked bidirectional gated recurrent unit network.

[0109] See also Figure 4 , Figure 4 The schematic diagram of the structure of the stacked bidirectional gated recurrent unit network (SBiGRU) is shown. The stacked bidirectional gated recurrent unit network is composed of two stacked bidirectional gated recurrent units (BiGRU), wherein the output of the previous bidirectional gated recurrent unit will be used as the input of the next bidirectional gated recurrent unit. The BiGRU neural network makes predictions based on the entire time series. It divides the hidden layer into two relative parts, forward and backward, to read the information of the past and future time respectively. Specifically, each bidirectional gated recurrent unit (BiGRU) includes a forward GRU network and a backward GRU network. The forward GRU network receives the sequence information of the current time series calculated by the fused features, and the backward GRU network receives the reverse sequence information of the same time series calculated by the fused features. The bidirectional gated recurrent unit outputs the weighted value of the sequence information and the reverse sequence information. Figure 4 As shown in After being processed by the forward GRU network and the backward GRU network of the bidirectional gated recurrent unit, the output sequence Then, the sequence It is further processed as input by the forward GRU network and the backward GRU network of the next bidirectional gated recurrent unit, and the output sequence is

[0110] In view of the problems in the related technologies in predicting the remaining service life of tools, such as the inability to automatically capture the nonlinear features in the data, poor dimensionality reduction and denoising capabilities, high computational complexity, and the ability to extract time-series dependencies, etc. In this regard, in this embodiment, starting from another aspect, the stacked bidirectional gated recurrent unit network is used to obtain multi-scale context information of tool processing data, adapt to the complex changes of tool wear during CNC machining, and solve the gradient problem caused by rough feature perception and network deepening.

[0111] In addition, in one embodiment, before the fused feature is input into the third training model, the fused feature is first adjusted to be suitable for the input data of the third training model. In one embodiment, a sliding window method is used to adjust the dimension of the fused feature to be suitable for the third training model.

[0112] In addition, in one embodiment, when evaluating whether the life prediction model obtained by combining the first training model, the second training model, and the third training model is optimal, the mean absolute percentage error MAPE (Mean Absolute Percentage Error), the mean square error MAE (Mean Absolute Error), the root mean square error RMSE (Root Mean Square Error), the determination coefficient R 2 (R-Square) and other evaluation indicators. Among them:

[0113]

[0114]

[0115] Among them, y i represents the true value, Represents the predicted value, and N represents the number of sample sets.

[0116] In summary, the method for predicting the remaining service life of a tool provided by the present invention makes several labels for the wear values ​​in the tool wear experimental data set; and randomly divides the several labels and the pre-processed vibration signal data and cutting force signal data into a training set, a validation set and a test set, and extracts the multi-domain features of the training set, the validation set and the test set. The multi-domain features are input into the first training model for training to extract nonlinear features; the multi-domain features are input into the second training model for training to extract linear features. By extracting nonlinear features and linear features separately, the diversity of feature representation is enhanced, the risk of overfitting caused by a single feature extraction method is reduced, various feature information in the data is captured more comprehensively, and the diversity, robustness and information content of the features are improved. In addition, the nonlinear features are further fused with the linear features, and the fused features are input into the third training model to output the prediction value of the remaining service life of the tool. The multi-scale context information of the tool processing data is obtained through the third training model, which adapts to the complex changes of tool wear in the CNC processing process, and solves the gradient problem caused by the rough feature perception and the deepening of the network. The method reduces the dependence on the learning process of artificially modifying related parameters and improves the prediction accuracy of tool life.

[0117] In addition, refer to Figure 5 , Figure 5 A tool remaining service life prediction device 300 is shown, which can achieve the following Figure 1 , Figure 3 The various processes of the tool remaining service life prediction method shown in FIG. A tool remaining service life prediction device 300, the device at least comprises:

[0118] The data cleaning and labeling module 301 is used to create a number of labels for the wear values ​​in the tool wear experimental data set, including: selecting the average wear value of the cutting edge of the tool back face as the prediction label of the remaining service life of the tool; and collecting the vibration signal data and cutting force signal data of the tool and pre-processing them.

[0119] The data set division module 302 randomly divides the plurality of labels and the pre-processed vibration signal data and cutting force signal data into a training set, a validation set and a test set.

[0120] The feature extraction module 303 is used to extract multi-domain features of the training set, the validation set and the test set.

[0121] The first training module 304 is used to input the multi-domain features into a first training model for training and output non-linear features.

[0122] The second training module 305 is used to input the multi-domain features into a second training model for training and output linear features.

[0123] The feature fusion module 306 is used to fuse the nonlinear feature with the linear feature to obtain a fused feature as a sensitive feature representation related to tool degradation information in the sample.

[0124] The third training module 307 is used to input the fusion feature into the third training model and output the prediction value of the remaining service life of the tool.

[0125] In addition, those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the module described can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here.

[0126] In addition, an embodiment of the present application also provides an electronic device, including a processor, a memory, and a program or instruction stored in the memory and executable on the processor. When the program or instruction is executed by the processor, the steps of the above-mentioned tool remaining service life prediction method are implemented and the same technical effect can be achieved.

[0127] The processor is the processor in the electronic device in the above embodiment. The readable storage medium includes a computer readable storage medium, such as a computer read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0128] In addition, an embodiment of the present application also provides a readable storage medium, on which a program or instruction is stored. When the program or instruction is executed by a processor, the steps of the above-mentioned tool remaining service life prediction method are implemented and the same technical effect can be achieved.

[0129] It should be noted that, in this article, the terms "comprise", "include" or any other variant thereof are intended to cover non-exclusive inclusion, so that the process, method, article or device including a series of elements includes not only those elements, but also includes other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "including one..." do not exclude the presence of other identical elements in the process, method, article or device including the element. In addition, it should be pointed out that the scope of the methods and devices in the embodiments of the present application is not limited to performing functions in the order shown or discussed, and may also include performing functions in a substantially simultaneous manner or in the opposite order according to the functions involved, for example, the described method may be performed in an order different from that described, and various steps may also be applied, omitted, or combined. In addition, the features described with reference to certain examples may be combined in other examples.

[0130] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method.

[0131] The embodiments of the present application are described above in conjunction with the accompanying drawings, but the present application is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the guidance of the present application, ordinary technicians in this field can also make many forms without departing from the purpose of the present application and the scope of protection of the claims, all of which are within the protection of the present application.

Claims

1. A method for predicting the remaining service life of a tool, characterized in that: Include: Several labels are made for the wear values ​​in the tool wear experimental data set, among which: the average wear value of the cutting edge of the tool flank is selected as the prediction label of the tool's remaining service life; Collect the vibration signal data and cutting force signal data of the tool and perform pre-processing; The labels and the preprocessed vibration signal data and cutting force signal data are randomly divided into a training set, a validation set and a test set; Extracting multi-domain features of the training set, validation set, and test set; Inputting the multi-domain features into a first training model for training, and outputting nonlinear features; The multi-domain features are input into a second training model for training, and linear features are output; The nonlinear feature is fused with the linear feature to obtain a fused feature as a sensitive feature representation related to tool degradation information in the sample; The fused feature is input into a third training model to output a prediction value of the remaining service life of the tool; wherein the third training model comprises a stacked bidirectional gated recurrent unit network and a fully connected network, and an input end of the fully connected network is connected to an output end of the stacked bidirectional gated recurrent unit network; The stacked bidirectional gated recurrent unit network is composed of two stacked bidirectional gated recurrent units. Each bidirectional gated recurrent unit includes a forward GRU network and a backward GRU network. The forward GRU network receives the sequence information of the current time series calculated by the fused features, and the backward GRU network receives the reverse sequence information of the same time series calculated by the fused features. The bidirectional gated recurrent unit outputs the weighted value of the sequence information and the reverse sequence information to obtain the predicted value of the remaining service life of the tool.

2. The method according to claim 1, characterized in that The first training model adopts an improved denoising autoencoder, comprising a noise module, an encoder module, and a decoder module; the input end of the encoder module is connected to the input end of the noise module and the output end of the decoder module, and the input end of the decoder module is connected to the output end of the encoder module; wherein: The noise module is used to dynamically adjust the type and intensity of added noise to generate noisy multi-domain features according to the training progress or the continuity of the input multi-domain features; The encoder module and the decoder module adopt a fully connected network with a symmetric number of neural units; wherein: Convolutional layers and recurrent layers are introduced into the encoder module and the decoder module to capture local features and temporal dependencies of data respectively; Increasing the number of layers of the encoder module and the decoder module, and adding a Dropout layer after each layer of the encoder module and the decoder module to randomly discard the output of a portion of neurons; A residual connection or a skip connection is added between the encoder module and the decoder module.

3. The method according to claim 2, characterized in that Dynamically adjust the type and intensity of added noise according to the training progress or the continuity of the input multi-domain features to generate noisy multi-domain features, including: In the early stage of training the first training model, injecting noise of relatively large intensity and multiple types; In the later stage of training of the first training model, gradually reducing the noise intensity; If the multi-domain feature is identified as a continuously changing feature, adding Gaussian noise is selected; If the multi-domain feature is identified as a discretized feature, salt and pepper noise and / or Poisson noise is added.

4. The method according to claim 2 or 3, characterized in that: The step of inputting the multi-domain features into the first training model for training comprises: The multi-domain features extracted from the training set and the validation set are input into the first training model, wherein: After the multi-domain feature passes through the noise module, the type and intensity of the added noise are dynamically adjusted according to the training progress or the continuity of the input multi-domain feature to generate a noisy multi-domain feature; The encoder module compresses the noisy multi-domain features into low-dimensional hidden layer representations through nonlinear transformation and outputs the noisy features; The noise-containing feature is further input into the decoder module to output a reconstructed feature, and the reconstructed feature is fed back to the input end of the encoder module; Calculating an error value between the reconstructed feature and the multi-domain feature to obtain a reconstruction error, and adjusting a network parameter of the first training model according to the reconstruction error; When it is determined that the training of the first training model is completed, The multi-domain features extracted from the training set, the validation set, and the test set are input into the trained first training model, and the encoder module outputs the nonlinear features.

5. The method according to claim 4, characterized in that Using binary cross entropy or the like as a loss function, the back propagation algorithm is combined with the AdamW algorithm to adjust the parameters of the encoder module and the decoder module so as to minimize the reconstruction error.

6. The method according to claim 4, characterized in that Determining whether the first training model is trained, including: Compare the loss of this round of the validation set with the loss of the previous round, and save the model parameters if the loss of this round is less than the loss of the previous round; It is determined whether the current number of network iterations meets the preset number value. If the preset number value is reached, the training of the first training model is completed, and the trained first training model is saved.

7. The method according to claim 1, characterized in that The second training model adopts a support vector regression model based on kernel principal component analysis, maps the multi-domain features to a high-dimensional space by introducing a kernel function, performs linear analysis in the high-dimensional space, and captures the linear features, including: Define the parameter grid for kernel principal component analysis; Using parameter grid search for cross validation, find the best dimension parameters by training the second training model on different dimensions and evaluating the performance; Using the optimal dimension parameters, the second training model is retrained.

8. The method according to claim 1, characterized in that Before inputting the fused feature into the third training model, the method further comprises: The fused features are adjusted to be suitable for input data of the third training model.

9. The method according to claim 8, characterized in that Using a sliding window method, the dimension of the fused feature is adjusted to be suitable for the third training model.

10. A tool remaining service life prediction device, characterized in that: The method for predicting the remaining service life of a tool according to any one of claims 1 to 9 is adopted, and the device comprises: A data cleaning and labeling module is used to create a number of labels for the wear values ​​in the tool wear experiment data set, wherein: the average wear value of the cutting edge of the tool back face is selected as the prediction label of the tool remaining service life; and Collect the vibration signal data and cutting force signal data of the tool and perform pre-processing; A data set partitioning module randomly divides the labels and the preprocessed vibration signal data and cutting force signal data into a training set, a validation set and a test set; A feature extraction module, used to extract multi-domain features of the training set, validation set and test set; A first training module, used for inputting the multi-domain features into a first training model for training, and outputting nonlinear features; A second training module is used to input the multi-domain features into a second training model for training and output linear features; A feature fusion module is used to fuse the nonlinear feature with the linear feature to obtain a fused feature as a sensitive feature representation related to tool degradation information in the sample; A third training module is used to input the fusion feature into a third training model to output a prediction value of the remaining service life of the tool; Wherein: the third training model comprises a stacked bidirectional gated recurrent unit network and a fully connected network, and the input end of the fully connected network is connected to the output end of the stacked bidirectional gated recurrent unit network; The stacked bidirectional gated recurrent unit network is composed of two stacked bidirectional gated recurrent units. Each bidirectional gated recurrent unit includes a forward GRU network and a backward GRU network. The forward GRU network receives the sequence information of the current time series calculated by the fused features, and the backward GRU network receives the reverse sequence information of the same time series calculated by the fused features. The bidirectional gated recurrent unit outputs the weighted value of the sequence information and the reverse sequence information to obtain the predicted value of the remaining service life of the tool.

Images