Cutter wear prediction method and device
By extracting and fusing the multi-domain features of CNC machine tools, and using multi-scale one-dimensional convolutional neural network and denoising automatic encoder model to integrate global and local features, the problem of low tool wear prediction accuracy in the existing technology is solved, and more efficient prediction and more scientific tool change decisions are achieved.
Patent Information
- Application Number
- CN202411727357.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-28
- Publication Date
- 2025-03-18
AI Technical Summary
The prior art is difficult to effectively predict the wear of CNC machine tools, resulting in reduced workpiece quality and sudden machine shutdown. The feature engineering based on traditional machine learning is complex, and the noise information cannot be processed, and the prediction performance is low.
A tool wear prediction method is adopted to extract multi-domain features and fusion by obtaining data such as vibration signals, cutting force signals and sound emission signals. A multi-domain feature is extracted and fused. A multi-scale one-dimensional convolutional neural network and a denoising automatic encoder model is used to integrate global and local features to improve prediction accuracy.
It achieves higher tool wear prediction accuracy, reduces the risk of overfitting, can capture feature information in the data more comprehensively, provide scientific tool change decision support, improves productivity and reduces losses.
Smart Images
Figure CN119646441A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to high-end equipment, and specifically relates to a tool wear prediction method and device. Background Art
[0002] As the number of machining times or time increases, the CNC machine tool tools will gradually wear out, resulting in a decrease in the surface quality of the workpiece and thus failing to meet the technical requirements of the workpiece. When to change the tool becomes a key issue. Changing the tool too early will waste the remaining service life of the tool, changing the tool too late will cause the workpiece quality to deteriorate, and the sudden failure of the tool will also cause the machine tool to suddenly stop, disrupting the production schedule. Predicting and monitoring tool wear information can provide scientific information for tool change decisions, thereby improving production efficiency and reducing losses. The data-driven method for tool wear prediction requires a large amount of historical data collected during the machining process, such as spindle vibration signals, machining cutting force signals, and acoustic emission signals, to train the tool wear monitoring model, and use the trained model to perform online tool wear monitoring. The indirect signal dimension is large, and the method based on multi-signal feature extraction and fusion can significantly improve the prediction performance. The machining data contains complex noise and has obvious non-stationary characteristics, which brings challenges to online monitoring of tool wear.
[0003] For multi-signal fusion, tool wear prediction based on feature engineering of traditional machine learning is a commonly used method, but manual feature engineering is computationally complex, requires strong domain knowledge, cannot process noise information, and mining of shallow features makes the model prediction performance low. However, the features extracted at the overall level of the sample can represent the global information of the sample. Summary of the invention
[0004] In view of the deficiencies of the prior art, the present invention discloses a tool wear prediction method and device, which can globally and locally integrate sensitive features related to tool degradation information to improve the prediction accuracy of tool wear.
[0005] In order to achieve the above object, the present invention provides a tool wear prediction method, comprising:
[0006] Acquire the vibration signal data, cutting force signal data, sound emission signal data, and tool wear value label data of the tool, construct a data set, and divide the data set into a training set, a validation set, and a test set;
[0007] Extracting multi-domain features of the training set, validation set, and test set respectively to obtain a first feature, a second feature, and a third feature;
[0008] Inputting the first feature, the second feature, and the third feature into a first training model for training, and outputting an intermediate feature;
[0009] The training set, the validation set, and the test set are input into a second training model for training, and multi-scale features are output;
[0010] The intermediate feature is fused with the multi-scale feature to form a training set and a test set of fused features;
[0011] The training set of fused features is input into the third training model for training, and the test set of fused features is input into the trained third training model to output the target predicted wear value.
[0012] In one embodiment, the second training model adopts a multi-scale one-dimensional convolutional neural network, including a sampling module, a multi-scale network, and a fully connected network; wherein: the sampling module is used to sample the data set to obtain original sample data with the same dimension as the data set; the multi-scale network includes: a wide-core convolution layer, connected to the output end of the sampling module, to reduce the dimension of the original sample data; a convolution kernel layer of three groups of parallel paths, connected to the output end of the wide-core convolution layer, each parallel path outputs the corresponding features of the path after three convolutions, pooling, batch normalization layers and Gaussian error linear unit activation functions; a fully connected network, connected to the output ends of the convolution kernel layers of the three groups of parallel paths, splicing and combining the features outputted by the three groups of parallel paths and outputting them to the fully connected network to obtain a first predicted wear value.
[0013] In one embodiment, the training set, the validation set, and the test set are input into a second training model for training, comprising:
[0014] The data set is input into the sampling module for sampling to obtain original sample data with the same dimension as the data set;
[0015] The original sample data is input into the multi-scale network and the fully connected network to obtain a first predicted wear value;
[0016] Calculating a first error value between the first predicted wear value and the actual wear value label, and using the first error value to adjust the network parameters of the second training model;
[0017] When it is determined that the training of the second training model is completed, the training set and the validation set are combined and input into the trained second training model to obtain a first local multi-scale feature representation;
[0018] Inputting the test set into the trained second training model to obtain a second local multi-scale feature representation;
[0019] The multi-scale feature includes a first local multi-scale feature representation and a second local multi-scale feature representation.
[0020] In one embodiment, using a second checkpoint of a second training model, comparing the current round loss of the validation set with the previous round loss, and saving the network parameters when the current round loss is less than the previous round loss;
[0021] It is determined whether the current number of network iterations meets a second preset number value. If the second preset number value is reached, the training of the second training model is completed, and the trained second training model is saved.
[0022] In one embodiment, the first training model includes a noise module, an encoder module, and a decoder module; wherein the noise module is used to generate Gaussian noise, 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, the input end of the decoder module is connected to the output end of the encoder module; and the encoder module and the decoder module adopt a fully connected network with a symmetric number of neural units.
[0023] In one embodiment, inputting the first feature, the second feature, and the third feature into the first training model for training includes:
[0024] The first feature is input into the first training model, and Gaussian noise is added after passing through the noise module, and the decoder module obtains the reconstructed feature on the verification set;
[0025] Calculating a second error value between the reconstructed feature and the first feature, and adjusting the network parameters of the first training model by back propagating the second error value;
[0026] When it is determined that the training of the first training model is completed, the first feature and the second feature are combined and input into the trained first training model, and the encoder module outputs a first intermediate dimensionality reduction representation;
[0027] The third feature is input into the trained first training model, and the encoder module outputs a second intermediate dimension reduction representation;
[0028] The intermediate feature includes the first intermediate reduced dimensionality representation and the second intermediate reduced dimensionality representation.
[0029] In one embodiment, determining whether the first training model has been trained includes:
[0030] Using the first checkpoint of the first training model, comparing the current round loss of the validation set with the previous round loss, and saving the network parameters of the first training model when the current round loss is less than the previous round loss;
[0031] It is determined whether the current number of network iterations meets a first preset number value. If the number of iterations reaches the first preset number value, the training of the first training model is completed, and the trained first training model is saved.
[0032] In one embodiment, the training set of fused features is input into a third training model for training to obtain a second predicted wear value;
[0033] Calculating a third error value between the second predicted wear value and the actual wear value label, and using the third error value to adjust the network parameters of the third training model;
[0034] Using the third checkpoint of the third training model, comparing the current round loss of the validation set with the previous round loss, and saving the network parameters of the third training model when the current round loss is less than the previous round loss;
[0035] It is determined whether the current number of network iterations meets a third preset number value. If the third preset number value is reached, the training of the third training model is completed, and the trained third training model is saved.
[0036] In one embodiment, a plurality of labels are produced for the wear values in the tool wear value label data, wherein: the average wear value of the cutting edge of the tool flank is selected as the prediction label of the tool wear.
[0037] Another aspect of the present invention further provides a tool wear prediction device, which adopts the above-mentioned tool wear prediction method and comprises:
[0038] A data set construction module is used to obtain the vibration signal data, cutting force signal data, sound emission signal data, and tool wear value label data of the tool, construct a data set, and divide the data set into a training set, a validation set, and a test set;
[0039] A feature extraction module is used to extract multi-domain features of the training set, the validation set and the test set to obtain a first feature, a second feature and a third feature;
[0040] A first training module is used to input the first feature, the second feature, and the third feature into a first training model for training, and output an intermediate feature;
[0041] A second training module is used to input the training set, the validation set and the test set into a second training model for training, and output multi-scale features;
[0042] A feature fusion module, used to fuse the intermediate feature with the multi-scale feature to form a training set and a test set of fused features;
[0043] The third training module is used to input the training set of fused features into the third training model for training, and input the test set of fused features into the trained third training model to output the target predicted wear value.
[0044] On the other hand, the present invention further 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 wear prediction method are implemented and the same technical effect can be achieved.
[0045] It can be seen from the above scheme that the advantages of the present invention are:
[0046] The tool wear prediction method provided by the present invention inputs the multi-domain features extracted from the training set, the validation set and the test set into the first training model for training, and outputs the intermediate features; inputs the training set, the validation set and the test set into the second training model for training, and outputs the multi-scale features; and fuses the intermediate features with the multi-scale features, and finally outputs the target predicted wear value by the third training model. The method can realize noise suppression of the input features and compress the overall information of the sample through the first training model, and can automatically extract multi-scale features through the second training model. The combination of the first training model and the second training model can integrate sensitive features related to tool degradation information from the global and local perspectives, 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, improve the multi-signal feature extraction capability, improve the prediction accuracy of tool wear, and realize error-free tool wear value prediction. BRIEF DESCRIPTION OF THE DRAWINGS
[0047] Figure 1 The overall process diagram of the tool wear prediction method provided by one embodiment of the present invention is shown;
[0048] Figure 2 1 is a schematic diagram showing the structure of the first training model;
[0049] Figure 3 A detailed schematic diagram of the process of step S3 is shown;
[0050] Figure 4 A schematic diagram of the structure of the second training model is shown;
[0051] Figure 5 A detailed schematic diagram of the process of step S4 is shown;
[0052] Figure 6 A detailed schematic diagram of the process of step S6 is shown;
[0053] Figure 7 The schematic diagram of the tool wear prediction device is shown.
[0054] in:
[0055] 300-Tool wear prediction device
[0056] 301-Dataset building module
[0057] 302-Feature Extraction Module
[0058] 303-First training module
[0059] 304-Second training module
[0060] 305-Feature Fusion Module
[0061] 306-Third Training Module
[0062] S1-S6, S31-S34, S41-S45, S61-S63: steps. DETAILED DESCRIPTION
[0063] 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.
[0064] An embodiment of the present invention discloses a tool wear prediction method. Figure 1-Figure 6 As shown. Figure 1 The overall process diagram of the tool wear prediction method is shown; Figure 2 1 is a schematic diagram showing the structure of the first training model; Figure 3 A detailed flow chart of step S3 is shown; Figure 4 A schematic diagram of the structure of the second training model is shown; Figure 5 A detailed schematic diagram of the process of step S4 is shown; Figure 6 A detailed flow chart of step S6 is shown.
[0065] Specifically, see Figure 1 As shown in , the tool wear prediction method includes at least the following steps:
[0066] Step S1, obtaining the vibration signal data, cutting force signal data, sound emission signal data, and tool wear value label data of the tool, constructing a data set, and dividing the data set into a training set, a validation set, and a test set.
[0067] In one embodiment, the average wear value of the cutting edge of the tool flank is selected as a prediction label for tool wear. The tool wear value label data includes experimental data of the entire life cycle. According to the national standard GB / T 16460-2016 "End Mill Life Test", when the average wear width VB of the cutting edge of the tool flank is selected to be 0.3 mm, 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 labels are screened, and the average wear value of the cutting edge of the tool flank is selected as the prediction label for tool wear.
[0068] In one embodiment, in addition to the above-mentioned tool wear value label data, vibration signal data, cutting force signal data, and sound emission signal data of the tool are further collected in real time to construct a data set together with the tool wear value label data.
[0069] In one embodiment, by detecting whether there are missing values, abnormal values or repeated values in the vibration signal data, cutting force signal data, and sound emission signal data of the tool, and performing corresponding preprocessing such as filling missing values, deleting abnormal values or repeated values, the invalid values of feed and retract at both ends of the sample signal are removed, and the outliers in the multi-channel signal data are removed, and the original data is standardized and transformed to meet the model assumptions or improve the model performance.
[0070] In one embodiment, the constructed data set is further divided into a training set, a validation set, and a test set. Specifically, in this embodiment, the training set selects a combination of data from at least one tool processing process in the data set, and the test set selects data from any new complete tool processing process in the data set, rather than randomly dividing the data of all tools in the data set into the training set, the validation set, and the test set.
[0071] Step S2, respectively extracting multi-domain features of the training set, validation set, and test set to obtain a first feature, a second feature, and a third feature, wherein the first feature is the multi-domain feature extracted from the training set, the second feature is the multi-domain feature extracted from the validation set, and the third feature is the multi-domain feature extracted from the test set.
[0072] In one embodiment, the extracted multi-domain features are features extracted from the time domain, frequency domain and time-frequency domain of the entire signal sample, and are used to characterize the overall information of the sample.
[0073] Step S3: input the first feature, the second feature, and the third feature into the first training model for training, and output the intermediate feature.
[0074] 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 S3 is shown.
[0075] See also Figure 2As shown in , the first training model is configured as a denoising autoencoder model (Denoising AutoEncoder, DAE), which is used to train the denoising autoencoder model to obtain the intermediate dimension reduction representation of a given sample. The first training model adds Gaussian noise to the input features to enhance the model denoising ability and the robustness of the training model. At the same time, a fully connected network with a symmetric number of neural units is used as the encoder and decoder of the denoising autoencoder network. The intermediate layer output of the encoder is a more compact feature representation of the original data, thereby realizing the learning of important features and achieving the effect of feature dimension reduction. Specifically, the first training model includes a noise module, an encoder module, and a decoder module, wherein the noise module is used to generate Gaussian noise, 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; and the encoder module and the decoder module use a fully connected network with a symmetric number of neural units.
[0076] In this embodiment, through the first training model, the feature representation of the data can be automatically learned without the need for manual feature design; the nonlinear features in the data can be better captured, and the dimensionality reduction and denoising of the data can be achieved, thereby extracting the most important information in the data and improving the accuracy and generalization ability of the feature representation.
[0077] Specific reference Figure 3 As shown in , in one embodiment, step S3 specifically includes the following steps:
[0078] First, the first feature is input into the first training model and Gaussian noise is added after passing through the noise module. At the same time, the reconstructed feature is obtained from the output of the decoder module by the verification set (S31), that is, the first feature with Gaussian noise added is used as input, and the second feature without noise is used as output to perform self-supervised pre-training on the first training model. At the same time, the reconstructed feature is fed back to the input end of the encoder module and used as the input of the encoder module together with the first feature with Gaussian noise added, so as to adjust the error between the first feature and the reconstructed feature.
[0079] The DAE model is self-supervised pre-trained by taking the sample feature data with Gaussian noise added as input and the sample feature data without noise added as output, that is, calculating the second error value between the reconstructed feature and the first feature, and adjusting the network parameters of the first training model by back propagation through the second error value (S32). In one embodiment, the second error value between the reconstructed feature and the first feature is calculated using a mean square error loss function. Then, when it is determined that the training of the first training model is completed, the first feature and the second feature are combined and input into the trained first training model, and the encoder module outputs the first intermediate dimensionality reduction representation (S33); at the same time, the third feature is input into the trained first training model, and the encoder module outputs the second intermediate dimensionality reduction representation (S34), that is, the intermediate feature can be obtained by the first intermediate dimensionality reduction representation and the second intermediate dimensionality reduction representation, and the intermediate representation includes the first intermediate dimensionality reduction representation and the second intermediate dimensionality reduction representation.
[0080] In addition, in step S34, as to how to determine whether the first training model has been trained, in one embodiment, by comparing the current round loss of the validation set with the previous round loss, if the current round loss is less than the previous round loss, the first training network parameters are saved, otherwise the first training model is continued to be trained. In addition, by judging whether the current network iteration number meets the first preset number value, if the first preset number value is reached, the first training model training is completed, and the trained first training model is saved, otherwise, if it is not reached, it returns to step S31 to continue iteration.
[0081] Step S4: input the training set, validation set and test set into a second training model for training, and output multi-scale features.
[0082] In this embodiment, the second training model uses a multi-scale one-dimensional convolutional neural network, which has a strong ability to extract features of local multi-scale information to obtain multi-scale local feature information of a given sample. Figure 4 , Figure 4The schematic diagram of the structure of the second training model is shown. The second training model includes a sampling module, a multi-scale network, and a fully connected network. Among them, the sampling module is used to sample the data set to obtain original sample data with the same dimension as the data set. The multi-scale network specifically includes: a wide core convolution layer, three groups of parallel path convolution kernel layers, wherein: the wide core convolution layer is connected to the output end of the sampling module to reduce the dimension of the original sample data; the convolution kernel layers of the three groups of parallel paths are connected to the output end of the wide core convolution layer, and each parallel path outputs the features corresponding to the path after three convolutions, pooling, batch normalization layers and Gaussian error linear unit activation functions. The features corresponding to the outputs of the three groups of parallel paths are spliced and combined and output to the fully connected network. The fully connected network is connected to the output ends of the convolution kernel layers of the three groups of parallel paths, and the features corresponding to the outputs of the three groups of parallel paths are spliced and combined and output to the fully connected network to obtain the first predicted wear value.
[0083] Specific reference Figure 5 As shown in , in one embodiment, step S4 specifically includes the following steps:
[0084] First, the data set is input into the sampling module for sampling to obtain the original sample data with the same dimension as the data set (S41). In one embodiment, the sampling method uses a sliding window to average, and each sample obtains the same dimension. For example, each sample forms 1000 rows of data of 7 signal channels after sampling, and each row is calculated by averaging the data points in the sliding window. Further, the original sample data is input into the multi-scale network and the fully connected network to obtain the first predicted wear value (S42). Then, the first error value between the first predicted wear value and the true wear value label is calculated, and the first error value is used to adjust the network parameters of the second training model (S43). In one embodiment, the first error value between the first predicted wear value and the true wear value label is calculated using the mean square error loss function. Finally, when it is determined that the training of the second training model is completed, the training set and the validation set are combined and input into the trained second training model to obtain the first local multi-scale feature representation of the overall training sample data (S44); at the same time, the test set is input into the trained second training model to obtain the second local multi-scale feature representation of the test set sample (S45), that is, the multi-scale feature can be obtained by the first local multi-scale feature representation and the second local multi-scale feature representation, and the multi-scale feature includes the first local multi-scale feature representation and the second local multi-scale feature representation.
[0085] In addition, in step S44, as to how to determine whether the training of the second training model is completed, in one embodiment, the second checkpoint of the second training model is used to compare the loss of the current round of the validation set with the loss of the previous round. If the loss of the current round is less than the loss of the previous round, the network parameters of the second training model are saved, otherwise the second training model is continued to be trained. In addition, by judging whether the current number of network iterations meets the second preset number value, if the second preset number value is reached, the training of the second training model is completed, and the trained second training model is saved. Otherwise, if it is not reached, it returns to step S41 to continue iteration.
[0086] In this embodiment, the first training model can realize noise suppression of input features and compress the overall information of the sample, and the second training model can automatically extract multi-scale features in the original signal, representing the local information related to tool degradation in the sample. By connecting these two methods, their advantages can be comprehensively utilized, and sensitive features related to tool degradation information can be integrated globally and locally, the diversity of feature representation can be enhanced, the overfitting risk caused by a single feature extraction method can be reduced, various feature information in the data can be captured more comprehensively, the multi-signal feature extraction capability can be improved, and error-free tool wear value prediction can be achieved.
[0087] Step S5: Fuse the intermediate feature with the multi-scale feature to form a training set and a test set of fused features.
[0088] In this embodiment, after the intermediate features are extracted in step S4 and the multi-scale features are extracted in step S4, the intermediate features are further fused with the multi-scale features to form a training set and a test set of fused features.
[0089] Step S6: input the training set of fused features into the third training model for training, and input the test set of fused features into the trained third training model to output the target predicted wear value.
[0090] In one embodiment, the third training model is a fully connected network, and a regularization layer (Dropout Layer) is added between each layer to prevent overfitting, and L1 regularization is added to reduce the complexity of the model, and the features are further reduced in dimension to screen effective features.
[0091] Specifically, see Figure 6 As shown in Figure 6A detailed flow chart of step S6 is shown. First, the training set of fused features is input into the third training model for training to obtain a second predicted wear value (S61). Then, the third error value between the second predicted wear value and the true wear value label is calculated, and the network parameters of the third training model are adjusted using the third error value (S62). In one embodiment, the third error value between the second predicted wear value and the true wear value label is calculated using a mean square error loss function. After determining that the training of the second training model is completed, the test set of fused features is input into the trained third training model to obtain a target predicted wear value, which is the final predicted wear value (S63).
[0092] In addition, in step S63, as to how to determine whether the third training model has been trained, in one embodiment, the third checkpoint of the third training model is used to compare the current round loss of the validation set with the previous round loss, and if the current round loss is less than the previous round loss, the network parameters of the third training model are saved, otherwise the third training model is continued to be trained. In addition, by judging whether the current number of network iterations meets the third preset number value, if the third preset number value is reached, the training of the third training model is completed, and the trained third training model is saved, otherwise, if it is not reached, it returns to step S61 to continue iteration.
[0093] In this embodiment, the tool wear value prediction model includes three parts: a first training model, a second training model, and a third training model. Among them, the first training model uses the root mean square between the reconstructed feature and the first feature as the training loss for back-propagation training; the second training model uses the root mean square of the first predicted wear value output by the fully connected network and the real wear value label as the training loss for back-propagation training. At the same time, the intermediate features after the dimension reduction of the input features output by the encoder of the first training model are spliced and fused with the multi-scale features output by the second training model, and trained by the third training model, wherein the second predicted wear value output by the third training model and the real wear value label are calculated as the training loss to back-propagate the third training model. The first training model, the second training model, and the third training model are tuned by the validation set to obtain the best tool wear value prediction model, and the tool wear value is predicted using the test set, and the mean square error MAE (Mean Absolute Error), root mean square error RMSE (Root Mean Square Error), and determination coefficient R 2 (R-Square) and other evaluation indicators to evaluate the model performance. Among them:
[0094]
[0095] Where m is the number of test samples, y i , They are the true wear value label, the predicted wear value, and the average of all true wear value labels.
[0096] In summary, the tool wear prediction method provided by the present invention uses the vibration signal data, cutting force signal data, sound emission signal data, and tool wear value label data of the tool to construct a data set, and divide the data set into a training set, a validation set, and a test set; and the multi-domain features extracted from the training set, the validation set, and the test set are input into the first training model for training, and the intermediate features are output; the training set, the validation set, and the test set are input into the second training model for training, and the multi-scale features are output; and the intermediate features are fused with the multi-scale features, and finally the target predicted wear value is output by the third training model. The method can realize noise suppression of the input features and compress the overall information of the sample through the first training model, and can automatically extract multi-scale features through the second training model. The first training model is combined with the second training model, and can integrate sensitive features related to tool degradation information from the global and local aspects, 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, improve the multi-signal feature extraction capability, improve the prediction accuracy of tool wear, and realize error-free tool wear value prediction.
[0097] In addition, refer to Figure 7 , Figure 7 A tool wear prediction device 300 is shown, which can be implemented as follows Figure 1 , Figure 3 , Figure 5 , Figure 6 The various processes of the tool wear prediction method shown in FIG. A tool wear prediction device 300, the device at least comprising:
[0098] The data set construction module 301 is used to obtain the vibration signal data, cutting force signal data, sound emission signal data, and tool wear value label data of the tool, construct a data set, and divide the data set into a training set, a validation set, and a test set.
[0099] The feature extraction module 302 is used to extract multi-domain features of the training set, the validation set and the test set respectively to obtain a first feature, a second feature and a third feature.
[0100] The first training module 303 is used to input the first feature, the second feature, and the third feature into the first training model for training, and output an intermediate feature.
[0101] The second training module 304 is used to input the training set, the validation set and the test set into a second training model for training and output multi-scale features.
[0102] The feature fusion module 305 is used to fuse the intermediate feature with the multi-scale feature to form a training set and a test set of fused features.
[0103] The third training module 306 is used to input the training set of fused features into the third training model for training, and input the test set of fused features into the trained third training model to output the target predicted wear value.
[0104] 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.
[0105] 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 wear prediction method are implemented and the same technical effect can be achieved.
[0106] 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.
[0107] 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 wear prediction method are implemented and the same technical effect can be achieved.
[0108] 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.
[0109] 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.
[0110] 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 tool wear prediction method, characterized in that: Include: Acquire the vibration signal data, cutting force signal data, sound emission signal data, and tool wear value label data of the tool, construct a data set, and divide the data set into a training set, a validation set, and a test set; Extracting multi-domain features of the training set, validation set, and test set respectively to obtain a first feature, a second feature, and a third feature; Inputting the first feature, the second feature, and the third feature into a first training model for training, and outputting an intermediate feature; The training set, validation set and test set are input into the second training model for training, and multi-scale features are output; wherein, the second training model adopts a multi-scale one-dimensional convolutional neural network, which includes a sampling module, a multi-scale network and a fully connected network; the sampling module is used to sample the data set to obtain original sample data with the same dimension as the data set; the multi-scale network includes: a wide-core convolution layer, connected to the output end of the sampling module, to reduce the dimension of the original sample data; a convolution kernel layer of three groups of parallel paths, connected to the output end of the wide-core convolution layer, each parallel path outputs the features corresponding to the path after three convolutions, pooling, batch normalization layers and Gaussian error linear unit activation functions; a fully connected network, connected to the output ends of the convolution kernel layers of the three groups of parallel paths, splicing and combining the features outputted by the three groups of parallel paths and outputting them to the fully connected network, to obtain a first predicted wear value; The intermediate feature is fused with the multi-scale feature to form a training set and a test set of fused features; The training set of fused features is input into the third training model for training, and the test set of fused features is input into the trained third training model to output the target predicted wear value.
2. The method according to claim 1, characterized in that The step of inputting the training set, the validation set, and the test set into the second training model for training comprises: The data set is input into a sampling module for sampling to obtain original sample data with the same dimension as the data set; the original sample data is input into the multi-scale network and the fully connected network to obtain a first predicted wear value; a first error value between the first predicted wear value and the true wear value label is calculated, and the network parameters of the second training model are adjusted using the first error value; When it is determined that the training of the second training model is completed, the training set and the validation set are combined and input into the trained second training model to obtain a first local multi-scale feature representation; the test set is input into the trained second training model to obtain a second local multi-scale feature representation; the multi-scale feature includes the first local multi-scale feature representation and the second local multi-scale feature representation.
3. The method according to claim 1 or 2, characterized in that: Using the second checkpoint of the second training model, comparing the current round loss of the validation set with the previous round loss, and saving the network parameters of the second training model when the current round loss is less than the previous round loss; It is determined whether the current number of network iterations satisfies a second preset number value. If the second preset number value is reached, the training of the second training model is completed, and the trained second training model is saved.
4. The method according to claim 1, characterized in that: The first training model comprises a noise module, an encoder module, and a decoder module; wherein the noise module is used to generate Gaussian noise, an input end of the encoder module is connected to an input end of the noise module and an output end of the decoder module, and an input end of the decoder module is connected to an output end of the encoder module; The encoder module and the decoder module use a fully connected network with symmetric numbers of neural units.
5. The method according to claim 4, characterized in that The step of inputting the first feature, the second feature, and the third feature into the first training model for training comprises: The first feature is input into the first training model, and Gaussian noise is added after passing through the noise module, and the decoder module obtains the reconstructed feature on the verification set; Calculating a second error value between the reconstructed feature and the first feature, and adjusting the network parameters of the first training model by back propagating the second error value; When it is determined that the training of the first training model is completed, the first feature and the second feature are combined and input into the trained first training model, and the encoder module outputs a first intermediate dimensionality reduction representation; The third feature is input into the trained first training model, and the encoder module outputs a second intermediate dimension reduction representation; The intermediate feature includes the first intermediate reduced dimensionality representation and the second intermediate reduced dimensionality representation.
6. The method according to claim 1, characterized in that Inputting the training set of fused features into a third training model for training to obtain a second predicted wear value; Calculating a third error value between the second predicted wear value and the actual wear value label, and using the third error value to adjust the network parameters of the third training model; Using the third checkpoint of the third training model, comparing the current round loss of the validation set with the previous round loss, and saving the network parameters of the third training model when the current round loss is less than the previous round loss; It is determined whether the current number of network iterations meets a third preset number of times. If the third preset number of times is reached, the training of the third training model is completed, and the trained third training model is saved.
7. The method according to claim 1, characterized in that A number of labels are made for the wear values in the tool wear value label data, wherein: the average wear value of the cutting edge of the tool back face is selected as the prediction label of the tool wear.
8. A tool wear prediction device, characterized in that: The tool wear prediction method according to any one of claims 1 to 7 is adopted, and the device comprises: A data set construction module is used to obtain the vibration signal data, cutting force signal data, sound emission signal data, and tool wear value label data of the tool, construct a data set, and divide the data set into a training set, a validation set, and a test set; A feature extraction module is used to extract multi-domain features of the training set, the validation set and the test set to obtain a first feature, a second feature and a third feature; A first training module is used to input the first feature, the second feature, and the third feature into a first training model for training, and output an intermediate feature; A second training module is used to input the training set, the validation set and the test set into a second training model for training, and output multi-scale features; A feature fusion module, used to fuse the intermediate feature with the multi-scale feature to form a training set and a test set of fused features; The third training module is used to input the training set of fused features into the third training model for training, and input the test set of fused features into the trained third training model to output the target predicted wear value.