Alloy machining machine tool operation monitoring management method based on digital twinning

Abstract

The application provides an alloy machining machine tool operation monitoring management method based on digital twinning, first constructs a digital twinning model for monitoring management, including a physical model and a twinning data model; the physical model includes a three-way cutting force sensor, a three-way vibration sensor and an acoustic emission sensor, a signal adjuster and an alloy machining tool, the sensors respectively collect three-way cutting force time domain signals, three-way vibration time domain signals and acoustic emission signals; the twinning data model is used for obtaining real-time monitoring data for monitoring management, and includes a trained multi-modal spatio-temporal feature fusion network and a tool wear amount prediction module; the multi-modal spatio-temporal feature fusion network comprehensively obtains residual multi-modal fusion features, and obtains an alloy machining tool flank wear amount prediction value through the tool wear amount prediction module. The application can comprehensively extract alloy machining tool features, has high model recognition stability, can discover potential wear problems in time, and improves production efficiency and alloy quality.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of alloy processing monitoring based on computer data processing, and particularly relates to an alloy processing machine tool operation monitoring and management method based on digital twinning. BACKGROUND

[0002] With the rapid development of intelligent manufacturing technology, the production efficiency, product quality and safety management level of alloy processing workshops, as important components of high-end manufacturing industry, have important influence on the overall production benefit. The most important equipment in alloy processing is an alloy processing machine tool, and the wear state of a tool of the processing machine tool is an important factor influencing the quality and processing efficiency of alloy processing. Therefore, whether the wear state of the tool can be monitored in a timely and accurate manner is an important prerequisite for guaranteeing the quality and safety of workpieces and machine tools.

[0003] A kind of milling cutter wear detection method based on matching wavelet depth migration learning is disclosed in Chinese invention patent with application number CN202310710028.3, first, vibration, acoustic emission, power sensor from machine tool milling plane processing process obtains original processing signal, extracts stationary cutting process signal from original processing signal, then one-dimensional fault signal is matched to dimension by Morse continuous wavelet, captures tiny changes, obtains visualized reinforced feature image;Second, effectively migrate the source domain model of depth migration network, the model has efficient image learning experience, can reduce the number of target domain training samples;Finally, according to the parameter optimization of process in model migration, limited data are carried out.

[0004] A kind of milling cutter wear monitoring method based on wavelet noise reduction and attention mechanism fusion GRU network is disclosed in Chinese invention patent with application number CN202210299396.9, first, the original processing signal in the process of machine tool milling plane processing is collected by sensor module, including external signal and internal signal, wherein, external signal includes vibration, current and acoustic emission, internal signal includes real-time position of machine tool internal moving shaft;Combined with the instruction in processing program and the position signal of machine tool internal moving shaft, stationary cutting process signal is extracted from original processing signal, then the signal is preprocessed by wavelet noise reduction, noise is removed and normalized operation is carried out, the denoised and normalized signal is analyzed and feature is extracted, to obtain multi-dimensional feature matrix on time series, combined with the corresponding wear real value to constitute sample set for model training;Then, a tool wear prediction model is established based on GRU network, the attention mechanism is integrated in time step and dimension for model training, the root mean square error between loss function, predicted value and real value is calculated in training process, to evaluate the model training effect;Finally, the tool wear prediction model after training is used to monitor the milling cutter wear.

[0005] However, the above technology only allows the neural network to learn the time domain features related to the tool wear value, without considering the frequency domain features, which can lead to the loss of information, because the frequency domain features often reveal the energy distribution and changes of the signal at different frequencies, which are closely related to the wear state of the tool. Through time-frequency conversion technology such as Fourier transform, time domain signals can be converted into frequency domain signals, so as to further extract frequency domain features and provide more comprehensive information for the identification of tool wear state. Moreover, the vibration signal, current and acoustic emission signal of the tool are data collected by different types of sensors, which have multi-modal properties. These different modal data each contain different aspects of information about the tool wear state, and there may be complementarity between these information. The above method does not consider the multi-modal properties of the training data, which may not fully utilize these complementary information, thereby affecting the accuracy and stability of the tool wear state identification. SUMMARY

[0006] In view of the above problems, the present application provides an alloy machining machine tool operation monitoring and management method based on digital twinning, comprising the following processes:

[0007] A digital twinning model for alloy machining machine tool operation monitoring and management is constructed, including an alloy machining tool physical model and a twinning data model;

[0008] The alloy machining tool physical model includes a three-way cutting force sensor, a three-way vibration sensor and an acoustic emission sensor, a signal conditioner and an alloy machining tool. The three-way cutting force sensor, the three-way vibration sensor and the acoustic emission sensor respectively collect three-way cutting force time domain signals, three-way vibration time domain signals and acoustic emission signals in the physical space where the alloy machining tool works during operation;

[0009] The twinning data model is used to obtain real-time monitoring data for monitoring and management, and includes a trained multi-modal spatio-temporal feature fusion network and a tool wear amount prediction module;

[0010] The multi-modal spatio-temporal feature fusion network includes a multi-scale convolution network, a spatial feature extraction network, a temporal feature extraction network and a multi-modal residual fusion network. The preprocessed three kinds of heterogeneous data are input into the multi-scale convolution network to obtain shallow features of different scales. The shallow features of different scales are input into the spatial feature extraction network to assign different weights to different features and extract key spatial features. At the same time, the shallow features of different scales are input into the temporal feature extraction network to capture the time sequence features in the three kinds of heterogeneous data, including the dynamic characteristics of tool wear over time. The spatial features and the temporal features are input into the multi-modal residual fusion network for cross-attention fusion to obtain residual multi-modal fusion features;

[0011] The tool wear amount prediction module obtains an alloy machining tool relief surface wear amount prediction value based on the input residual multi-modal fusion features.

[0012] Preferably, the multi-scale convolutional network comprises a multi-scale convolutional layer, a max pooling layer, an up-sampling layer and a feature concatenation layer; the multi-scale convolutional layer is composed of a 1*1 depth separable convolutional layer, a 3*3 depth separable convolutional layer and a 5*5 depth separable convolutional layer; the cutting force data, the vibration signal data and the acoustic emission data are respectively input into the multi-scale convolutional network to obtain shallow features of different scales, specifically as follows:

[0013] The cutting force data [A X A Y A Z ] of the alloy tool along the X axis, the Y axis and the Z axis in the working process, the vibration signal data [B X B Y B Z ] of the alloy tool along the X axis, the Y axis and the Z axis in the working process and the acoustic emission data C are respectively input into the multi-scale convolutional layer, and the specific calculation formula is as follows:

[0014] F A1 =DSConv 1*1 ([A X A Y A Z ])

[0015] F A3 =DSConv 3*3 ([A X A Y A Z ])

[0016] F A5 =DSConv 5*5 ([A X A Y A Z ])

[0017] F B1 =DSConv 1*1 ([B X B Y B Z ])

[0018] F B3 =DSConv 3*3 ([B X B Y B Z ])

[0019] F B5 =DSConv5*5 ([B X B Y B Z ])

[0020] F C1 = DSConv 1*1 (C)

[0021] F C3 = DSConv 3*3 (C)

[0022] F C5 = DSConv 5*5 (C)

[0023] wherein DSConv 1*1 (·), DSConv 3*3 (·) and DSConv 5*5 (·) represent deep separable convolution with kernel size of 1*1, 3*3 and 5*5 respectively; F A1 , F A3 and F A5 are the cutting force data features output by the 1*1 deep separable convolution layer, the 3*3 deep separable convolution layer and the 5*5 deep separable convolution layer respectively; F B1 , F B3 and F B5 are the vibration signal data features output by the 1*1 deep separable convolution layer, the 3*3 deep separable convolution layer and the 5*5 deep separable convolution layer respectively; F C1 , F C3 and F C5 are the acoustic emission data features output by the 1*1 deep separable convolution layer, the 3*3 deep separable convolution layer and the 5*5 deep separable convolution layer respectively;

[0024] F A1 , F A3 and F A5 are sequentially input into the up-sampling layer, and the resolution of different scale features is unified before being input into the feature concatenation layer to obtain the cutting force multi-scale feature MF A ; F B1 , F B3 and F B5 are sequentially input into the up-sampling layer, and the resolution of different scale features is unified before being input into the feature concatenation layer to obtain the vibration signal multi-scale feature MF B ; F C1 , F C3 and F C5 are sequentially input into the up-sampling layer, and the resolution of different scale features is unified before being input into the feature concatenation layer to obtain the acoustic emission multi-scale feature;

[0025] MFA , MF B and MF C respectively input the maximum pooling layer to reduce the resolution of the feature map.

[0026] Preferably, the spatial feature extraction network is composed of a channel attention mechanism layer and a spatial attention mechanism layer, which pool the cutting force data feature, the vibration signal data feature and the acoustic emission data feature RF A , RF B and RF C input the spatial feature extraction network, and the specific processing process is as follows:

[0027] RF A , RF B and RF C are sequentially input into the channel attention mechanism layer, which explicitly models the dependency between channels, and adaptively recalibrates the feature response of the channel direction, and the specific calculation formula is as follows:

[0028] SEF A = SE(RF A )

[0029] SEF5= SE(RF B )

[0030] SEF C = SE(RF C )

[0031] Where SF(·) is the channel attention layer, SEF A , SEF B and SEF C are the cutting force data channel attention feature, the vibration signal data channel attention feature and the acoustic emission data channel attention feature respectively.

[0032] SF A , SF B and SF C are sequentially input into the spatial attention layer, which captures the key area in the feature map, thereby extracting more accurate spatial features SAF A , SAF B and SAF C , and the SAF A , SAF B and SAF C are the cutting force data spatial attention feature, the vibration signal data spatial attention feature and the acoustic emission data spatial attention feature respectively.

[0033] Preferably, the time feature extraction network comprises a feature embedding layer, a convolution layer, an expanded convolution layer and a gating unit, and the time feature extraction network and the spatial feature extraction network are in a parallel structure. A B C The time feature extraction network is inputted with the RF

[0034] The RF A , the RF B and the RF C are inputted into the feature embedding layer and then into the convolution layer in sequence, the feature embedding layer is responsible for converting the input feature map into a low-dimensional feature vector representation, and the convolution layer is responsible for extracting the local time dependence of the feature map, and the specific calculation formula is as follows:

[0035] CEF A = Conv (Embe (RF A ))

[0036] CEF B = Conv (Embe (RF B ))

[0037] CEF C = Conv (Embe (RF C ))

[0038] Where Embe(·) is the feature embedding layer, Conv(·) is the convolution layer, CEF A , CEF B and CEF C are the local features of the cutting force data, the vibration signal data and the acoustic emission data respectively.

[0039] The CEF A , the CEF B and the CEF C are inputted into the expanded convolution layer and then into the gating unit in sequence to capture the long-term dependence in the time series data, and the specific calculation formula is as follows:

[0040] GEF A = GRU (Econv (CEF A ))

[0041] GEF B = GRU (Econv (CEF B ))

[0042] GEF C = GRU (Econv (CEF C )) ​​

[0043] where Econv(·) is an expanded convolutional layer, GRU(·) is a Gated Unit, GEF A , GEF B , and GEF C are the cutting force data time attention feature, the vibration signal data time attention feature, and the acoustic emission data time attention feature, respectively.

[0044] Preferably, the multi-modal residual fusion network comprises a feature concatenation layer, a feature summation layer, a multi-head attention mechanism layer, and a random dropout layer, and the input and output of the multi-modal spatio-temporal feature fusion network adopt a residual connection. The cutting force data spatial attention feature SAF A , the vibration signal data spatial attention feature SAF B , and the acoustic emission data spatial attention feature SAF, as well as the cutting force data time attention feature GEF A , the vibration signal data time attention feature GEF B , and the acoustic emission data time attention feature GEF C are input into the multi-modal residual fusion network to obtain multi-modal residual fusion features, and the specific processing process is as follows:

[0045] SAF A , SAF B , and SAF C are input into the multi-head attention mechanism layer to obtain multi-head attention weights, capturing the interaction and dependency between different modalities; at the same time, GEF A , GEF B , and GEF C are input into the multi-head attention mechanism layer to obtain multi-head attention weights, capturing the interaction and dependency between different modalities:

[0046] Att1 = Multihead(SAF A )

[0047] Att2 = Multihead(SAF B )

[0048] Att3 = Multihead(SAF C )

[0049] Att4 = Multihead(GEF A )

[0050] Att5 = Multihead(GEF B )

[0051] Att6 = Multihead(GEF C )

[0052] wherein Multihead(-) represents a multi-head attention mechanism, Att1, Att2, Att3, Att4, Att5 and Att6 are multi-head attention weight coefficients of SAF A , SAF B , SAF C , GEF A , GEF B and GEF C respectively;

[0053] The SAF A , SAF B , SAF C and their multi-head attention coefficients are input into a feature summation layer to obtain spatial multi-modal fusion features; meanwhile, the GEF B , GEF C and their multi-head attention coefficients are input into the feature summation layer to obtain temporal multi-modal fusion features:

[0054] SAF = Feasum(Att1*SAF A + Att2*SAF5+ Att3*SAF C )

[0055] CEF = Feasum(Att4*GEF A + Att5*GEF B + Att6*GEF C )

[0056] wherein Feasum(·) represents a feature summation layer, SAF is spatial multi-modal fusion features, and CEF is temporal multi-modal fusion features;

[0057] The SAF and CEF are input into a feature concatenation layer and then into a random inactivation layer to obtain multi-modal fusion features. The purpose of adding the random inactivation layer to the multi-modal residual fusion network is to inactivate part of the neurons in the network and inhibit the overfitting phenomenon in the training process.

[0058] MDF = Dropout(Concat(SAF; CEF))

[0059] wherein Dropout(·) is a random inactivation layer, and MDF is multi-modal fusion features.

[0060] The input SAF A , SAF B , SAF C , GEF A , GEF B and GEF C of the multi-modal residual fusion network are input into a feature concatenation layer to obtain multi-modal fusion features.The output MDF of the multimodal residual fusion network is connected to the feature summation layer to obtain the multimodal residual fusion feature RMDF.

[0061] Preferably, the tool wear prediction module includes a frequency domain feature recognition module and a tool wear prediction network;

[0062] The frequency domain feature recognition module is used to perform time-frequency conversion on the triaxial cutting force time-domain signal, the triaxial vibration time-domain signal, and the acoustic emission signal, and further extract the frequency domain signal feature HF; the tool wear prediction network outputs the predicted tool wear value based on the multimodal residual fusion feature RMDF obtained by the input frequency domain feature recognition module HF and the multimodal spatiotemporal feature fusion network.

[0063] Preferably, the frequency domain feature recognition module specifically includes a Fourier transform module, a principal component analysis module, and a convolutional layer;

[0064] The time-domain signal A of the three-axis cutting force X A Y A Z The three-dimensional vibration time-domain signal B X B Y B Z The acoustic emission signal C is sequentially input into the Fourier transform module, which converts the cutting force time-domain signal, vibration time-domain signal, and acoustic emission time-domain signal into the cutting force frequency-domain signal HA. X HA Y HA Z Vibration frequency domain signal HB X HB Y and HB Z Harmony signal HC in radio frequency domain;

[0065] HA X HA Y HAZ, HB X HB Y HB Z The HC input is used to obtain the dimensionality-reduced one-dimensional principal component H from the principal component analysis module. The specific calculation formula is as follows:

[0066] H = PCA([HA) X HA Y HA Z HB X HB Y HB Z HC])

[0067] PCA(·) is the principal component analysis module;

[0068] Input H into the convolutional layer to obtain the frequency domain signal characteristics HF.

[0069] Preferably, the tool wear amount prediction network is composed of stacked full connection layers, linear activation function layers and feature concatenation layers, the stacked full connection layers adopt a three-layer full connection layer structure, residual connections are used between the three-layer full connection layers to enhance the generalization ability of the network, the frequency domain signal feature HF and the obtained multi-modal residual fusion feature RMDF are input into the tool wear amount prediction network, and the specific process is as follows:

[0070] After the HF and the RMDF are input into the feature concatenation layer for feature concatenation, the stacked full connection layer is input, and the calculation formula is as follows:

[0071] MDH = Fullcon (Concat (HF; RMDF))

[0072] Where Fullcon(·) is the stacked full connection layer, and MDH is the output of the full connection layer.

[0073] The MDH is input into the linear activation function layer to obtain the tool wear amount prediction value The calculation formula is as follows:

[0074]

[0075] Where LAF(·) is the linear activation function.

[0076] Compared with the prior art, the present application has the following beneficial effects:

[0077] 1. Comprehensive alloy machining tool feature extraction: the multi-modal spatio-temporal feature fusion network can capture shallow features of vibration signals at different scales, while the spatial feature extraction network and the temporal feature extraction network can give different features different weights and extract key spatial and temporal features. This multi-scale, multi-dimensional feature extraction method can more comprehensively reflect the characteristics of tool three-direction cutting force time domain signals, three-direction vibration time domain signals and acoustic emission time domain signals. The multi-modal spatio-temporal feature fusion network can efficiently fuse spatial and temporal features through a cross-attention fusion mechanism to form more representative fusion features, providing strong support for subsequent wear state recognition.

[0078] 2. High precision of alloy machining tool wear state recognition: the frequency domain feature recognition module and the advanced feature recognition network analyze tool three-direction cutting force time domain signals, three-direction vibration time domain signals and acoustic emission time domain signals from different angles, improving the accuracy and comprehensiveness of wear state recognition.

[0079] 3. High stability of alloy machining tool wear state recognition: the adaptive smoothing layer smoothes the output results, which helps to eliminate accidental errors and noise interference and improve the stability and reliability of recognition.

[0080] 4. Alloy machining tool wear state real-time monitoring: a digital twin model of alloy machining machine tool is constructed, real-time monitoring and diagnosis of three-direction cutting force time domain signal, three-direction vibration time domain signal and acoustic emission time domain signal of the tool can be realized, and potential wear problems can be found in time. Through timely maintenance, the alloy machining quality decline and equipment downtime caused by tool wear can be reduced, and the production efficiency and alloy quality can be improved. BRIEF DESCRIPTION OF DRAWINGS

[0081] Figure 1 The overall method flowchart of the application.

[0082] Figure 2 The multi-modal spatio-temporal feature fusion network structure diagram of the application.

[0083] Figure 3 The tool wear amount prediction module network structure diagram of the application.

[0084] Figure 4 The digital twin model structure diagram constructed by the application.

[0085] Figure 5 The comparison curve diagram of the predicted value of the twin data model and the real tool flank wear amount in the embodiment of the application. DETAILED DESCRIPTION

[0086] The application proposes an alloy machining machine tool operation monitoring and management method based on digital twinning, and the overall process is as shown in Figure 1 The alloy machining tool wear state is monitored in real time through the alloy machining tool wear state information provided by the alloy machining tool digital twin model, and the production efficiency and equipment safety are improved.

[0087] The route of the present application comprises: firstly, collecting three kinds of heterogeneous data of three-direction cutting force time domain signal, three-direction vibration time domain signal and acoustic emission time domain signal of the tool in the alloy machining process, and dividing into training set and test set; secondly, constructing a multi-modal spatio-temporal feature fusion network to extract the features of the three kinds of heterogeneous data, the multi-modal spatio-temporal feature fusion network includes a multi-scale convolution network, a spatial feature extraction network, a temporal feature extraction network and a multi-modal residual fusion network, the preprocessed three kinds of heterogeneous data are input into the multi-scale convolution network to obtain shallow features of different scales, the shallow features of different scales are input into the spatial feature extraction network to give different weights to different features and extract key spatial features; at the same time, the shallow features of different scales are input into the temporal feature extraction network to capture the time sequence features in the three kinds of heterogeneous data, especially the dynamic characteristics of tool wear changing with time, the features output by the spatial feature extraction network and the temporal feature extraction network are input into the multi-modal residual fusion network for cross attention fusion to obtain multi-modal fusion features; thirdly, constructing a tool wear amount prediction module, inputting the three kinds of heterogeneous data and the fusion features into the module to obtain alloy machining tool wear state recognition results, the tool wear amount prediction module includes a frequency domain feature recognition module, a high-level feature recognition network, a softmax activation function layer and an adaptive smoothing layer, the three kinds of heterogeneous data are input into the frequency domain feature recognition module, the fusion features are input into the high-level feature recognition network, the output results of the three networks are weighted and averaged and then input into the adaptive smoothing layer for smoothing processing, the output of the adaptive smoothing layer is weighted and summed and then input into the softmax activation function layer to obtain the tool wear state recognition result; fourthly, training and testing the multi-modal spatio-temporal feature fusion module and the tool wear amount prediction module by using the training set and the test set to obtain the trained multi-modal spatio-temporal feature fusion network and the tool wear amount prediction module; finally, constructing an alloy machining machine tool digital twin model, inputting the three-direction cutting force time domain signal, the three-direction vibration time domain signal and the acoustic emission time domain signal obtained by the digital twin model into the trained multi-scale spatio-temporal feature fusion module and the tool wear amount prediction module to obtain the wear state diagnosis result.

[0088] According to the tool wear prediction value fed back by the alloy machining machine tool digital twin model, the machine tool in the alloy machining workshop is managed; for example, if the tool flank wear value fed back by the tool digital twin model is greater than the safety threshold when the alloy machining machine tool is working, the work is stopped, and the tool is replaced for the machine tool to continue to start.

[0089] The application will be further described below in combination with specific embodiments.

[0090] I. Alloy machining tool data set making

[0091] The alloy machining tool data set includes sample D and label L; a three-way cutting force sensor is installed below the alloy machining tool holder, a three-way vibration sensor is installed above the alloy machining tool holder and close to the front end of the tool, and an acoustic emission sensor is installed on the side of the alloy machining tool holder and close to the front end of the tool; the three-way cutting force sensor is used to sample the cutting force of the alloy tool along the X axis, the Y axis and the Z axis in the working process respectively, and the sampling frequency of each channel is 50Hz; the three-way vibration sensor is used to sample the vibration signal of the alloy tool along the X axis, the Y axis and the Z axis in the working process respectively, and the sampling frequency of each channel is 50Hz; the acoustic emission sensor is used to sample the acoustic emission signal of the alloy tool in the working process, and the sampling frequency is 50Hz, and the above data is used as a sample, so the sample D is:

[0092]

[0093] Wherein A X , A Y and A Z are the cutting force sample data of the alloy tool along the X axis, the Y axis and the Z axis in the working process, B X , B Y and B Z are the vibration signal sample data of the alloy tool along the X axis, the Y axis and the Z axis in the working process, C is the acoustic emission sample data, and N is the sample data size; the alloy machining tool relief surface wear L is used as a label, a three-dimensional microscope is used to measure the alloy machining tool relief surface wear, and the measurement interval is represented by the tool travel distance; in order to ensure the accuracy of the data, the measurement interval is set to 200mm, that is, the alloy machining tool relief surface wear is measured once every 200mm of tool travel:

[0094]

[0095] After D and L are combined, they are divided into a training set Train and a test set Test according to a ratio of 7:3.

[0096] II. Constructing an alloy machining tool digital twin model

[0097] The wear of the alloy machining tool is the result of the thermal and force coupling in the cutting process. The cutting parameters, tool materials, tool installation, and the physical and mechanical properties of the workpiece materials comprehensively affect the cutting state and wear of the tool. Therefore, the wear state evolution of the tool in the cutting process has the characteristics of complexity, dynamics and randomness. In order to realize accurate monitoring and prediction of the wear state of the tool, a digital twin model consistent with the physical alloy machining tool and synchronously virtual and real needs to be established. The digital twin model constructed by the present application includes an alloy machining tool physical model and a twin data model, and the specific structure is as shown in Figure 4 .

[0098] The alloy machining tool physical model is composed of a three-way cutting force sensor, a three-way vibration sensor and an acoustic emission sensor, a signal conditioner and an alloy machining tool, and the three-way cutting force sensor, the three-way vibration sensor and the acoustic emission sensor are responsible for sampling the three-way cutting force time domain signal, the three-way vibration time domain signal and the acoustic emission signal of the physical space where the alloy machining tool works during operation, respectively;

[0099] The twin data model is used to obtain real-time monitoring data for monitoring management, and includes a trained multi-modal spatio-temporal feature fusion network and a tool wear amount prediction module; the three-way cutting force time domain signal, the three-way vibration time domain signal and the acoustic emission signal obtained by the alloy machining tool physical model are input into the multi-modal spatio-temporal feature fusion network to obtain residual multi-modal fusion features, and then the residual multi-modal fusion features are input into the tool wear amount prediction module to obtain the alloy machining tool flank wear amount prediction value.

[0100] III. Construction of multi-modal spatio-temporal feature fusion network

[0101] The multi-modal spatio-temporal feature fusion network includes a multi-scale convolution network, a spatial feature extraction network, a temporal feature extraction network and a multi-modal residual fusion network, and the specific network structure is as shown in Figure 2

[0102] 1. Multi-scale convolution network

[0103] The multi-scale convolution network includes a multi-scale convolution layer, a maximum pooling layer, an up-sampling layer and a feature concatenation layer, wherein the multi-scale convolution layer is composed of a 1*1 deep separable convolution layer, a 3*3 deep separable convolution layer and a 5*5 deep separable convolution layer. The cutting force sample data, the vibration signal sample data and the acoustic emission sample data are input into the multi-scale convolution network to obtain shallow features of different scales, specifically as follows:

[0104] 1) The three modal sample data [A X A Y A Z ], [B X B Y B Z ] and C contained in the sample data D are input into the multi-scale convolution layer, and the specific calculation formula is as follows:

[0105] F A1 =DSConv 1*1 ([A X A Y A Z ])

[0106] F A3 =DSConv 3*3 ([A​X A Y A Z ])

[0107] F A5 =DSConv 5*5 ([A X A Y A Z ])

[0108] F B1 =DSConv 1*1 ([B X B Y B Z ])

[0109] F B3 =DSConv 3*3 ([B X B Y B Z ])

[0110] F B5 =DSConv 5*5 ([B X B Y B Z ])

[0111] F C1 =DSConv 1*1 (C)

[0112] F C3 =DSConv 3*3 (C)

[0113] F C5 =DSConv 5*5 (C)

[0114] Among them DSConv 1*1 (·), DSConv 3*3 (·) and DSConv 5*5 (·) represent depthwise separable convolutions with kernel sizes of 1*1, 3*3, and 5*5, respectively. The purpose of using depthwise separable convolutions to construct multi-scale convolutional layers in this application is to reduce the computational cost of convolutions and prevent gradient explosion during training; F A1 F A3 and F A5 The cutting force sample data features output by 1*1 depth separable convolutional layers, 3*3 depth separable convolutional layers, and 5*5 depth separable convolutional layers are respectively; F B1 F B3 and F B5are the sample data features of the vibration signal output by the 1*1 depth separable convolution layer, the 3*3 depth separable convolution layer and the 5*5 depth separable convolution layer respectively. C1 、F C3 and F C5 are the sample data features of the acoustic emission output by the 1*1 depth separable convolution layer, the 3*3 depth separable convolution layer and the 5*5 depth separable convolution layer respectively.

[0115] 2) input F A1 , F A3 and F A5 into the up-sampling layer in sequence, unify the resolutions of different scale features, and then input into the feature concatenation layer to obtain the multi-scale feature MF A of the cutting force sample. B1 , F B3 and F B5 are input into the up-sampling layer in sequence, the resolutions of different scale features are unified, and then input into the feature concatenation layer to obtain the multi-scale feature MF B of the vibration signal sample. C1 , F C3 and F C5 are input into the up-sampling layer in sequence, the resolutions of different scale features are unified, and then input into the feature concatenation layer to obtain the multi-scale feature MF C of the acoustic emission sample, and the specific calculation formula is as follows:

[0116] MF A =Concat(Up(F A1 ; F A3 ; F A5 ))

[0117] MF B =Concat(Up(F B1 ; F B3 ; F B5 ))

[0118] MF C =Concat(Up(F C1 ; F C3 ; F C5 ))

[0119] Where Up(·) is the up-sampling layer, and Concat(·) is the feature concatenation layer.

[0120] 3) input MF A , MF B and MF C into the max-pooling layer to reduce the resolution of the feature map, and then reduce the size and calculation amount of the feature map, and the specific calculation formula is as follows:

[0121] RF A= Maxpool(MF A )

[0122] RF B = Maxpool(MF B )

[0123] RF C = Maxpool(MF C )

[0124] wherein RF A , RF B and RF C are the cutting force sample data pooled feature, the vibration signal sample data pooled feature and the acoustic emission sample data pooled feature, respectively.

[0125] 2. Spatial feature extraction network

[0126] The spatial feature extraction network is composed of a channel attention mechanism layer and a spatial attention mechanism layer, and RF A , RF B and RF C are input into the spatial feature extraction network, and the specific steps are as follows:

[0127] RF A , RF B and RF C are input into the channel attention mechanism layer in turn, the dependency between channels is explicitly modeled, and the feature response of the channel direction is adaptively recalibrated, and the specific calculation formula is as follows:

[0128] SEF A = SE(RF A )

[0129] SEF B = SE(RF B )

[0130] SEF C = SE(RF C )

[0131] wherein SF(·) is the channel attention layer, SEF A , SEF B and SEF C are the cutting force sample data channel attention feature, the vibration signal sample data channel attention feature and the acoustic emission sample data channel attention feature, respectively.

[0132] SF A , SF B and SF CThe spatial attention layer is sequentially inputted to capture the key region in the feature map, so as to extract more accurate spatial features, and the specific calculation formula is as follows:

[0133] SAF A A )

[0134] SAF B B )

[0135] SAF C C )

[0136] Wherein, SA(·) is a spatial attention layer, SAF A , SAF B and SAF C are spatial attention features of cutting force sample data, vibration signal sample data and acoustic emission sample data respectively.

[0137] 3. Time feature extraction network

[0138] The time feature extraction network includes a feature embedding layer, a convolution layer, a dilated convolution layer and a gating unit. The time feature extraction network and the spatial feature extraction network are parallel structures. RF A , RF S and RF C are inputted into the time feature extraction network, and the specific steps are as follows:

[0139] RF A , RF B and RF C are sequentially inputted into the feature embedding layer and then into the convolution layer. The feature embedding layer is responsible for converting the input feature map into a low-dimensional feature vector representation to better represent the time-dependent information of the input feature map. The convolution layer is responsible for extracting the local time dependence of the feature map, and the specific calculation formula is as follows:

[0140] CEF A = Conv(Embe(RF A ))

[0141] CEF B = Conv(Embe(RF B ))

[0142] CEF C = Conv(Embe(RF C ))

[0143] Wherein Embe(·) is a feature embedding layer, Conv(·) is a convolution layer, CEF A ​​​CEF B and CEF C are local features of the cutting force sample data, the vibration signal sample data and the acoustic emission sample data, respectively.

[0144] CEF A , CEF B and CEF C are sequentially input into the dilated convolution layer and then input into the gate unit to capture long-term dependencies in the time series data, and the specific calculation formula is as follows:

[0145] GEF A = GRU (Econv (CEF A ))

[0146] GEF B = GRU (Econv (CEF B ))

[0147] GEF C = GRU (Econv (CEF C ))

[0148] wherein Econv(·) is the dilated convolution layer, GRU(·) is the gate unit, GEF A , GEF B and GEF C are time attention features of the cutting force sample data, the vibration signal sample data and the acoustic emission sample data, respectively.

[0149] 4. Multimodal residual fusion network

[0150] The multimodal residual fusion network comprises a feature concatenation layer, a feature summation layer, a multi-head attention mechanism layer and a random deactivation layer, and the input and output of the multimodal spatio-temporal feature fusion network adopt a residual connection, which allows the gradient to flow directly, thereby alleviating the problem of gradient disappearance, and the gradient is not easy to explode in the propagation process, further improving the stability of the network; the obtained SAF A , SAF B and SAF C and the obtained GEF A , GEF B and GEF C are input into the multimodal residual fusion network to obtain multimodal residual fusion features, and the specific calculation steps are as follows:

[0151] 1) SAF A , SAF B and SAF CThe input multi-head attention mechanism layer obtains multi-head attention weights, and captures the interaction and dependency between different modalities. A , GEF B and GEF C The input multi-head attention mechanism layer obtains multi-head attention weights, and captures the interaction and dependency between different modalities.

[0152] Att1=Multihead(SAF A )

[0153] Att2=Multihead(SAF B )

[0154] Att3=Multihead(SAF C )

[0155] Att4=Multihead(GEF A )

[0156] Att5=Multihead(GEF B )

[0157] Att6=Multihead(GEF C )

[0158] wherein Multihead(·) represents a multi-head attention mechanism, Att1, Att2, Att3, Att4, Att5 and Att6 are multi-head attention weight coefficients of SAF A , SAF B , SAF C , GEF A , GEF B and GEF C respectively.

[0159] 2) The input feature summation layer obtains spatial multi-modal fusion features of SAF A , SAF B and SAF C and their multi-head attention coefficients; meanwhile, the input feature summation layer obtains temporal multi-modal fusion features of GEF A , GEF B and GEF C and their multi-head attention coefficients.

[0160] SAF=Feasum(Att1*SAF A +Att2*SAF B +Att3*SAF C )

[0161] CEF = Feasum(Att4*GEF A + Att5*GEF B + Att6*GEF C )

[0162] Wherein Feasum(·) represents the feature summation layer, SAF is the spatial multi-modal fusion feature, and CEF is the temporal multi-modal fusion feature.

[0163] 3) input the SAF and CEF into the feature concatenation layer and then into the random inactivation layer to obtain the multi-modal fusion feature, and the purpose of adding the random inactivation layer to the multi-modal residual fusion network is to inactivate part of the neurons in the network and inhibit the overfitting phenomenon in the training process;

[0164] MDF = Dropout(Concat(SAF; CEF))

[0165] Wherein Dropout(·) is the random inactivation layer, and MDF is the multi-modal fusion feature.

[0166] 4) the input SAF A , SAF B , SAF C , GEF A , GEF B and GEF C of the multi-modal residual fusion network are connected in residual connection with the output MDF of the multi-modal residual fusion network through the feature summation layer to obtain the multi-modal residual fusion feature RMDF, and the specific calculation formula is as follows:

[0167] RMDF = Feasum(SAF A + SAF B + SAF C + GEF A + GEF B + GEF C + MDF).

[0168] Four, construct a tool wear prediction module

[0169] The tool wear prediction module includes a frequency domain feature recognition module and a tool wear prediction network, and the specific structure is as shown in Figure 3 The cutting force signal, vibration signal and acoustic emission signal generated by the alloy machining tool during the working process are essentially superimpositions of different frequency waves, and it is crucial to consider the characteristics of different frequency waves when analyzing the cutting force signal, vibration signal and acoustic emission signal. Therefore, the present application constructs a frequency domain feature recognition module to directly extract three modal sample data [A X A Y A Z ], [B XB Y B Z The frequency domain features of A, B and C reveal the change characteristics of the three modal sample data in the frequency domain, and provide more comprehensive information for the tool wear amount prediction network.

[0170] 1. A frequency domain feature identification module

[0171] The frequency domain feature identification module comprises a Fourier transform module, a principal component analysis module and a convolution layer.

[0172] A X , A Y , A Z , B X , B Y , B Z and C are sequentially input into the Fourier transform module to convert the cutting force time domain signal, the vibration time domain signal and the acoustic emission time domain signal into the cutting force frequency domain signal, the vibration frequency domain signal and the acoustic emission frequency domain signal. The specific calculation formula is as follows:

[0173] HA X =FT(A X )

[0174] HA Y =FT(A Y )

[0175] HA Z =FT(A Z )

[0176] HB X =FT(B X )

[0177] HB Y =FT(B Y )

[0178] HB Z =FT(B Z )

[0179] HC=FT(C)

[0180] Wherein FT(·) is the Fourier transform module, HA X , HA Y and HA Z are the cutting force frequency domain signals of the alloy tool along the X-axis, Y-axis and Z-axis during the working process, HB X , HB Y and HB Z are the vibration frequency domain signals of the alloy tool along the X-axis, Y-axis and Z-axis during the working process, and HC is the acoustic emission frequency domain signal of the alloy tool during the working process

[0181] The HA X , the HA Y , the HA Z , the HB X , the HB Y , the HB Z , and the HC are input into a principal component analysis module to obtain one-dimensional principal components H after dimension reduction. The principal component analysis is a commonly used data dimension reduction technique. It projects the cutting force frequency domain signal, the vibration frequency domain signal, and the acoustic emission frequency domain signal into a new coordinate system through linear transformation, retains the direction with the largest variance in the cutting force frequency domain signal, the vibration frequency domain signal, and the acoustic emission frequency domain signal, removes the directions with smaller variances, thereby effectively reducing the dimension of the data, which helps to reduce the computational complexity and storage requirements, and obtains the main features in the cutting force frequency domain signal, the vibration frequency domain signal, and the acoustic emission frequency domain signal. The specific calculation formula is as follows:

[0182] H = PCA ([HA X HA Y HA Z HB X HB Y HB Z HC])

[0183] Wherein PCA(·) is the principal component analysis module;

[0184] The H is input into a convolution layer to obtain a frequency domain signal feature HF. The specific calculation formula is as follows:

[0185] HF = Conv(H)

[0186] 2. Tool wear amount prediction network

[0187] The tool wear amount prediction network is composed of stacked fully connected layers, linear activation function layers, and feature concatenation layers. The stacked fully connected layers adopt a three-layer fully connected layer structure, and residual connections are used between the three-layer fully connected layers to enhance the generalization ability of the network. The frequency domain signal feature HF and the obtained multi-modal residual fusion feature RMDF are input into the tool wear amount prediction network. The specific calculation formula is as follows:

[0188] After the HF and the RMDF are input into the feature concatenation layer for feature concatenation, they are input into the stacked fully connected layer. The specific calculation formula is as follows:

[0189] MDH = Fullcon(Concat(HF; RMDF))

[0190] Wherein Fullcon(·) is the stacked fully connected layer, and MDH is the output of the fully connected layer.

[0191] The MDH is input into the linear activation function layer to obtain the tool wear amount prediction value The specific calculation formula is as follows:

[0192]

[0193] wherein LAF(·) is a linear activation function.

[0194] V. Regarding model training

[0195] Both the multi-modal spatio-temporal feature fusion network and the tool wear amount prediction network are trained in an end-to-end manner, so that the model can autonomously learn the tool wear amount time-domain features, frequency-domain features and multi-modal features from the training data. The mean square error (MSE) is used as the loss function during training:

[0196]

[0197] wherein N represents the number of samples, y i is the corresponding label, y i is the output of the model.

[0198] Model saving: according to the loss function value of the validation set, when the loss cannot be reduced for 50 consecutive iterations, the training process is terminated, and the best multi-modal spatio-temporal feature fusion network and tool wear amount prediction network model parameters are saved, and the obtained model is used as the final deployment model.

[0199] VI. Analysis of experimental results

[0200] The present application selects SNMG120408-PZ hard alloy tool, Kistler 9129AA three-way cutting force sensor, Kistler 8763B050BB three-way vibration sensor, Kistler 8152C0050502 acoustic emission sensor and Kistler 5152C signal conditioning instrument to construct an alloy machining tool physical model for three-way cutting force signal, three-way vibration signal and acoustic emission signal acquisition, and to make a training set and a test set to train and test the multi-modal spatio-temporal feature fusion network and tool wear amount prediction module constructed by the present application, and the trained multi-modal spatio-temporal feature fusion network and tool wear amount prediction module constitute a twin data model. The 40CrNiMoA alloy steel is selected for alloy machining experiment, and the comparison curve of the predicted value of the twin data model and the real tool flank wear amount is as shown in Figure 5 The error between the measured value and the predicted value obtained by the digital twin model constructed by the present application is small, which proves that the multi-modal spatio-temporal feature fusion network and tool wear amount prediction module proposed by the present application has high prediction accuracy.

[0201] The above merely describes preferred embodiments of the present application and is not intended to limit the present application. Any modifications, equivalent replacements, improvements, and the like made within the principle and technical scope of the present application should be included in the protection scope of the present application.

[0202] Although the specific embodiments of the present application are described above, the present application is not limited to the above, and various modifications or changes can be made by those skilled in the art without creative efforts based on the technical solutions of the present application, and still fall within the protection scope of the present application.

Claims

1. A method for monitoring and managing the operation of alloy machining machine tools based on digital twins, characterized in that, The process includes the following: Construct a digital twin model for monitoring and managing the operation of alloy machining machine tools, including a physical model of the alloy machining tool and a twin data model; The physical model of the alloy machining tool includes a triaxial cutting force sensor, a triaxial vibration sensor, an acoustic emission sensor, a signal modulator, and an alloy machining tool. The triaxial cutting force sensor, triaxial vibration sensor, and acoustic emission sensor respectively collect three heterogeneous data: triaxial cutting force time-domain signal, triaxial vibration time-domain signal, and acoustic emission signal in the physical space where the alloy machining tool is working. The twin data model is used to acquire real-time monitoring data for monitoring and management, and includes a trained multimodal spatiotemporal feature fusion network and a tool wear prediction module. The multimodal spatiotemporal feature fusion network includes a multi-scale convolutional network, a spatial feature extraction network, a temporal feature extraction network, and a multimodal residual fusion network. The three types of preprocessed heterogeneous data are input into the multi-scale convolutional network to obtain shallow features at different scales. The shallow features at different scales are then input into the spatial feature extraction network to assign different weights to different features and extract key spatial features. Simultaneously, shallow features at different scales are input into the temporal feature extraction network to capture time series features in three heterogeneous datasets, including the dynamic characteristics of tool wear over time; spatial and temporal features are input into the multimodal residual fusion network for cross-attention fusion to obtain residual multimodal fusion features. The tool wear prediction module obtains the predicted value of the flank wear of the alloy machining tool based on the input residual multimodal fusion features; the tool wear prediction module includes a frequency domain feature recognition module and a tool wear prediction network; The frequency domain feature recognition module is used to perform time-frequency conversion on the triaxial cutting force time-domain signal, the triaxial vibration time-domain signal, and the acoustic emission signal, and further extract frequency domain signal features. The tool wear prediction network is based on the characteristics of the input frequency domain signal. Multimodal residual fusion features obtained by multimodal spatiotemporal feature fusion network Predicted output tool wear value ; The frequency domain feature recognition module specifically includes a Fourier transform module, a principal component analysis module, and a convolutional layer; Time-domain signal of three-dimensional cutting force , , Three-dimensional vibration time-domain signal , , and acoustic emission signals The signals are sequentially input into the Fourier transform module, which converts the cutting force time-domain signal, vibration time-domain signal, and acoustic emission time-domain signal into the cutting force frequency-domain signal. , , Vibration frequency domain signal , and Harmony radio frequency domain signal ; Will , , , , , and Input the principal component analysis module to obtain the dimensionality-reduced one-dimensional principal components. The specific calculation formula is as follows: ； in Principal component analysis module; Will Frequency domain signal features are obtained by inputting the convolutional layer. ; The tool wear prediction network consists of stacked fully connected layers, linear activation function layers, and feature splicing layers. The stacked fully connected layers employ a three-layer fully connected layer structure, with residual connections between the three fully connected layers to enhance the network's generalization ability and incorporate frequency domain signal features. and the obtained multimodal residual fusion features The tool wear prediction network is input as follows: Will and The input feature concatenation layer performs feature concatenation, and then the concatenated features are input into the stacked fully connected layers. The calculation formula is as follows: ； in For stacked fully connected layers, This is the output of the fully connected layer; Will The predicted tool wear value is obtained by inputting a linear activation function layer. The calculation formula is as follows: ； in It is a linear activation function.

2. The method for monitoring and managing the operation of alloy machining machine tools based on digital twins as described in claim 1, characterized in that: The multi-scale convolutional network includes multi-scale convolutional layers, max pooling layers, upsampling layers, and feature concatenation layers. The multi-scale convolutional layers consist of 1x1 depthwise separable convolutional layers, 3x3 depthwise separable convolutional layers, and 5x5 depthwise separable convolutional layers. Cutting force data, vibration signal data, and acoustic emission data are input into the multi-scale convolutional network to obtain shallow features at different scales, as detailed below: The cutting force data of the carbide cutting tool along the X, Y, and Z axes during operation. Vibration signal data of carbide cutting tools along the X, Y, and Z axes during operation. Harmony emission data The specific calculation formulas for inputting multi-scale convolutional layers are as follows: ； ； ； ； ； ； ； ； ； in , and These represent depthwise separable convolutions with kernel sizes of 1*1, 3*3, and 5*5, respectively. , and The cutting force data features output by the 1*1 depth separable convolutional layer, the 3*3 depth separable convolutional layer, and the 5*5 depth separable convolutional layer are respectively. , and The vibration signal data characteristics output by the 1*1 depth separable convolutional layer, the 3*3 depth separable convolutional layer, and the 5*5 depth separable convolutional layer are respectively. , and The acoustic emission data features output by the 1*1 depth separable convolutional layer, the 3*3 depth separable convolutional layer, and the 5*5 depth separable convolutional layer are respectively. Will , and The features are sequentially input into the upsampling layer, and after unifying the resolution of features at different scales, they are input into the feature concatenation layer to obtain multi-scale features of cutting force. ;Will , and The upsampling layer is input sequentially, and after unifying the resolution of features at different scales, it is input into the feature stitching layer to obtain multi-scale features of the vibration signal. ;Will , and The acoustic emission multi-scale features are obtained by sequentially inputting the upsampling layer, unifying the resolution of features at different scales, and then inputting them into the feature concatenation layer. ; Will , and Input max pooling layers separately to reduce the resolution of the feature maps.

3. The method for monitoring and managing the operation of alloy machining machine tools based on digital twins as described in claim 1, characterized in that: The temporal feature extraction network includes feature embedding layers, convolutional layers, dilated convolutional layers, and gating units. The temporal feature extraction network and the spatial feature extraction network are in parallel structure, pooling features from cutting force data, vibration signal data, and acoustic emission data. , and The specific steps for inputting the temporal feature extraction network are as follows: Will , and The input features are sequentially fed into the feature embedding layer and then into the convolutional layer. The feature embedding layer is responsible for transforming the input feature map into a low-dimensional feature vector representation, and the convolutional layer is responsible for extracting the local temporal dependencies of the feature map. The specific calculation formula is as follows: ； ； ； in For feature embedding layer, It is a convolutional layer. , and These are local features of cutting force data, local features of vibration signal data, and local features of acoustic emission data, respectively. Will , and The data is sequentially fed into dilated convolutional layers and then into gated units to capture long-term dependencies in temporal data. The specific calculation formula is as follows: ； ； ； in It is a dilated convolutional layer. For gating units, , and These are the time attention features of cutting force data, vibration signal data, and acoustic emission data, respectively.

4. The method for monitoring and managing the operation of alloy machining machine tools based on digital twins as described in claim 1, characterized in that: The multimodal residual fusion network comprises a feature splicing layer, a feature summation layer, a multi-head attention mechanism layer, and a random deactivation layer. The input and output of the multimodal spatiotemporal feature fusion network are connected using residual connections, integrating spatial attention features of the cutting force data. Spatial attention characteristics of vibration signal data Spatial attention features of harmonic emission data and the time attention features of cutting force data Time attention characteristics of vibration signal data Harmonious emission data temporal attention features The multimodal residual fusion network is input to obtain multimodal residual fusion features. The specific processing steps are as follows: Will , and Input the multi-head attention mechanism layer to obtain multi-head attention weights and capture the interactions and dependencies between different modalities; simultaneously... , and Input the multi-head attention mechanism layer to obtain the multi-head attention weights and capture the interactions and dependencies between different modalities: ； ； ； ； ； ； in This indicates a multi-head attention mechanism. , , , , and They are respectively , , , , and Multi-head attention weighting coefficient; Will , and It also uses a multi-head attention coefficient input feature summation layer to obtain spatial multimodal fusion features; At the same time , and And its multi-head attention coefficient input feature summation layer, to obtain temporal multimodal fusion features: ； ； in Indicates the feature summation layer. For spatial multimodal fusion features, For temporal multimodal fusion features; Will and After the input feature concatenation layer, a random deactivation layer is input to obtain multimodal fusion features. The purpose of adding a random deactivation layer to the multimodal residual fusion network is to deactivate some neurons in the network and suppress overfitting during training. ； in For random deactivation layer, This is a multimodal fusion feature; Input to the multimodal residual fusion network , , , , and Output of the multimodal residual fusion network Multimodal residual fusion features are obtained by performing residual connections through a feature summation layer. .

Images