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

Abstract

The invention provides an alloy processing machine tool operation monitoring management method based on digital twinning, and the method comprises the steps: firstly constructing a digital twinning model used for monitoring management, and the digital twinning model comprises a physical model and a twinning data model; the physical model comprises a three-way cutting force sensor, a three-way vibration sensor, an acoustic emission sensor, a signal modulator and an alloy machining tool, and the sensors collect three-way cutting force time-domain signals, three-way vibration time-domain signals and acoustic emission signals respectively; the twin data model is used for acquiring real-time monitoring data for monitoring management, and comprises a trained multi-modal spatial-temporal feature fusion network and a tool wear loss prediction module; and the multi-modal spatial-temporal feature fusion network comprehensively obtains residual multi-modal fusion features, and obtains a predicted value of the abrasion loss of the rear tool face of the alloy machining tool through a tool abrasion loss prediction module. According to the method, alloy machining tool feature extraction can be comprehensively carried out, the model recognition stability is high, potential abrasion problems are found in time, and the production efficiency and the alloy quality are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of alloy processing monitoring based on computer data processing, and in particular relates to an alloy processing machine tool operation monitoring and management method based on digital twins. Background Art

[0002] With the rapid development of intelligent manufacturing technology, alloy processing workshops, as an important part of high-end manufacturing, have a significant impact on the overall production efficiency, product quality and safety management level. The most important equipment in alloy processing is the alloy processing machine tool, and the wear state of the machining tool is an important factor affecting the quality and 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 ensuring the quality and safety of workpieces and machine tools.

[0003] The Chinese invention patent with application number CN202310710028.3 discloses a milling tool wear detection method based on matched wavelet deep transfer learning. First, the original processing signal is obtained from the machine tool milling plane processing process through vibration, acoustic emission, and power sensors, and the smooth cutting process signal is extracted from the original processing signal. Then, the one-dimensional fault signal is matched and upgraded through Morse continuous wavelet to capture small changes and obtain a visualized enhanced feature image; secondly, the source domain model of the deep transfer network is effectively migrated. The model has efficient image learning experience and can reduce the number of training samples in the target domain; finally, the process parameters are optimized according to limited data in the model migration.

[0004] The Chinese invention patent with application number CN202210299396.9 discloses a milling tool wear monitoring method based on wavelet denoising and attention mechanism fused GRU network. First, the original processing signals in the milling plane processing process of the machine tool are collected through the sensor module, including external signals and built-in signals, wherein the external signals include vibration, current and acoustic emission, and the built-in signals include the real-time position of the built-in moving axis of the machine tool; combined with the instructions in the processing program and the position signal of the built-in moving axis of the machine tool, the stable cutting process signal is extracted from the original processing signal, and then the signal is pre-processed by wavelet denoising to remove noise and perform normalization operations, and the denoised and normalized signals are analyzed and feature extracted to obtain a multi-dimensional feature matrix on the time series, and the corresponding wear real values ​​are combined to form a sample set for model training; then a tool wear prediction model is established based on the GRU network, and the attention mechanism is integrated in the time step and dimension for model training. During the training process, the loss function and the root mean square error between the predicted value and the real value are calculated to judge the model training effect; finally, the trained tool wear prediction model is used to monitor the tool wear of milling processing.

[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 may lead to information loss, because the frequency domain features can often reveal the energy distribution and changes of the signal at different frequencies, and these changes are closely related to the wear state of the tool. Through time-frequency conversion technology, such as Fourier transform, the time domain signal can be converted into a frequency domain signal, so as to further extract the frequency domain features and provide more comprehensive information for the identification of the tool wear state. In addition, the vibration signal, current and acoustic emission signal of the tool are data collected by different types of sensors and have multimodal properties. These different modal data each contain different aspects of the tool wear state information, and there may be complementarity between these information. The above method does not consider the multimodal properties of the training data, and may not be able to fully utilize these complementary information, thus affecting the accuracy and stability of tool wear state identification. Summary of the invention

[0006] In view of the above problems, the present invention proposes a method for monitoring and managing alloy processing machine tools based on digital twins, which includes the following processes:

[0007] Build a digital twin model for monitoring and management of alloy processing machine tools, including the physical model of alloy processing tools and the twin data model;

[0008] The alloy processing tool physical model includes a three-dimensional cutting force sensor, a three-dimensional vibration sensor and an acoustic emission sensor, a signal conditioner and an alloy processing tool, wherein the three-dimensional cutting force sensor, the three-dimensional vibration sensor and the acoustic emission sensor respectively collect three kinds of heterogeneous data of a three-dimensional cutting force time domain signal, a three-dimensional vibration time domain signal and an acoustic emission signal in the physical space where the alloy processing tool is working;

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

[0010] 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 pre-processed heterogeneous data are input into the multi-scale convolutional network to obtain shallow features of different scales, and 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 heterogeneous data, including the dynamic characteristics of tool wear changing over time; the spatial features and temporal features are input into the multimodal residual fusion network for cross-attention fusion to obtain residual multimodal fusion features;

[0011] The tool wear prediction module obtains a predicted value of the flank wear of the alloy machining tool based on the input residual multimodal fusion feature.

[0012] Preferably, the multi-scale convolutional network includes a multi-scale convolutional layer, a maximum pooling layer, an upsampling layer and a feature splicing layer; the multi-scale convolutional layer consists 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-level features of different scales, as follows:

[0013] The cutting force data of the alloy tool along the X-axis, Y-axis and Z-axis during the working process [A X A Y A Z ], the vibration signal data of the alloy tool along the X-axis, Y-axis and Z-axis during the working process [B X B Y B Z ] and acoustic emission data C are input into the multi-scale convolution layer respectively, 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] Where DSConv 1*1 (·), DSConv 3*3 (·) and DSConv 5*5 (·) represents the depthwise separable convolution with kernel sizes of 1*1, 3*3, and 5*5 respectively; F A1 、F A3 and F A5 They are the cutting force data features output by 1*1 depthwise separable convolution layer, 3*3 depthwise separable convolution layer and 5*5 depthwise separable convolution layer respectively; F B1 、F B3 and F B5 They are the vibration signal data features output by 1*1 depthwise separable convolution layer, 3*3 depthwise separable convolution layer and 5*5 depthwise separable convolution layer respectively; F C1 、F C3 and F C5 They are the acoustic emission data features output by the 1*1 depthwise separable convolutional layer, the 3*3 depthwise separable convolutional layer, and the 5*5 depthwise separable convolutional layer;

[0024] F A1 、F A3 and F A5 The upsampling layer is input in sequence to unify the resolution of features of different scales and then input into the feature concatenation layer to obtain the multi-scale feature MF of cutting force. A ; F B1 、F B3 and F B5 The upsampling layer is input in sequence to unify the resolution of features of different scales and then input into the feature concatenation layer to obtain the multi-scale feature MF of the vibration signal. B ; F C1 、F C3 and F C5 The upsampling layers are input in sequence to unify the resolution of features at different scales and then input into the feature concatenation layer to obtain multi-scale features of acoustic emission;

[0025] Will MFA MF B and MF C Input the maximum pooling layer separately 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 pools the cutting force data features, the vibration signal data features and the acoustic emission data features. A ,RF B and RF C Input spatial feature extraction network, the specific processing process is:

[0027] RF A ,RF B and RF C The channel attention mechanism layer is input sequentially, and the feature response in the channel direction is adaptively recalibrated by explicitly modeling the dependencies between channels. 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 They are the attention features of the cutting force data channel, the attention features of the vibration signal data channel, and the attention features of the acoustic emission data channel;

[0032] SF A , SF B and SF C Input the spatial attention layer in sequence to capture the key areas in the feature map, thereby extracting more accurate spatial features SAF A 、SAF B and SAF C , the SAF A 、SAF B and SAF C They are the spatial attention features of cutting force data, the spatial attention features of vibration signal data and the spatial attention features of acoustic emission data respectively.

[0033] Preferably, the temporal feature extraction network comprises a feature embedding layer, a convolution layer, an expansion convolution layer and a gating unit. The temporal feature extraction network and the spatial feature extraction network are parallel structures. The cutting force data pooling feature, the vibration signal data pooling feature and the acoustic emission data pooling feature RF A ,RF B and RF C Enter the temporal feature extraction network. The specific steps are as follows:

[0034] RF A ,RF B and RF C The feature embedding layer is input in turn and then the convolution layer. The feature embedding layer is responsible for converting the input feature map into a low-dimensional feature vector representation. The convolution layer is responsible for extracting the local time dependency of the feature map. 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] Among them, Embe(·) is the feature embedding layer, Conv(·) is the convolutional layer, CEF A , CEF B and CEF C They are the local features of cutting force data, the local features of vibration signal data and the local features of acoustic emission data;

[0039] CEF A , CEF B and CEF C The dilated convolutional layers are input into the gated unit in turn to capture the long-term dependencies in the time series data. The specific calculation formula is as follows:

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

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

[0042] GEF C =GRU(Econv(SEF C ))

[0043] Among them, Econv(·) is the dilated convolutional layer, GRU(·) is the gated unit, GEF A ,GEF B and GEF C They are the time attention features of cutting force data, the time attention features of vibration signal data and the time attention features of acoustic emission data respectively.

[0044] Preferably, the multimodal residual fusion network comprises a feature concatenation layer, a feature summation layer, a multi-head attention mechanism layer and a random inactivation layer. The input and output of the multimodal spatiotemporal feature fusion network are connected by residual connection. The cutting force data spatial attention feature SAF A , vibration signal data spatial attention feature SAF B and the spatial attention feature SAF of acoustic emission data and the temporal attention feature GEF of cutting force data A , vibration signal data temporal attention feature GEF B Harmony emission data temporal attention feature GEF C Input the multimodal residual fusion network to obtain the multimodal residual fusion features. The specific processing process is as follows:

[0045] SAF A 、SAF B and SAF C Input the multi-head attention mechanism layer to obtain the multi-head attention weights to capture the interactions and dependencies between different modalities; at the same time, GEF A ,GEF B and GEF C Enter the multi-head attention mechanism layer to obtain the multi-head attention weights to capture the interactions and dependencies 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] Among them, Multihead(-) represents the multi-head attention mechanism, Att1, Att2, Att3, Att4, Att5 and Att6 are SAF A 、SAF B 、SAF C ,GEF A ,GEF B and GEF C The multi-head attention weight coefficient;

[0053] SAF A 、SAF B and SAF C and its multi-head attention coefficients are input into the feature summation layer to obtain spatial multimodal fusion features; at the same time, GEFA, GEF B and GEF C And its multi-head attention coefficient input feature summation layer to obtain the temporal multimodal fusion feature:

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

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

[0056] Where Feasum(·) represents the feature summation layer, SAF is the spatial multimodal fusion feature, and CEF is the temporal multimodal fusion feature;

[0057] SAF and CEF are input into the feature concatenation layer and then into the random dropout layer to obtain multimodal fusion features. The purpose of adding a random dropout layer to the multimodal residual fusion network is to inactivate some neurons in the network and suppress overfitting in the training process.

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

[0059] Among them, Dropout(·) is a random dropout layer, and MDF is a multimodal fusion feature;

[0060] The input SAF of the multimodal residual fusion network A 、SAF B 、SAF C ,GEF A ,GEF B and GEF CThe output MDF of the multimodal residual fusion network is residually connected through 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 three-dimensional cutting force time domain signal, the three-dimensional 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 tool wear prediction value based on the input frequency domain feature recognition module HF and the multimodal residual fusion feature RMDF obtained by 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 convolution layer;

[0064] The three-dimensional cutting force time domain signal A X , A Y , A Z , three-way vibration time domain signal B X , B Y , B Z , and the acoustic emission signal C are sequentially input into the Fourier transform module to convert 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 emission frequency domain signal HC;

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

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

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

[0068] Input H into the convolution layer to obtain the frequency domain signal feature HF.

[0069] Preferably, the tool wear prediction network is composed of a stacked fully connected layer, a linear activation function layer and a feature splicing layer. The stacked fully connected layer adopts a three-layer fully connected layer structure. Residual connections are used between the three fully connected layers to enhance the generalization ability of the network. The frequency domain signal feature HF and the obtained multimodal residual fusion feature RMDF are input into the tool wear prediction network. The specific process is as follows:

[0070] HF and RMDF are input into the feature concatenation layer for feature concatenation and then input into the stacked fully connected layer. The calculation formula is as follows:

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

[0072] Where Fullcon(·) is the stacked fully connected layer, and MDH is the output of the fully connected layer;

[0073] Input MDH into the linear activation function layer to obtain the tool wear 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 invention has the following beneficial effects:

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

[0078] 2. High accuracy in identifying the wear status of alloy processing tools: The frequency domain feature recognition module and the advanced feature recognition network analyze the tool's three-dimensional cutting force time domain signal, three-dimensional vibration time domain signal, and acoustic emission time domain signal from different angles, thereby improving the accuracy and comprehensiveness of wear status recognition.

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

[0080] 4. Real-time monitoring of alloy processing tool wear: Building a digital twin model of alloy processing machine tools can achieve real-time monitoring and diagnosis of the tool's three-dimensional cutting force time domain signal, three-dimensional vibration time domain signal, and acoustic emission time domain signal, and timely discover potential wear problems. Through timely maintenance, the decline in alloy processing quality and equipment downtime caused by tool wear can be reduced, and production efficiency and alloy quality can be improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0081] Figure 1 The figure is a flow chart of the overall method of the present invention.

[0082] Figure 2 Schematic diagram of the multimodal spatiotemporal feature fusion network structure of the present invention.

[0083] Figure 3 This is a schematic diagram of the network structure of the tool wear prediction module of the present invention.

[0084] Figure 4 Schematic diagram of the digital twin model structure constructed for the present invention.

[0085] Figure 5 This is a comparison curve between the predicted value of the twin data model and the actual wear amount of the tool flank in an embodiment of the present invention. DETAILED DESCRIPTION

[0086] The present invention proposes a method for monitoring and managing alloy processing machine tools based on digital twins. The overall process is as follows: Figure 1 As shown, the wear status information of the alloy processing machine tool tools provided by the digital twin model of the alloy processing machine tool tools can realize real-time monitoring of the wear status of the alloy processing tool tools, thereby improving production efficiency and equipment safety.

[0087] The technical route of the present invention includes: first, collecting three kinds of heterogeneous data, namely, three-dimensional cutting force time domain signals, three-dimensional vibration time domain signals and acoustic emission time domain signals of the tool during alloy processing, and dividing them into training sets and test sets; second, constructing a multimodal spatiotemporal feature fusion to extract the features of the three kinds of heterogeneous data, and 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, inputting the preprocessed three kinds of heterogeneous data into the multi-scale convolutional network to obtain shallow features of different scales, inputting the shallow features of different scales into the spatial feature extraction network to assign different weights to different features, and extracting key spatial features; at the same time, inputting the shallow features of different scales into the temporal feature extraction network to capture the time series features in the three kinds of heterogeneous data, especially the dynamic characteristics of tool wear changing with time, and inputting the features output by the spatial feature extraction network and the temporal feature extraction network into the multimodal residual fusion network for cross-attention fusion to obtain multimodal fusion features; third, constructing a tool wear prediction module to combine the three kinds of heterogeneous data into the multi-scale convolutional network, and obtaining the three kinds of heterogeneous data. The data and fusion features are input into the module to obtain the wear state recognition result of the alloy processing tool. The tool wear prediction module includes a frequency domain feature recognition module, an advanced feature recognition network, a softmax activation function layer and an adaptive smoothing layer. The three heterogeneous data are input into the frequency domain feature recognition module, and the fusion features are input into the advanced feature recognition network. The output results of the three networks are weighted averaged and input into the adaptive smoothing layer for smoothing. The output of the adaptive smoothing layer is weighted summed and input into the softmax activation function layer to obtain the tool wear state recognition result; again, the multimodal spatiotemporal feature fusion module and the tool wear prediction module are trained and tested using the training set and the test set to obtain the trained multimodal spatiotemporal feature fusion network and the tool wear prediction module; finally, a digital twin model of the tool for alloy processing machine tools is constructed, and the real-time three-dimensional cutting force time domain signal, three-dimensional vibration time domain signal and acoustic emission time domain signal obtained by the digital twin model are input into the trained multi-scale spatiotemporal feature fusion module and the tool wear prediction module to obtain the wear state diagnosis result.

[0088] The machine tools in the alloy processing workshop are managed based on the tool wear prediction value fed back in real time by the digital twin model of the alloy processing machine tool tools. For example, when the alloy processing machine tool is working, if the tool back surface wear fed back by the digital twin model of the tool is greater than the safety threshold, the operation is stopped and the machine tool is started again after the tool is replaced.

[0089] The invention will be further described below in conjunction with specific embodiments.

[0090] 1. Preparation of alloy machining tool data set

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

[0092]

[0093] Among them A X , A Y and A Z They are the cutting force sample data of the alloy tool along the X-axis, Y-axis and Z-axis during the working process, B X , B Y and B Z are the vibration signal sample data of the alloy tool along the X-axis, Y-axis and Z-axis during operation, C is the acoustic emission sample data, and N is the sample data scale; the flank wear L of the alloy machining tool is used as a label, and a three-dimensional microscope is used to measure the flank wear of the alloy machining tool. The measurement interval is characterized by the tool travel distance. In order to ensure the accuracy of the data, the measurement interval is set to 200mm, that is, the flank wear of the alloy machining tool is measured once every 200mm of tool travel:

[0094]

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

[0096] 2. Building a Digital Twin Model of Alloy Processing Machine Tools

[0097] The wear of alloy processing machine tool tools is the result of thermal-mechanical coupling during the cutting process. Cutting parameters, tool materials, tool installation, physical and mechanical properties of the workpiece material, etc. have a comprehensive impact on the cutting state and wear of the tool. Therefore, the evolution of the tool wear state during the cutting process is complex, dynamic, and random. To accurately monitor and predict the tool wear state, it is necessary to establish a digital twin model that is consistent with the physical alloy processing tool and synchronized with the real and virtual. The digital twin model constructed by the present invention includes a physical model of the alloy processing tool and a twin data model. The specific structure is as follows: Figure 4 shown.

[0098] The physical model of the alloy processing tool consists of a three-dimensional cutting force sensor, a three-dimensional vibration sensor and an acoustic emission sensor, a signal conditioner and an alloy processing tool. The three-dimensional cutting force sensor, the three-dimensional vibration sensor and the acoustic emission sensor are responsible for collecting the three-dimensional cutting force time domain signal, the three-dimensional vibration time domain signal and the acoustic emission signal of the physical space where the alloy processing tool is working.

[0099] The twin data model is used to obtain real-time monitoring data for monitoring and management, and includes a trained multimodal spatiotemporal feature fusion network and a tool wear prediction module; the three-dimensional cutting force time domain signal, three-dimensional vibration time domain signal and acoustic emission signal obtained by the physical model of the alloy processing tool are input into the multimodal spatiotemporal feature fusion network to obtain the residual multimodal fusion feature, and then the residual multimodal fusion feature is input into the tool wear prediction module to obtain the predicted value of the back face wear of the alloy processing tool.

[0100] 3. Constructing a multimodal spatiotemporal feature fusion network

[0101] 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 specific network structure is as follows: Figure 2 shown.

[0102] 1. Multi-scale Convolutional Network

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

[0104] 1) The three modal sample data [A X A Y A Z ]、[B X B Y B Z ] and C are input into the multi-scale convolution layer respectively, 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 ([AX 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] Where 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 convolution layers in this application proposal is that the depthwise separable convolution layers can reduce the amount of convolution operations and prevent gradient explosion during training; F A1 、F A3 and F A5 The cutting force sample data features output by 1*1 depthwise separable convolution layer, 3*3 depthwise separable convolution layer and 5*5 depthwise separable convolution layer respectively; F B1 、F B3 and F B5They are the vibration signal sample data features output by 1*1 depthwise separable convolution layer, 3*3 depthwise separable convolution layer and 5*5 depthwise separable convolution layer respectively; F C1 、F C3 and F C5 These are the acoustic emission sample data features output by the 1*1 depthwise separable convolutional layer, the 3*3 depthwise separable convolutional layer, and the 5*5 depthwise separable convolutional layer.

[0115] 2) F A1 、F A3 and F A5 The upsampling layer is input in sequence to unify the resolution of features of different scales and then input into the feature concatenation layer to obtain the multi-scale features MF of the cutting force sample. A , F B1 、F B3 and F B5 The samples are input into the upsampling layer in sequence, and the resolution of features of different scales is unified and then input into the feature concatenation layer to obtain the multi-scale features MF of the vibration signal samples. B ; F C1 、F C3 and F C5 The upsampling layer is input in sequence to unify the resolution of features of different scales and then input into the feature concatenation layer to obtain the multi-scale features MF of the acoustic emission samples. C , 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] Among them, Up(·) is the upsampling layer, and Concat(·) is the feature concatenation layer.

[0120] 3) Change MF A MF B and MF C The maximum pooling layer is input separately to reduce the resolution of the feature map, thereby reducing the size and calculation amount of the feature map. 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] Among them, RF A ,RF B and RF C They are cutting force sample data pooling features, vibration signal sample data pooling features and acoustic emission sample data pooling features respectively.

[0125] 2. Spatial feature extraction network

[0126] The spatial feature extraction network consists of a channel attention mechanism layer and a spatial attention mechanism layer. A ,RF B and RF C Input spatial feature extraction network, the specific steps are as follows:

[0127] RF A ,RF B and RF C The channel attention mechanism layer is input sequentially, and the feature response in the channel direction is adaptively recalibrated by explicitly modeling the dependencies between channels. 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] Where SF(·) is the channel attention layer, SEF A 、SEF B and SEF C They are the cutting force sample data channel attention features, vibration signal sample data channel attention features and acoustic emission sample data channel attention features.

[0132] SF A , SF B and SF CInput the spatial attention layer in sequence to capture the key areas in the feature map, thereby extracting more accurate spatial features. The specific calculation formula is as follows:

[0133] SAF A =SA(SEF A )

[0134] SAF B =SA(SEF B )

[0135] SAF C =SA(SEF C )

[0136] Among them, SA(·) is the spatial attention layer, SAF A 、SAF B and SAF C They are the spatial attention features of cutting force sample data, the spatial attention features of vibration signal sample data, and the spatial attention features of acoustic emission sample data.

[0137] 3. Temporal feature extraction network

[0138] The temporal feature extraction network includes a feature embedding layer, a convolutional layer, an expanded convolutional layer, and a gating unit. The temporal feature extraction network and the spatial feature extraction network are parallel structures. A ,RF S and RF C Enter the temporal feature extraction network. The specific steps are as follows:

[0139] RF A ,RF B and RF C The feature embedding layer is input into the convolution layer in turn. The feature embedding layer is responsible for converting the input feature map into a low-dimensional feature vector representation to better characterize the time dependency information of the input feature map. The convolution layer is responsible for extracting the local time dependency of the feature map. 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] Among them, Embe(·) is the feature embedding layer, Conv(·) is the convolutional layer, CEF A, CEF B and CEF C They are the local characteristics of cutting force sample data, the local characteristics of vibration signal sample data and the local characteristics of acoustic emission sample data.

[0144] CEF A , CEF B and CEF C The dilated convolutional layers are input into the gated unit in turn to capture the long-term dependencies in the time series data. The specific calculation formula is as follows:

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

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

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

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

[0149] 4. Multimodal residual fusion network

[0150] The multimodal residual fusion network includes a feature concatenation layer, a feature summation layer, a multi-head attention mechanism layer, and a random inactivation layer. The input and output of the multimodal spatiotemporal feature fusion network use residual connections. The residual connection allows the gradient to flow directly, thereby alleviating the problem of gradient disappearance. The gradient is not prone to explosion during propagation, further improving the stability of the network. The obtained SAF A 、SAF B and SAF C and GEF received A ,GEF B and GEF C Input the multimodal residual fusion network to obtain the multimodal residual fusion features. The specific calculation steps are as follows:

[0151] 1) SAF A 、SAF B and SAF CInput the multi-head attention mechanism layer to obtain the multi-head attention weights to capture the interactions and dependencies between different modalities; at the same time, GEF A ,GEF B and GEF C Input the multi-head attention mechanism layer to obtain the multi-head attention weights to capture the interactions and dependencies 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] Where Multihead(·) represents the multi-head attention mechanism, Att1, Att2, Att3, Att4, Att5 and Att6 are SAFs respectively. A 、SAF B 、SAF C ,GEF A ,GEF B and GEF C The multi-head attention weight coefficient.

[0159] 2) SAF A 、SAF B and SAF C and its multi-head attention coefficients are input into the feature summation layer to obtain spatial multimodal fusion features; at the same time, GEF A ,GEF B and GEF C And its multi-head attention coefficient input feature summation layer to obtain temporal multimodal fusion features;

[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] Where Feasum(·) represents the feature summation layer, SAF is the spatial multimodal fusion feature, and CEF is the temporal multimodal fusion feature.

[0163] 3) SAF and CEF are input into the feature concatenation layer and then into the random dropout layer to obtain multimodal fusion features. The purpose of adding a random dropout layer to the multimodal residual fusion network is to inactivate some neurons in the network and suppress the overfitting phenomenon in the training process.

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

[0165] Dropout(·) is a random dropout layer, and MDF is a multimodal fusion feature.

[0166] 4) Input SAF of the multimodal residual fusion network A 、SAF B 、SAF C ,GEF A ,GEF B and GEF C The output MDF of the multimodal residual fusion network is residually connected through the feature summation layer to obtain the multimodal residual fusion feature RMDF. The specific calculation formula is as follows:

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

[0168] 4. Constructing tool wear prediction module

[0169] The tool wear prediction module includes a frequency domain feature recognition module and a tool wear prediction network. The specific structure is as follows: Figure 3 As shown in the figure, the cutting force signal, vibration signal and acoustic emission signal generated by the alloy processing tool during the working process are essentially the superposition of waves of different frequencies. When analyzing the cutting force signal, vibration signal and acoustic emission signal, it is very important to consider the characteristics of the waves of different frequencies. Therefore, this application proposes to construct a frequency domain feature recognition module to directly extract the three modal sample data [A X A Y A Z ]、[B XB Y B Z ] and C’s frequency domain characteristics, revealing the changing characteristics of the three modal sample data in the frequency domain, and providing more comprehensive information for the tool wear prediction network.

[0170] 1. Frequency domain feature recognition module

[0171] The frequency domain feature recognition module includes 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 transform the cutting force time domain signal, vibration time domain signal and acoustic emission time domain signal into cutting force frequency domain signal, vibration frequency domain signal and 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] Where FT(·) is the Fourier transform module, HA X , HA Y and HA Z They are the cutting force frequency domain signals of the alloy tool along the X-axis, Y-axis and Z-axis during operation, and HB X HB Y and HB Z They 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] HA X , HA Y , HA Z HB X HB Y HB Z The principal component analysis module is used to obtain the one-dimensional principal component H after dimensionality reduction. Principal component analysis is a commonly used data dimensionality reduction technology. It projects the cutting force frequency domain signal, vibration frequency domain signal and 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, vibration frequency domain signal and acoustic emission frequency domain signal, and removes those directions with smaller variance, thereby effectively reducing the dimension of the data. This helps to reduce the computational complexity and storage requirements, and obtain the most important features of the cutting force frequency domain signal, vibration frequency domain signal and 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] Among them, PCA(·) is the principal component analysis module;

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

[0185] HF=Conv(H)

[0186] 2. Tool wear prediction network

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

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

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

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

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

[0192]

[0193] Where LAF(·) is the linear activation function.

[0194] 5. About model training

[0195] The multimodal spatiotemporal feature fusion network and the tool wear prediction network are both trained end-to-end. Through end-to-end training, the model can autonomously learn the time domain features, frequency domain features and multimodal features of tool wear from the training data. The mean square error (MSE) is used as the loss function during training:

[0196]

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

[0198] Model preservation: According to the loss function value of the validation set, when the loss cannot be reduced after 50 consecutive iterations, the training process is terminated, the best multimodal spatiotemporal feature fusion network and tool wear prediction network model parameters are saved, and the obtained model is used as the final deployment model.

[0199] 6. Experimental Results Analysis

[0200] The present invention selects SNMG120408-PZ carbide tool, Kistler 9129AA three-way cutting force sensor, Kistler8763B050BB three-way vibration sensor, Kistler8152C0050502 acoustic emission sensor and Kistler5152C signal conditioner to construct a physical model of alloy processing tool to collect three-way cutting force signal, three-way vibration signal and acoustic emission signal, and prepares training set and test set to train and test the multi-modal spatiotemporal feature fusion network and tool wear prediction module constructed by the present invention. The trained multi-modal spatiotemporal feature fusion network and tool wear prediction module constitute a twin data model. 40CrNiMoA alloy steel was selected for alloy processing experiment. The comparison curve between the predicted value of the twin data model and the actual wear of the tool back face is as shown in the figure. Figure 5 As shown, the error between the measured value and the predicted value obtained by the digital twin model constructed by the present invention is small, which proves that the multimodal spatiotemporal feature fusion network and tool wear prediction module proposed in the present invention have high prediction accuracy.

[0201] The above description is only the preferred embodiment of the present application and is not intended to limit the present application. For those skilled in the art, the present application may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

[0202] Although the above describes the specific implementation methods of the present invention, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without creative work are still within the scope of protection of the present invention.

Claims

1. A method for monitoring and managing alloy processing machine tools based on digital twins, characterized in that: The process includes: Build a digital twin model for monitoring and management of alloy processing machine tools, including the physical model of alloy processing tools and the twin data model; The alloy processing tool physical model includes a three-dimensional cutting force sensor, a three-dimensional vibration sensor and an acoustic emission sensor, a signal conditioner and an alloy processing tool, wherein the three-dimensional cutting force sensor, the three-dimensional vibration sensor and the acoustic emission sensor respectively collect three kinds of heterogeneous data of a three-dimensional cutting force time domain signal, a three-dimensional vibration time domain signal and an acoustic emission signal in the physical space where the alloy processing tool is working; The twin data model is used to obtain real-time monitoring data for monitoring and management, and includes a trained multi-modal 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 pre-processed heterogeneous data are input into the multi-scale convolutional network to obtain shallow features of different scales, and the shallow features of different scales are input into the spatial feature extraction network to assign different weights to different features, thereby extracting key spatial features; At the same time, shallow features of different scales are input into the temporal feature extraction network to capture the time series features in the three heterogeneous data, including the dynamic characteristics of tool wear over time; spatial features 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 a predicted value of the flank wear of the alloy machining tool based on the input residual multimodal fusion feature.

2. The method for monitoring and managing alloy processing machine tools based on digital twins according to claim 1, characterized in that: The multi-scale convolutional network includes a multi-scale convolutional layer, a maximum pooling layer, an upsampling layer and a feature splicing layer; the multi-scale convolutional layer consists 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-level features of different scales, as follows: The cutting force data of the alloy tool along the X-axis, Y-axis and Z-axis during the working process [A X A Y A Z ], the vibration signal data of the alloy tool along the X-axis, Y-axis and Z-axis during the working process [B X B Y B Z ] and acoustic emission data C are input into the multi-scale convolution layer respectively, and the specific calculation formula is as follows: F A1 =DSConv 1*1 ([A X A Y A Z ]) F A3 =DSConv 3*3 ([A X A Y A Z ]) F A5 =DSConv 5*5 ([A X A Y A Z ]) F B1 =DSConv 1*1 ([B X B Y B Z ]) F B3 =DSConv 3*3 ([B X B Y B Z ]) F B5 =DSConv 5*5 ([B X B Y B Z ]) F C1 =DSConv 1*1 (C) F C3 =DSConv 3*3 (C) F C5 =DSConv 5*5 (C) Where DSConv 1*1 (·), DSConv 3*3 (·) and DSConv 5*5 (·) represents the depthwise separable convolution with kernel sizes of 1*1, 3*3, and 5*5 respectively; F A1 , FA3 and F A5 They are the cutting force data features output by 1*1 depthwise separable convolution layer, 3*3 depthwise separable convolution layer and 5*5 depthwise separable convolution layer respectively; F B1 、F B3 and F B5 They are the vibration signal data features output by 1*1 depthwise separable convolution layer, 3*3 depthwise separable convolution layer and 5*5 depthwise separable convolution layer respectively; F C1 、F C3 and Fc5 are the acoustic emission data features output by the 1*1 depthwise separable convolutional layer, the 3*3 depthwise separable convolutional layer, and the 5*5 depthwise separable convolutional layer, respectively; Will F A1 、F A3 and F A5 The upsampling layer is input in sequence to unify the resolution of features of different scales and then input into the feature concatenation layer to obtain the multi-scale feature MF of cutting force. A ; F B1 、F B3 and F B5 The upsampling layer is input in sequence to unify the resolution of features of different scales and then input into the feature concatenation layer to obtain the multi-scale feature MF of the vibration signal. B ; F C1 、F C3 and F C5 The upsampling layers are input in sequence to unify the resolution of features at different scales and then input into the feature concatenation layer to obtain multi-scale features of acoustic emission; MFA, MF B and MF C Input the maximum pooling layer separately to reduce the resolution of the feature map.

3. The method for monitoring and managing alloy processing machine tools based on digital twins according to claim 1, characterized in that: The spatial feature extraction network consists of a channel attention mechanism layer and a spatial attention mechanism layer, which pools the cutting force data features, the vibration signal data features and the acoustic emission data features. A ,RF B and RF C Input spatial feature extraction network, the specific processing process is: RF A ,RF B and RF C The channel attention mechanism layer is input sequentially, and the feature response in the channel direction is adaptively recalibrated by explicitly modeling the dependencies between channels. The specific calculation formula is as follows: SEF A =SE(RF A ) SEF B =SE(RF B ) SEF C =SE(RF C ) Where SF(·) is the channel attention layer, SEF A 、SEF B and SEF C They are the cutting force data channel attention features, vibration signal data channel attention features and sound Emit data channel attention features; SFA, SF B and SF C The spatial attention layer is input sequentially to capture the key areas in the feature map, thereby extracting more accurate spatial features SAFA and SAF B and SAF C , the SAFA, SAF B and SAF C They are the spatial attention features of cutting force data, the spatial attention features of vibration signal data and the spatial attention features of acoustic emission data respectively.

4. The method for monitoring and managing alloy processing machine tool operations based on digital twins according to claim 1, characterized in that: The temporal feature extraction network comprises a feature embedding layer, a convolution layer, an expansion convolution layer and a gating unit. The temporal feature extraction network and the spatial feature extraction network are parallel structures. The cutting force data pooling feature, the vibration signal data pooling feature and the acoustic emission data pooling feature RF A ,RF B and RF C Enter the temporal feature extraction network. The specific steps are as follows: RF A ,RF B and RF C The feature embedding layer is input in turn and then the convolution layer. The feature embedding layer is responsible for converting the input feature map into a low-dimensional feature vector representation. The convolution layer is responsible for extracting the local time dependency of the feature map. The specific calculation formula is as follows: CEF A =Conv(Embe(RFA)) CEF B =Conv(Embe(RFB)) CEF C =Conv(Embe(RF C )) Among them, Embe(·) is the feature embedding layer, Conv(·) is the convolutional layer, CEFA, CEF B and CEF C They are the local features of cutting force data, the local features of vibration signal data and the local features of acoustic emission data; CEF A , CEF B and CEF C The dilated convolutional layers are sequentially input into the gated unit to capture the long-term dependencies in the time series data. The specific calculation formula is as follows: GEF A =GRU(Econv(SEFA)) GEF B =GRU(Econv(SEF B )) GEF C =GRU(Econv(SEF c )) Among them, Econv(·) is the dilated convolution layer, GRU(·) is the gated unit, GEFA, GEF B and GEF C They are the time attention features of cutting force data, the time attention features of vibration signal data and the time attention features of acoustic emission data respectively.

5. The method for monitoring and managing alloy processing machine tool operations based on digital twins according to claim 1, characterized in that: The multimodal residual fusion network includes a feature concatenation layer, a feature summation layer, a multi-head attention mechanism layer and a random inactivation layer. The input and output of the multimodal spatiotemporal feature fusion network adopt residual connection to combine the cutting force data spatial attention feature SAFA and the vibration signal data spatial attention feature SAF B and the spatial attention feature SAF of acoustic emission data, the temporal attention feature GEFA of cutting force data, and the temporal attention feature GEF of vibration signal data B Harmony emission data temporal attention feature GEF C Input the multimodal residual fusion network to obtain the multimodal residual fusion features. The specific processing process is as follows: SAF A 、SAF B and SAF C Input the multi-head attention mechanism layer to obtain the multi-head attention weights to capture the interactions and dependencies between different modalities; at the same time, GEFA, GEF B and GEF C Enter the multi-head attention mechanism layer to obtain the multi-head attention weights to capture the interactions and dependencies between different modalities: Att1=Multihead(SAFA) Att2=Multihead(SAF B ) Att3=Multihead(SAF C ) Att4=Multihead(GEF A ) Att5=Multihead(GEF B ) Att6=Multihead(GEF C ) Where Multihead(·) represents the multi-head attention mechanism, Att1, Att2, Att3, Att4, Att5 and Att6 are SAFs respectively. A 、SAF B 、SAF C ,GEF A ,GEF B and GEF C The multi-head attention weight coefficient; SAF A 、SAF B and SAF C and its multi-head attention coefficients are input into the feature summation layer to obtain spatial multimodal fusion features; at the same time, GEFA, GEF B and GEF C And its multi-head attention coefficient input feature summation layer to obtain the temporal multimodal fusion feature: <h2 style=";text-align:left;direction:ltr">SAF = Fe(Att1*SAF)<h2 style=";text-align:left;direction:ltr"> A <h2 style=";text-align:left;direction:ltr"> +Att2*SAF<h2 style=";text-align:left;direction:ltr"> B <h2 style=";text-align:left;direction:ltr"> +Att3*SAF<h2 style=";text-align:left;direction:ltr"> C <h2 style=";text-align:left;direction:ltr"> ) CEF=Feasum(Att4*GEF A +Att5*GEF B +Att6*GEF C ) Where Feasum(·) represents the feature summation layer, SAF is the spatial multimodal fusion feature, and CEF is the temporal multimodal fusion feature; SAF and CEF are input into the feature concatenation layer and then into the random dropout layer to obtain multimodal fusion features. The purpose of adding a random dropout layer to the multimodal residual fusion network is to inactivate some neurons in the network and suppress overfitting in the training process. MDF=Dropout(Concat(SAF;CEF)) Among them, Dropout(·) is a random dropout layer, and MDF is a multimodal fusion feature; The input of the multimodal residual fusion network is SAFA, SAF B 、SAF C , GEFA, GEF B and GEF C The output MDF of the multimodal residual fusion network is residually connected through the feature summation layer to obtain the multimodal residual fusion feature RMDF.

6. The method for monitoring and managing alloy processing machine tool operations based on digital twins according to claim 1, characterized in that: 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 three-dimensional cutting force time domain signal, the three-dimensional 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 tool wear prediction value based on the input frequency domain feature recognition module HF and the multimodal residual fusion feature RMDF obtained by the multimodal spatiotemporal feature fusion network 7. The method for monitoring and managing alloy processing machine tool operations based on digital twins according to claim 6, characterized in that: The frequency domain feature recognition module specifically includes a Fourier transform module, a principal component analysis module and a convolution layer; The three-dimensional cutting force time domain signal A X , A Y , A Z , three-way vibration time domain signal B X , B Y , B Z , and the acoustic emission signal C are sequentially input into the Fourier transform module to convert 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 emission frequency domain signal HC; HA X , HA Y , HA Z HB X HB Y HB Z The HC is input into the principal component analysis module to obtain the one-dimensional principal component H after dimensionality reduction. The specific calculation formula is as follows: H=PCA([HA X HA Y HA Z HB X HB Y HB Z HC]) Among them, PCA(·) is the principal component analysis module; Input H into the convolution layer to obtain the frequency domain signal feature HF.

8. The method for monitoring and managing alloy processing machine tool operations based on digital twins according to claim 6, characterized in that: The tool wear prediction network is composed of a stacked fully connected layer, a linear activation function layer and a feature splicing layer. The stacked fully connected layer adopts a three-layer fully connected layer structure. Residual connections are used between the three fully connected layers to enhance the generalization ability of the network. The frequency domain signal feature HF and the obtained multimodal residual fusion feature RMDF are input into the tool wear prediction network. The specific process is as follows: HF and RMDF are input into the feature concatenation layer for feature concatenation and then input into the stacked fully connected layer. The calculation formula is as follows: MDH=Fullcon(Concat(HF;RMDF)) Where Fullcon(·) is the stacked fully connected layer, and MDH is the output of the fully connected layer; Input MDH into the linear activation function layer to obtain the tool wear prediction value The calculation formula is as follows: Where LAF(·) is the linear activation function.

Images