Online monitoring method and system for milling surface roughness considering real-time tool status

Abstract

This invention discloses an online monitoring method and system for milling surface roughness considering the real-time tool status, relating to the field of thin-walled part cutting process monitoring technology. It utilizes a service model to extract global and local features of the signal and calculates feature behavior indices; it filters input channels, sequentially selecting the feature vectors of each input channel with the largest behavior index to obtain a filtering feature matrix; it calculates the fused feature matrix based on an adaptive fusion algorithm of global and local features; it establishes and trains an observation model and a system model; and it uses the trained service model, observation model, and system model to monitor edge surface roughness online. This invention incorporates real-time tool status information into surface roughness monitoring through the observation model and system model, effectively improving monitoring accuracy. This simplifies the system model structure, ensuring that system training can be completed on edge devices; and it can quickly adapt to adjustments in working conditions.

Description

Technical Field

[0001] This invention relates to the field of monitoring technology for the cutting process of thin-walled parts, and in particular to a method and system for online monitoring of milling surface roughness that takes into account the real-time status of the tool. Background Technology

[0002] To ensure the service performance of critical components in extreme environments, the milling manufacturing process for thin-walled workpieces, such as aerospace impellers and blades, requires high precision. Among these, the surface roughness of the machined surface, a major component of the integrity of the machined surface, is an important indicator for evaluating workpiece machining quality and a key factor determining the service performance of components. Research shows that the geometric parameters and wear state of the cutting tool are among the main factors affecting the surface roughness of the machined surface. Therefore, introducing tool wear state parameters into surface quality prediction models is considered an effective way to improve the accuracy of machining quality identification. However, due to the influence of cutting fluid, oil mist, and chips at the machining site, direct online in-situ detection of tool condition or workpiece surface roughness is impractical. Surface roughness can only be measured after machining is completed and the machine is stopped, which greatly inconveniences the optimization and control of surface machining quality.

[0003] Cutting process monitoring technology based on advanced sensing and intelligent data pattern recognition makes it possible to perceive surface quality in real time. Existing technologies already exist that predict the remaining life of cutting tools and predict tool wear status based on measured surface roughness. For example, CN202010660771.9 discloses a method for predicting the remaining tool life considering tool wear and surface roughness, which achieves tool life prediction based on degradation indices by measuring tool wear and workpiece surface roughness at the same time. CN202210367064.X discloses a method for simultaneously predicting surface roughness and tool wear, which achieves simultaneous prediction of surface roughness and tool wear by establishing a stacked denoising autoencoder network and a multi-task model.

[0004] However, in the machining of thin-walled aerospace parts, the coupling between tool-workpiece vibration and faults such as tool wear, dulling, and chipping makes it difficult for existing methods to effectively predict the tool state during the cutting process, let alone guarantee the machining quality of thin-walled parts based on tool state monitoring. Furthermore, the time-varying dynamic characteristics of thin-walled workpiece machining determine the complexity of the state recognition model structure. When machining conditions change, both the tool state recognition model and the surface roughness recognition model need to be retrained, which is clearly detrimental to improving production efficiency. Therefore, the difficulty in obtaining the real-time tool state in surface roughness monitoring of thin-walled parts, and the complex adjustment process for different machining conditions, have become key technical bottlenecks restricting the development of thin-walled workpiece machining quality monitoring technology. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the purpose of this invention is to provide an online monitoring method and system for milling surface roughness that considers the real-time status of the cutting tool. By introducing real-time tool status information into surface roughness monitoring through observation models and system models, the monitoring accuracy is effectively improved. This simplifies the structure of the system model, ensuring that the training of the system structure can be completed on edge devices, and enabling rapid adaptation to changes in working conditions.

[0006] To achieve the above objectives, the present invention is implemented through the following technical solution:

[0007] In a first aspect, embodiments of the present invention provide an online monitoring method for milling surface roughness that takes into account the real-time state of the tool, including:

[0008] Acquire multi-channel processed data and build and train service models;

[0009] The global and local features of the signal are extracted using the service model. The global and local features are then concatenated to obtain the observation feature matrix and the system feature matrix. Feature behavior indicators are then calculated.

[0010] Calculate the channel behavior index and filter out the input channels; sequentially filter out the feature vectors of each input channel that have the maximum behavior index in the time domain, frequency domain, time-frequency domain and local features to obtain the filtering feature matrix;

[0011] An adaptive fusion algorithm based on global-local features is used to obtain feature components according to feature behavior indicators. The feature components are then superimposed to obtain a fusion feature matrix of time domain, frequency domain, time-frequency domain and local features.

[0012] An observation model is established and trained based on the filtered feature matrix and the fused feature matrix. The system model is trained using the output data of the observation model, the filtered feature data, and the surface roughness label data.

[0013] The surface roughness at the edge is monitored online using the trained service model, observation model, and system model.

[0014] As a further implementation, the service model includes a global feature extraction module and a local feature extraction module. The global feature extraction module calculates the statistical information of each original signal data sample in the time domain, frequency domain, and time-frequency domain. The local feature extraction module compresses the original signal time series data through the learning of a deep convolutional neural network to extract local features of the signal.

[0015] As a further implementation, the feature behavior index corresponding to each feature vector in the observation feature matrix is ​​calculated to obtain the observation feature behavior vector; the observation feature behavior vector is used to characterize the behavioral pattern of the observation features extracted by the service model.

[0016] As a further implementation, the characteristic behavior indicators include Spearman correlation coefficient, mutual information, and maximum information coefficient.

[0017] As a further implementation, the observation feature behavior vector and system feature behavior vector corresponding to all channels are calculated; the channel behavior indices of the observation model and system model are calculated, and the input channels are selected respectively.

[0018] As a further implementation, based on the input channels of the selected observation model and system model, the feature vectors with the largest behavioral indicators in the time domain, frequency domain, time-frequency domain and local features of each input channel are selected in turn. The feature vector with the best performance in each domain is selected to obtain the filter feature matrix.

[0019] The feature components of the fusion vector are calculated based on the filtered feature matrix. The feature components are then superimposed to calculate the fusion vector, resulting in a fusion feature matrix of time-domain, frequency-domain, time-frequency-domain features, and local features.

[0020] As a further implementation, an observation model is established through a convolutional block attention module; and a recognition network with multiple bidirectional GRU networks and multiple linear layers is used as the system model.

[0021] Secondly, embodiments of the present invention also provide an online monitoring system for milling surface roughness that takes into account the real-time state of the tool, including:

[0022] The service model building module is used to acquire multi-channel processed data and build and train service models.

[0023] The feature behavior index calculation module is used to extract global and local features of the signal using the service model, concatenate the global and local features to obtain the observation feature matrix and the system feature matrix, and calculate the feature behavior index.

[0024] The filter feature matrix acquisition module is used to calculate channel behavior indicators and filter out input channels; it sequentially filters out the feature vectors of each input channel that have the maximum behavior indicators in the time domain, frequency domain, time-frequency domain, and local features to obtain the filter feature matrix;

[0025] The fusion feature matrix acquisition module is used for the adaptive fusion algorithm based on global-local features. It obtains feature components according to feature behavior indicators, and then superimposes the feature components to obtain the fusion feature matrix of time domain, frequency domain, time-frequency domain and local features.

[0026] The system model training module is used to build and train the observation model based on the filtered feature matrix and the fused feature matrix, and to train the system model using the output data of the observation model, the filtered feature data and the surface roughness label data.

[0027] The online monitoring module is used to monitor the surface roughness of the edge end online using the trained service model, observation model, and system model.

[0028] Thirdly, embodiments of the present invention also provide an electronic device, including a memory and a processor, wherein the memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to perform the online monitoring method for milling surface roughness considering the real-time state of the tool.

[0029] Fourthly, embodiments of the present invention also provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the online monitoring method for milling surface roughness considering the real-time state of the tool.

[0030] The beneficial effects of this invention are as follows:

[0031] (1) This invention comprises three parts: a service model, an observation model, and a system model. By introducing real-time tool status information into roughness monitoring, the dimension of the input matrix of the system model is reduced, and the structure of the system model is simplified. Thus, the system model can be trained directly on the edge device. When the cutting conditions change, the system model only needs to be retrained at the edge, without repeatedly transmitting data and models to the cloud, and without retraining the service model and observation model in the cloud. This improves the efficiency of model deployment when the cutting conditions change and realizes agile roughness monitoring for flexible manufacturing.

[0032] (2) In the milling of thin-walled workpieces that are difficult to machine, the present invention addresses the problem of variable feature evolution due to the time-varying dynamic characteristics of the cutting system. By considering feature behavior indicators in the extraction and fusion of features, the invention achieves accurate monitoring of the real-time wear of the tool flank. By introducing the real-time tool wear value obtained from the monitoring into the system model, the invention achieves online monitoring of the surface roughness of the machined thin-walled workpiece considering the real-time tool state, which effectively improves the monitoring accuracy and generalization ability.

[0033] (3) The adaptive fusion algorithm of global-local features proposed in this invention can comprehensively learn the multidimensional information in the feature matrix, ensuring the robustness of the monitoring algorithm. At the same time, the feature behavior index proposed to determine the adaptive fusion feature components considers multiple parameters such as the correlation between the feature and the monitoring target vector, the maximum information coefficient, and mutual information. The feature weights obtained based on the multidimensional behavior index can accurately characterize the behavior of the feature. The selection of the input channel based on the channel behavior index improves the computational efficiency. Attached Figure Description

[0034] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0035] Figure 1 This is a flowchart of the present invention according to one or more embodiments;

[0036] Figure 2 This is a distribution diagram of the behavioral indicators of all features when the cutting data of the T1 tool is used as the test set according to one or more embodiments of the present invention;

[0037] Figure 3 This is a schematic diagram of the average value of the channel behavior index of the cutting data of three cutting tools T1, T2, and T3 according to one or more embodiments of the present invention;

[0038] Figure 4 This is a schematic diagram illustrating the calculation results of the characteristic behavior indicators of the channel selected according to one or more embodiments of the present invention;

[0039] Figure 5 This is a schematic diagram of the monitoring results of surface roughness Ra value when the cutting data of tool T1 is used as the test set according to one or more embodiments of the present invention;

[0040] Figure 6 This is a comparison chart of the surface roughness and the actual surface roughness trends after the working conditions are changed according to one or more embodiments of the present invention. Detailed Implementation

[0041] Example 1:

[0042] Existing thin-walled part machining quality monitoring models suffer from low surface roughness monitoring accuracy, poor algorithm robustness, and weak adaptability to working conditions due to neglecting the real-time status and time-varying dynamic characteristics of the tool. This embodiment provides an online monitoring method for milling surface roughness that considers the real-time status of the tool. It establishes a distributed model to incorporate real-time tool status information and combines it with cloud manufacturing technology; it includes:

[0043] Acquire multi-channel processed data and build and train service models;

[0044] The global and local features of the signal are extracted using the service model. The global and local features are then concatenated to obtain the observation feature matrix and the system feature matrix. Feature behavior indicators are then calculated.

[0045] Calculate the channel behavior index and filter out the input channels; sequentially filter out the feature vectors of each input channel that have the maximum behavior index in the time domain, frequency domain, time-frequency domain and local features to obtain the filtering feature matrix;

[0046] An adaptive fusion algorithm based on global-local features is used to obtain feature components according to feature behavior indicators. The feature components are then superimposed to obtain a fusion feature matrix of time domain, frequency domain, time-frequency domain and local features.

[0047] An observation model is established and trained based on the filtered feature matrix and the fused feature matrix. The system model is trained using the output data of the observation model, the filtered feature data, and the surface roughness label data.

[0048] The surface roughness at the edge is monitored online using the trained service model, observation model, and system model.

[0049] Specifically, such as Figure 1 As shown, it includes the following steps:

[0050] Step 1. Establish a service model, an observation model, and a system model; whereby the service model is used to extract features of the original machining signal, the observation model is used to monitor tool wear during the machining process and provide real-time tool status information to the system model, and the system model is used to identify the surface roughness of the workpiece in real time.

[0051] During the machining of thin-walled parts, cutting signals are acquired in real time using multi-channel sensors for acceleration, force, and noise to obtain the input signal matrix S. m×p×q =[c1,c2,c3,…,c p There are p channels in total, each channel has m signal samples, and each signal sample is time series data related to the sampling frequency, with a total of q sampling points.

[0052] Step 2. Extract signal features in real time through the service model; the service model includes a global feature extraction module and a local feature extraction module. The global feature extraction module calculates the statistical information of each original signal data sample in the time domain, frequency domain, and time-frequency domain; the local feature extraction module extracts local signal features by learning and compressing the original signal time series data through a 1D deep convolutional neural network.

[0053] Before using the service model to extract local features, the local feature extraction module is first trained based on the training set data. When the training data label of the service model is set to the tool wear value obtained by experimental measurement, local observation features can be extracted. When the training data label of the service model is set to the measured surface roughness value of the milled workpiece, local system features can be extracted.

[0054] The 1D deep convolutional neural network consists of four layers: the first layer is a 1dCNN layer with 8 units, the second layer is a 1dCNN layer with 64 units, the third layer is a linear layer with 128 units, and the last layer is a linear layer with 1 unit. The first three layers are activated using the ReLU function. A pooling layer with 2 kernels is placed between the convolutional and linear layers. Dropout = 0.6 is set before the last two linear layers. Notably, by changing the number of units in the first layer, local features at different scales can be obtained.

[0055] Based on this, by directly inputting the raw signal data into the service model, both global and local features can be extracted. Concatenating the global and local features yields the observation feature matrix and the system feature matrix, which can be represented as:

[0056]

[0057]

[0058] In equations (1) and (2), and Let f represent the observation feature matrix and system feature matrix of the j-th channel, respectively, where m is the number of original signal samples and n is the total number of all feature vectors; t(j) f f(j) and f tf(j) Let f represent the multidimensional time-domain feature vector, frequency-domain feature vector, and time-frequency-domain feature vector obtained from the statistical information of the original signal data based on the j-th channel; l(j) observe and f l(j) system Let represent the local observation feature vector and the local system feature vector of the j-th channel, respectively.

[0059] Step 3. For all eigenvectors in the observation feature matrix and system feature matrix obtained in Step 2, calculate their corresponding feature behavior indices in sequence; these indices consist of the Spearman correlation coefficient, mutual information, and maximum information coefficient between the eigenvector and the target vector after nonlinear activation; the feature behavior index b corresponding to any eigenvector in the feature matrix. i(j) The calculation process is as follows:

[0060] b i(j) =σ(Ac i(j) +Mic i(j) +Mi i(j) (3)

[0061] In equation (3), σ(·) represents the Sigmoid activation function, Ac i(j) f represents the i-th feature vector in the j-th channel.i(j) The activated Spearman correlation coefficient between the target vector t and the monitored target vector can be expressed as:

[0062]

[0063] In equation (4), the correlation parameter representing the feature negatively correlated with the monitoring target is directly set to zero using the ReLU activation function, d i(j) (u) f represents the i-th eigenvector of the j-th channel. i(j) The level difference between the u-th feature data and the u-th target element in the monitoring target vector;

[0064] In equation (3), Mic i(j) f represents the i-th eigenvector of the j-th channel. i(j) The maximum information coefficient of the monitored target vector t, according to the literature Reshef DN, Reshef YA, Finucane HK, et al. Detecting Novel Associations in Large Data Sets. Science 2011; 334:1518-24, can be obtained by the following formula:

[0065]

[0066] In equation (5), B is a variable related to the sample size.

[0067] In this embodiment, B is set to B(m) = m 0.6 I* represents the calculation of mutual information between two variables.

[0068] Furthermore, in equation (3), Mi i(j) f represents the i-th eigenvector of the j-th channel. i(j) The mutual information between the target vector t and the monitored vector, which is the normalized result, can be obtained by the following formula:

[0069]

[0070] In the formula, H(·) represents the information entropy of the calculated feature vector and the monitored target vector.

[0071] Based on equations (3) to (6), the observed feature matrix F (j) observe The characteristic behavior index corresponding to each feature vector can be calculated, thus forming an observation feature behavior vector, which represents the behavioral pattern of the observation features extracted by the service model. The observation feature behavior vector can be represented as:

[0072]

[0073] In equation (7), since the global features input in the observation model and the system model are the same, they are both results extracted by the global feature extraction module in the service model; therefore, the feature behavior index b of the local observation feature vector in equation (7) is used. (j) observe Replace with the feature behavior index b of the local system feature vector l(j) system This allows us to obtain the system characteristic behavior vector b. (j) system .

[0074] Step 4. For the observation model and the system model, the input channels are selected by calculating the channel behavior index.

[0075] Based on step 3, the observed characteristic behavior vector and system characteristic behavior vector corresponding to each channel can be calculated respectively; therefore, according to equation (8), the behavior characterization index B of the j-th channel can be calculated. j :

[0076]

[0077] Therefore, based on the original signal data with p channels, the observed channel behavior vector B of the cutting signal can be obtained. observe = (B (1) observe B (2) observe ,…B (j) observe ,…B (p) observe ) T Similarly, the system channel behavior vector B can be obtained. system = (B (1) system B (2) system ,…B (j) system ,…B (p) system ) T Based on the ranking of observed channel behavior indicators, the top two channel c values ​​can be selected from the input of the observation model. (1’) observe and c (2’) observe The calculation process can be expressed as follows:

[0078]

[0079]

[0080] In the formula, f(c j B observe ) indicates the calculation of channel c j Observation channel behavior vector B observe B * observe The observed channel behavior vector, excluding the maximum behavior index, is represented as follows:

[0081] B * observe ={B j observe |B j observe ∈B observe B j observe ≠max(B observe (11)

[0082] Using the same method, the top two channels c can be selected based on the input of the system model. (1’) system and c (2’) system This completes the separate selection of input channels for the observation model and the system model.

[0083] Step 5. Based on the input channels selected in Step 4 for the observation model and system model, sequentially select the feature vectors with the largest behavioral indicators in the time domain, frequency domain, time-frequency domain and local features of each input channel. Select the best performing feature vector in each domain to obtain the filtered feature matrix.

[0084] Taking the input channel of the observation-oriented model as an example, the calculation process of the filtering feature matrix is ​​as follows:

[0085]

[0086] In equation (12), the subscript c∈{1',2'} represents the first and second ranked input channels, respectively; f(f t(c) ,b t(c) ) represents the computation of the time-domain eigenvector f t(c) Characteristic behavioral indicators b t(c) f t(c) f f(c) f tf(c) and f l(c) These represent the multidimensional time-domain characteristics, frequency-domain characteristics, time-frequency-domain characteristics, and local characteristics of the input channel c, respectively. Similarly, the filtering feature matrix of the system model can be obtained by observing the filtering feature matrix of the model.

[0087] Step 6. After obtaining the filtering feature matrix, the feature components of the fusion vector can be calculated. The purpose is to adaptively fuse the features in each domain according to their behavioral indicators, thereby obtaining a fusion vector from the time-domain features, frequency-domain features, time-frequency-domain features, and local features of the selected channel. The feature components corresponding to each feature in the selected channel can be calculated according to the following formula:

[0088]

[0089] In equation (13), β is a hyperparameter, which is an odd number greater than 1, representing the fusion level. In this embodiment, it is set to 25 by default. The weight of the feature component obtained by equation (13) is related to the behavior index of the feature component. It can automatically assign corresponding coefficients to the original feature according to the size of the feature behavior index, thereby obtaining the feature component.

[0090] Step 7. By superimposing the feature components obtained in Step 6, the fusion vector can be calculated. Thus, a fusion vector can be obtained from the time domain, frequency domain, time-frequency domain features, and local features, and then the fusion feature matrix is ​​obtained. In this way, the adaptive fusion of global and local features is realized.

[0091] Taking the observation fusion matrix of the selected channel c as an example, its calculation process is as follows:

[0092]

[0093] In equation (14), M, N, P, and K represent the dimensions of the time-domain, frequency-domain, time-frequency-domain features, and local feature vectors, respectively. Similarly, by replacing the observed local feature components in equation (14) with the system local feature components, we can obtain the system fusion matrix A of the selected channel c. (c) system .

[0094] Step 8. By concatenating the filter feature matrices of the two channels selected in Step 5 and the fusion feature matrices of the two channels obtained in Step 7, the input feature matrix of the observation model can be obtained, which is represented as follows:

[0095]

[0096] Furthermore, to improve the accuracy of surface roughness identification, the input feature matrix of the system model can be obtained by concatenating the output vector of the observation model with the filtering feature matrix of the system model, which can be expressed as:

[0097]

[0098] According to equation (12), F (c)observe and F (c) system Both are 4-dimensional filtering feature matrices; according to equation (14), observe the fusion matrix A. (c) observe System fusion matrix A (c) system Also 4-dimensional, therefore the input matrices for the observation model and the system model are 16-dimensional and 9-dimensional, respectively.

[0099] Step 9. Observe the model building process as follows:

[0100] An observation model with a parallel heterogeneous structure is established using a convolutional block attention module, comprising four sub-models: sub-model A is a two-layer bidirectional LSTM, sub-model B is a two-layer bidirectional GRU, sub-model C is a two-layer unidirectional LSTM, and sub-model D is a two-layer unidirectional GRU network. The first and second layers of these four sub-models each have 64 units. The outputs of the four sub-models are stacked and then sequentially passed through the channel attention mechanism and spatial attention mechanism of the convolutional block attention module. Finally, the real-time tool wear monitoring results are output through two linear layers with 128 and 1 units respectively. To prevent overfitting, Dropout = 0.6 is set between the two linear layers.

[0101] It should be noted that the calculation principle formula of the convolutional block attention module is described in detail in the paper S.Woo, J.Park, J.-Y.Lee, ISKweon, CBAM: Convolutional Block Attention Module, in: V.Ferrari, M.Hebert, C.Sminchisescu, Y.Weiss, Computer Vision-ECCV 2018, Springer International Publishing, Cham, 2018, pp.3-19. The calculation principle of LSTM and GRU is described in detail in the paper Wang J, Yan J, Li C, Gao RX, Zhao R. Deep heterogeneous GRU model for predictive analytics in smart manufacturing: application to tool wear prediction. Computers in Industry 2019; 111:1-14.

[0102] Step 10. By introducing real-time tool status information, a system model with a simple structure can be established to achieve accurate monitoring of machining quality; the process of establishing the system model is as follows:

[0103] A recognition network consisting of two bidirectional GRU layers and two linear layers is established as the system model; the number of units in the first bidirectional GRU layer is 16, the number of units in the second bidirectional GRU layer is 32, the number of units in the first fully connected layer is 128, and the number of units in the last fully connected layer is 1.

[0104] Step 11. Train and deploy the observation model and system model; by sequentially inputting cutting signals and tool wear label data into the service model and observation model, the tool state is output, and the observation model is trained based on the training set data; then, according to step 8, the output of the trained observation model and the filtered features are input into the system model, and the system model is trained based on the training set data and surface roughness label data.

[0105] In this embodiment, the training of the observation model, system model, and service model are all completed on cloud devices. After the models are trained, the service model, observation model, and system model are jointly deployed to the edge device in the workshop to achieve online monitoring of surface roughness at the edge. When the operating conditions change, it is no longer necessary to repeatedly transmit models and data to the cloud to retrain the service model and observation model. Only the system model needs to be retrained on the edge device. Since the system model has a simple structure, the performance of the edge device is fully capable of meeting the requirements for training the system model. In this way, the model can be quickly adjusted when the operating conditions change, thereby achieving agile monitoring for flexible manufacturing.

[0106] Example 2:

[0107] To verify the online monitoring method described in Example 1, this example uses a high-speed milling experiment on thin-walled parts to collect milling signals, tool wear data tags, and workpiece surface roughness data tags. A full life-cycle cutting experiment was conducted using three identical end mills (denoted as T1 to T3). The cutting tool was a 14mm diameter insert end mill, the workpiece was a 5mm thick rectangular titanium alloy plate, the cutting speed was 351.85m / min, the feed per tooth was 0.08mm / t, the depth of cut was 4mm, and the width of cut was 0.2mm.

[0108] During the cutting process, accelerometers and inductive rotary force-measuring tool holders are used to collect workpiece vibration, milling force, and cutting bending moment, respectively. This yields vibration signals in three mutually perpendicular directions, milling force signals in three mutually perpendicular directions, and axial cutting bending moment signals, totaling seven channels of cutting signal data. One cutting stroke along the length of the thin plate constitutes one cutting operation. Each cutting operation corresponds to a tool wear data tag, a workpiece surface roughness data tag, and a set of seven channels of cutting signals.

[0109] After each machine stop, the cutting inserts are disassembled, and the average wear width of the insert's flank face is measured using a digital microscope. Between each measurement, nonlinear interpolation is used to determine the tool wear amount for each milling stroke, thus obtaining the tool wear vector for each milling cutter, which serves as the data label for the observation model. Similarly, after each machine stop, the workpiece is disassembled, and the surface roughness of the machined workpiece is measured using a Keyence VK-X2503D laser confocal microscope. Between each measurement, nonlinear interpolation is used to determine the surface roughness for each milling stroke, thus obtaining the workpiece surface roughness vector corresponding to all cutting strokes of each milling cutter.

[0110] First, according to step 1, a seven-channel input signal matrix S can be obtained for each tool. m×p×q In this example, each tool performs 100 cuts, so there are m = 100 samples in the input signal matrix, channel p = 7, the sampling frequency is set to 5kHz during the processing, and each time series data has a total of 32,000 sampling points.

[0111] Next, global and local features are extracted based on the service model in step 2. When extracting global features, for each channel, time-domain features, frequency-domain features, and time-frequency-domain features are extracted sequentially. The time-domain features include 12 features: average amplitude, maximum amplitude, standard deviation, root square amplitude, root mean square amplitude, peak-to-peak value, skewness, kurtosis, peak factor, margin factor, waveform factor, and impulse index, denoted as f1 to f2. 12 There are 12 frequency domain features, denoted as f. 13 ~f 24 The expression can be found in Table 1 of the paper "High-Performance Milling Tool Condition Monitoring Based on Feature Adaptive Fusion and Ensemble Learning [J / OL]. Journal of Mechanical Engineering 2023". The time-frequency domain features are obtained by decomposing 8 wavelet packet energy features through 3-layer wavelet packet decomposition, denoted as f. 25 ~f 32 .

[0112] When extracting local features, a 1D deep convolutional neural network model is first trained. The input data for extracting local observation features are the input signal matrix and tool wear vector obtained in step 1. For extracting local system features, the label data of the input signal matrix is ​​the surface roughness vector. During training, cutting data from tools T2 and T3 are used as the training set, and cutting data from tool T1 is used as the test set. The learning rate is set to 0.001, the batch size is 32, and a total of 100 training rounds are performed. Based on this, by setting the number of units in the first layer of the model to 8, 16, 32, and 64 respectively, four types of local features can be extracted from both the observation model and the system model. Thus, four types of local features can be extracted for each channel, denoted as f. 33 ~f 36.

[0113] After feature extraction, the behavior index corresponding to each feature is calculated according to step 3. For each channel of the cutting data of the three tools T1 to T3, the observed feature behavior vector and the system feature behavior vector can be calculated.

[0114] Taking the machining data of tool T1 as the test set as an example, Figure 2 The distribution of observed characteristic behavior indicators and system characteristic behavior indicators of cutting signals across various channels is shown, and the calculation of the behavior indicator for each feature lays the foundation for calculating channel behavior indicators in S4.

[0115] Following step 4, the input channels can be selected by calculating channel behavior indicators. Figure 3 The table shows the channel behavior characterization index of the cutting data calculated according to equation (8). To complete the selection of input channels, the average value of the channel behavior index of the three tools is calculated here. Thus, the input channels can be selected for the observation model and the system model respectively. The two input channels of the observation model are the cutting force in the x-direction and the cutting bending moment in the axial direction, respectively represented by f. x and m z This indicates that the two input channels of the system model are the cutting force in the y-direction and the cutting bending moment in the axial direction, respectively denoted by f. y and m z express.

[0116] Furthermore, according to step 5, by selecting the best-performing feature vector from the time domain, frequency domain, time-frequency domain, and local features, the filter feature matrix can be obtained for both the observation model and the system model. Figure 4 The figure shows the magnitude of the average behavior index corresponding to each feature in the input channel of the cutting signals of the three cutting tools for the observation model and the system model.

[0117] according to Figure 4 It can be obtained that, in this embodiment, F (1’) observe =[f3,f 16 ,f 25 ,f 35 ],F (2’) observe =[f1,f 14 ,f 25 ,f 33 ],F (1’) system =[f3,f 17 ,f 32 ,f 35 ],F (2’) system =[f8,f 17,f 25 ,f 34 ].

[0118] Then, according to equation (13) in step 6, by substituting all the original features and their corresponding behavioral indicators in sequence, the feature components corresponding to each feature can be calculated sequentially. Following step 7, by superimposing the feature components, a fusion vector is obtained in the time domain, frequency domain, time-frequency domain, and local features, thus obtaining the observation fusion matrix and the system fusion matrix. According to step 8, by concatenating the filter feature matrix of the observation model and the observation fusion matrix, the input feature matrix of the observation model can be obtained.

[0119] After obtaining the input feature matrix of the observation model, the observation model can be established according to step 9 and trained based on the training set data. During model training, the dataset partitioning is consistent with the training process of the local feature extraction model to complete the model training. When training the observation model, the input label data is the tool wear vector composed of the collected tool wear values. By inputting all feature data into the observation model, the monitored tool wear state can be output in real time. Thus, the input feature matrix of the system model can be obtained according to formula (16) in step 8.

[0120] Subsequently, a system model is established according to step 10, and the training of the system model is completed according to step 11.

[0121] Similarly, during system model training, after the model training is completed, the cutting data of the T1 tool is used to verify the monitoring effect on surface roughness. Figure 5 The image shows the monitoring results of surface roughness Ra values ​​for cutting data using tool T1, where T2 and T3 are the training set, and the cutting data from T1 is the test set. The image also calculates the Ra value between the predicted roughness and the actual surface roughness in the test set. 2 The R-value, used as a metric to evaluate predictive performance, is... 2 The closer the value is to 1, the higher the prediction accuracy.

[0122] pass Figure 5 As can be seen, the method proposed in Example 1 achieves a surface roughness monitoring accuracy of 91.9% for T1, and can accurately identify the surface roughness during the processing in real time.

[0123] Furthermore, in order to verify the ability to adapt to changes in operating conditions, Figure 6 The image shows a comparison of the surface roughness trends monitored and the actual surface roughness when using the T2 tool as the test set. At this point, only retraining the system model is required; the observation model and the system model remain unchanged. According to... Figure 6The proposed method accurately monitored the surface roughness of thin-walled parts after tool replacement, achieving a monitoring accuracy of 90.6%, which is essentially consistent with the accuracy before tool replacement. Furthermore, it eliminates the need for repeated training of the observation model and system model; the training of the system model can be directly completed on the edge device, thus verifying the feasibility of the proposed method.

[0124] Example 3:

[0125] To implement the method corresponding to Embodiment 1 above and achieve the corresponding functions and technical effects, an online monitoring system for milling surface roughness considering the real-time status of the tool is provided below. This system includes:

[0126] The service model building module is used to acquire multi-channel processed data and build and train service models.

[0127] The feature behavior index calculation module is used to extract global and local features of the signal using the service model, concatenate the global and local features to obtain the observation feature matrix and the system feature matrix, and calculate the feature behavior index.

[0128] The filter feature matrix acquisition module is used to calculate channel behavior indicators and filter out input channels; it sequentially filters out the feature vectors of each input channel that have the maximum behavior indicators in the time domain, frequency domain, time-frequency domain, and local features to obtain the filter feature matrix;

[0129] The fusion feature matrix acquisition module is used for the adaptive fusion algorithm based on global-local features. It obtains feature components according to feature behavior indicators, and then superimposes the feature components to obtain the fusion feature matrix of time domain, frequency domain, time-frequency domain and local features.

[0130] The system model training module is used to build and train the observation model based on the filtered feature matrix and the fused feature matrix, and to train the system model using the output data of the observation model, the filtered feature data and the surface roughness label data.

[0131] The online monitoring module is used to monitor the surface roughness of the edge end online using the trained service model, observation model, and system model.

[0132] Example 4:

[0133] This embodiment provides an electronic device, including a memory and a processor. The memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to perform the online monitoring method for milling surface roughness considering the real-time status of the tool as described in Embodiment 1.

[0134] The aforementioned electronic device may be a server.

[0135] Example 5:

[0136] This embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the online monitoring method for milling surface roughness considering the real-time status of the tool as described in Embodiment 1.

[0137] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.

[0138] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for on-line monitoring of milling surface roughness taking into account the real-time state of the tool, characterized in that, The method comprises the following steps: acquiring multi-channel processing data, establishing and training a service model; extracting global features and local features of signals by using the service model, splicing the global features and the local features to obtain an observation feature matrix and a system feature matrix, and calculating a feature behavior index; calculating a channel behavior index and screening input channels; sequentially screening feature vectors with the largest behavior index in time domain, frequency domain, time-frequency domain and local features of each input channel to obtain a filtered feature matrix; based on an adaptive fusion algorithm of global-local features, obtaining feature components according to the feature behavior index, and superimposing the feature components to obtain a fusion feature matrix of the time domain, the frequency domain, the time-frequency domain and the local features; establishing and training an observation model based on the filtered feature matrix and the fusion feature matrix, and training a system model by using output data of the observation model, screened feature data and surface roughness label data; using the trained service model, the observation model and the system model to monitor the surface roughness of the edge online; wherein the service model is used to extract features of original processing signals, the observation model is used to monitor tool wear during processing and provide real-time tool state information for the system model, and the system model is used to identify the surface roughness of the processed workpiece in real time; the service model comprises a global feature extraction module and a local feature extraction module, the global feature extraction module calculates statistical information of each original signal data sample in time domain, frequency domain and time-frequency domain, and the local feature extraction module compresses original signal time series data through learning of a deep convolutional neural network to extract local features of the signals; calculating observation feature behavior vectors and system feature behavior vectors corresponding to all channels; calculating channel behavior indexes of the observation model and the system model, and screening input channels; based on the screened input channels of the observation model and the system model, sequentially screening feature vectors with the largest behavior index in time domain, frequency domain, time-frequency domain and local features of each input channel, screening one optimal feature vector in each domain to obtain a filtered feature matrix; based on the filtered feature matrix, calculating feature components of fusion vectors, superimposing the feature components to calculate the fusion vectors, and obtaining fusion feature matrices of time domain, frequency domain, time-frequency domain features and local features; establishing an observation model through a convolution block attention module; and taking a recognition network with a multi-layer bidirectional GRU network and a multi-layer linear layer as a system model.

2. The method of claim 1, wherein the method is characterized by, The feature behavior index is calculated for each feature vector in the observation feature matrix to obtain an observation feature behavior vector; the observation feature behavior vector is used to represent the behavior law of the observation features extracted by the service model.

3. The method of claim 1 or 2, wherein the method is characterized by, The feature behavior index includes Spearman correlation coefficient, mutual information and maximum information coefficient.

4. A milling surface roughness on-line monitoring system considering real-time state of a tool, characterized by, The method comprises the following steps: a service model establishment module is configured to acquire multi-channel processing data, establish and train a service model; a feature behavior index calculation module is configured to extract global features and local features of signals by using the service model, splice the global features and the local features to obtain an observation feature matrix and a system feature matrix, and calculate a feature behavior index; The filtering feature matrix acquisition module is configured to calculate a channel behavior index and filter out input channels; the filtering feature matrix acquisition module is further configured to sequentially filter out feature vectors with the maximum behavior index in the time domain, the frequency domain, the time-frequency domain, and the local feature of each input channel, and obtain a filtering feature matrix; The fusion feature matrix acquisition module is configured to perform adaptive fusion algorithm based on global-local features, and obtain feature components according to feature behavior indexes, and obtain a fusion feature matrix of the time domain, the frequency domain, the time-frequency domain, and the local feature by superimposing the feature components; The system model training module is configured to establish and train an observation model based on the filtering feature matrix and the fusion feature matrix, and train a system model by using output data of the observation model, filtered feature data, and surface roughness label data. The online monitoring module is configured to perform online monitoring of edge surface roughness by using the trained service model, the observation model, and the system model.

5. An electronic device, comprising: The electronic device comprises a memory and a processor, the memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to perform the online monitoring method of the milling surface roughness considering the real-time state of the tool.

6. A computer-readable storage medium, characterized in that, The electronic device stores a computer program, and the computer program is executed by the processor to implement the online monitoring method of the milling surface roughness considering the real-time state of the tool.

Images