Blade milling profile tolerance error prediction method driven by spatio-temporal feature fusion

Abstract

The invention discloses a spatio-temporal feature fusion driven blade milling profile tolerance error prediction method. According to the method, GAT, ECA Net and CNN are combined to form a blade profile error prediction model, the model is applied to a scene in which the profile error of a machine tool processing blade is predicted, static process parameters and dynamic process parameters are combined, a graph data driving model is constructed, time features and space features in blade processing data can be extracted through an algorithm, and the precision of the blade processing data is improved. And the two features are fused. Wherein the introduced GAT can learn the spatial relationship between the blade process parameters and the dynamic process parameters. The ECA Net can capture the time feature weight in the dynamic process parameters, so that the time sequence analysis capability of the model is improved. And finally, fusion extraction is performed on the features extracted by the GAT and the ECA Net through CNN, so that the time features and the space features are fused, and the prediction precision of the model on the blade milling profile tolerance error is effectively improved.

Description

Technical Field

[0001] The present invention relates to the field of intelligent manufacturing, and in particular to a blade milling profile error prediction method driven by spatiotemporal feature fusion. Background Art

[0002] In the field of intelligent manufacturing, complex surface parts represented by blades are difficult to process, have high processing costs, and require high manufacturing precision. Profile error is an important parameter to measure the quality of blade processing, which will directly affect the performance of the blade, including the distribution of airflow on the blade surface, vibration and noise, life and reliability. Therefore, it is particularly important to accurately predict the profile error of the blade.

[0003] In order to solve the problems existing in blade profile error prediction, deep learning has gradually become an effective tool. Compared with traditional methods, deep learning models can extract complex nonlinear relationships in blade processing data through multi-layer neural networks, which is particularly suitable for processing large amounts of multidimensional data in multi-axis processing, providing reliable technical support for high-quality control of blade manufacturing.

[0004] At present, the prediction methods for blade contour error mainly have the following problems: First, the model is not deep enough and the structure is unreasonable, which makes it impossible to accurately predict the complex mapping relationship between processing data, resulting in poor final prediction effect; Second, the feature engineering is insufficient. For the working conditions of blade multi-axis processing, there is a lack of model-driven mode driven by the fusion of static process parameters and dynamic process parameters. The prediction model cannot comprehensively learn the static process parameters and processing parameters of the blade, resulting in inaccurate final model prediction results; Third, the degree of fusion of spatiotemporal features is insufficient. The degree of fusion of temporal features and spatial features in the blade processing process data is insufficient, resulting in deficiencies in the rationality of data application and prediction effect.

[0005] Therefore, a blade profile error prediction method with a reasonable model structure and depth, sufficient feature engineering processing, and highly integrated temporal and spatial features is needed to accurately and efficiently predict blade processing. Summary of the invention

[0006] In view of the deficiencies in the prior art, the technical problem that the present invention intends to solve is to provide a blade milling profile error prediction method driven by spatiotemporal feature fusion.

[0007] The technical solution of the present invention to solve the technical problem is to provide a blade milling profile error prediction method driven by spatiotemporal feature fusion, the method comprising the following steps: Step 1: Before machining, obtain the CNC machining program according to the designed process parameters and import it into the machine tool, and install the blade blank; Step 2: Install the accelerometer probe to measure the vibrations of the spindle and the blade during the machining process; adjust the direction of the accelerometer probe so that the accelerometer coordinate system is consistent with the space coordinate system; Step 3, start processing, and collect acceleration signals on the spindle and the blade during the processing; after the blade processing is completed, under the premise that the state of the processed blade remains unchanged, use the machine tool probe to measure the profile error of the blade on the machine, and use the profile error as label data; Step 4, performing the same preprocessing on the acceleration signals on the main shaft and the blades obtained in step 3 to obtain preprocessed acceleration signals; Step 5, generating graph data according to the preprocessed acceleration signal obtained in step 4 and the process parameters in step 1, and setting the number of model iterations; Step 6: Input the graph data obtained in step 5 into GAT, extract the spatial features of the graph data through the combination of self-attention mechanism and multi-head attention mechanism, and obtain graph data with spatial features; Step 7: Input the graph data with spatial features obtained in step 6 into ECA Net. When allocating attention weights in the time dimension, ECA Net can dynamically adjust the channel weights of the input data and effectively capture the time features to obtain data with spatiotemporal features. Step 8: Use CNN to further fuse and extract the data with spatiotemporal features obtained in step 7, and output the predicted value of contour error; Step 9: Compare the contour error prediction value obtained in step 8 with the label data obtained in step 3, and perform back propagation iteration according to the principle of minimizing the loss function; repeatedly adjust the internal parameters of GAT, ECA Net and CNN according to the number of model iterations set in step 5, and the loss function value is determined by the square of the difference between the contour error prediction value and the label data; after stopping the iteration, obtain the contour error prediction value with the minimum loss function value within the number of model iterations.

[0008] Compared with the prior art, the present invention has the following beneficial effects: (1) The present invention combines static process parameters with dynamic process parameters to construct a graph data-driven model. The algorithm can extract the temporal features and spatial features in the blade processing data and fuse the two features. Among them, the introduced GAT can learn the spatial relationship between the blade process parameters and the dynamic process parameters. ECA Net can capture the temporal feature weights in the dynamic process parameters, thereby improving the model's timing analysis capability. Finally, the features extracted by GAT and ECA Net are fused and extracted through CNN, so that the temporal features and spatial features can be fused, effectively improving the model's prediction accuracy for blade milling profile errors. This deep learning network model structure is more suitable for processing complex nonlinear relationships between blade processing data, and helps to improve the prediction accuracy of blade processing.

[0009] (2) The present invention integrates the use of static process parameters and dynamic process parameters. These data include acceleration sensor information of the workpiece in the X, Y, and Z directions and acceleration sensor information of the spindle in the X, Y, and Z directions, as well as process parameters such as spindle speed, feed speed, path spacing, cutting depth, and tool roll angle, and generate graph data through processing.

[0010] (3) The present invention uses GAT to extract the spatial features of the sequence data of the blade processing process. GAT dynamically adjusts the weights between different nodes through the attention mechanism, which can capture complex nonlinear relationships, thereby more accurately reflecting the mutual influence between different process parameters and sensor data (i.e., preprocessed acceleration signals) during the blade processing process. In addition, GAT can take into account the importance of each node when aggregating node features, effectively improving the accuracy of feature extraction.

[0011] (4) The ECA Net of the present invention learns the weights of the time dimension of sensor data and process parameters. Through global pooling operations, ECA Net can effectively integrate the global information of each channel, thereby automatically learning the importance of each channel at different time points. This mechanism allows the model to dynamically adjust the weights of each channel and highlight key time series features.

[0012] (5) The present invention extracts spatiotemporal fusion features through CNN. CNN can efficiently capture local spatiotemporal features in input data through local connections and shared weight mechanisms, reducing the number of parameters and computational complexity. Secondly, the multi-layer structure of CNN enables it to extract and combine low-level to high-level features layer by layer, thereby capturing complex spatiotemporal patterns.

[0013] (6) The present invention combines GAT, ECA Net and CNN to form a blade profile error prediction model, and applies it to the scenario of predicting the profile error of machine tool blades. The advantages of each component are combined through a multi-stage processing flow to ensure that spatiotemporal features are gradually extracted, learned and fused from the initial data, and finally a more accurate analysis and prediction of the blade profile error is achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 It is a schematic diagram of the overall process of the present invention; Figure 2 It is a schematic diagram of the overall structure of the blade profile error prediction model of the present invention; Figure 3 It is a schematic diagram of the feature extraction process of the blade profile error prediction model of the present invention; Figure 4 It is a schematic diagram of the principle of ECA Net of the present invention; Figure 5 It is the true value-predicted value regression fitting diagram of Example 1 of the present invention. DETAILED DESCRIPTION

[0015] The specific embodiments of the present invention are given below. The specific embodiments are only used to further illustrate the present invention in detail and do not limit the protection scope of the present invention.

[0016] The present invention provides a blade milling profile error prediction method driven by spatiotemporal feature fusion (hereinafter referred to as method), the method comprising the following steps: Step 1: Before processing, obtain the CNC machining program according to the designed process parameters and import it into the machine tool, and install the blade blank: Preferably, step 1 specifically includes: before processing, performing CNC programming according to the process parameters designed before processing to obtain a CNC processing program; then importing the CNC processing program into the machine tool; clamping the blade blank according to the requirements of the process card, using the machine tool probe to position the blade blank and aligning the machine tool C-axis to the correct position; installing the auxiliary clamping device to ensure that the axis of the blade blank is always perpendicular to the turntable of the machining center during the processing.

[0017] Preferably, in step 1, the process parameters include spindle speed, feed speed, path spacing, cutting depth and tool rake angle.

[0018] Step 2: Install the accelerometer probe to measure the vibrations of the spindle and blade during the machining process; adjust the direction of the accelerometer probe so that the accelerometer coordinate system is consistent with the space coordinate system to ensure the accuracy of the measurement data; Preferably, in step 2, the vibration generated by cutting during the machining process will affect the contour of the finished blade. Therefore, an acceleration sensor probe is installed on the machine tool spindle near the milling cutter and the relative position is kept unchanged. The collected signal reflects the vibration generated by the spindle during the machining process; an acceleration sensor probe is installed on the auxiliary clamping device of the blade near the blade and the relative position is kept unchanged. The collected signal reflects the vibration generated by the blade during the machining process.

[0019] Preferably, in step 2, the installation position of the acceleration sensor probe should be as close as possible to the measurement position and the relative position should be kept unchanged to ensure the accuracy of the collected signal.

[0020] Step 3, run the CNC machining program of step 1, start machining, and collect acceleration signals on the spindle and the blade during machining to reflect vibration information during machining; after the blade machining is completed, under the premise that the state of the machined blade remains unchanged (that is, the fixture is not disassembled and the position of the blade is not changed), use the machine tool probe of step 1 to measure the profile error of the blade on the machine, and use the profile error as label data; Step 4: Perform the same preprocessing on the acceleration signals on the main shaft and the blade obtained in step 3 to improve the quality of the data, so that the blade profile error prediction model can more accurately extract key features and make predictions, and obtain the preprocessed acceleration signal; Preferably, in step 4, affected by actual processing conditions, the collected acceleration signal may have problems such as noise, redundant data, and abnormal signals.

[0021] Preferably, in step 4, preprocessing includes: denoising, downsampling and removing outliers on the collected acceleration signal, effectively reducing noise interference, removing redundant information, and making the data more standardized, thereby improving the accuracy of the prediction results.

[0022] Step 5: Generate graph data according to the preprocessed acceleration signal obtained in step 4 and the process parameters in step 1, and set the number of model iterations; the significance of generating graph data is that by combining data from different sources, it is conducive to more comprehensively capturing the complex relationship between time and space in the processing process; the significance of setting the number of model iterations is that by reasonably controlling the training process, the model can fully learn the features without overfitting, thereby optimizing the accuracy and generalization ability of the model; Preferably, in step 5, the process of generating graph data includes in sequence: reading and normalizing the preprocessed acceleration signal, reading the label data, dividing the preprocessed acceleration signal into samples, dividing each sample into multiple groups of data, generating a graph data for each group of data, and returning a graph data list.

[0023] Preferably, in step 5, the graph data is generated by calling the Gen_graph function through the python software platform, and the specific steps are as follows: S51, converting data into a graph structure: constructing the preprocessed acceleration signal obtained in step 4, the process parameters obtained in step 1, and the label data obtained in step 3 into a signal data matrix X, whose shape is N×F, where N is the number of nodes (a node corresponds to a sampling point or a feature point in the signal data. For example, different channel data of the acceleration signal of a blade in a certain period of time during the processing), and F is the feature dimension of each node; S52, generate adjacency matrix: construct the adjacency matrix A of the graph through the RadiusGraph graph generation strategy; S53, construct graph data: construct graph data by using the signal data matrix X and the graph adjacency matrix A, expressed as .

[0024] Step 6: Input the graph data obtained in step 5 into GAT (Graph Attention Network). Aiming at the nonlinear relationship between the data in the blade processing process, the self-attention mechanism and the multi-head attention mechanism are combined to effectively learn and capture the complex spatial features of the data in the blade processing process, extract the spatial features of the graph data, and obtain graph data with spatial features. Preferably, in step 6, the self-attention mechanism receives the feature matrix and adjacency matrix of each node through adaptive weight learning, calculates the importance of each neighbor node relative to the target node, dynamically adjusts the weight according to the relationship between the node features (that is, the initial weight of the self-attention mechanism changes with the change of the input), assigns different weights to different neighbors when aggregating the features of the neighbor nodes, and then performs weighted summation on the information of the neighbor nodes; specifically: in the prediction of the milling profile error of the blade, the processing process of the blade involves complex interactions between multiple influencing factors (such as process parameters, processing steps, processing operations, etc.), and the data corresponding to the influencing factors that are relatively important to the final contour shape of the blade are assigned higher weights; the self-attention coefficient β obtained by adaptive weight learning ij The importance of neighbor node j to target node i is measured as shown in formula (1), which quantifies the influence of each neighbor node in the feature aggregation of the target node: (1) In formula (1), LeakyReLU is the activation function; q is the attention weight vector; G is the weight matrix; f i is the feature vector of target node i, f jis the feature vector of neighbor node j; N(i) is the set of neighbor nodes of target node i, including all nodes connected to target node i; k is the index used to traverse the neighbor node set N(i) of target node i; Represents a concatenation operation.

[0025] Preferably, in step 5, the multi-head attention mechanism repeats the ordinary attention operation for multiple times, and concatenates or averages the results, which further enhances the expression ability of the model, better identifies and emphasizes the key features of different spatial positions in the processing process, and analyzes the key features of each spatial position to more comprehensively capture the spatial features in the blade processing data, thereby improving the prediction accuracy of the profile error. The multi-head attention mechanism is shown in formula (2): (2) In formula (2), is the final output feature of the target node i, D is the number of attention heads, σ is the activation function, is the attention coefficient of target node i to neighbor node j calculated by the dth attention head, G d is the weight matrix of the d-th attention head, f j is the feature vector of the neighbor node; j∈N(i) means that the neighbor node j is an element in the neighbor node set N(i) of the target node i.

[0026] Step 7: Input the graph data with spatial features obtained in step 6 into ECA Net (Efficient Channel Attention Network). When allocating attention weights in the time dimension, ECA Net can dynamically adjust the channel weights of the input data and effectively capture the time features, highlight and amplify important time features, so that the features of the data in the time dimension can be obtained more accurately, thereby enhancing the model's ability to understand and extract time series information, and obtaining data with rich spatiotemporal features, that is, data with spatial and temporal features; Preferably, in step 7, the ECA Net is composed of a global average pooling, a one-dimensional convolution, a Sigmoid activation function and an attention weight connected in sequence; Preferably, step 7 specifically includes: firstly, performing global average pooling (GAP) on the feature map of each input channel (i.e., the graph data with spatial features) to obtain a global description vector of each channel; then, performing a one-dimensional convolution (conv1d) operation to realize local interaction between channels, and then using a Sigmoid activation function to activate the result of the one-dimensional convolution; finally, using the attention weight to recalibrate the feature map of each channel to obtain an enhanced feature map (i.e., data with spatiotemporal features); For the input feature map , W represents the width of the feature map, H represents the height of the feature map, C represents the number of channels, and a vector of size C is obtained by global average pooling. The value of the global average pooling value of the cth channel in the feature map is Z c As shown in formula (3): (3) In formula (3), M c (t,r) is the pixel value of the cth channel in the feature map M at the tth row and rth column.

[0027] The expression of Sigmoid activation function is shown in formula (4): (4) In formula (4), x is the data input into the activation function in the previous step.

[0028] The enhanced feature map is shown in formula (5): (5) In formula (5), a c is the attention weight of channel c; is the Hadamard product, U is the feature output by formula (4); is a weighted feature.

[0029] Step 8: Use CNN (convolutional neural network) to further fuse and extract the data with spatiotemporal features obtained in step 7. After multiple layers of convolution and maximum pooling operations, and then through the fully connected layer, the contour error prediction value is output.

[0030] Preferably, in step 8, the CNN is composed of multiple layers of convolution, one layer of maximum pooling layer and two layers of fully connected layers connected in sequence, and can extract high-level features of spatiotemporal features layer by layer; the multi-layer convolution is composed of multiple layers of convolution layers; The data with spatiotemporal features obtained in step 7 is first subjected to multi-layer convolution operations (i.e., multiple convolution operations). After each convolution operation, the dimension of the data will change; after the convolution operation, the maximum pooling operation is used to reduce the data dimension while retaining important features; finally, the output of the blade milling profile error prediction value is completed through the fully connected layer. This hierarchical feature extraction method enables CNN to extract independent temporal features and spatial features together, extract higher-level features layer by layer, and capture complex spatiotemporal patterns and correlations. After the spatiotemporal feature data processed by GAT and ECA Net is input into CNN, the convolution and pooling operations will further integrate and optimize these features, allowing the model to better understand and utilize complex patterns in the data, thereby improving prediction performance.

[0031] Step 9. Compare the contour error prediction value obtained in step 8 with the label data obtained in step 3, and perform back propagation iteration according to the principle of minimizing the loss function; repeatedly adjust the internal parameters of the blade contour error prediction model according to the number of model iterations set in step 5, and the loss function value is determined by the square of the difference between the contour error prediction value and the label data; after stopping the iteration, obtain the contour error prediction value with the smallest loss function value within the number of model iterations to complete the prediction.

[0032] Preferably, in step 9, the loss function adopts a mean square error loss function, as shown in formula (6): (6) In formula (6), n is the number of samples, y m is the label data of the mth sample, is the predicted value of the contour error of the mth sample.

[0033] Embodiment 1: In step 1, the machine tool uses a five-axis machining center, specifically a Jingdiao CNC five-axis machining center GJSK400M; the CNC programming software is JDSoft SurfMill. The blade blank material used is aluminum alloy.

[0034] In step 2, the acceleration sensor used is Kistler's three-axis waterproof accelerometer-8763B.

[0035] In step 3, the sampling frequency of the acceleration sensor is 200000 Hz, and a total of 18 groups of blade data are processed and collected, with 128 pieces of profile error data collected in each group.

[0036] In step 4, each group of collected acceleration signal sequences is processed to 4000.

[0037] In steps 5 to 8, the blade profile error prediction model is based on the PyTorch framework, the learning rate of the training model is set to 0.0002, the number of model iterations is set to 100, the input sequence length is set to 500, and the Adam optimizer is used to update the network parameters through the batch gradient descent algorithm.

[0038] During the training process, the root mean square error RMSE and R 2 The coefficient is used as an evaluation indicator for model training.

[0039] The training results are as follows Figure 5 As shown: Figure 5 In the figure, the horizontal axis represents the true value, the vertical axis represents the predicted value, and a 45° line is drawn as a reference. Figure 5 It can be seen that the scattered points are densely distributed near the 45° line (that is, most of the scattered points are close to or fall on the 45° line), indicating that the error between the model's predicted value and the true value is small, and the model can better fit the mapping relationship in the data. The results are as follows: 2 The coefficient is 0.89, the RMSE is 0.017, and the prediction effect is relatively ideal, which shows the superiority of the present invention.

[0040] Any matters not described in the present invention are applicable to the prior art.

Claims

1. A blade milling profile error prediction method driven by spatiotemporal feature fusion, characterized in that: The method comprises the following steps: Step 1: Before machining, obtain the CNC machining program according to the designed process parameters and import it into the machine tool, and install the blade blank; Step 2: Install the accelerometer probe to measure the vibrations of the spindle and the blade during the machining process; adjust the direction of the accelerometer probe so that the accelerometer coordinate system is consistent with the space coordinate system; Step 3, start processing, and collect acceleration signals on the spindle and the blade during the processing; after the blade processing is completed, under the premise that the state of the processed blade remains unchanged, use the machine tool probe to measure the profile error of the blade on the machine, and use the profile error as label data; Step 4, performing the same preprocessing on the acceleration signals on the main shaft and the blades obtained in step 3 to obtain preprocessed acceleration signals; Step 5, generating graph data according to the preprocessed acceleration signal obtained in step 4 and the process parameters in step 1, and setting the number of model iterations; Step 6: Input the graph data obtained in step 5 into GAT, extract the spatial features of the graph data through the combination of self-attention mechanism and multi-head attention mechanism, and obtain graph data with spatial features; Step 7: Input the graph data with spatial features obtained in step 6 into ECA Net. When allocating attention weights in the time dimension, ECA Net can dynamically adjust the channel weights of the input data and effectively capture the time features to obtain data with spatiotemporal features. Step 8: Use CNN to further fuse and extract the data with spatiotemporal features obtained in step 7, and output the predicted value of contour error; Step 9: Compare the contour error prediction value obtained in step 8 with the label data obtained in step 3, and perform back propagation iterations according to the principle of minimizing the loss function; repeatedly adjust the internal parameters of GAT, ECANet and CNN according to the number of model iterations set in step 5, and the loss function value is determined by the square of the difference between the contour error prediction value and the label data; after stopping the iteration, obtain the contour error prediction value with the minimum loss function value within the number of model iterations.

2. The blade milling profile error prediction method driven by spatiotemporal feature fusion according to claim 1 is characterized in that: Step 1 is as follows: before machining, CNC programming is performed according to the designed process parameters to obtain the CNC machining program; the CNC machining program is then imported into the machine tool; the blade blank is clamped according to the requirements of the process card, the blade blank is positioned using the machine tool probe, and the C axis of the machine tool is adjusted to the correct position; Install the auxiliary clamping device to ensure that the axis of the blade blank is always perpendicular to the turntable of the machining center during the machining process.

3. The blade milling profile error prediction method driven by spatiotemporal feature fusion according to claim 1 or 2 is characterized in that: In step 1, the process parameters include spindle speed, feed rate, path spacing, cutting depth and tool rake angle.

4. The blade milling profile error prediction method driven by spatiotemporal feature fusion according to claim 1 is characterized in that: In step 2, an acceleration sensor probe is installed on the machine tool spindle near the milling cutter and the relative position is kept unchanged. The collected signal reflects the vibration generated by the spindle during the machining process; an acceleration sensor probe is installed on the auxiliary clamping device of the blade near the blade and the relative position is kept unchanged. The collected signal reflects the vibration generated by the blade during the machining process.

5. The blade milling profile error prediction method driven by spatiotemporal feature fusion according to claim 1 is characterized in that: In step 4, preprocessing includes: denoising, downsampling and removing outliers on the collected acceleration signal.

6. The blade milling profile error prediction method driven by spatiotemporal feature fusion according to claim 1 is characterized in that: In step 5, the process of generating graph data includes: reading and normalizing the preprocessed acceleration signal, reading the label data, dividing the preprocessed acceleration signal into samples, dividing each sample into multiple groups of data, generating a graph data for each group of data, and returning a graph data list.

7. The blade milling profile error prediction method driven by spatiotemporal feature fusion according to claim 6 is characterized in that: In step 5, the generated graph data is obtained by calling the Gen_graph function through the python software platform. The specific steps are as follows: S51, converting data into a graph structure: constructing the preprocessed acceleration signal obtained in step 4, the process parameters obtained in step 1, and the label data obtained in step 3 into a signal data matrix X, whose shape is N×F, where N is the number of nodes and F is the feature dimension of each node; S52, generate adjacency matrix: construct the adjacency matrix A of the graph through the RadiusGraph graph generation strategy; S53, construct graph data: construct graph data by using the signal data matrix X and the graph adjacency matrix A, expressed as .

8. The blade milling profile error prediction method driven by spatiotemporal feature fusion according to claim 1 is characterized in that: In step 6, the self-attention mechanism receives the feature matrix and adjacency matrix of each node through adaptive weight learning, calculates the importance of each neighbor node relative to the target node, dynamically adjusts the weight according to the relationship between node features, assigns different weights to different neighbors when aggregating neighbor node features, and then performs weighted summation of the information of neighbor nodes; in the prediction of blade contour error, the data corresponding to the factors that are relatively important to the final contour shape of the blade are assigned higher weights; Self-attention coefficient β obtained by adaptive weight learning ij As shown in formula (1): （1） In formula (1), LeakyReLU is the activation function; q is the attention weight vector; G is the weight matrix; fi is the feature vector of target node i, and f j is the feature vector of neighbor node j; N(i) is the set of neighbor nodes of target node i, including all nodes connected to target node i; k is the index used to traverse the neighbor node set N(i) of target node i; Represents a splicing operation; The multi-head attention mechanism is shown in formula (2): （2） In formula (2), is the final output feature of the target node i, D is the number of attention heads, σ is the activation function, is the attention coefficient of target node i to neighbor node j calculated by the dth attention head, G d is the weight matrix of the d-th attention head, f j is the feature vector of the neighbor node; j∈N(i) means that the neighbor node j is an element in the neighbor node set N(i) of the target node i.

9. The blade milling profile error prediction method driven by spatiotemporal feature fusion according to claim 1 is characterized in that: In step 7, ECA Net consists of global average pooling, one-dimensional convolution, sigmoid activation function and attention weight connected in sequence; Step 7 is as follows: first, perform global average pooling on the feature map of each input channel to obtain the global description vector of each channel; then, perform a one-dimensional convolution operation to realize the local interaction between channels, and then use the Sigmoid activation function to activate the result of the one-dimensional convolution; finally, use the attention weight to recalibrate the feature map of each channel to obtain the enhanced feature map; For the input feature map , W represents the width of the feature map, H represents the height of the feature map, C represents the number of channels, and a vector of size C is obtained by global average pooling. The value of the global average pooling value of the cth channel in the feature map is Z c As shown in formula (3): （3） In formula (3), M c (t,r) is the pixel value of the cth channel in the feature map M at the tth row and rth column; The expression of Sigmoid activation function is shown in formula (4): （4） In formula (4), x is the data input into the activation function in the previous step; The enhanced feature map is shown in formula (5): （5） In formula (5), a c is the attention weight of channel c; is the Hadamard product, U is the feature output by formula (4); is a weighted feature.

10. The blade milling profile error prediction method driven by spatiotemporal feature fusion according to claim 1 is characterized in that: In step 8, CNN is composed of multiple layers of convolution, one layer of maximum pooling and two layers of fully connected layers connected in sequence; multiple layers of convolution are composed of multiple layers of convolution layers; The data with spatiotemporal characteristics obtained in step 7 is first subjected to multi-layer convolution operations. After each convolution operation, the dimension of the data will change. After the convolution operation, the maximum pooling operation is used to reduce the data dimension while retaining important features. Finally, the contour error prediction value is output through the fully connected layer.

Images