Geometric feature-based aero-engine blade milling location degree prediction method

Abstract

The method for predicting the milling position degree of the aero-engine blade based on the geometrical characteristics comprises the steps that key machining parameters are determined, and current data in the machining process are collected; importing the blade model into CAD software, parameterizing the curved surface of the blade, establishing a UV coordinate system, uniformly arranging detection points, detecting the location degree of each detection point, and calculating the curvature of each detection point along a V-direction streamline through the CAD software; calculating the residual height h of each detection point, and forming a data set by the current data corresponding to different key processing parameters, the curvature of the detection point along the V-direction streamline, the residual height of the detection point and the location degree of the corresponding detection point; constructing a neural network model; and training a neural network model by using the data set to predict the milling position degree of the aero-engine blade. According to the method, the complex nonlinear relation can be learned from the processing data, and the generalization ability and the prediction precision of the model can be improved through the geometric features.

Description

Technical Field

[0001] The invention belongs to the field of intelligent manufacturing and relates to a method for predicting the milling position of an aero-engine blade based on geometric features. Background Art

[0002] As a key component of an aero-engine, the machining accuracy of aero-engine blades is crucial to the overall performance of the engine. Due to the complex structure, diverse shapes, and stringent dimensional requirements of blades, the position deviation during machining directly affects the aerodynamic performance, combustion efficiency, and service life of the engine. Traditional methods for predicting blade machining position deviations rely on a large number of experiments or empirical rules, which are difficult to accurately respond to various factors in the complex surface machining process, such as the shape of the workpiece, tool wear, and the dynamic characteristics of the machine tool.

[0003] With the development of intelligent manufacturing technology, research based on data-driven methods has become an effective means to solve the problem of blade machining position prediction. By utilizing a large amount of historical machining data, the neural network prediction model can automatically learn and capture the hidden laws in the machining process, and achieve accurate prediction of position deviation. However, purely data-driven models often require a large amount of historical data, and the prediction accuracy is limited when faced with small sample data; and only using machining data ignores the influence of the bending and torsional characteristics of the workpiece on the position deviation, further limiting the prediction accuracy of the model. To this end, this patent proposes a method for predicting the position of aircraft engine blade milling based on geometric features.

[0004] By incorporating geometric features into intelligent algorithms, not only the size of the data set is expanded, but the network model can also learn the bending and torsional characteristics of the blades, thereby improving the overall efficiency and quality of aero-engine manufacturing. Summary of the invention

[0005] In view of the shortcomings of the prior art, the purpose of the present invention is to provide a method for predicting the position of aero-engine blade milling based on geometric features. The method fully considers the bending and torsion characteristics of aero-engine blades, calculates the curvature at each point, calculates the residual height based on the curvature and processing parameters and integrates it into the neural network as domain knowledge, expands the scale of the data set, and improves the prediction accuracy of the network.

[0006] To achieve the above object, the technical solution of the present invention is: A method for predicting the position of aero-engine blade milling based on geometric features, the prediction method comprising the following contents: Determine four key processing parameters, namely: spindle speed, feed speed, cutting depth, and path spacing. Each key processing parameter is changed at different levels to form an orthogonal experimental table. Processing is carried out according to the orthogonal experimental table, and current data during the processing is collected; The blade model is imported into the CAD software, the blade surface is parameterized, and a UV coordinate system is established, in which the V direction is developed along the long axis direction of the blade, and the U direction is developed along the chord direction of the blade. The detection points are evenly arranged, and the position at each detection point is detected; the V-direction streamline at each detection point is generated by the surface feature function in the CAD software, and the curvature of the V-direction streamline at each detection point is calculated by the CAD software; The residual height h at each detection point is calculated according to the following formula:

[0007] Wherein, h is the residual height at the detection point, ρ is the curvature of the V-direction streamline at the detection point, R is the radius of the ball-end milling cutter for milling, and I is the path spacing; The current data corresponding to different key processing parameters, the curvature of the V-direction streamline at the detection point, the residual height of the detection point and the position of the corresponding detection point are combined into a data set; Build a neural network model: The neural network model includes a CNN-LSTM network and a DNN network. The current data is used as the input of the CNN-LSTM network, and the residual height is used as the input of the DNN network. The output of the CNN-LSTM network is spliced ​​and fused with the key processing parameters and the curvature along the V-direction streamline, and then processed and predicted by the first fully connected layer to obtain the first feature result; the residual height is used as the domain knowledge and the second feature result is obtained after the DNN network is upgraded in dimension. The first feature result and the second feature result are spliced ​​and then processed and predicted by the second fully connected layer to obtain the final position; The neural network model is trained using the data set to obtain a trained neural network model for position prediction in aero-engine blade milling.

[0008] Furthermore, the CNN-LSTM network is composed of a CNN and an LSTM connected in series, the current data is input into the CNN to obtain a low-dimensional local feature representation, and then the time-dependent feature n1 in the sequence is extracted through the LSTM in the time dimension; Then, the stitching layer is used to merge n1 with the key processing parameters and the curvature along the V-direction streamline to obtain a new feature n2; Pass n2 through a fully connected layer with 10 neurons for preliminary prediction, and get the first feature result n3; After the residual height is subjected to nonlinear transformation through a DNN network with three hidden layers, the second feature result n4 is obtained; the first feature result n3 and the second feature result n4 are concatenated through a Concatnate layer to obtain n5; n5 is calculated through a fully connected layer with 10 neurons to obtain the final result.

[0009] Furthermore, the numbers of neurons in the three hidden layers are 100, 100, and 10 respectively.

[0010] Furthermore, the data set is divided into a training set and the root mean square error RMSE is used as the loss function when training the neural network model.

[0011] Among them, RMSE is the root mean square error, y predi is the position value of the i-th detection point predicted by the neural network model, y truei is the actual position value of the i-th detection point, and n is the number of detection points; Draw the loss curve during training and observe the changing trend of the loss value to determine whether the model has converged and determine the appropriate number of training rounds.

[0012] The present invention also protects a computer-readable storage medium having a computer program stored thereon, which can implement the steps of the method when executed by a processor.

[0013] Compared with the prior art, the present invention has the following beneficial effects: By combining geometric features with neural networks, the network model of the present invention can not only learn complex nonlinear relationships from processing data, but also improve the generalization ability and prediction accuracy of the model through geometric features. The residual height is processed by the deep learning network DNN and effectively integrated with the relevant features of the extracted current data to generate a position evaluation model with high prediction accuracy. This model shows strong adaptability and robustness under various processing conditions.

[0014] The present invention fully considers the differences between different points, calculates the curvature along the V-direction streamline at each point, calculates the residual height based on the curvature and key processing parameters as prior knowledge and integrates it into the neural network, expands the scale of the data set, and improves the prediction accuracy of the network.

[0015] The method of the present invention not only improves the predictive reliability of blade processing accuracy, but also enhances the accuracy control capability under complex working conditions to adapt to complex working conditions or small sample scenarios, and can provide strong support and guarantee for the high-precision manufacturing of aircraft engine blades. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 It is a structural schematic diagram of the neural network model in the method for predicting the position of aero-engine blade milling based on geometric features of the present invention.

[0017] Figure 2 This is a diagram showing the distribution of detection points and the establishment of a coordinate system in the present invention.

[0018] Figure 3 This is a diagram of the training loss convergence process in the embodiment.

[0019] Figure 4 This is a comparison chart of the prediction results accuracy.

[0020] Figure 5 To evaluate the overall effect of the prediction results. DETAILED DESCRIPTION

[0021] In order to make the technical solution of the present invention clearer, the present invention is further described below in conjunction with the accompanying drawings.

[0022] The method for predicting the position accuracy of aero-engine blade milling based on geometric features of the present invention is specifically implemented according to the following steps: Step 1: Design the experiment. Determine the four key machining parameters: spindle speed, feed rate, cutting depth, and path spacing. Each key machining parameter is changed at different levels to form an orthogonal experimental table. The orthogonal experimental method is used to evaluate the influence of each key machining parameter on the five-axis milling machining accuracy of aircraft engine blades.

[0023] a. Determine four key machining parameters: spindle speed, feed rate, cutting depth, and path spacing.

[0024] b. Set different levels for each selected key processing parameter to fully cover the possible working range. According to the principle of orthogonal experimental design, construct an orthogonal experimental matrix and ensure that all key processing parameter combinations are fully and reasonably investigated.

[0025] Step 2: Data collection. Perform processing according to the orthogonal experimental table in the first step; collect current data during processing; arrange detection points and detect the position at each point; calculate the curvature along the V-direction streamline at each detection point and calculate the residual height at the detection point.

[0026] (1) Carry out processing one by one according to the designed experimental combination to ensure strict control of each processing parameter.

[0027] (2) Collect current data.

[0028] a. Determine the sampling frequency of current data to ensure the timeliness and accuracy of the data. The sampling frequency should not be less than five times per second. The current data of each blade during the processing should be collected based on the CNC system of the machine tool.

[0029] b. Define the collection time periods during each blade processing, including the beginning, middle and end of processing.

[0030] (3) Collecting position data. Use 8 U-direction streamlines and 16 V-direction streamlines to evenly divide the blade surface, where the U-direction streamline is the tool feed direction. The intersection of the U-direction streamline and the V-direction streamline is the position detection point. For each blade, at least 64 detection points are evenly selected on the blade basin (concave surface) and the blade back (convex surface) to obtain the actual processing results of the blade. In the embodiment, 64 detection points are distributed on the blade basin and the blade back of each group of blades, and a total of 128 points are detected.

[0031] (4) Import the blade model. Import the blade model into the CAD software. Common formats include STEP, IGES, STL, etc. to ensure that the model can be used in the CAD software.

[0032] (5) Determine the UV coordinate system. The blade surface is usually a complex three-dimensional surface. In order to calculate the streamlines and curvature of the surface, the surface is parameterized and a UV coordinate system is established. The V direction is developed along the long axis of the blade, and the U direction is developed along the chord direction of the blade. The established UV coordinate system is as follows: Figure 2 As shown in the figure, the intersection of U direction and V direction is the detection point of position.

[0033] The actual machining results of the blade are compared with its corresponding 3D CAD model to obtain the position of these inspection points under the same key machining parameters.

[0034] (6) Generate V-direction streamlines. Generate V-direction streamlines along each inspection point using the surface feature function in the CAD software.

[0035] (7) The curvature of each detection point along the V-direction streamline is calculated by CAD software. In this embodiment, nine blades with a total of 1152 detection points are constructed. The curvature matrix of the 1152 detection points is expressed as [ρ1, ρ2, …, ρ 1152 ].

[0036] (8) Calculation of residual height.

[0037] The residual height h at each detection point is calculated according to the following formula: the residual height matrix [h1,h2,…h 1152 ].

[0038]

[0039] Wherein, h is the residual height at the detection point, ρ is the curvature of the V-direction streamline at the detection point, R is the radius of the ball-end milling cutter for milling, and I is the path spacing; The current data corresponding to different key processing parameters, the curvature of the detection point along the V-direction streamline, the residual height of the detection point and the position of the corresponding detection point are combined into a data set; the key processing parameters, current data, curvature along the V-direction streamline, residual height and position corresponding to each detection point constitute a sample, and the key processing parameters and current data of the same blade are the same.

[0040] The third step is to build a neural network model, including a CNN-LSTM network and a DNN network. The current data is used as the input of the CNN-LSTM network, and the residual height calculated according to the curvature along the V-direction streamline is used as the input of the DNN network. The output of the CNN-LSTM network is spliced ​​and fused with the key processing parameters and the curvature along the V-direction streamline, and then processed and predicted by the first fully connected layer (at least one layer) to obtain the first feature result; the residual height is used as domain knowledge and is upgraded through the DNN network to obtain the second feature result. The first feature result and the second feature result are spliced ​​and then processed and predicted by a second fully connected layer to obtain the final position.

[0041] The number of hidden layers of the DNN network in the present invention is at least two, preferably 3-5.

[0042] The present invention takes into account the problem of small amount of data in aero-engine blade milling processing, calculates the residual height at each detection point as domain knowledge and embeds it into a neural network, thereby enhancing the interpretability of the neural network and improving the prediction accuracy of the neural network.

[0043] The specific data processing process of the neural network model is: (1) Feature extraction module: CNN-LSTM is used to extract features from current data.

[0044] a. The input current data first passes through the convolutional neural network (CNN) to obtain a low-dimensional local feature representation.

[0045] b. Use the output of CNN as the input of LSTM, and extract the time-dependent feature n1 in the sequence through LSTM in the time dimension.

[0046] (2) Feature fusion. The key processing parameters and the blade morphology have an important influence on the processing quality. Therefore, the features extracted from the current data are fused with the key processing parameters and the curvature along the V-direction streamline using a stitching layer, and a preliminary prediction is made.

[0047] a. Splice the key machining parameters (spindle speed, feed rate, cutting depth, path spacing) and the curvature along the V-direction streamline into n1 to obtain a new feature n2.

[0048] b. The concatenated new feature n2 is preliminarily predicted through a fully connected layer with 10 neurons to obtain the preliminary prediction result n3 (3) Embedding domain knowledge. The nonlinear transformation of the residual height is realized by using a deep neural network. The features after nonlinear transformation are combined with the features to make them linearly differentiable in the high-dimensional output space, which greatly improves the ability to handle complex nonlinear problems. The residual height after nonlinear transformation is spliced ​​with the preliminary prediction results to prepare for the final prediction.

[0049] a. The residual height in the first step is transformed nonlinearly through a DNN network with three hidden layers to obtain feature n4. The number of neurons in the three hidden layers is 100, 100, and 10 respectively.

[0050] b. Concatenate the initial prediction result n3 and the residual height result n4 after nonlinear transformation through the Concatnate layer to obtain n5 (4) Calculate the final result. Pass n5 through a fully connected layer with 10 neurons to obtain the final result.

[0051] The fourth step is to divide the data set into training set and test set according to a certain ratio, and train the neural network model in the third step.

[0052] a. Define the loss function: For the position prediction task, the root mean square error (RMSE) is selected as the loss function, which is the square root of the difference between the model prediction value and the actual value.

[0053]

[0054] The loss function is used to measure the difference between the model's predictions and the actual results.

[0055] RMSE is the root mean square error, y predi is the position value of the i-th detection point predicted by the neural network model, y truei is the actual position value of the i-th detection point, n is the number of detection points, i.e. the number of samples; b. Select an optimization algorithm: Select a suitable optimization algorithm to update the parameters of the model. Common optimization algorithms include stochastic gradient descent (SGD), Adam, Adagrad, etc. Select a suitable optimization algorithm based on the actual situation and data characteristics. This embodiment uses the Adam optimization algorithm to update the parameters of the model.

[0056] c. Divide the dataset: Divide the dataset into training set, validation set and test set. The training set is used to update the model parameters, the validation set is used to adjust the hyperparameters, and the test set is used to evaluate the model performance.

[0057] d. Adjust hyperparameters: Tune the model's hyperparameters, including learning rate, batch size, number of hidden layer nodes, etc. You can use grid search, random search, or Bayesian optimization to search for the best hyperparameter combination.

[0058] e. Training model: Use the training set to train the model and monitor the change in loss value on the validation set. Adjust the hyperparameters based on the loss value on the validation set until the best hyperparameter combination is found.

[0059] f. Evaluate the model: Evaluate the performance of the model on the test set, and calculate at least one of the MSE and MAE values ​​of the position prediction as the evaluation criteria. Judge the accuracy of the model based on the evaluation results, and make necessary adjustments and improvements.

[0060] g. Visualize the loss curve: Draw the loss curve of the model during training and observe the changing trend of the loss value to determine whether the model has converged and determine the appropriate number of training rounds.

[0061] Example: The steps of the method for predicting the position accuracy of aero-engine blade milling based on geometric features in this embodiment are as follows: Step 1: Design the experiment. Determine the four key machining parameters: spindle speed, feed rate, cutting depth, and path spacing. Each key machining parameter is changed at different levels to form an orthogonal experimental table. The orthogonal experimental method is used to evaluate the influence of each key machining parameter on the five-axis milling machining accuracy of aircraft engine blades.

[0062] a. Determine four key machining parameters: spindle speed, feed rate, cutting depth and path spacing.

[0063] b. Set different levels for each selected key machining parameter to fully cover the possible working range. The spindle speed range is [13000-15000] rpm, the feed speed range is [1000-2000] mmpm, the cutting depth range is [0.12-0.14] mm, and the path spacing range is [0.07-0.09]. Design a four-factor three-level orthogonal experiment. The orthogonal experiment table is shown in Table 1.

[0064]

[0065] Step 2: Data collection.

[0066] (1) Carry out processing one by one according to the experimental combinations designed according to the orthogonal experimental table to ensure strict control of each processing parameter.

[0067] (2) Collect current data.

[0068] a. Determine the sampling frequency of current data to ensure the timeliness and accuracy of the data. Sampling is performed five times per second. The current data of each blade during the processing is collected based on JDMon_V1.5.

[0069] b. Define the collection time periods during each blade processing, including the beginning, middle and end of processing.

[0070] (3) Collecting position data. Use 8 U-direction streamlines and 16 V-direction streamlines to evenly divide the blade surface. The U-direction streamline is the tool feed direction. The intersection of the U-direction streamline and the V-direction streamline is the position detection point. Each group of blades has 64 detection points distributed on the blade base and the blade back, for a total of 128 points.

[0071] (4) The curvature of the V-direction streamline at each test point is calculated using CAD software; the residual height at the 128 test points is calculated using the following formula: ; The third step is to build a neural network model. First, the features extracted by CNN-LSTM are integrated with key processing parameters and the curvature along the V-direction streamline. After several layers of fully connected layers, a feature result is predicted. Then, the residual height is used as domain knowledge and the second feature result is obtained after the deep neural network DNN is upgraded. The two feature results are spliced ​​to predict the final position.

[0072] (1) Feature extraction module: CNN-LSTM is used to extract features from current data.

[0073] a. The input current data is first passed through the convolutional neural network to obtain a low-dimensional local feature representation. CNN is constructed using Tensorflow. According to experiments, the best effect is achieved when filters=1, kernel_size=4, and pool_size=3.

[0074] b. Use the output of CNN as the input of LSTM, and extract the time-dependent feature n1 in the sequence through the LSTM network in the time dimension. Use Tensorflow to build LSTM, and when lstm_units=4, it has the best effect.

[0075] (2) Feature fusion. Considering only the current data ignores the impact of key processing parameters and blade morphology on processing quality. Therefore, the features extracted from the current data are fused with the key processing parameters and the curvature along the V-direction streamline using a stitching layer, and a preliminary prediction is made.

[0076] a. Splice the vector P containing key machining parameters (spindle speed, feed rate, cutting depth, path spacing) and the curvature along the V-direction streamline into n1 to obtain a new feature n2. Use the concatenate function in the Keras functional API to merge P and n1, and set axis=1 to splice the two vectors along the first dimension.

[0077]

[0078] b. The concatenated new feature n2 is preliminarily predicted through a fully connected layer with 10 neurons to obtain the preliminary prediction result n3 (3) Embedding domain knowledge. The nonlinear transformation of the residual height is realized by using a deep neural network. The features after nonlinear transformation are combined with the features to make them linearly differentiable in the high-dimensional output space, which greatly improves the ability to handle complex nonlinear problems. The residual height after nonlinear transformation is spliced ​​with the preliminary prediction results to prepare for the final prediction.

[0079] a. The residual height in the first step is subjected to a nonlinear transformation operation through a deep neural network DNN with three hidden layers. The three hidden layers contain 100, 100, and 10 neurons respectively, and feature n4 is obtained through DNN.

[0080] b. Concatenate n3 and n4 through the Concatnate layer to get n5

[0081] (4) Calculate the final result. Pass n5 through a fully connected layer with 10 neurons to obtain the final result.

[0082] The fourth step is position prediction.

[0083] The 1152 samples were divided into training set and test set according to the ratio of 9:1 to train the neural network model. The following two evaluation indicators were used to evaluate the predictive effectiveness of the model: MSE is the mean square error and MAE is the mean absolute error.

[0084]

[0085]

[0086] from Figure 4 It can be seen that the predicted value and the true value are very close. In most cases, the two curves almost overlap, which shows that the prediction results of the model are very consistent with the actual situation, indicating that the model can accurately capture the characteristics and laws of the data; and at most data points, the error between the predicted value and the true value is very small, indicating that the model has high stability and reliability. Figure 3The different lines in the figure represent the changes in training loss on the training set and validation set. It can be seen that the training loss is decreasing with the increase in batches and is basically stable after 60 times, indicating that the neural network model of the present invention can converge quickly, verifying the effectiveness of the model design. Figure 5 It can be seen that the MSE of the model of the present invention is 0.0016 and the MAE is 0.025. In comparison, the MSE of the CNN-LSTM network that only considers current data and processing parameters is 0.003 and the MAE is 0.0038, which proves that the model of the present invention has high accuracy and can achieve accurate estimation of the accuracy of different inspection points, which is conducive to timely adjustment of areas with large local deviations in the later stage.

[0087] Any matters not described in the present invention are applicable to the prior art.

Claims

1. A method for predicting the position of aero-engine blade milling based on geometric features, characterized in that: The prediction method Includes the following: Determine four key processing parameters, namely: spindle speed, feed speed, cutting depth, and path spacing. Each key processing parameter is changed at different levels to form an orthogonal experimental table. Processing is carried out according to the orthogonal experimental table, and current data during the processing is collected; The blade model is imported into the CAD software, the blade surface is parameterized, and a UV coordinate system is established, in which the V direction is developed along the long axis direction of the blade, and the U direction is developed along the chord direction of the blade. The detection points are evenly arranged, and the position at each detection point is detected; the V-direction streamline at each detection point is generated by the surface feature function in the CAD software, and the curvature of the V-direction streamline at each detection point is calculated by the CAD software; The residual height h at each detection point is calculated according to the following formula: Wherein, h is the residual height at the detection point, ρ is the curvature of the V-direction streamline at the detection point, R is the radius of the ball-end milling cutter for milling, and I is the path spacing; The current data corresponding to different key processing parameters, the curvature of the V-direction streamline at the detection point, the residual height of the detection point and the position of the corresponding detection point are combined into a data set; Build a neural network model: The neural network model includes a CNN-LSTM network and a DNN network. The current data is used as the input of the CNN-LSTM network, and the residual height is used as the input of the DNN network. The output of the CNN-LSTM network is spliced ​​and fused with the key processing parameters and the curvature along the V-direction streamline, and then processed and predicted by the first fully connected layer to obtain the first feature result; the residual height is used as the domain knowledge and the second feature result is obtained after the DNN network is upgraded in dimension. The first feature result and the second feature result are spliced ​​and then processed and predicted by the second fully connected layer to obtain the final position; The neural network model is trained using the data set to obtain a trained neural network model for position prediction in aero-engine blade milling.

2. The method according to claim 1, characterized in that: The CNN-LSTM network is composed of CNN and LSTM in series. The current data is input into CNN to obtain a low-dimensional local feature representation, and then the time-dependent feature n1 in the sequence is extracted through LSTM in the time dimension; Then, the stitching layer is used to merge n1 with the key processing parameters and the curvature along the V-direction streamline to obtain a new feature n2; Pass n2 through a fully connected layer with 10 neurons for preliminary prediction, and get the first feature result n3; After the residual height is subjected to nonlinear transformation through a DNN network with three hidden layers, the second feature result n4 is obtained; the first feature result n3 and the second feature result n4 are concatenated through a Concatnate layer to obtain n5; n5 is calculated through a fully connected layer with 10 neurons to obtain the final result.

3. The method according to claim 2, characterized in that The number of neurons in the three hidden layers are 100, 100, and 10 respectively.

4. The method according to claim 1, characterized in that: When dividing the data set into a training set and training the neural network model, the root mean square error RMSE is used as the loss function. Among them, RMSE is the root mean square error, y predi is the position value of the i-th detection point predicted by the neural network model, y truei is the actual position value of the i-th detection point, and n is the number of detection points; Draw the loss curve during training and observe the changing trend of the loss value to determine whether the model has converged and determine the appropriate number of training rounds.

5. The method according to claim 1, characterized in that The prediction accuracy of the neural network model is: MSE is 0.0016, MAE is 0.

025.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of any method described in claims 1-5 can be implemented.

Images