Prediction Method for Burr at the Milling Exit of Turbine Blade Root Groove Based on Data Inference

Abstract

The present invention discloses a method for predicting the burr at the milling exit of a turbine blade root groove based on data reasoning. Currently, there are not many studies on burr length prediction, and the data used are all single signals or one-dimensional data, resulting in low prediction accuracy for burr length and large errors. The present invention evaluates the milling cutter state through the characteristics of the milling process voice signal and the vibration signal, and predicts the burr length based on the milling cutter state and process parameters; through the joint reasoning of the mechanism model and the data model, the robustness and accuracy of the prediction method are synchronously optimized; through the transfer learning of this prediction method, it can be applied to the burr prediction of the milling of other complex components.

Description

Technical Field

[0001] The present invention relates to a method for predicting burrs in the milling of blade root grooves, and particularly to a method for predicting outlet burrs in the milling of turbine blade root grooves based on data reasoning. Background Art

[0002] With the rapid development of mechanical manufacturing science and technology, cutting technology is developing towards precision and ultra-precision machining, automated machining and intelligent machining, and the requirements for the machining quality (including edge quality) of parts are also getting higher and higher. This poses a severe challenge to the understanding that the influence of burrs generated on the edges, corners and edges of workpieces (or components) in traditional machining can be ignored.

[0003] According to the position where burrs are formed in face milling, burrs are divided into inlet burrs and outlet burrs. Compared with outlet burrs, the influence of inlet burrs can be ignored. Among all burr parameters, burr length is the most important because it shows the maximum size of the burr and directly determines the deburring time of the deburring process on the production line and the wear amount of the brush for removing burrs. Currently, there are not many studies on burr length prediction, and the data used are all single signals or one-dimensional data, resulting in low prediction accuracy and large errors for burr length. Summary of the Invention

[0004] Aiming at the problem that in the milling process of turbine blade root grooves, due to the influence of non-linear factors such as tool and machine tool states, it is impossible to effectively predict the machining burrs formed at the milling outlet, which not only affects the subsequent deburring process but also affects the assembly performance of components, the present invention provides a method for predicting turbine blade root groove milling outlet burrs based on data reasoning. This method is a prediction method for evaluating the tool state through the characteristics of the milling process voice signal and vibration signal; it is a prediction method for burr length through the tool state and process parameters; it is a prediction method for burr length through the characteristics of the voice signal.

[0005] The present invention is achieved through the following solutions:

[0006] The present invention relates to a method for predicting turbine blade root groove milling outlet burrs based on data reasoning, which mainly includes the following steps:

[0007] Step 1: Use a sound sensor and a three-axis acceleration sensor to collect sound signals and vibration signals (signals at multiple moments), perform denoising processing on the obtained data, then use the discrete Fourier transform to obtain the short-time energy of the signals, and at the same time perform time-frequency analysis.

[0008] Step 2: Perform multi-layer maximum overlap discrete wavelet packet transform (MODWPT) on the time-frequency analyzed signal to obtain the corresponding decomposed frequency band energy characteristics, and construct an adaptive weight allocation filter bank using the obtained decomposed frequency band energy characteristics.

[0009] Step 3: Add the obtained short-time energy to the adaptive weight allocation filter bank to obtain the maximum overlap discrete wavelet packet transform cepstrum (MODWPTC).

[0010] Step 4: Use the maximum overlap discrete wavelet packet transform cepstrum data as input and input it into the established multi-feature convolutional neural network for processing to obtain a set of continuous tool wear rate data.

[0011] Step 5: Compare the error between the predicted tool wear rate and the true label of the tool wear rate, select the cross-entropy loss function as the damage function of the dual-feature fusion convolutional neural network (DFCNN), and further optimize the multi-feature convolutional neural network model through backpropagation.

[0012] Step 6: Convert the tool wear rate and the six types of data including the rake angle, spindle speed, cutting feed, cutting depth, and cutting speed that have a greater impact on burrs in the experiment into multi-dimensional data through feature crossing. Use the FM model for feature crossing.

[0013] Step 7: Input the multi-dimensional data into the multi-dimensional residual convolutional network to obtain the burr length.

[0014] Step 8: Compare the obtained results with the actually measured burr length (compare the data at each moment in the results with the measured burr length at the corresponding moment one by one, that is, traverse all moments of collecting sound signals and vibration signals). After comparison, update the multi-dimensional residual convolutional network structure through backpropagation, and finally obtain an optimal multi-dimensional residual convolutional network model. Here, the cross-entropy function is used as the loss function of the multi-dimensional residual convolutional network, the set learning rate is 0.05, and the number of iterations epoch is 100.

[0015] Preferably, the calculation formula for the maximum overlap discrete wavelet packet transform cepstrum of the j-th decomposition layer in Step 3 is as follows:

[0016]

[0017] where X(k) is the spectral line energy of the k-th spectral line, 0 ≤ k ≤ N, and M j (k) is the filter of the k-th spectral line in the j-th decomposition layer.

[0018] Preferably, step four is specifically as follows: The maximum overlap discrete wavelet packet transform cepstrum data is input into three different scale convolutional layers of a multi-feature convolutional neural network to obtain features of different scales, and the features of different scales are fused to obtain multi-level depth features; then the multi-level depth features are input into a convolutional module with a scale of 5*5*64 to extract abstract features, and finally the tool wear rate is output through a fully connected layer operation.

[0019] Preferably, the multi-level depth feature F M in the multi-feature convolutional neural network is obtained by the following formula:

[0020] F M = concate[ReLU(w1X + b1), ReLU(w2X + b2), ReLU(w3X + b3)] where: X represents a randomly selected maximum overlap discrete wavelet packet transform cepstrum data sample, w1, w2, w3 represent weight matrices of three different scales, b1, b2, b3 represent bias vectors of three different convolutional layers; ReLU(·) represents the rectified linear unit, which is a commonly used activation function in neural networks; concate(·) is used to splice multiple features together.

[0021] More preferably, the scales of the three different convolutional layers for extracting features of different scales are 3*3*16, 5*5*16, and 7*7*16 respectively.

[0022] Preferably, two residual structures are set in the multi-dimensional residual convolutional network, and two convolutional layers with inconsistent convolutional kernel sizes are set in each residual structure; the results of the two convolutional layers and the features are fused through the first residual structure, and then input into the second residual structure after passing through the Relu activation function. The second residual structure fuses the results of the two convolutional layers and the features, and then transmits them to the last convolutional layer after passing through the Relu activation function. After being processed by the ReLu activation function, it is input into the fully connected layer, and the predicted burr length value is mapped through the fully connected layer operation. Finally, it is output by the output layer.

[0023] More preferably, the predicted burr length value is mapped through the fully connected layer operation, and the specific calculation is as follows:

[0024] Y i = f(X i ) = f(x1, x2, x3, x4, x5, x6)

[0025] where the input matrix X i=(x1, x2, x3, x4, x5, x6), where x1, x2, x3, x4, x5, and x6 respectively represent the data corresponding to the wear rate, rake angle, spindle speed, cutting feed, cutting depth, and cutting speed output by the second residual structure, f(*) represents the operation of the fully connected layer; Yi represents the i-th data output by the multi-dimensional residual convolution network.

[0026] The beneficial effects of the present invention are as follows:

[0027] A method for predicting the burr at the exit of the root slot milling of a turbine blade based on data reasoning according to the present invention evaluates the milling cutter state through the characteristics of the voice signal and vibration signal during the milling process, and predicts the burr length based on the milling cutter state and process parameters, that is, the burr of the root slot milling can be effectively predicted through the real-time state of the machine tool and process parameters. The present invention synchronously optimizes the robustness and accuracy of the prediction method through the joint reasoning of the mechanism model and the data model; through the transfer learning of the present prediction method, it can be applied to the burr prediction of the milling of other complex components. Description of the Drawings

[0028] Figure 1 It is the prediction flow chart of the tool wear rate in the present invention.

[0029] Figure 2 It is the network structure diagram for predicting the burr length value in the present invention. Detailed Embodiments

[0030] The present invention will be further described below with reference to the drawings.

[0031] As Figure 1 shown, the method for predicting the burr at the exit of the root slot milling of a turbine blade based on data reasoning is as follows:

[0032] Step 1: Install a sound sensor and a three-axis acceleration sensor on the workbench and fixture of the numerical control machine tool, and use a face milling cutter to perform milling processing on the workpiece to obtain the corresponding sound signal and vibration signals in three directions.

[0033] Step 2: Denoise the collected sound signal and vibration signal, and then perform pre-emphasis, framing, and windowing preprocessing respectively. Use the discrete Fourier transform to obtain the short-time energy of the signal, and then perform time-domain analysis on the denoised sound signal and vibration signal.

[0034] Step 3: Perform multi-layer MODWPT processing on the sound signal and vibration signal to obtain the corresponding decomposed frequency band energy characteristics, and use the obtained frequency band energy characteristics to construct an adaptive weight distribution filter bank. The weight calculation formula is:

[0035]

[0036] Among them, when j = 1, set W j,n,t = 1; n is the frequency band number index value of the j-th decomposition layer, 1 ≤ j ≤ N, n = j - 1, t = j - 1, N is the sample length. When the value of n mod 4 is 0 or 3, assign a scaling filter to the l-th wave, and the coefficient r n,l takes the value of the scaling filter length L. When the value of n mod 4 is 1 or 2, assign a wavelet filter to the l-th wave, and the coefficient r n,l takes the value of the wavelet filter length of the l-th wave.

[0037] Step Four: Add the obtained short-time energy to the adaptive weight assignment filter bank to obtain the maximum overlap discrete wavelet packet transform cepstrum (MODWPTC). The calculation formula for the maximum overlap discrete wavelet packet transform cepstrum of the j-th decomposition layer is as follows:

[0038]

[0039] Among them, X(k) is the spectral line energy of the k-th spectral line, 0 ≤ k ≤ N, M j (k) is the filter of the k-th spectral line in the j-th decomposition layer, corresponding to the filter assigned to the l-th wave in the MODWPT processing.

[0040] Step Five: Design of multi-feature convolutional neural network: Design convolutional layers with different receptive field sizes to obtain richer feature representations. After each convolutional layer, use standard batch normalization and the RELU activation function. The convolutional layer modules with different sizes set here are convolutional layer modules of 3*3*16, 5*5*16, and 7*7*16 respectively. Calculate multi-level depth features through cascading, and the specific formula is as follows:

[0041] F M = concate [ReLU (w1X + b1), ReLU (w2X + b2), ReLU (w3X + b3)] (3) Where: X represents a randomly selected maximum overlap discrete wavelet packet transform cepstrum data sample, w1, w2, w3 represent weight matrices of 3 different scales, b1, b2, b3 represent bias vectors of 3 different convolutional layers; RELU(·) represents the rectified linear unit, which is a commonly used activation function in neural networks; concate(·) is used to splice multiple features together.

[0042] Step Six: Input the MODWPTC data obtained in Step Four into different convolutional layers designed by the multi-feature convolutional neural network to obtain features of different scales, and fuse the features of different scales to obtain multi-level depth features.

[0043] Step 7: Input the multi-level depth features into the designed convolutional module with a scale of 5*5*64 to extract abstract features, and finally output the tool wear rate through the operation of the fully connected layer.

[0044] Step 8: Compare the error between the predicted tool wear rate and the true label of the tool wear rate, select the cross-entropy loss function as the damage function of the DFCNN (Dual Feature Fusion Convolutional Neural Network), and further optimize the multi-feature convolutional neural network model through backpropagation.

[0045] Step 9: As Figure 2 shown, introduce a multi-dimensional residual convolutional network, set two residual structures, and set two convolutional layers with inconsistent convolutional kernel sizes in each residual structure. The activation function is Relu.

[0046] Step 10: Convert the wear rate output in Step 8 and the data of the rake angle, spindle speed, cutting feed, cutting depth, and cutting speed in the experiment into multi-dimensional data through feature crossing to provide richer network features for the multi-dimensional residual convolutional network. Here, the process parameters that have a greater impact on the burr are obtained according to the mechanism model of the relationship between the burr length and the process parameters as the network input.

[0047] Step 11: Take the converted multi-dimensional data as the input and input it into the multi-dimensional residual convolutional network. Specifically, fuse the results of the two convolutional layers with the features through the first residual structure, then input it into the second residual structure after passing through the Relu activation function. The second residual structure fuses the results of the two convolutional layers with the features, and then transmits it to the last convolutional layer after passing through the Relu activation function. After being processed by the ReLu activation function, it is input into the fully connected layer (in this embodiment, the fully connected layer Fc1 and the fully connected layer Fc2 are set). After the operation of the fully connected layer, the predicted burr length value is mapped, and finally it is output by the output layer. Among them, the first convolutional layer of the first residual structure needs to go through a pooling layer pool-1 operation, and the third convolutional layer of the second residual structure needs to go through a pooling layer pool-2 operation; Equations (4), (5), and (6) are the defined models of the multi-dimensional residual convolutional network:

[0048] y l (j) = F k [x l-1 (j), w l-1 , b l-1 (4)

[0049]

[0050]

[0051] where x l-1(j) represents the input of the j-th neuron in the (l-1)-th layer (where the layer number represents the depth of the convolutional layer), represents the input of the j-th neuron in the i-th data of the (l-1)-th layer, F k represents a convolution with a convolution kernel size of k, y l (j) represents the input of the j-th neuron in the l-th layer, w l-1 and b l-1 represent the weight and bias of the (l-1)-th layer respectively, f is the activation function, represents the weight of the i-th data in the (l-1)-th layer, q l-1 (j) represents the output of the j-th neuron in the (l-1)-th layer, w is the width of the pooling layer, P l (j) is the output of the j-th neuron in the l-th layer.

[0052] The fully connected layer operation maps to obtain the predicted burr length value. The specific calculation formula is shown in Equation (7):

[0053] Y i = f(X i ) = f(x1, x2, x3, x4, x5, x6) (7)

[0054] Among them, the input matrix X i = (x1, x2, x2, x4, x5, x6), x1, x2, x3, x4, x5, x6 respectively represent the data corresponding to the wear rate, rake angle, spindle speed, cutting feed, cutting depth, and cutting speed output by the second residual structure, f(*) represents the operation of the fully connected layer; Yi represents the i-th data output by the multi-dimensional residual convolution network.

[0055] Step Twelve: Compare the predicted burr length with the actually measured burr length. After comparing the error, update the structure of the multi-dimensional residual convolution network through the backpropagation network, and continuously optimize the prediction model of the multi-dimensional residual convolution network. When training, the loss function of the multi-dimensional residual convolution network uses the cross-entropy loss function, the learning rate is 0.05, and the number of iterations epoch is 100.

[0056] The above specific embodiments can be locally adjusted by those skilled in the art in different ways without departing from the principles and purposes of the present invention. The protection scope of the present invention is subject to the claims and is not limited by the above specific embodiments. All implementation schemes within its scope are subject to the constraints of the present invention.

Claims

1. A method for predicting the burr at the milling exit of the root slot of a turbine blade based on data reasoning, characterized in that: Step 1: Use a sound sensor and a three-axis acceleration sensor to collect the sound signal and vibration signal during the milling of the root slot of the turbine blade. Denoise the obtained data, and then use the discrete Fourier transform to obtain the short-time energy of the signal and perform time-frequency analysis simultaneously; Step 2: Perform multi-layer maximum overlap discrete wavelet packet transform on the signal after time-frequency analysis to obtain the corresponding decomposed frequency band energy characteristics, and construct an adaptive weight allocation filter bank using the obtained decomposed frequency band energy characteristics; Step 3: Add the obtained short-time energy to the adaptive weight allocation filter bank to obtain the maximum overlap discrete wavelet packet transform cepstrum; Step 4: Use the maximum overlap discrete wavelet packet transform cepstrum data as input and input it into the established multi-feature convolutional neural network for processing to obtain a set of continuous tool wear rate data; Step 5: Compare the error between the predicted tool wear rate and the true label of the tool wear rate, select the cross-entropy loss function as the damage function of the dual-feature fusion convolutional neural network, and further optimize the multi-feature convolutional neural network model through backpropagation; Step 6: Convert the tool wear rate, rake angle, spindle speed, cutting feed, cutting depth, and cutting speed data into multi-dimensional data through feature crossing; Step 7: Input the multi-dimensional data into a multi-dimensional residual convolutional network to obtain the burr length; Step 8: Compare the obtained result with the actually measured burr length, and then update the multi-dimensional residual convolutional network structure through backpropagation. Finally, obtain an optimal multi-dimensional residual convolutional network model; Two residual structures are set in the multi-dimensional residual convolutional network, and two convolutional layers with different kernel sizes are set in each residual structure; the results of the two convolutional layers are fused with the features through the first residual structure, and then input into the second residual structure after passing through the Relu activation function. The second residual structure fuses the results of the two convolutional layers with the features, and then transmits them to the last convolutional layer after passing through the Relu activation function. After passing through the ReLu activation function, it is input into the fully connected layer. After the fully connected layer operation, the predicted burr length value is mapped, and finally output by the output layer; The fully connected layer operation maps to obtain the predicted burr length value, and the specific calculation is as follows: Y i = f(X i ) = f(x 1 , x 2 , x 3 , x 4 , x 5 , x 6 ) Among them, the input matrix X i =(x1, x2, x3, x4, x5, x6), where x1, x2, x3, x4, x5, and x6 respectively represent the data corresponding to the wear rate, rake angle, spindle speed, cutting feed, cutting depth, and cutting speed output by the second residual structure, f(*) represents the operation of the fully connected layer; Yi represents the i-th data output by the multi-dimensional residual convolution network.

2. The method for predicting the burr at the milling exit of the root slot of a turbine blade based on data reasoning according to claim 1, characterized in that: The calculation formula for the maximum overlap discrete wavelet packet transform cepstrum of the j-th decomposition layer in Step 3 is as follows: Among them, X(k) is the spectral line energy of the k-th spectral line, where 0 ≤ k ≤ N, and M j (k) is the filter of the k-th spectral line in the j-th decomposition layer.

3. The method for predicting the burr at the milling exit of the root slot of a turbine blade based on data reasoning according to claim 1, characterized in that: Step 4 is specifically as follows: The maximum overlap discrete wavelet packet transform cepstrum data is input into three different scale convolutional layers of the multi-feature convolutional neural network to obtain different scale features, and the different scale features are fused to obtain multi-level depth features; then the multi-level depth features are input into the convolutional module to extract abstract features, and finally the tool wear rate is output after the fully connected layer operation.

4. The method for predicting the burr at the milling exit of the root slot of a turbine blade based on data reasoning according to claim 3, characterized in that: Obtaining formula for multi-level deep feature F in multi-feature convolutional neural network M is as follows: F M = concate[ReLU(w1X + b1), ReLU(w2X + b2), ReLU(w3X + b3)] where: X represents a randomly selected maximum overlap discrete wavelet packet transform cepstrum data sample, w1, w2, w3 represent weight matrices of three different scales, b1, b2, b3 represent bias vectors of three different convolutional layers; RELU(·) represents the rectified linear unit; concate(·) is used to concatenate multiple features together.

5. The method for predicting the burr at the milling exit of the root slot of a turbine blade based on data reasoning according to claim 4, characterized in that: The three different convolutional layer scales for extracting different scale features are 3*3*16, 5*5*16, and 7*7*16 respectively.

Images