Aero-engine blade location degree prediction method based on multi-source data deep fusion

Abstract

The invention relates to an aero-engine blade location degree prediction method based on multi-source data depth fusion. The method comprises the following steps of collecting five process parameters including the spindle rotating speed, the feeding speed, the cutting depth, the path spacing and the cutter roll angle and acceleration signal data of a clamp and a spindle in the X direction, the Y direction and the Z direction as multi-source input data; obtaining the location degree on the blade profile, and taking the maximum value as a label; missing value filling, mean value down-sampling, directional elimination and normalization are carried out on the acceleration signal data, and a multi-source data set is obtained after preprocessing; constructing a deep learning framework RLTN; and training the RLTN by using the multi-source data set. According to the RLTN, technological parameters are introduced in an information mode and a condition mode in the process of deeply fusing clamp and main shaft acceleration signals, efficient prediction of the position degree of the aero-engine blade is achieved, and powerful support is provided for aero-engine blade milling quality prediction.

Description

Technical Field

[0001] The invention relates to the field of intelligent monitoring of aircraft engine blade milling, and in particular to an aircraft engine blade position prediction method that integrates multi-source information. Background Art

[0002] Aerospace engine blades are one of the key components of the engine, and their manufacturing accuracy directly affects the overall performance and reliability of the engine. In recent years, with the rapid development of sensor technology and data processing technology, a large amount of sensor data can be used to monitor and predict the machining position of blades. By summarizing and analyzing these data, the position of aerospace engine blades at different positions can be accurately obtained, which is of great significance for improving the production efficiency of aerospace engine blades.

[0003] At present, the prediction models of aero-engine blade position can be mainly divided into three categories: statistical methods, machine learning methods and deep learning methods. In contrast, deep learning models can directly accept raw data input and build an "end-to-end" prediction model, which greatly reduces the complexity of feature engineering. On large-scale data sets, deep learning models show higher accuracy and robustness. For example, Ya Kunjiang et al. proposed a new method combining deep learning and reinforcement learning to achieve effective prediction and compensation of contour errors. Using the internal data of the CNC system, the tracking error of the feed axis was modeled as a nonlinear autoregressive long short-term memory (NAR-LSTM) network to achieve effective prediction and compensation of contour errors. Zhang Yun et al. proposed a contour error prediction model based on a long short-term memory (LSTM) neural network. Taking hexagonal groove processing as an example, the influence of factors such as power, vibration and temperature signals on contour errors was considered. The LSTM model can accurately predict the contour error of processed parts.

[0004] However, existing deep learning prediction methods often consider spatial information and temporal information separately, fail to fully consider the spatial dimension and the spatial correlation that changes over time, that is, the integration of space and time, and ignore or fail to consider the impact of different process parameters on position. Therefore, it is necessary to develop a new prediction technology that can not only process time series data, but also capture spatial features and their correlation over time, and consider the impact of different process parameters on position, so as to achieve higher accuracy prediction. Summary of the invention

[0005] In view of the deficiencies in the prior art, an object of the present invention is to provide an aero-engine blade position prediction method that fuses multi-source data.

[0006] In order to achieve the above object, the technical solution adopted by the present invention is as follows: A method for predicting the position of an aeroengine blade by fusing multi-source data, the method comprising the following steps: Step 1: Collect five process parameters including spindle speed, feed rate, cutting depth, path spacing and tool tilt angle and the acceleration signal data of the fixture and spindle in the three directions of X, Y and Z as multi-source input data; Step 2: Preprocess the data, where only the acceleration signal and process parameters are normalized, and the label value is not normalized; Step 3: After preprocessing, a multi-source data set is obtained; the data set is divided into a training set, a validation set, and a test set; Step 4: Build the deep learning framework RLTN; The RLTN includes a spatial feature extraction part, a temporal feature extraction part, a spatiotemporal feature extraction part, an adaptive average pooling operation, and a fully connected layer; The spatial feature extraction part is implemented by a multi-core ResNet network, and the structure of the multi-core ResNet network is: the process parameters are spliced ​​to each time step of the multivariate time series x as the input R1input, and R1input is saved as another copy and recorded as identity1; then the input Rlinput is processed by two convolution branches with 64 channels respectively, the results of the two convolution branches are added and then downsampled, and then added with identity1, and the output result is a spatial feature sequence R1output with a feature dimension of 64; R1output is then processed by two convolution branches with 128 channels respectively, the outputs of the two convolution branches with 128 channels are added and then downsampled, and then added with R1output to obtain a spatial feature sequence R2output with a feature dimension of 128; the convolution kernel size and step size of the two convolution branches are different; The temporal feature extraction part adopts LSTM network, and the output R2output of the spatial feature extraction part is used as the input of LSTM network. After being processed by LSTM network, a spatial-temporal feature sequence of dimension 512 is obtained, which is recorded as Loutput; The spatiotemporal feature extraction part adopts the iTransformer model to realize spatiotemporal integrated feature extraction, and adds Linear and GRU modules to realize the introduction of process parameters as conditions. The iTransformer model includes a transposition operation and multiple stacked units connected in series. Each stacked unit has the same structure and includes a cross attention module. The specific structure is: take the output Loutput of the temporal feature extraction part as input, perform a transposition operation on Loutput, and copy two copies as T1 and T2, and save another copy as identity2; embed the process parameters as conditions into the last dimension of each token of T1; In the cross attention module: the features after embedding the process parameters are processed by the linear layer and the GRU module to obtain the static features TLG of the fused process parameters; at the same time, T2 is processed by multi-head attention to obtain the multi-head attention output, and the static feature TLG and the multi-head attention output are multiplied to obtain the output Attitionoutput1 of the cross attention module; Attitionoutput1 and identity2 are added and layer normalized and forward propagated. The layer normalized input and the forward propagated output are added and layer normalized again to obtain the output of a stacked unit. The final output results of multiple series-connected stacking units are then transposed to obtain the spatiotemporal feature sequence Toutput; The spatiotemporal feature sequence Toutput is then processed by an adaptive average pooling operation and a fully connected layer to output the prediction result; This completes the construction of the deep learning framework RLTN; Step 5: Use multi-source datasets to train the deep learning framework RLTN for the prediction of aero-engine blade position; Furthermore, the process parameters include spindle speed, feed speed, cutting depth, path spacing and tool rake angle.

[0007] Furthermore, the preprocessing includes: performing absolute value processing on the data points, then filling in missing values, performing mean downsampling, and normalizing the acceleration signal and process parameters.

[0008] Furthermore, one convolution branch includes a convolution kernel of size 5 and stride 2, applying layer normalization and ReLU activation function, a convolution kernel of size 1 and stride 1, applying layer normalization and ReLU activation function, and an adaptive average pooling operation, which are connected in sequence; the other convolution branch has the same structure except that the size of the first convolution kernel is 3 and the stride is 1.

[0009] Furthermore, the number of the stacking units is 4 or 8 layers.

[0010] The present invention also protects a computer-readable storage medium on which a computer program is stored, and when the program is executed by a processor, the steps of the prediction method can be implemented.

[0011] Compared with the prior art, the present invention has the following beneficial effects: 1. The present invention creatively constructs a deep learning framework RLTN for multivariate time series prediction with static parameters. In the process of deep data fusion, different methods are used to introduce static parameters into the model as information (enhancing the model's understanding ability and information enhancement) and conditions (better understanding of the static characteristics in the input data, preparing for subsequent process parameter optimization), which improves the prediction effect of the model.

[0012] 2. The RLTN in the present invention is composed of a multi-core ResNet network, an LSTM network and an iTransformer model, which effectively integrates different types of features (spatial features, temporal features, and process parameters) to ensure that information is efficiently transmitted between different components to take advantage of the advantages of the three.

[0013] 3. On real processing data, RLTN shows better stability and accuracy than traditional models such as RNN, LSTM, and Transformer, especially in regression prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 It is a schematic diagram of the position measurement punctuation points of the present invention.

[0015] Figure 2 The present invention is a flow chart of an aero-engine blade position prediction method integrating multi-source data.

[0016] Figure 3 It is a structural diagram of the iTransformer model in the present invention.

[0017] Figure 4 Schematic diagram of the structure of the cross-attention module in the present invention.

[0018] Figure 5 Schematic diagram of the structure of the multi-core ResNet network in the present invention.

[0019] Figure 6 It is a structural diagram of the LSTM network in the present invention. DETAILED DESCRIPTION

[0020] The following will be combined with the accompanying drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments. The following description of at least one exemplary embodiment is actually only illustrative and is by no means intended to limit the present invention and its application or use.

[0021] The present invention integrates the multi-source data to predict the position of the aeroengine blade (see Figure 2 ), including the following steps: Step 1: Record the five process parameters of spindle speed, feed rate, cutting depth, path spacing and tool tilt angle, and use sensors to collect acceleration signal data of the fixture and tool spindle in the three directions of X, Y and Z as multi-source input data; use Renishaw probe to measure and obtain the position on each path, and take the maximum value as the label; fill missing values, reduce the mean, eliminate directionality and normalize the data; Step 2: Data preprocessing, which includes: filling missing values, reducing the mean, and eliminating directional data on the acceleration signal data, and normalizing the acceleration signal and process parameters, but not normalizing the label value (position).

[0022] Step 3: After preprocessing, a multi-source data set is obtained, and the multi-source data set is divided into a training set, a validation set, and a test set with a division ratio of 8:1:1; Step 4: Build a deep learning framework - RLTN The RLTN includes a spatial feature extraction part, a temporal feature extraction part, a spatiotemporal feature extraction part, an adaptive average pooling operation, and a fully connected layer; The spatial feature extraction part (see Figure 5 ) is implemented using a multi-core ResNet network. The structure of the multi-core ResNet network is: the process parameters are spliced ​​to each time step of the multivariate time series x as the input R1input, and R1input is saved as another copy and recorded as identity1; then the input Rlinput is processed by two convolution branches with 64 channels respectively, the results of the two convolution branches are added and down-sampled, and then added with identity1, and the output result is a spatial feature sequence R1output with a feature dimension of 64; R1output is then processed by two convolution branches with 128 channels respectively, the outputs of the two convolution branches with 128 channels are added and down-sampled, and then added with R1output to obtain a spatial feature sequence R2output with a feature dimension of 128; One convolution branch includes a convolution kernel of size 5 and stride 2, applying layer normalization and ReLU activation function, a convolution kernel of size 1 and stride 1, applying layer normalization and ReLU activation function, and an adaptive average pooling operation; the other convolution branch has the same structure except that the size of the first convolution kernel is 3 and the stride is 1.

[0023] The input of the spatial feature extraction part is the preprocessed multivariate time series and process parameters. The process parameters are used as additional features of each time step of the multivariate time series to perform convolution feature extraction together, and the output data is a 128-dimensional multivariate time series.

[0024] The temporal feature extraction part adopts LSTM network, and the output R2output of the spatial feature extraction part is used as the input of LSTM network. After being processed by LSTM network, a spatial-temporal feature sequence of dimension 512 is obtained, which is recorded as Loutput; The input of the temporal feature extraction part is the output of the spatial feature extraction part, and the output is a feature sequence after long-term dependencies are captured by multi-layer LSTM units, denoted as Loutput; The spatiotemporal feature extraction part adopts the iTransformer model to realize spatiotemporal integrated feature extraction, and adds Linear and GRU modules to realize the introduction of process parameters as conditions. The iTransformer model includes a transposition operation and multiple stacked units connected in series. Each stacked unit has the same structure and includes a cross attention module. The specific structure (see Figure 3 and Figure 4 ) is: take the output Loutput of the time feature extraction part as input, transpose Loutput, and copy two copies as T1 and T2, and save another copy as identity2; embed the process parameters into the last dimension of each token of T1 as conditions, and in the cross attention module: process the features embedded with the process parameters through the linear layer and the GRU module to obtain the static feature TLG combined with the process parameters; at the same time, process T2 with multi-head attention to obtain the multi-head attention output, and multiply the static feature TLG and the multi-head attention output to obtain the output Attitionoutput1 of the cross attention module; Attitionoutput1 and identity2 are added and layer normalized and forward propagated. The layer normalized input and the forward propagated output are added and layer normalized again to obtain the output of a stacked unit. The final output results of multiple series-connected stacking units are then transposed to obtain the spatiotemporal feature sequence Toutput; The spatiotemporal feature sequence Toutput is then processed by an adaptive average pooling operation and a fully connected layer to output the prediction result; This completes the construction of the deep learning framework RLTN.

[0025] Step 5: Use multi-source datasets to train the deep learning framework RLTN for the prediction of aero-engine blade position.

[0026] Furthermore, the number of the stacking units is 4 or 8 layers.

[0027] Example Assume that a single sample is represented as , t represents the time step contained in a single sample, and the n variables at time step t are expressed as , n represents the number of types of multivariate data acquisition variables, and in this embodiment, the value is 6; the process parameters are expressed as , m represents the number of process parameters, and in this embodiment, the value is 5. The prediction target Y is the position degree.

[0028] Based on the above assumptions, the forecast targets are as follows:

[0029] in It is a mapping function that needs to be learned, which is used to map the multivariate time series data X and the process parameter S to the predicted value of the profile error, that is, RLTN; the semicolon here indicates that the process parameter S is used as an additional input together with the multivariate time series data for prediction, and Y is the prediction result.

[0030] The present invention collects five process parameters including spindle speed, feed speed, cutting depth, path spacing and tool roll angle and acceleration signal data of a fixture and a spindle in three directions of X, Y and Z as multi-source input data.

[0031] Position measurement points are set on the blade, and the position of each point on each path is measured using a detector. The maximum value is taken as the label, and the missing values ​​of the data points are filled. Then, mean downsampling and directional elimination are performed, and the acceleration signal and process parameters are normalized to ensure that all data are at the same level. The label value (position) is not normalized.

[0032] In view of the huge amount of acceleration signal data collected during the monitoring of aircraft engine blade processing, the mean downsampling method is adopted to simplify the data collected during each blade processing process to 4000 data points (each data point contains six acceleration signals and each path contains 500 data points) to eliminate the influence of signal directionality and ensure the consistency and comparability of the data.

[0033] Building a deep learning framework - RLTN like Figure 4 As shown in Figure 1, a multi-core ResNet network is used to extract spatial features from real-time sensor data. The multi-core ResNet network can help further refine these high-dimensional features. The specific processing process of the multi-core ResNet network is: Step 4.1: Concatenate the process parameter s to each time step of the multivariate time series x to obtain the input R1input of the spatial feature extraction model, with a feature dimension of 11 (six acceleration signals + five process parameters); Step 4.2: Save another copy of R1input as identity1; Step 4.3: Perform a convolution operation on R1input using a convolution kernel of size 5 and stride 2; Step 4.3.1: Apply layer normalization (BatchNorm1d) and activation function (ReLU); Step 4.3.2: Perform a convolution operation on R1input using a convolution kernel of size 1 and stride 1; Step 4.3.3: Apply layer normalization (BatchNorm1d) and activation function (ReLU); Step 4.4: Perform a convolution operation on R1input using a convolution kernel of size 3 and stride 1. Step 4.4.1: Apply layer normalization (BatchNorm1d) and activation function (ReLU); Step 4.4.2: Perform a convolution operation on R1input using a convolution kernel of size 1 and stride 1; Step 4.4.3: Apply layer normalization (BatchNorm1d) and activation function (ReLU); Step 4.5: Adaptively perform pooling on the above steps 4.3.3 and 4.4.3 and sum them; Step 4.6: Downsample the summation result according to the identity1 dimension, which is the same as the dimension of the original input identity1; Step 4.7: Add the down-sampling result to identity1 to obtain the spatial feature sequence R1output with a feature dimension of 64; Step 4.8: Repeat the above steps 4.2-4.7 with R1output as input to obtain the spatial feature sequence R2output with a feature dimension of 128.

[0034] The LSTM network is used to track dynamic changes in the time series. In this embodiment, the Hidden-Size of the LSTM network is set to 512 and the number of layers is 4.

[0035] R2output is input into the LSTM network, and after passing through the four-layer LSTM network, a spatial-temporal feature sequence of dimension 512 is obtained, which is recorded as Loutput.

[0036] like Figure 2 As shown, the iTransformer model is used to realize spatiotemporal integrated feature extraction, and Linear and GRU are added to introduce process parameters as conditions. While revealing the implicit correlation in the spatial and temporal dimensions, the static characteristics in the input data are better understood, preparing for the subsequent process parameter optimization.

[0037] The iTransformer model treats the entire time series of each variable output by the LSTM network as a whole, rather than splitting it into tokens at a single time point. This means that the entire time series of each variable is independently embedded into a token, rather than the traditional one for each time point. The process parameters are embedded into the last dimension of each token as conditions to achieve the fusion of process parameters and time series data. The adjustment information combined with the process parameters is generated through the linear layer and the GRU module, and the output of the cross-attention module Attitionoutput1 is obtained by multiplying it with the multi-head attention output. Finally, the time series is processed by forward propagation (FNN) to capture dependencies across time and variables. This processing method helps to maintain the integrity of the timing information and avoids the information confusion that may occur in traditional methods. In this embodiment, the number of stacked units of the iTransformer model is set to 4, and the internal dimension is 64. The specific process is: Step 6.1: Transpose Loutput, make two copies, named T1 and T2, and save another copy as identity2; Step 6.2: Embed the process parameters as conditions into the last dimension of each token of T1; Step 6.2.1: Generate adjustment information TLG combined with process parameters through linear layer and GRU module, which is static feature; Step 6.3.1: Use linear projection of multiple heads to obtain the values ​​of Q, K, and V for T2; Step 6.3.2: Calculate the Softmax score; Step 6.3.3: Multiply the Softmax score by T2; Step 6.4: Multiply the outputs of steps 6.2 and 6.3 above to get the attention output Attitionoutput1; Step 6.5: Add Attitionoutput1 and identity2, and save another copy as identity3; Step 6.6: Layer Normalization (LayerNorm); Step 6.7: Forward propagation, the result of forward propagation is added to identity3; Step 6.8: Layer Normalization (LayerNorm); Step 6.9: Repeat steps 6.2-6.8 above; Step 6.10: Transpose to obtain the spatiotemporal feature sequence Toutput; The input of the spatiotemporal attention model is the output of the temporal feature extraction model, and the output is a multivariate sequence of the same dimension.

[0038] The output layer of RLTN uses adaptive flat pooling operation to obtain a one-dimensional feature representation; then the linear layer (fully connected layer) is used to fit the final output Y.

[0039] The dataset is divided into training set, validation set and test set with a ratio of 8:1:1. The learning rate is adjusted using the StepLR scheduler, and the learning rate is multiplied by 0.95 after every 10 epochs. The initial learning rate is set to 0.0001, and the number of training iterations is 200 epochs. The model performance is evaluated by calculating the MSE, MAE and R² on the validation set. The experiment is repeated and the average is taken to reduce the error caused by randomness. The superiority of RLTN in predicting the position of aircraft engine blades is verified by comparative experiments compared with traditional methods (such as CNN, LSTM, ResNet-LSTM, iTransformer and Transformer). The specific experimental results are shown in Table 1.

[0040]

[0041] Table 2 is a statistical table of various performance indicators of the ablation experiment, where no parameter information is introduced, which means that the process parameters are not spliced ​​in the spatial feature extraction part; no parameter conditions are introduced, which means that the process parameters are not used as conditions in the spatiotemporal feature extraction part, that is, the Linear and GRU modules are not added; no parameters are introduced, which means that the above-mentioned parameter information and parameter conditions are not introduced.

[0042]

[0043] The present invention uses a deep learning framework, RLTN (Residual Temporal Long-term Network), to improve the production efficiency of aircraft engine blade manufacturing and reduce production costs. The spatial characteristics are captured by a multi-core ResNet network, the dynamic changes in the time series are tracked using LSTM, and the implicit associations in the time and space dimensions are revealed using iTransformer. In addition, the process parameters are introduced into the cross attention as background information and adjusted through the gating unit (GRU module) to identify the interaction between the process parameters and other factors. The present invention introduces process parameters in the form of information and conditions in the process of deep fusion of the fixture and spindle acceleration signals, thereby enriching feature extraction and constraint feature fusion, achieving efficient prediction of the position of aircraft engine blades, and providing strong support for improving the efficiency of aircraft engine blade milling.

[0044] Any matters not described in the present invention are applicable to the prior art.

Claims

1. A method for predicting the position of an aero-engine blade based on deep fusion of multi-source data, characterized in that: The prediction method comprises the following steps: Step 1: Collect five process parameters including spindle speed, feed speed, cutting depth, path spacing and tool roll angle and the acceleration signal data of the fixture and spindle in the three directions of X, Y and Z as multi-source input data; at the same time, obtain the position on the blade surface and take the maximum value as the label; Step 2: Preprocess the data, where only the acceleration signal and process parameters are normalized, and the label value is not normalized; Step 3: After preprocessing, a multi-source dataset is obtained, and the multi-source dataset is divided into a training set, a validation set, and a test set; Step 4: Construct a multi-source data fusion deep learning framework RLTN; The RLTN includes a spatial feature extraction part, a temporal feature extraction part, a spatiotemporal feature extraction part, an adaptive average pooling operation, and a fully connected layer; The spatial feature extraction part is implemented by a multi-core ResNet network, and the structure of the multi-core ResNet network is: the process parameters are spliced ​​to each time step of the multivariate time series x as the input R1input, and R1input is saved as another copy and recorded as identity1; then the input Rlinput is processed by two convolution branches with 64 channels respectively, the results of the two convolution branches are added and then downsampled, and then added with identity1, and the output result is a spatial feature sequence R1output with a feature dimension of 64; R1output is then processed by two convolution branches with 128 channels respectively, the outputs of the two convolution branches with 128 channels are added and then downsampled, and then added with R1output to obtain a spatial feature sequence R2output with a feature dimension of 128; the convolution kernel size and step size of the two convolution branches are different; The temporal feature extraction part adopts LSTM network, and the output R2output of the spatial feature extraction part is used as the input of LSTM network. After being processed by LSTM network, a spatial-temporal feature sequence of dimension 512 is obtained, which is recorded as Loutput; The spatiotemporal feature extraction part adopts the iTransformer model to realize spatiotemporal integrated feature extraction, and adds Linear and GRU modules to realize the introduction of process parameters as conditions. The iTransformer model includes a transposition operation and multiple stacked units connected in series. Each stacked unit has the same structure and includes a cross attention module. The specific structure is: take the output Loutput of the temporal feature extraction part as input, perform a transposition operation on Loutput, and copy two copies as T1 and T2, and save another copy as identity2; embed the process parameters as conditions into the last dimension of each token of T1; In the cross attention module: the features embedded with the process parameters are processed through the linear layer and the GRU module to obtain the static features TLG combined with the process parameters; at the same time, T2 is processed using multi-head attention to obtain the multi-head attention output, and the static feature TLG and the multi-head attention output are multiplied to obtain the output Attitionoutput1 of the cross attention module; Attitionoutput1 and identity2 are added and layer normalized and forward propagated. The layer normalized input and the forward propagated output are added and layer normalized again to obtain the output of a stacked unit. The final output results of multiple series-connected stacking units are then transposed to obtain the spatiotemporal feature sequence Toutput; The spatiotemporal feature sequence Toutput is then processed by an adaptive average pooling operation and a fully connected layer to output the prediction result; This completes the construction of RLTN; Step 5: Use the multi-source data set to train the multi-source data fusion deep learning framework RLTN for the prediction of aero-engine blade position.

2. The prediction method according to claim 1, characterized in that: The preprocessing includes: filling missing values, downsampling the mean, eliminating directions and normalizing the acceleration signal data, while normalizing the process parameters and not normalizing the label values.

3. The prediction method according to claim 1, characterized in that: One convolution branch includes a convolution kernel of size 5 and stride 2, applying layer normalization and ReLU activation function, a convolution kernel of size 1 and stride 1, applying layer normalization and ReLU activation function, and an adaptive average pooling operation; the other convolution branch has the same structure except that the size of the first convolution kernel is 3 and the stride is 1.

4. The prediction method according to claim 1, characterized in that: The number of the stacking units is 4 or 8 layers.

5. 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 prediction method described in claims 1-4 can be implemented.

Images