A multi-sensor data fusion method for compressor blisk machining

Abstract

The present invention relates to the technical field of aero-engines and discloses a multi-sensor data fusion method for compressor integral blade disc processing. The present invention filters and time-aligns data signals of each sensor and normalizes the data. A BiLSTM feature extraction module with dual-stream input in the time-frequency domain is then used to extract features of vibration signals and power signals. A BiLSTM feature extraction module with a differential structure is used to extract features of temperature signals. The extracted features are fused with processing parameters. Based on the feature fusion results, processing deformation is predicted using a processing deformation prediction model, thereby solving the problem of difficulty in fusing multi-sensor signals.

Description

Technical Field

[0001] The invention relates to the technical field of aero-engines and discloses a multi-sensor data fusion method for compressor integral blade disc processing. Background Art

[0002] In the high-precision machining of the integral blade disk of the aircraft engine compressor, the real-time data fusion and machining deformation prediction of multi-source heterogeneous sensors (including high-frequency vibration sensors, spindle power monitoring units, distributed temperature probes and CNC system machining parameters) are the core technical links to ensure the machining quality of the blade disk. These sensors have different sampling frequencies and response characteristics (vibration signal Level response, temperature signal The effective integration of dynamic machining process information (including thermal inertia and thermal inertia) directly determines the accuracy of machining status monitoring and the control level of the final form and position tolerances of the blisk. Currently, achieving precise time alignment of multi-source signals, physically meaningful feature fusion, and deformation prediction consistent with the laws of material mechanics has become a key technical bottleneck in improving the machining accuracy and process stability of the blisk.

[0003] However, existing technologies face certain challenges: on the one hand, the coupling relationship between multiple physical quantity data collected by multiple sensors is complex, making it difficult to accurately extract and fuse features; on the other hand, due to differences in sampling frequency and physical response delays, traditional time alignment methods lead to feature fusion errors. Summary of the Invention

[0004] The purpose of the present invention is to provide a multi-sensor data fusion method for compressor integral blade disk processing, so as to solve the problem that multi-sensor signals are difficult to fuse during the processing of engine compressor integral blade disk.

[0005] In order to achieve the above technical effects, the technical solution adopted by the present invention is:

[0006] A multi-sensor data fusion method for compressor blisk machining, comprising:

[0007] During the blade processing of the compressor integral blade disk, the temperature signals, vibration signals, and power signals corresponding to the test points on the blade surface are analyzed by a filtering algorithm and a time alignment module to obtain the time-aligned temperature signals, vibration signals, and power signals. The data are then normalized to obtain normalized data of the temperature signals, vibration signals, and power signals. The temperature signal is the temperature measurement data of the test points on the blade disk, the vibration signal is characterized by analyzing the tool speed measurement data and cutting force measurement data of the machine tool, and the power signal is the spindle current measurement data of the machine tool.

[0008] According to the normalized data of temperature signal, vibration signal and power signal, the vibration signal and power signal are analyzed by using the BiLSTM feature extraction module with dual stream input in time and frequency domain, and the temperature signal is analyzed by using the BiLSTM feature extraction module with differential structure to obtain the temperature feature vector , vibration eigenvector and power eigenvector ;

[0009] According to the temperature eigenvector , vibration eigenvector and power eigenvector , and the processing parameters of the machine tool, through feature fusion model analysis, to obtain the fusion feature The machining parameters of the machine tool include axial milling depth, radial milling depth, feed speed and machine tool spindle speed;

[0010] According to the fusion features ,The processing deformation prediction model is adopted to analyze and obtain the ,deformation feature vector of the processing deformation assessment point.

[0011] Furthermore, the method for achieving time alignment of the temperature signal, the vibration signal, and the power signal through the time alignment module is: taking the time starting point of the vibration signal as the reference starting point, analyzing the time deviation values ​​of the starting points of the temperature signal and the power signal from the reference starting point respectively;

[0012] If the time deviation between the time starting point of the power signal and the reference starting point is less than or equal to the first time difference threshold, it is determined that the time starting point of the power signal is aligned with the reference starting point; otherwise, the time starting point of the power signal is shifted toward the reference starting point and aligned;

[0013] If the time deviation between the time starting point of the temperature signal and the time starting point of the power signal is less than or equal to the second time difference threshold, it is determined that the time starting point of the temperature signal and the time starting point of the power signal are aligned; otherwise, the time starting point of the temperature signal is shifted toward the time starting point of the power signal and aligned.

[0014] Furthermore, the method for analyzing the vibration signal by the BiLSTM feature extraction module using the dual-stream input in the time-frequency domain is:

[0015] According to the time domain data of the vibration signal, the corresponding frequency domain data is converted by short-time Fourier transform;

[0016] Through the time domain BiLSTM structure of time domain convolution unit and double-layer hidden unit, the local time series features in the time domain data are extracted to obtain the characteristic components corresponding to the time evolution pattern Through the frequency domain BiLSTM structure of frequency domain convolution unit and single layer hidden unit, the local time series features in the frequency domain data are extracted to obtain the frequency distribution feature components ;

[0017] pass Normalization converts the time evolution pattern to the characteristic component and frequency distribution characteristic components Fusion, get the vibration eigenvector .

[0018] Furthermore, the method for analyzing the power signal by the time-frequency domain dual-stream input BiLSTM feature extraction module is:

[0019] According to the time domain data of the power signal, the corresponding frequency domain data is converted by short-time Fourier transform;

[0020] Through the time domain BiLSTM structure of time domain convolution unit and single layer hidden unit, the local time series features in the time domain data are extracted to obtain the characteristic components corresponding to the time evolution pattern. ;

[0021] Through the frequency domain BiLSTM structure of frequency domain convolution unit and double-layer hidden unit, the local time series features in the frequency domain data are extracted to obtain the frequency distribution feature components ;

[0022] pass Normalization converts the time evolution pattern to the characteristic component and frequency distribution characteristic components Fusion, get the power feature vector .

[0023] Furthermore, the method for analyzing the temperature signal using the BiLSTM feature extraction module with a differential structure includes:

[0024] According to the temperature signal data, the first-order difference eigenvector is constructed through the first-order difference and second-order difference analysis algorithms. and the second-order difference eigenvector ;

[0025] According to the first-order difference eigenvector and the second-order difference eigenvector , through feature differential encoding, and then through the linear sub-layer and LeakyReLU (0.1) activation function, to obtain the fusion coding feature ;

[0026] According to the fusion coding feature , a deep delay compensation BiLSTM module with three hidden units is used for analysis to obtain the time series feature vector; the deep delay compensation BiLSTM module has a thermal inertia compensation module, through The thermal inertia compensation is obtained by analysis, where is the fusion coding feature after net compensation, The initial weight is 0.1, and the constraint ;

[0027] According to the time series feature vector, the temperature feature vector is obtained by analyzing the adaptive pooling module. .

[0028] Furthermore, the fusion feature Obtained through the following analysis formula:

[0029]

[0030] in, For 、 、 、 The related linear normalization fusion function, For The related normalized exponential function, is the characteristic component corresponding to the time evolution pattern extracted from the vibration signal, is the frequency distribution characteristic component extracted from the vibration signal, is the characteristic component corresponding to the time evolution pattern extracted from the power signal, is the frequency distribution characteristic component extracted from the power signal, is the net-compensated fusion coding feature extracted from the temperature signal, is the axial milling depth in the machining parameters, is the radial milling depth in the machining parameters, is the feed rate in the processing parameters, is the machine tool spindle speed in the processing parameters, is the dimension alignment matrix of the processing parameters, is the weight matrix, is the weight matrix The element in the mth row and nth column of is the fusion feature obtained by fusion.

[0031] Furthermore, the method for analyzing the machining deformation prediction model to obtain the deformation feature vector of the machining deformation assessment point is:

[0032] According to the input fusion features ,pass Normalization, obtain standardized feature vector;

[0033] Analyzing the normalized feature vector using a convolution operation to obtain a feature vector after dimensionality reduction;

[0034] The reduced eigenvector is analyzed through a dual-path residual mapping architecture to obtain the residual enhanced eigenvector.

[0035] The residual enhancement feature vector is globally aggregated through the adaptive average pooling algorithm to generate the deformation feature vector of the machining deformation assessment point.

[0036] Furthermore, the temperature signal is filtered using a Savitzky-Golay filtering method, and the vibration signal and the power signal are filtered using a sliding average filtering method.

[0037] Compared with the prior art, the present invention has the following beneficial effects:

[0038] The present invention filters and time-aligns the data signals of each sensor and normalizes the data. It then extracts features from the vibration and power signals through a BiLSTM feature extraction module with dual stream input in the time and frequency domains. It also uses a BiLSTM feature extraction module with a differential structure to extract features from the temperature signal. The extracted features are fused with the processing parameters, and based on the feature fusion results, a processing deformation prediction model is used to predict processing deformation, thus solving the problem of multi-sensor signal fusion being difficult. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 This is a flow chart of a multi-sensor data fusion method for compressor blisk processing in an embodiment;

[0040] Figure 2 Schematic diagram of the feature extraction module and feature fusion model corresponding to the temperature signal, vibration signal and power signal in the embodiment;

[0041] Figure 3 Schematic diagram of the structure of the machining deformation prediction model in the embodiment. DETAILED DESCRIPTION

[0042] The present invention will be described in further detail below with reference to the embodiments and accompanying drawings. However, this should not be construed as limiting the scope of the present invention to the following embodiments, as all technologies implemented based on the present invention fall within the scope of the present invention.

[0043] Example 1

[0044] See also Figure 1-3 A multi-sensor data fusion method for compressor blisk machining, comprising:

[0045] During the blade processing of the integral impeller of the compressor, the filtering algorithm and the time alignment module are used for analysis in succession according to the temperature signals, vibration signals and power signals corresponding to the assessment points on the blade surface to obtain the time-aligned temperature signals, vibration signals and power signals, and then the data are normalized respectively to obtain the normalized data of the temperature signals, vibration signals and power signals; the temperature signal is the temperature measurement data of the assessment points on the impeller, the vibration signal is characterized by analyzing the tool speed measurement data and cutting force measurement data of the machine tool, and the power signal is the spindle current measurement data of the machine tool; it should be noted that the tool speed measurement data and cutting force measurement data of the machine tool are analyzed by existing methods such as machine learning to obtain data that can characterize the vibration signal.

[0046] According to the normalized data of temperature signal, vibration signal and power signal, the vibration signal and power signal are analyzed by using the BiLSTM feature extraction module with dual stream input in time and frequency domain, and the temperature signal is analyzed by using the BiLSTM feature extraction module with differential structure to obtain the temperature feature vector , vibration eigenvector and power eigenvector ;

[0047] According to the temperature eigenvector , vibration eigenvector and power eigenvector , and the processing parameters of the machine tool, through feature fusion model analysis, to obtain the fusion feature The machining parameters of the machine tool include axial milling depth, radial milling depth, feed speed and machine tool spindle speed;

[0048] According to the fusion features ,The processing deformation prediction model is adopted to analyze and obtain the ,deformation feature vector of the processing deformation assessment point.

[0049] This embodiment first solves the problem of different sensor signals with different forms and complex error factors, and the difficulty of accurate alignment by filtering and time-aligning the temperature signal, vibration signal and power signal, thus ensuring the consistency and accuracy of the data. Then, based on the characteristics of the temperature signal, vibration signal and power signal, a BiLSTM feature extraction module with dual-stream input in the time-frequency domain and a BiLSTM feature extraction module with a differential structure are used to accurately extract their respective key feature vectors. Finally, the temperature feature vector, vibration feature vector and power feature vector are fused with the processing parameters of the machine tool to obtain the fused feature vector. , fusion features A variety of influencing factors are comprehensively considered to make the subsequent machining deformation prediction more accurate and reliable. Finally, the machining deformation prediction model is used to fusion feature This analysis obtains deformation feature vectors of machining deformation assessment points, solving the difficulty of multi-sensor signal fusion and providing strong support for compressor blisk machining. This embodiment uses a multi-sensor data fusion method to accurately predict deformation assessment points during compressor blisk machining.

[0050] Example 2

[0051] See also Figure 1-3 A multi-sensor data fusion method for compressor blisk machining, comprising:

[0052] Step 1: Determine the assessment points on the compressor blisk blades where machining deformation prediction is required;

[0053] Four sites are evenly selected on a certain side of the blade arc. The blade arc has two sides, so there are eight corresponding sites. These eight assessment points are the sites where machining deformation prediction is required.

[0054] Step 2: During the blade processing of the compressor integral blade disk, the temperature signals, vibration signals, and power signals corresponding to the test points on the blade surface are analyzed by a filtering algorithm and a time alignment module to obtain the time-aligned temperature signals, vibration signals, and power signals. The data are then normalized to obtain normalized data of the temperature signals, vibration signals, and power signals. The temperature signal is the temperature measurement data of the test points on the blade disk, the vibration signal is characterized by analyzing the tool speed measurement data and cutting force measurement data of the machine tool, and the power signal is the spindle current measurement data of the machine tool.

[0055] Specifically, first, the temperature signal, vibration signal and power signal corresponding to the machining deformation assessment point are filtered respectively.

[0056] (1) The temperature signal is filtered using the Savitzky-Golay filtering method, which includes the following steps:

[0057] First, a Savitzky-Golay filter is used to perform local curve fitting analysis on the data near the center of the window based on 21 consecutive temperature measurements within the sliding window using a third-order polynomial model to obtain a fitting polynomial. Assuming the center of the window is t = 0, the fitting polynomial is:

[0058]

[0059] in is the coefficient to be determined, is the relative time within the window.

[0060] Secondly, the fitting polynomial and the temperature measurement value are analyzed by the least squares estimation algorithm to obtain the polynomial coefficients The optimal numerical solution is calculated as follows: ,

[0061] in is the signal data within the window, W is a Vandermonde matrix, element .

[0062] Finally, according to the polynomial coefficients and the pre-generated Savitzky-Golay convolution coefficients , using the center point weighted calculation formula to analyze the measured values ​​in the sliding window and obtain the window center point filtering value , the calculation formula is ,in .

[0063] Finally, according to the slowly changing characteristics of the temperature signal and the filter value of the window center point , through the first-order inertia compensation formula Analyze the changing trend of the initial smoothing value to obtain the final filtered temperature output value after compensation, where is the compensation time constant.

[0064] (2) The vibration signal adopts the sliding average filtering method, the window length is 7 points, and the weight coefficient is , to balance noise suppression and feature preservation.

[0065] (3) The power signal adopts the sliding average filtering method, the window length is 5 points, and the weight coefficient is .

[0066] Then, based on the filtered temperature signal, vibration signal and power signal, the time starting point of each type of signal is analyzed and obtained. Taking the time starting point of the vibration signal as the reference starting point, the time deviations of the starting points of the temperature signal and the power signal and the reference starting point are analyzed respectively. The starting points of the temperature signal and the power signal are aligned with the reference starting point to obtain the time-aligned temperature signal, vibration signal and power signal.

[0067] The time starting point judgment method of the vibration signal is: for any time point have: And continue for 10 sampling points time, time point is the starting point, for Momentary energy, , is the amplitude.

[0068] The time starting point of the power signal is determined by using the first-order derivative of the power signal on the basis of the filtered power signal. Analyze, where I is the spindle current monitoring value, t For time, when a certain moment , its first-order derivative Requirements:

[0069] 1) Exceeding the maximum no-load fluctuation ;

[0070] 2) Continue for 10 sampling points .

[0071] It is considered that the moment is the starting point of the power signal.

[0072] The method for determining the time starting point of the temperature signal is as follows: based on the filtered temperature signal, the first-order derivative of the temperature signal is used to determine the time starting point of the temperature signal. Analyze, when a certain moment , and satisfy the following conditions:

[0073] 1) Satisfy the derivative For the first time interval, where 、 are the mean and variance of the temperature signal respectively;

[0074] 2) Continue for 10 sampling points until .

[0075] It is considered that the moment is the starting point of the temperature signal.

[0076] After determining the time starting points of the temperature signal, vibration signal and power signal respectively, the time starting point of the vibration signal is used as the reference starting point, and the time deviation values ​​of the starting points of the temperature signal and the power signal and the reference starting point are analyzed respectively; if the time deviation between the time starting point of the power signal and the reference starting point is less than or equal to a first time difference threshold, it is determined that the time starting point of the power signal is aligned with the reference starting point, otherwise, the time starting point of the power signal is shifted toward the reference starting point and aligned; if the time deviation between the time starting point of the temperature signal and the time starting point of the power signal is less than or equal to a second time difference threshold, it is determined that the time starting point of the temperature signal is aligned with the time starting point of the power signal, otherwise, the time starting point of the temperature signal is shifted toward the time starting point of the power signal and aligned.

[0077] For example, if the time deviation between the time starting point of the vibration signal and the time starting point of the power signal is less than or equal to 20ms, then it is determined that the time starting point of the power signal is aligned with the reference starting point. Otherwise, the time starting point of the power signal is shifted toward the reference starting point and aligned. If the time deviation between the time starting point of the temperature signal and the time starting point of the power signal is less than or equal to 500ms, then it is determined that the time starting point of the temperature signal is aligned with the time starting point of the power signal. Otherwise, the time starting point of the temperature signal is shifted toward the time starting point of the power signal and aligned.

[0078] ,

[0079] ,

[0080] in , They are the time series corresponding to vibration signal and temperature signal respectively.

[0081] Finally, the time-aligned temperature signal, vibration signal, and power signal are normalized separately. Then, through zero-mean normalization processing, the three types of signals are uniformly mapped to the [-1, 1] interval to eliminate the dimensional differences and obtain the normalized data of the three types of signals.

[0082] The temperature signal, vibration signal and power signal are respectively normalized by the following formulas:

[0083] ,

[0084] ,

[0085] ,

[0086] in , , They are the data before normalization of vibration signal, power signal and temperature signal, , , They are the normalized data of vibration signal, power signal and temperature signal respectively. , , are the baseline values ​​of vibration signal, power signal and temperature signal respectively. , , are the signal variances of vibration signal, power signal and temperature signal respectively, express time.

[0087] In step 2, by filtering the multi-sensor signals, the time starting point of each signal is calculated on this basis, and time alignment is performed based on the time starting point of the vibration signal. Methods such as data normalization processing solve the problem that different sensor signals have different forms, complex error factors, and are difficult to align accurately.

[0088] Step 3: Based on the normalized data of temperature signal, vibration signal and power signal, the vibration signal and power signal are analyzed using the time-frequency domain dual-stream input BiLSTM feature extraction module, and the temperature signal is analyzed using the BiLSTM feature extraction module with a differential structure to obtain the temperature feature vector , vibration eigenvector and power eigenvector ; The temperature characteristic vector By first-order difference characteristics , second-order difference characteristics After fusion coding and deep temporal conversion, the vibration feature vector is obtained through adaptive aggregation; The characteristic components corresponding to the time evolution pattern and frequency distribution characteristic components The power eigenvector is fused The characteristic components corresponding to the time evolution pattern and frequency distribution characteristic components Fusion.

[0089] Specifically, because vibration signals have the characteristics of nonlinearity and high-frequency transients, a BiLSTM feature extraction module with dual-channel input in the time and frequency domains is used for feature extraction. Figure 2 The method of analyzing the vibration signal by the time-frequency domain dual-stream input BiLSTM feature extraction module is:

[0090] According to the time domain data of the vibration signal, the time domain data is the normalized data mentioned above, and the time domain data is transformed into the normalized data by short-time Fourier transform. Convert to obtain the corresponding frequency domain data, where is the time domain signal, is the Hamming window function, H =64 is frame shift, M is the frame index, K is the frequency index;

[0091] Through the time domain BiLSTM structure with a 5×1 convolution kernel, a time domain convolution unit with a step size of 2 and a double-layer 64×2 hidden unit, the output channel of the convolution layer is 16-dimensional to extract the local time series features in the time domain data, and then obtain the characteristic components corresponding to the time evolution pattern ;

[0092] Through the frequency domain BiLSTM structure including 64×1 convolution kernel, frequency domain convolution unit with a step size of 16 and a single layer of 32×2 hidden units, the output channel of the convolution layer is 16-dimensional to extract the local time series features in the frequency domain data and then obtain the frequency distribution feature components ;

[0093] pass Normalization converts the time evolution pattern to the characteristic component and frequency distribution characteristic components , and obtain the vibration eigenvector , .

[0094] Specifically, because the power signal has the characteristics of strong periodicity and high-frequency transients, a time-frequency domain dual-stream input BiLSTM feature extraction module with dual-channel input in the time-frequency domain is used for feature extraction. Figure 2 The method for analyzing the power signal by the time-frequency domain dual-stream input BiLSTM feature extraction module is:

[0095] According to the time domain data of the power signal, the time domain data is the normalized data mentioned above, and the time domain data is transformed by short-time Fourier transform. Convert to obtain the corresponding frequency domain data, where is the time domain signal, is the Hamming window function, H =64 is frame shift, M is the frame index, K is the frequency index;

[0096] Through the time domain BiLSTM structure with a 5×1 convolution kernel, a time domain convolution unit with a step size of 2, and a single layer of 32×2 hidden units, the output channel of the convolution layer is 8-dimensional, extracting the local time series features in the time domain data, and then obtaining the characteristic components corresponding to the time evolution pattern ;

[0097] Through the time domain BiLSTM structure with a 32×1 convolution kernel, a time domain convolution unit with a step size of 8, and a double-layer 32×2 hidden unit, the output channel of the convolution layer is 8-dimensional, extracting the local time series features in the frequency domain data, and then obtaining the frequency distribution feature components ;

[0098] pass Normalization converts the time evolution pattern to the characteristic component and frequency distribution feature component fusion , and obtain the power eigenvector , .

[0099] Specifically, because the temperature signal has the characteristics of slow change and large thermal inertia, a deeper LSTM structure is used to extract long time series features and perform thermal delay compensation. Figure 2 The method for analyzing the temperature signal using the BiLSTM feature extraction module with a differential structure includes:

[0100] According to the temperature signal data, the first-order difference eigenvector is constructed through the first-order difference and second-order difference analysis algorithms. and the second-order difference eigenvector ; First-order difference eigenvector and the second-order difference eigenvector pass , The structure is obtained, where for t Time temperature, for t -1 moment temperature, for t The first-order difference of temperature at time t, for t -1 time first order difference of temperature, for t Second-order difference of temperature at each moment;

[0101] According to the first-order difference eigenvector and the second-order difference eigenvector , through feature differential encoding, and then through the linear sub-layer and LeakyReLU (0.1) activation function, output fusion encoding features , ;

[0102] According to the fusion coding feature , a deep delay compensation BiLSTM module with 3 layers of 64×2 hidden units is used to analyze and obtain the time series feature vector; the deep delay compensation BiLSTM module is equipped with a thermal inertia compensation module, and the thermal inertia compensation is performed by The analysis obtained is the fusion coding feature after net compensation, The initial weight is 0.1, and the constraint ;

[0103] According to the time series feature vector, the temperature feature vector is obtained by analyzing the adaptive pooling module , .

[0104] In step 3, a structurally optimized bidirectional long short-term memory network (BiLSTM) feature extraction module was constructed for each type of signal. The feature extraction modules of the three signals constitute a multi-channel feature extraction architecture. Compared with the general model, this multi-channel feature extraction module architecture can more accurately capture the temporal dynamic characteristics of each channel signal.

[0105] In the BiLSTM feature extraction module for vibration and power signals, a dual-stream input structure is implemented, with the original signal's time-domain waveform and corresponding spectrogram fed simultaneously as two independent input branches. The time-domain branch and the frequency-domain branch utilize different convolutional network structures combined with a bidirectional long short-term memory (BiLSTM) network for feature extraction. After parallel processing, the two branches are concatenated and fused to form a more representative initial feature representation. Compared to traditional methods that rely solely on a single branch in the time or frequency domain, this dual-stream input structure significantly enhances the model's ability to perceive signal changes under complex operating conditions.

[0106] For temperature signals, this embodiment builds a model that includes a differential structure, BiLSTM, and thermal error compensation for feature extraction. Compared with traditional CNN or unidirectional RNN, BiLSTM not only improves the model's memory capacity and gradient propagation efficiency, but also better captures long-term dependency patterns in the signal, thereby improving the robustness of feature expression.

[0107] In summary, step 3 effectively compensates for the shortcomings of existing technologies in processing multi-source heterogeneous signals, such as imprecise modeling and weak feature expression capabilities, through differentiated modeling of signal channels, a time-frequency dual-stream input mechanism, BiLSTM deep time series modeling, and a multimodal feature fusion strategy. It significantly improves the accuracy and generalization capability of feature extraction and is suitable for intelligent monitoring and predictive maintenance systems in complex industrial environments.

[0108] Step 4: Based on the temperature eigenvector , vibration eigenvector and power eigenvector The temperature characteristics, the time-frequency characteristics of the vibration signal, the time-frequency characteristics of the power signal and the processing parameters of the CNC system are combined through the weight matrix Align, then analyze through feature fusion model to obtain fusion features The machining parameters of the CNC system include axial milling depth, radial milling depth, feed speed and spindle speed;

[0109] The fusion features It is obtained by analyzing the following formula:

[0110]

[0111] in, For 、 、 、 The related linear normalization fusion function, For The related normalized exponential function, is the characteristic component corresponding to the time evolution pattern extracted from the vibration signal, is the frequency distribution characteristic component extracted from the vibration signal, is the characteristic component corresponding to the time evolution pattern extracted from the power signal, is the frequency distribution characteristic component extracted from the power signal, ' is the net compensated fusion coding feature extracted from the temperature signal, is the axial milling depth in the machining parameters, is the radial milling depth in the machining parameters, is the feed rate in the processing parameters, is the machine tool spindle speed in the processing parameters, is the dimension alignment matrix of the processing parameters, is the weight matrix, is the weight matrix The element in the mth row and nth column of To fuse the fusion features, the dimension alignment matrix of the processing parameters , weight matrix and the weight matrix are all learning parameters in the neural network.

[0112] After completing the feature extraction of each channel in step 3, the outputs of different modalities are integrated through feature fusion in step 4 to obtain a unified fusion feature vector. This fusion mechanism fully considers the complementary information between multi-source signals and avoids the information redundancy or feature conflict problems caused by simple splicing in traditional methods.

[0113] Step 5: Based on the fusion features ,The processing deformation prediction model is adopted to analyze and obtain the ,deformation feature vector of the assessment point during the blisk processing.

[0114] See also Figure 3 The machining deformation prediction model is a lightweight model based on residual network and convolutional neural network. The machining deformation prediction model analyzes the fusion features ,The method to obtain the deformation feature vector of the machining deformation assessment point is:

[0115] According to the input 512-dimensional fusion features ,pass Normalize to obtain a standardized 512-dimensional feature vector;

[0116] A one-dimensional convolution with a 7×1 one-dimensional convolution kernel, a stride of 2, and 64 output channels is used to analyze the standardized 512-dimensional feature vector to obtain a 64-dimensional feature vector after dimensionality reduction.

[0117] The reduced feature vector is analyzed using a dual-path residual mapping architecture. The first path is a direct transfer path that uses 1×1 convolution (channel number 64→128) to increase the 64-dimensional feature vector to 128 dimensions. The second path is a nonlinear transformation path through multiple layers of convolution and activation functions. The complex nonlinear mapping is achieved through the combination of 3×1 convolution (channel number 64→128) → batch normalization (BN) → ReLU activation → 3×1 convolution (channel number 128→128) to obtain a 128-dimensional residual enhanced feature vector.

[0118] The adaptive aggregation output layer globally aggregates the 128-dimensional residual enhancement feature vector output by the dual-path residual block group layer to generate a processed deformation feature vector of size 64×1.

[0119] In step 5, the fusion features are firstly fused through the input layer Perform preprocessing to ensure the consistency and validity of input data. The input dimension is 512 dimensions, including 128 dimensions for vibration signal, 128 dimensions for power signal, and 128 dimensions for process parameters. The fusion feature is normalized by LayerNorm. Perform normalization processing to obtain a standardized 512-dimensional feature vector.

[0120] Following the input layer, the initial convolutional layer uses one-dimensional convolution as the initial processing unit, with a kernel size of 7×1, a stride of 2, and 64 output channels. This dimensionality reduction process effectively reduces the dimensionality of high-dimensional features while preserving critical time series information, thereby reducing computational complexity and avoiding overfitting. Compared to traditional fully connected layer dimensionality reduction, one-dimensional convolution can better capture local correlations between features, further improving the model's expressiveness.

[0121] To further enhance the model's deep learning capabilities and feature extraction, the machining deformation prediction model introduces a dual-path residual block layer. Each residual block consists of two paths: a direct transfer path and a nonlinear transformation path involving multiple layers of convolution and activation functions. Specifically, the first path uses a 1×1 Conv (64→128) for dimensionality increase, while the second path implements a complex nonlinear mapping through a combination of Conv (64→128)→BN→ReLU→3×1 Conv (128→128). This dual-path design not only alleviates the vanishing gradient problem of deep networks but also enables better learning of high-level abstract features through the residual connection mechanism.

[0122] The machining deformation prediction model then combines adaptive average pooling technology with the adaptive aggregation output layer to globally aggregate the outputs of the dual-path residual block grouping layer, ultimately generating a 64×1 feature vector representing the deformation feature vector of the assessment point during blisk machining. Compared to traditional fixed-size pooling methods, adaptive average pooling in this step automatically adjusts the pooling window size based on the dynamic changes in input features, better adapting to the feature distribution in different scenarios and improving the model's generalization capabilities.

[0123] Therefore, this machining deformation prediction model significantly improves the accuracy and robustness of machining deformation prediction through model structure design and multi-level feature processing. Compared with existing technologies, it has significant advantages in feature dimensionality reduction, deep learning capabilities, and global feature aggregation, and can more effectively handle machining deformation prediction tasks in complex industrial environments.

[0124] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A multi-sensor data fusion method for compressor blade processing, characterized by: include: During the blade processing of the compressor integral blade disk, the temperature signals, vibration signals, and power signals corresponding to the test points on the blade surface are analyzed by a filtering algorithm and a time alignment module to obtain the time-aligned temperature signals, vibration signals, and power signals. The data are then normalized to obtain normalized data of the temperature signals, vibration signals, and power signals. The temperature signal is the temperature measurement data of the test points on the blade disk, the vibration signal is characterized by analyzing the tool speed measurement values ​​and cutting force measurement values ​​of the machine tool, and the power signal is the spindle current measurement data of the machine tool. According to the normalized data of temperature signal, vibration signal and power signal, the vibration signal and power signal are analyzed by using the BiLSTM feature extraction module with dual stream input in time and frequency domain, and the temperature signal is analyzed by using the BiLSTM feature extraction module with differential structure to obtain the temperature feature vector , vibration eigenvector and power eigenvector ; According to the temperature eigenvector , vibration eigenvector and power eigenvector , and the processing parameters of the machine tool, through feature fusion model analysis, to obtain the fusion feature ; The machining parameters of the machine tool include axial milling depth, radial milling depth, feed speed and machine spindle speed; According to the fusion features ,The deformation feature vector of the machining deformation assessment point is obtained by analyzing the machining deformation prediction model; The method for the time-frequency domain dual-stream input BiLSTM feature extraction module to analyze the vibration signal and power signal is: According to the time domain data of the vibration signal and the power signal, corresponding frequency domain data is converted by short-time Fourier transform; Through the time domain convolution unit and the time domain BiLSTM structure, the local time series features in the time domain data are extracted, and the characteristic components corresponding to the time evolution pattern corresponding to the vibration signal are obtained respectively. and the characteristic components corresponding to the time evolution pattern of the power signal ; Through the frequency domain convolution unit and frequency domain BiLSTM structure, the local time series features in the frequency domain data are extracted to obtain the frequency distribution feature components corresponding to the vibration signal. and the frequency distribution characteristic component corresponding to the power signal ; pass Normalization converts the time evolution pattern to the characteristic component and frequency distribution characteristic components Fusion, get the vibration eigenvector ;pass Normalization converts the time evolution pattern to the characteristic component and frequency distribution characteristic components Fusion, get the power feature vector ; Among them, the characteristic component corresponding to the time evolution pattern The corresponding time domain BiLSTM structure is a time domain BiLSTM structure with two hidden units; and the frequency distribution feature component The corresponding frequency domain BiLSTM structure is a frequency domain BiLSTM structure with a single layer of hidden units; Characteristic components corresponding to the time evolution pattern The corresponding time domain BiLSTM structure is a time domain BiLSTM structure with a single layer of hidden units, which is consistent with the frequency distribution feature component The corresponding frequency domain BiLSTM structure of the two-layer hidden unit is the frequency domain BiLSTM structure; The method for analyzing the temperature signal by the BiLSTM feature extraction module with a differential structure includes: According to the temperature signal data, the first-order difference eigenvector is constructed through the first-order difference and second-order difference analysis algorithms. and the second-order difference eigenvector ; According to the first-order difference eigenvector and the second-order difference eigenvector , through feature differential encoding, and then through the linear sub-layer and LeakyReLU (0.1) activation function, to obtain the fusion coding feature ; According to the fusion coding feature , a deep delay compensation BiLSTM module with three hidden units is used for analysis to obtain the time series feature vector; the deep delay compensation BiLSTM module has a thermal inertia compensation module, through The thermal inertia compensation is obtained by analysis, where is the fusion coding feature after net compensation, The initial weight is 0.1, and the constraint ; According to the time series feature vector, the temperature feature vector is obtained by analyzing the adaptive pooling module. , .

2. The multi-sensor data fusion method according to claim 1, characterized in that: The method for achieving time alignment of the temperature signal, the vibration signal, and the power signal through the time alignment module is as follows: taking the time starting point of the vibration signal as the reference starting point, analyzing the time deviation values ​​of the starting points of the temperature signal and the power signal from the reference starting point respectively; If the time deviation between the time starting point of the power signal and the reference starting point is less than or equal to the first time difference threshold, it is determined that the time starting point of the power signal is aligned with the reference starting point; otherwise, the time starting point of the power signal is shifted toward the reference starting point and aligned; If the time deviation between the time starting point of the temperature signal and the time starting point of the power signal is less than or equal to the second time difference threshold, it is determined that the time starting point of the temperature signal and the time starting point of the power signal are aligned; otherwise, the time starting point of the temperature signal is shifted toward the time starting point of the power signal and aligned.

3. The multi-sensor data fusion method according to claim 1, characterized in that: The fusion features Obtained through the following analysis formula: in, For 、 、 、 The related linear normalization fusion function, For The related normalized exponential function, is the characteristic component corresponding to the time evolution pattern extracted from the vibration signal, is the frequency distribution characteristic component extracted from the vibration signal, is the characteristic component corresponding to the time evolution pattern extracted from the power signal, is the frequency distribution characteristic component extracted from the power signal, is the net-compensated fusion coding feature extracted from the temperature signal, is the axial milling depth in the machining parameters, is the radial milling depth in the machining parameters, is the feed rate in the processing parameters, is the machine tool spindle speed in the processing parameters, is the dimension alignment matrix of the processing parameters, is the weight matrix, is the weight matrix The element in the mth row and nth column of is the fusion feature obtained by fusion.

4. The multi-sensor data fusion method according to claim 3, characterized in that: The method for analyzing the machining deformation prediction model to obtain the deformation feature vector of the machining deformation assessment point is: According to the input fusion features ,pass Normalization, obtain standardized feature vector; Analyzing the normalized feature vector using a convolution operation to obtain a feature vector after dimensionality reduction; The reduced eigenvector is analyzed through a dual-path residual mapping architecture to obtain the residual enhanced eigenvector. The residual enhancement feature vector is globally aggregated through the adaptive average pooling algorithm to generate the deformation feature vector of the machining deformation assessment point.

5. The multi-sensor data fusion method according to claim 1, characterized in that: The temperature signal is filtered using a Savitzky-Golay filtering method, and the vibration signal and the power signal are filtered using a sliding average filtering method.

Images