A flutter intelligent monitoring method and system for integral bladed disk manufacturing

Abstract

This invention relates to the field of intelligent manufacturing technology and discloses a chatter intelligent monitoring method and system for integral bladed disk (IBD) machining. By performing dimensionality reduction processing on multi-source heterogeneous signals during IBD machining, and then employing a cascaded architecture of dual-path feature fusion, transfer learning mapping, and Weibull three-parameter prediction, the method simultaneously captures millisecond-level chatter mutations and long-term trend changes through parallel dual-path feature fusion, achieving short-term and long-term feature complementarity. This constructs a Weibull three-parameter mapping model based on a multilayer perceptron model for predicting the remaining machining time of the tool, and an attention-based linear regression analysis model for predicting surface roughness Ra, significantly improving the model's ability to characterize complex working conditions. Furthermore, by combining transfer learning mapping to achieve training adaptation for different cutting processes, and using neural network fitting to precisely constrain the Weibull three-parameter network, the remaining stable machining time is directly output.

Description

Technical Field

[0001] This invention relates to the field of intelligent manufacturing technology, and discloses a flutter intelligent monitoring method and system for integral bladed disk machining. Background Technology

[0002] As a key component of high-end equipment such as aero-engines, the machining quality and efficiency of integral bladed disks directly affect the performance and production cost of the equipment. Traditional integral bladed disk machining relies primarily on manual experience and conventional inspection methods for monitoring the machining process, which suffers from problems such as untimely and inaccurate monitoring, and the inability to provide real-time feedback on machining status to optimize machining parameters. Meanwhile, tool chatter is a common self-excited vibration phenomenon in machining (such as turning, milling, and drilling), referring to the periodic relative vibration generated between the tool and the workpiece due to dynamic force coupling. This vibration is not caused by external periodic excitation but is maintained by the dynamic characteristics of the cutting process itself (such as the interaction between cutting force fluctuations and system vibration), significantly impacting machining quality, efficiency, and equipment lifespan.

[0003] Traditional flutter detection and suppression techniques have the following drawbacks:

[0004] Limitations of a single signal: Existing technologies rely on a single sensor (such as cutting force or vibration signal), which makes it difficult to fully reflect the complex dynamic characteristics of chatter. In particular, signal noise interference increases significantly in the machining of integral bladed disks with variable parameters.

[0005] Feature engineering relies on human experience: traditional methods require manual extraction of time-domain / frequency-domain features (such as mean, variance, and wavelet energy), which has low sensitivity to nonlinear wear processes, and feature redundancy leads to insufficient generalization ability of flutter recognition models.

[0006] The predictive models have poor adaptability: existing physical-driven models (such as gamma processes and Wiener processes) rely on prior degradation laws, while data-driven models (such as CNN and LSTM) are prone to performance drops under multiple operating conditions (such as variable speed and variable materials), and cannot meet the high dynamic requirements of real-time flutter suppression in the overall bladed disk manufacturing process. Summary of the Invention

[0007] The purpose of this invention is to provide a flutter intelligent monitoring method and system for integral bladed disk machining, which can predict surface roughness and remaining stable machining time, and provide an important reference indicator for on-site flutter intelligent adaptive adjustment strategy.

[0008] To achieve the above-mentioned technical effects, the technical solution adopted by the present invention is as follows:

[0009] A flutter intelligent monitoring method for integral bladed disk machining includes:

[0010] Step 1: Perform time axis alignment, filtering, and normalization preprocessing on the heterogeneous signals from multiple sensors during the integral bladed disk milling process to obtain dimensionless signal data; the heterogeneous signals include the triaxial vibration signal, spindle power signal, and surface roughness data of the integral bladed disk machining position after the tool passes through during the integral bladed disk machining process.

[0011] Step 2: Using the sliding window sampling method, extract the 7-dimensional vibration time-domain statistics, the 8-dimensional vibration frequency band energy ratio, and the 3-dimensional power statistics in the main shaft power signal from the triaxial vibration signal within each sampling window to form a dataset. Pair the measured surface roughness Ra values ​​at the center time of the corresponding sampling window in the dataset as a label set.

[0012] Step 3: Assign corresponding weights to all samples in the dataset and calculate the weighted covariance matrix of the dataset; perform symmetric eigenvalue decomposition on the weighted covariance matrix, and sort the decomposed eigenvalues ​​in descending order; calculate the ratio of the sum of the first K eigenvalues ​​to the sum of all eigenvalues; extract the first K eigenvalues ​​whose ratio is greater than or equal to a preset ratio threshold, and reduce the dimension of the dataset to the subspace corresponding to the first K eigenvalues ​​to obtain the dimension-reduced dataset;

[0013] Step 4: Perform wavelet packet decomposition on the triaxial vibration signal and the main shaft power signal, and use a one-dimensional convolutional neural network with the first preset parameter configuration to extract features from the decomposed data. After global average pooling processing on the extracted features, the features after global average pooling processing are fused with the dimensionality-reduced dataset through the dynamic gate weighted fusion method to obtain the first comprehensive feature vector.

[0014] Step 5: Map the first comprehensive feature vector using a transfer learning mapping model to obtain the transferred second comprehensive feature vector. Use a one-dimensional convolutional neural network and an LSTM network with the second preset parameter configuration to extract features from the transferred second comprehensive feature vector to obtain the corresponding third comprehensive feature vector.

[0015] Step 6: Construct a Weibull three-parameter mapping model based on a multilayer perceptron model, and use the third comprehensive feature vector as input and the corresponding surface roughness Ra as output to construct a linear regression analysis model based on the attention mechanism.

[0016] Step 7: Acquire the three-axis vibration signal and spindle power signal during the integral bladed disk milling process, and perform time axis alignment, filtering and normalization preprocessing to obtain dimensionless signal data. Use the methods in steps 2 to 5 to obtain the corresponding third comprehensive feature vector. Based on the obtained third comprehensive feature vector, use the linear regression analysis model to analyze and obtain the surface roughness data of the integral bladed disk machining position after the tool passes, and use the Weibull three-parameter mapping model to analyze and obtain the predicted value of the remaining machining time of the tool.

[0017] Further, in step two, the 7-dimensional vibration time-domain statistics in the triaxial vibration signal include: mean, variance, root mean square, kurtosis, peak factor, impulse factor, and margin factor of the triaxial vibration signal modulus; the 8-dimensional vibration frequency band energy proportions in the triaxial vibration signal include: the energy proportions of the main spindle power signal in the eight sub-bands of 0–250Hz, 250–500Hz, 500–750Hz, 750–1000Hz, 1000–1250Hz, 1250–1500Hz, 1500–1750Hz, and 1750–2000Hz; the 3-dimensional power statistics in the main spindle power signal include the root mean square, peak value, and peak factor of the main spindle power signal.

[0018] Further, in step three, the weighted covariance matrix is ​​subjected to symmetric eigenvalue decomposition, and the eigenvalues ​​obtained from the decomposition are sorted in descending order. The ratio of the sum of the first K eigenvalues ​​to the sum of all eigenvalues ​​is calculated. The method for extracting the first K eigenvalues ​​whose ratio is greater than or equal to a preset ratio threshold includes:

[0019] The weighted covariance matrix of the dataset Perform symmetric eigenvalue decomposition, such that... ,in The matrix consists of orthogonal eigenvectors. , The first diagonal line of the diagonal matrix after symmetric eigenvalue decomposition. One element, , The dimensions of the diagonal matrix are arranged in descending order. ;

[0020] The first K eigenvalues ​​are selected if the ratio of the sum of the first K eigenvalues ​​to the sum of all eigenvalues ​​is greater than or equal to 90%.

[0021] Furthermore, in step three, the dataset is reduced in dimensionality to the subspace corresponding to the first K feature values, resulting in the dimensionality-reduced dataset. ,in For the dataset, for The weighted mean, for The matrix formed by the first K column vectors.

[0022] Furthermore, in step five, the method for mapping the first comprehensive feature vector using a transfer learning mapping model to obtain the transferred second comprehensive feature vector includes:

[0023] When the input first comprehensive feature vector and the training samples belong to the same process, the output is the same as the input without any transformation. The training samples are a set composed of the dataset and the label set. When the input first comprehensive feature vector and the training samples do not belong to the same process, the mapping parameters are calculated by the training dataset formed by the preset lower limit sample size in the process where the input data is located and the dataset through a preset formula.

[0024] Furthermore, in step six, the method for constructing the Weibull three-parameter mapping model based on the multilayer perceptron model includes:

[0025] Using the third comprehensive feature vector from step five as input, and employing a multilayer perceptron model as the fitting function, the shape factor in the Weibull three-parameter mapping model is analyzed and obtained. Scale factor Location factor This leads to the formation of a Weibull three-parameter mapping model for predicting the remaining machining time of the tool. ;in This is the predicted remaining machining time for the cutting tool. Given a failure probability threshold, The current machining time of the tool, shape factor The output of a multilayer perceptron model with a logistic function as the activation function, and the scaling factor. The output of a multilayer perceptron model with an exponential activation function, location factor The output is from a multilayer perceptron model that uses a linear rectified function as its activation function.

[0026] To achieve the above-mentioned technical effects, the present invention also provides a flutter intelligent monitoring system for integral bladed disk machining, used to implement the aforementioned flutter intelligent monitoring method, comprising:

[0027] The data preprocessing module is used to perform time axis alignment, filtering and normalization preprocessing on the heterogeneous signals from multiple sensors during the integral bladed disk milling process to obtain dimensionless signal data; the heterogeneous signals include triaxial vibration signals, spindle power signals and surface roughness data of the integral bladed disk machining position after the tool passes through during the machining process.

[0028] The data extraction module is used to extract the 7-dimensional vibration time-domain statistics, the 8-dimensional vibration frequency band energy ratio, and the 3-dimensional power statistics in the main shaft power signal from the triaxial vibration signal in each sampling window using the sliding window sampling method to form a dataset, and pair the measured surface roughness Ra values ​​at the center time of the corresponding sampling window in the dataset as a label set;

[0029] The data dimensionality reduction module is used to assign corresponding weights to all samples in the dataset and calculate the weighted covariance matrix of the dataset; perform symmetric eigenvalue decomposition on the weighted covariance matrix, sort the decomposed eigenvalues ​​in descending order, calculate the ratio of the sum of the first K eigenvalues ​​to the sum of all eigenvalues; extract the first K eigenvalues ​​whose ratio is greater than or equal to a preset ratio threshold, and reduce the dimensionality of the dataset to the subspace corresponding to the first K eigenvalues ​​to obtain the dimensionality-reduced dataset;

[0030] The first feature extraction module is used to perform wavelet packet decomposition on the triaxial vibration signal and the main shaft power signal, and to extract features from the decomposed data using a one-dimensional convolutional neural network with a first preset parameter configuration. After global average pooling processing of the extracted features, the features after global average pooling processing are fused with the dimensionality-reduced dataset using a dynamic gate weighted fusion method to obtain the first comprehensive feature vector.

[0031] The second feature extraction module is used to map the first comprehensive feature vector using a transfer learning mapping model to obtain the transferred second comprehensive feature vector, and to extract features from the transferred second comprehensive feature vector using a one-dimensional convolutional neural network and an LSTM network with second preset parameters to obtain the corresponding third comprehensive feature vector.

[0032] The model building module is used to build a Weibull three-parameter mapping model based on a multilayer perceptron model, and to build a linear regression analysis model based on an attention mechanism with the third comprehensive feature vector as input and the corresponding surface roughness Ra as output.

[0033] The analysis and prediction module is used to obtain the third comprehensive feature vector during the milling process of the integral bladed disk, and based on the obtained third comprehensive feature vector, to analyze and obtain the surface roughness data of the integral bladed disk machining position after the tool passes through the linear regression analysis model, and to analyze and obtain the predicted value of the remaining machining time of the tool using the Weibull three-parameter mapping model.

[0034] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention utilizes a cascaded architecture of dual-path feature fusion, transfer learning mapping, and Weibull three-parameter prediction. Parallel dual-path feature fusion simultaneously captures millisecond-level chatter abrupt changes and long-term trend variations, achieving short-term and long-term feature complementarity. This constructs a Weibull three-parameter mapping model based on a multilayer perceptron model for predicting remaining tool machining time, and an attention-based linear regression analysis model for predicting surface roughness Ra, significantly improving the model's ability to characterize complex working conditions. Furthermore, transfer learning mapping enables training and adaptation for different cutting processes, and a Weibull three-parameter network is precisely constrained by neural network fitting, directly outputting the remaining stable machining time. This provides an important reference indicator for predicting surface roughness and remaining stable machining time, enabling intelligent adaptive adjustment strategies for chatter in the field. Attached Figure Description

[0035] Figure 1 This is a flowchart of a flutter intelligent monitoring method for integral bladed disk machining in Example 1 or 2.

[0036] Figure 2 This is a structural block diagram of a flutter intelligent monitoring system for integral bladed disk machining in Example 1;

[0037] The module comprises: 1. Data preprocessing module; 2. Data extraction module; 3. Data dimensionality reduction module; 4. First feature extraction module; 5. Second feature extraction module; 6. Model building module; and 7. Analysis and prediction module. Detailed Implementation

[0038] The present invention will now be described in further detail with reference to the embodiments and accompanying drawings. However, this should not be construed as limiting the scope of the above-described subject matter of the present invention to the following embodiments; all technologies implemented based on the content of the present invention fall within the scope of the present invention.

[0039] Example 1

[0040] See Figure 1 and Figure 2 A flutter intelligent monitoring method for integral bladed disk machining includes:

[0041] Step 1: Perform time axis alignment, filtering, and normalization preprocessing on the heterogeneous signals from multiple sensors during the integral bladed disk milling process to obtain dimensionless signal data; the heterogeneous signals include the triaxial vibration signal, spindle power signal, and surface roughness data of the integral bladed disk machining position after the tool passes through during the integral bladed disk machining process.

[0042] Step 2: Using the sliding window sampling method, extract the 7-dimensional vibration time-domain statistics, the 8-dimensional vibration frequency band energy ratio, and the 3-dimensional power statistics in the main shaft power signal from the triaxial vibration signal within each sampling window to form a dataset. Pair the measured surface roughness Ra values ​​at the center time of the corresponding sampling window in the dataset as a label set.

[0043] Step 3: Assign corresponding weights to all samples in the dataset and calculate the weighted covariance matrix of the dataset; perform symmetric eigenvalue decomposition on the weighted covariance matrix, and sort the decomposed eigenvalues ​​in descending order; calculate the ratio of the sum of the first K eigenvalues ​​to the sum of all eigenvalues; extract the first K eigenvalues ​​whose ratio is greater than or equal to a preset ratio threshold, and reduce the dimension of the dataset to the subspace corresponding to the first K eigenvalues ​​to obtain the dimension-reduced dataset;

[0044] Step 4: Perform wavelet packet decomposition on the triaxial vibration signal and the main shaft power signal, and use a one-dimensional convolutional neural network with the first preset parameter configuration to extract features from the decomposed data. After global average pooling processing on the extracted features, the features after global average pooling processing are fused with the dimensionality-reduced dataset through the dynamic gate weighted fusion method to obtain the first comprehensive feature vector.

[0045] Step 5: Map the first comprehensive feature vector using a transfer learning mapping model to obtain the transferred second comprehensive feature vector. Use a one-dimensional convolutional neural network and an LSTM network with the second preset parameter configuration to extract features from the transferred second comprehensive feature vector to obtain the corresponding third comprehensive feature vector.

[0046] Step 6: Construct a Weibull three-parameter mapping model based on a multilayer perceptron model, and use the third comprehensive feature vector as input and the corresponding surface roughness Ra as output to construct a linear regression analysis model based on the attention mechanism.

[0047] Step 7: Acquire the three-axis vibration signal and spindle power signal during the integral bladed disk milling process, and perform time axis alignment, filtering and normalization preprocessing to obtain dimensionless signal data. Use the methods in steps 2 to 5 to obtain the corresponding third comprehensive feature vector. Based on the obtained third comprehensive feature vector, use the linear regression analysis model to analyze and obtain the surface roughness data of the integral bladed disk machining position after the tool passes, and use the Weibull three-parameter mapping model to analyze and obtain the predicted value of the remaining machining time of the tool.

[0048] In this embodiment, by preprocessing the multi-source heterogeneous signals during the machining process of the overall bladed disk, relevant information is extracted to form a dataset and a label set. The weighted covariance matrix of the dataset is calculated by weight allocation to achieve dimensionality reduction of the dataset. Then, through a cascaded architecture of dual-path feature fusion, transfer learning mapping, and Weibull three-parameter prediction, the parallel dual-path feature fusion simultaneously captures millisecond-level chatter abrupt changes and long-term trend changes, achieving short-term and long-term feature complementarity. This constructs a Weibull three-parameter mapping model based on a multilayer perceptron model for predicting the remaining machining time of the tool, and a linear regression analysis model based on an attention mechanism for predicting surface roughness Ra, which can significantly improve the model's ability to characterize complex working conditions. Furthermore, the transfer learning mapping is combined to achieve training and adaptation for different cutting processes, and the Weibull three-parameter network is precisely constrained by neural network fitting to directly output the remaining stable machining time. This provides an important reference indicator for on-site intelligent adaptive adjustment strategies for chatter, which is conducive to truly achieving accurate identification of chatter, reducing Ra over-limit rate, early warning shutdown, and realizing a system that is universal for multiple processes and plug-and-play on-site.

[0049] Based on the same inventive concept, this embodiment also provides a flutter intelligent monitoring system for integral bladed disk machining, used to implement the aforementioned flutter intelligent monitoring method, including:

[0050] The data preprocessing module 1 is used to perform time axis alignment, filtering and normalization preprocessing on the heterogeneous signals from multiple sensors during the integral bladed disk milling process to obtain dimensionless signal data; the heterogeneous signals include the triaxial vibration signal, spindle power signal and surface roughness data of the integral bladed disk machining position after the tool passes through during the integral bladed disk machining process.

[0051] Data extraction module 2 is used to extract the 7-dimensional vibration time domain statistics, the 8-dimensional vibration frequency band energy ratio, and the 3-dimensional power statistics in the main shaft power signal from the triaxial vibration signal in each sampling window using the sliding window sampling method to form a dataset, and pair the measured surface roughness Ra values ​​at the center time of the corresponding sampling window in the dataset as a label set;

[0052] Data dimensionality reduction module 3 is used to assign corresponding weights to all samples in the dataset and calculate the weighted covariance matrix of the dataset; perform symmetric eigenvalue decomposition on the weighted covariance matrix, sort the decomposed eigenvalues ​​in descending order, calculate the ratio of the sum of the first K eigenvalues ​​to the sum of all eigenvalues; extract the first K eigenvalues ​​whose ratio is greater than or equal to a preset ratio threshold, and reduce the dimensionality of the dataset to the subspace corresponding to the first K eigenvalues ​​to obtain the dimensionality-reduced dataset;

[0053] The first feature extraction module 4 is used to perform wavelet packet decomposition on the triaxial vibration signal and the main shaft power signal, and to extract features from the decomposed data using a one-dimensional convolutional neural network with first preset parameters. After global average pooling processing of the extracted features, the features after global average pooling processing are fused with the dimensionality-reduced dataset using a dynamic gate weighted fusion method to obtain the first comprehensive feature vector.

[0054] The second feature extraction module 5 is used to map the first comprehensive feature vector using a transfer learning mapping model to obtain the transferred second comprehensive feature vector, and to extract features from the transferred second comprehensive feature vector using a one-dimensional convolutional neural network and an LSTM network with second preset parameters to obtain the corresponding third comprehensive feature vector.

[0055] Model building module 6 is used to build a Weibull three-parameter mapping model based on a multilayer perceptron model, and to build a linear regression analysis model based on an attention mechanism with the third comprehensive feature vector as input and the corresponding surface roughness Ra as output.

[0056] The analysis and prediction module 7 is used to obtain the third comprehensive feature vector in the overall bladed disk milling process, and based on the obtained third comprehensive feature vector, to analyze and obtain the surface roughness data of the overall bladed disk machining position after the tool passes through using the linear regression analysis model, and to analyze and obtain the predicted value of the remaining machining time of the tool using the Weibull three-parameter mapping model.

[0057] Example 2

[0058] See Figure 1 A flutter intelligent monitoring method for integral bladed disk machining includes:

[0059] Step 1: Perform time axis alignment, filtering, and normalization preprocessing on the heterogeneous signals from multiple sensors during the integral bladed disk milling process to obtain dimensionless signal data; the heterogeneous signals include the triaxial vibration signal, spindle power signal, and surface roughness data of the integral bladed disk machining position after the tool passes through during the integral bladed disk machining process.

[0060] In this embodiment, the clocks of the vibration sampling card and the power sensor used to collect triaxial vibration signals have a drift of ≤1ms. If they are not interpolated and aligned, mismatches will occur during subsequent fusion—the vibration from the previous second will be assigned to the current from the next second. Therefore, it is necessary to align the heterogeneous signals from different sensors using timestamps. Even with the aligned signals, there is still interference from power frequency, inconsistent dimensions, and occasional overshoot spikes. Furthermore, the fundamental frequency and second harmonic of the workshop power grid will drown out low-frequency cutting characteristics. Therefore, further filtering and normalization preprocessing are required to transform them into a strictly time-aligned, clean, dimensionless "standard format" before feeding them into the subsequent deep learning network.

[0061] In this embodiment, two 4th-order Butterworth band-stop filters can be designed to filter the fundamental frequency first, and then filter the second harmonic. The two filters are performed serially (the fundamental frequency is determined according to the actual situation). After filtering, a signal with unchanged shape and the power frequency fundamental and second harmonic peaks in the spectrum are removed is obtained.

[0062] In this embodiment, the mean is calculated for each signal channel. and standard deviation ,according to Data within each channel Perform normalization to obtain normalized data values. This solves the problem that when the numerical ranges of different units (acceleration m / s², power W) differ by a factor of 10, the entire network's attention is diverted to adjusting the bias.

[0063] In this embodiment, occasional spikes caused by data anomalies or interference, if left untreated, may be mistakenly identified as valid shocks by the model. This is addressed by scanning the normalized signal: if... The abnormal points are marked and replaced with 5 normal points before and after them using linear interpolation. The final output is a clean, aligned, dimensionless 4-channel signal.

[0064] Step 2: Using the sliding window sampling method, extract the 7-dimensional vibration time-domain statistics, the 8-dimensional vibration frequency band energy ratio, and the 3-dimensional power statistics in the main shaft power signal from the triaxial vibration signal within each sampling window to form a dataset. Pair the measured surface roughness Ra values ​​at the center time of the corresponding sampling window in the dataset as a label set.

[0065] In this embodiment, directly feeding the original dimensionless signal data sequence into the LSTM results in too many parameters and excessive noise. A two-step purification process, "coarse first, then fine," is required.

[0066] 1. Divide the sample into N sampling windows with a window length of 2048 and a step size of 256;

[0067] 2. Extract features from each window to form a dataset, and associate it with the surface roughness Ra value as a label set;

[0068] Each window extracts a total of 18 features, including:

[0069] The seven time-domain statistics of triaxial vibration signals include: mean, variance, root mean square (RMS), kurtosis, peak factor, impulse factor, and margin factor of the triaxial vibration signal modulus.

[0070] The energy proportion of the 8-dimensional vibration frequency bands in the triaxial vibration signal includes the energy proportion of the main spindle power signal in the eight sub-bands of 0-250Hz, 250-500Hz, 500-750Hz, 750-1000Hz, 1000-1250Hz, 1250-1500Hz, 1500-1750Hz, and 1750-2000Hz.

[0071] The three-dimensional power statistics in the spindle power signal include the root mean square (RMS), peak value, and crest factor of the spindle power signal.

[0072] Step 3: Assign corresponding weights to all samples in the dataset and calculate the weighted covariance matrix of the dataset; perform symmetric eigenvalue decomposition on the weighted covariance matrix, and sort the decomposed eigenvalues ​​in descending order; calculate the ratio of the sum of the first K eigenvalues ​​to the sum of all eigenvalues; extract the first K eigenvalues ​​whose ratio is greater than or equal to a preset ratio threshold, and reduce the dimension of the dataset to the subspace corresponding to the first K eigenvalues ​​to obtain the dimension-reduced dataset;

[0073] The method for performing symmetric eigenvalue decomposition on the weighted covariance matrix, sorting the decomposed eigenvalues ​​in descending order, calculating the ratio of the sum of the first K eigenvalues ​​to the sum of all eigenvalues, and extracting the first K eigenvalues ​​whose ratio is greater than or equal to a preset ratio threshold includes:

[0074] The weighted covariance matrix of the dataset Perform symmetric eigenvalue decomposition, such that... ,in The matrix consists of orthogonal eigenvectors. , The first diagonal line of the diagonal matrix after symmetric eigenvalue decomposition. One element, , The dimensions of the diagonal matrix are arranged in descending order. ;

[0075] Take the ratio of the sum of the first K eigenvalues ​​to the sum of all eigenvalues. The top K features that are greater than or equal to 90%.

[0076] The dimensionality of the dataset is reduced to the subspace corresponding to the first K feature values ​​to obtain the dimensionality-reduced dataset. ,in For the dataset, for The weighted mean, for The matrix formed by the first K column vectors.

[0077] Step 4: Perform wavelet packet decomposition on the triaxial vibration signal and the main shaft power signal, and use a one-dimensional convolutional neural network with the first preset parameter configuration to extract features from the decomposed data. After global average pooling processing on the extracted features, the features after global average pooling processing are fused with the dimensionality-reduced dataset through the dynamic gate weighted fusion method to obtain the first comprehensive feature vector.

[0078] In this embodiment, band-limited wavelet packet decomposition is used to process the triaxial vibration signal and the principal axis power signal.

[0079] Use pywt's db4 function to perform a 3-level decomposition, retaining the high-frequency detail subband at level 2.

[0080] The establishment of a one-dimensional convolutional neural network (1-DCNN) requires strict matching of input and output data channels; other hyperparameters can be specified as needed. The first preset parameters include the number of input channels, the number of output channels, and the kernel size. In this embodiment, the one-dimensional convolutional neural network is configured with 64 one-dimensional convolutional kernels of length 5; each kernel simultaneously convolves four input time-series signals (triaxial vibration signal + main axis power signal), and the four results are summed to form one output; two zeros are padded at both ends of each time axis to make the output sequence length equal to the input sequence length, resulting in a total of 64 output feature sequences. This is used to capture transient impact patterns within a local time window.

[0081] Global average pooling (GAP) averages along the time axis to output 64-dimensional local features. .

[0082] Then, the gating weights of the dynamic fusion gate are calculated using a multilayer perceptron with a logistic function as the activation function. Dimensionally reduced dataset After affine transformation and weighted fusion with local features, the first comprehensive feature vector is obtained. ,in , This is an affine transformation function.

[0083] Step 5: Map the first comprehensive feature vector using a transfer learning mapping model to obtain the transferred second comprehensive feature vector. Use a one-dimensional convolutional neural network and an LSTM network with the second preset parameter configuration to extract features from the transferred second comprehensive feature vector to obtain the corresponding third comprehensive feature vector.

[0084] In this embodiment, the 64-dimensional first comprehensive feature vector output by the dynamic fusion gate... To integrate the intermediate representation of local impact features and global degradation trends, a transfer adaptation processing layer is constructed. This layer primarily performs feature transfer and adaptation processing on data from different processing scenarios or operating conditions. Transfer learning is carried out according to a set process: first, source and target domain data are selected, and after feature extraction and adaptation processing, the first comprehensive feature vector set of the source domain data is obtained. The first comprehensive feature vector set of the target domain data Then, the source domain model knowledge is transferred to the target domain task. Through training and optimization, the transfer learning mapping model effectively utilizes transferred knowledge to improve performance in the target scenario, enhancing prediction or monitoring performance in cross-scenario applications and resolving model adaptability issues caused by differences in working conditions. The transfer adaptation processing layer of the transfer learning mapping model performs a linear transformation. ,in, These are the features after transfer alignment. It is a transformation matrix. It is an offset vector, derived from , It is concluded that, among them It is the identity matrix. yes The mean, It is the first comprehensive feature vector set of the source domain data obtained through feature extraction and adaptation processing. covariance; It is the first comprehensive feature vector set of the target domain data. The mean, yes The covariance.

[0085] When the input first comprehensive feature vector and the training samples belong to the same process, the output (second) is the same as the input without any transformation. The training samples are a set composed of the dataset and the label set. When the input first comprehensive feature vector and the training samples do not belong to the same process, the mapping parameters are calculated by the training dataset formed by the preset lower limit sample size in the process where the input data is located and the dataset through a preset linear transformation formula.

[0086] The second preset parameters in this embodiment include the number of input channels, the number of output channels, the convolution kernel size, and the padding size. These parameters also need to be strictly matched with the input and output data channels. Other hyperparameters can be specified as needed. The specific settings of the second preset parameters are well known to those skilled in the art and can be set or adjusted as needed. They will not be elaborated here.

[0087] Step 6: Construct a Weibull three-parameter mapping model based on a multilayer perceptron model, and use the third comprehensive feature vector as input and the corresponding surface roughness Ra as output to construct a linear regression analysis model based on the attention mechanism.

[0088] In this embodiment, the shape factor in the Weibull three-parameter mapping model is obtained by using the third comprehensive feature vector from step five as input and a multilayer perceptron model as the fitting function. Scale factor Location factor This leads to the formation of a Weibull three-parameter mapping model for predicting the remaining machining time of the tool. ;in This is the predicted remaining machining time for the cutting tool. Given a failure probability threshold, The current machining time of the tool, shape factor The output of a multilayer perceptron model with a logistic function as the activation function, and the scaling factor. The output of a multilayer perceptron model with an exponential activation function, location factor The output is from a multilayer perceptron model that uses a linear rectified function as its activation function.

[0089] Step 7: Acquire the three-axis vibration signal and spindle power signal during the integral bladed disk milling process, and perform time axis alignment, filtering and normalization preprocessing to obtain dimensionless signal data. Use the methods in steps 2 to 5 to obtain the corresponding third comprehensive feature vector. Based on the obtained third comprehensive feature vector, use the linear regression analysis model to analyze and obtain the surface roughness data of the integral bladed disk machining position after the tool passes, and use the Weibull three-parameter mapping model to analyze and obtain the predicted value of the remaining machining time of the tool.

[0090] The above are merely 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 within the protection scope of the present invention.

Claims

1. A flutter intelligent monitoring method for integral bladed disk machining, characterized in that, include: Step 1: Perform time axis alignment, filtering, and normalization preprocessing on the heterogeneous signals from multiple sensors during the integral bladed disk milling process to obtain dimensionless signal data; the heterogeneous signals include the triaxial vibration signal, spindle power signal, and surface roughness data of the integral bladed disk machining position after the tool passes through during the integral bladed disk machining process. Step 2: Using the sliding window sampling method, extract the 7-dimensional vibration time-domain statistics, the 8-dimensional vibration frequency band energy ratio, and the 3-dimensional power statistics in the main shaft power signal from the triaxial vibration signal within each sampling window to form a dataset. Pair the measured surface roughness Ra values ​​at the center time of the corresponding sampling window in the dataset as a label set. Step 3: Assign corresponding weights to all samples in the dataset and calculate the weighted covariance matrix of the dataset; perform symmetric eigenvalue decomposition on the weighted covariance matrix, sort the eigenvalues ​​obtained by decomposition in descending order, and calculate the ratio of the sum of the first K eigenvalues ​​to the sum of all eigenvalues. Extract the top K feature values ​​whose ratio is greater than or equal to a preset ratio threshold, and reduce the dimensionality of the dataset to the subspace corresponding to the top K feature values ​​to obtain the dimensionality-reduced dataset; Step 4: Perform wavelet packet decomposition on the triaxial vibration signal and the principal axis power signal, and use a one-dimensional convolutional neural network with the first preset parameter configuration to extract features from the decomposed data. After global average pooling of the extracted features, a multilayer perceptron with a logistic function as the activation function is used to calculate the gating weights of the dynamic fusion gate. The dimensionality-reduced dataset undergoes affine transformation and is then weighted and fused with local features to obtain the first comprehensive feature vector. Step 5: Map the first comprehensive feature vector using a transfer learning mapping model to obtain the transferred second comprehensive feature vector. Then, use a one-dimensional convolutional neural network and an LSTM network with second preset parameter configuration to extract features from the transferred second comprehensive feature vector, obtaining the corresponding third comprehensive feature vector. The transfer adaptation layer of the transfer learning mapping model performs a linear transformation. ,in, These are the features after transfer alignment. It is a transformation matrix. It is an offset vector, derived from , It is concluded that, among them It is the identity matrix. yes The mean, It is the first comprehensive feature vector set of the source domain data obtained through feature extraction and adaptation processing. covariance; It is the first comprehensive feature vector set of the target domain data. The mean, yes The covariance; when the input first comprehensive feature vector and the training sample belong to the same process, the output second comprehensive feature vector is the same as the input without any transformation, and the training sample is a set composed of the dataset and the label set; when the input first comprehensive feature vector and the training sample do not belong to the same process, the mapping parameter is calculated by the training dataset formed by the preset lower limit sample size in the process where the input data is located and the dataset through a linear transformation formula. Step 6: Construct a Weibull three-parameter mapping model based on a multilayer perceptron model. ;in This is the predicted remaining machining time for the cutting tool. Given a failure probability threshold, The current machining time of the tool, shape factor The output of a multilayer perceptron model with a logistic function as the activation function, and the scaling factor. The output of a multilayer perceptron model with an exponential activation function, location factor This method uses the output of a multilayer perceptron model with a linear rectified function as the activation function, and the third comprehensive feature vector as the input. It employs the multilayer perceptron model as the fitting function to analyze and obtain the Weibull three-parameter mapping model. , , And using the third comprehensive feature vector as input and the corresponding surface roughness Ra as output, a linear regression analysis model based on the attention mechanism is constructed. Step 7: Acquire the three-axis vibration signal and spindle power signal during the integral bladed disk milling process, and perform time axis alignment, filtering and normalization preprocessing to obtain dimensionless signal data. Use the methods in steps 2 to 5 to obtain the corresponding third comprehensive feature vector. Based on the obtained third comprehensive feature vector, use the linear regression analysis model to analyze and obtain the surface roughness data of the integral bladed disk machining position after the tool passes, and use the Weibull three-parameter mapping model to analyze and obtain the predicted value of the remaining machining time of the tool.

2. The flutter intelligent monitoring method according to claim 1, characterized in that, In step two, the 7-dimensional vibration time-domain statistics in the triaxial vibration signal include: mean, variance, root mean square, kurtosis, peak factor, impulse factor, and margin factor of the triaxial vibration signal modulus; the 8-dimensional vibration frequency band energy proportions in the triaxial vibration signal include: the energy proportions of the main spindle power signal in the eight sub-bands of 0–250Hz, 250–500Hz, 500–750Hz, 750–1000Hz, 1000–1250Hz, 1250–1500Hz, 1500–1750Hz, and 1750–2000Hz; the 3-dimensional power statistics in the main spindle power signal include the root mean square, peak value, and peak factor of the main spindle power signal.

3. The flutter intelligent monitoring method according to claim 1, characterized in that, In step three, the weighted covariance matrix is ​​subjected to symmetric eigenvalue decomposition, and the eigenvalues ​​obtained from the decomposition are sorted in descending order. The ratio of the sum of the first K eigenvalues ​​to the sum of all eigenvalues ​​is calculated. The method for extracting the first K eigenvalues ​​whose ratio is greater than or equal to a preset ratio threshold includes: The weighted covariance matrix of the dataset Perform symmetric eigenvalue decomposition, such that... ,in The matrix consists of orthogonal eigenvectors. , The first diagonal line of the diagonal matrix after symmetric eigenvalue decomposition. One element, , The dimensions of the diagonal matrix are arranged in descending order. ≥ ≥…≥ ≥0; The first K eigenvalues ​​are selected if the ratio of the sum of the first K eigenvalues ​​to the sum of all eigenvalues ​​is greater than or equal to 90%.

4. The flutter intelligent monitoring method according to claim 3, characterized in that, In step three, the dataset is reduced in dimensionality to the subspace corresponding to the first K feature values, resulting in the dimensionality-reduced dataset. ,in For the dataset, for The weighted mean, for The matrix formed by the first K column vectors.

5. The flutter intelligent monitoring method according to claim 1, characterized in that, In step five, the method for mapping the first comprehensive feature vector using a transfer learning mapping model to obtain the transferred second comprehensive feature vector includes: When the input first comprehensive feature vector and the training samples belong to the same process, the output is the same as the input without any transformation. The training samples are a set composed of the dataset and the label set. When the input first comprehensive feature vector and the training samples do not belong to the same process, the mapping parameters are calculated by the training dataset formed by the preset lower limit sample size in the process where the input data is located and the dataset through a preset formula.

6. A flutter intelligent monitoring system for integral bladed disk machining, used to implement the flutter intelligent monitoring method according to any one of claims 1-5, characterized in that, include: The data preprocessing module is used to perform time axis alignment, filtering and normalization preprocessing on the heterogeneous signals from multiple sensors during the integral bladed disk milling process to obtain dimensionless signal data; the heterogeneous signals include triaxial vibration signals, spindle power signals and surface roughness data of the integral bladed disk machining position after the tool passes through during the machining process. The data extraction module is used to extract the 7-dimensional vibration time-domain statistics, the 8-dimensional vibration frequency band energy ratio, and the 3-dimensional power statistics in the main shaft power signal from the triaxial vibration signal in each sampling window using the sliding window sampling method to form a dataset, and pair the measured surface roughness Ra values ​​at the center time of the corresponding sampling window in the dataset as a label set; The data dimensionality reduction module is used to assign corresponding weights to all samples in the dataset and calculate the weighted covariance matrix of the dataset; perform symmetric eigenvalue decomposition on the weighted covariance matrix, sort the eigenvalues ​​obtained by decomposition in descending order, and calculate the ratio of the sum of the first K eigenvalues ​​to the sum of all eigenvalues. Extract the top K feature values ​​whose ratio is greater than or equal to a preset ratio threshold, and reduce the dimensionality of the dataset to the subspace corresponding to the top K feature values ​​to obtain the dimensionality-reduced dataset; The first feature extraction module is used to perform wavelet packet decomposition on the triaxial vibration signal and the main shaft power signal, and to extract features from the decomposed data using a one-dimensional convolutional neural network with a first preset parameter configuration. After global average pooling processing of the extracted features, the features after global average pooling processing are fused with the dimensionality-reduced dataset using a dynamic gate weighted fusion method to obtain the first comprehensive feature vector. The second feature extraction module is used to map the first comprehensive feature vector using a transfer learning mapping model to obtain the transferred second comprehensive feature vector, and to extract features from the transferred second comprehensive feature vector using a one-dimensional convolutional neural network and an LSTM network with second preset parameters to obtain the corresponding third comprehensive feature vector. The model building module is used to build a Weibull three-parameter mapping model based on a multilayer perceptron model, and to build a linear regression analysis model based on an attention mechanism with the third comprehensive feature vector as input and the corresponding surface roughness Ra as output. The analysis and prediction module is used to obtain the third comprehensive feature vector during the milling process of the integral bladed disk, and based on the obtained third comprehensive feature vector, to analyze and obtain the surface roughness data of the integral bladed disk machining position after the tool passes through the linear regression analysis model, and to analyze and obtain the predicted value of the remaining machining time of the tool using the Weibull three-parameter mapping model.

Images