A method for condition assessment of aerospace manufacturing equipment

Abstract

A method for assessing the condition of aerospace manufacturing equipment is proposed. Sensors are installed between the workpiece and the platform to collect and process signals. The Whale Optimization Algorithm is used to optimize the parameters of a support vector machine model. Multi-dimensional comprehensive optimization is performed on the characteristics of the collected and processed signals in different time, frequency, and time-frequency domains to obtain a WOA-SVM regression model and a classification model. This method addresses the technical problem that many existing models have poor performance in assessing tool wear condition. The method has the characteristics of fast convergence speed, strong global search capability, and improved accuracy of mechanical fault assessment.

Description

Technical Field

[0001] The present invention relates to the technical field of aerospace processing, and in particular to a method for evaluating the state of aerospace manufacturing equipment. Background Art

[0002] Aerospace components require high precision and tight geometric tolerances, and utilize a large number of materials with excellent corrosion and high-temperature resistance. These materials, characterized by high plasticity, high viscosity, and difficulty in machining, place extremely high demands on tool quality and performance. The condition of aerospace manufacturing equipment significantly impacts the quality of high-precision aerospace parts, making condition assessment and monitoring of manufacturing equipment of great significance.

[0003] Currently, aerospace manufacturing equipment operating status monitoring is generally divided into direct and indirect methods. For example, direct methods determine the current status of cutting tools by monitoring changes in the manufacturing equipment's dimensions or surface texture. This method is complex and difficult to implement online. Indirect methods, however, utilize variable parameters related to tool wear during the cutting process to indirectly determine tool wear status, making it more convenient to implement real-time tool wear monitoring in actual aerospace production.

[0004] Cutting tools are the direct actuators of CNC machine tools. Due to direct contact with the workpiece during machining, they are subject to intense friction and impact. Consequently, worn cutting tools significantly impact part surface quality. Research shows that tool wear accounts for 20% of all machine downtime. Effective tool condition monitoring methods can improve machining efficiency by 10% to 50% and reduce production costs by 10% to 40%.

[0005] Chinese patent publication CN 112036296 A discloses a motor bearing fault diagnosis method based on a generalized S transform and WOA-SVM. The S transform is performed on the input bearing vibration signal to obtain a sample set of time-domain and frequency-domain feature vectors, which are then input into the WOA-SVM for testing and fault type determination. This overcomes the drawback of the S transform's Gaussian window function, which lacks flexibility due to its inability to adjust with frequency, and offers improved time-frequency analysis capabilities.

[0006] Chinese patent document CN 108319962 B discloses a tool wear monitoring method based on a convolutional neural network. By collecting the original vibration signal of the workpiece, extracting the energy spectrum through wavelet packets, and inputting it into the convolutional neural network for model training and testing, it completely overcomes the shortcomings of shallow models in convergence speed and optimization methods.

[0007] Currently, most manufacturing equipment wear status detection methods are based on the fusion of collected signal features, without multi-dimensional comprehensive optimization of features in different time domains, frequency domains, and time-frequency domains, resulting in low evaluation accuracy. Summary of the Invention

[0008] In order to overcome the defects of the above-mentioned prior art, the purpose of the present invention is to provide a method for evaluating the status of aerospace manufacturing equipment, which adopts the whale optimization algorithm to optimize the support vector machine model parameters to obtain the WOA-SVM regression model and classification model, so as to solve the technical problem that many existing models have poor evaluation effect when applied to tool wear status.

[0009] In order to achieve the above object, the technical solution of the present invention is:

[0010] A method for evaluating the condition of aerospace manufacturing equipment comprises the following steps:

[0011] Step 1: Install a three-axis cutting force sensor between the workpiece and the platform to measure the cutting force signals in the x, y, and z directions. Place three vibration sensors in the z direction of the spindle and in the x and y directions perpendicular to the spindle on the side of the workpiece near the bottom to measure the vibration signals in the x, y, and z directions respectively. Install an acoustic emission signal sensor on the side of the workpiece near the bottom to measure the sound signal.

[0012] Step 2: Collect three-dimensional cutting force signals, vibration signals, and acoustic emission signals between the manufacturing equipment and the workpiece under different wear conditions. The signals are evenly divided into multiple samples at a sampling frequency of 30 kHz to 60 kHz. Each manufacturing equipment wear state has a data set consisting of m sampling points. The data set is then labeled in ascending order of wear as 1, 2, 3, ..., n, for a total of n channels of data.

[0013] Step 3: Preprocess the collected signal, including effective data interception, Hampel filtering to remove burrs, and wavelet noise reduction to obtain a relatively pure signal;

[0014] Among them, the effective data interception method is as follows: the values ​​of each channel are arranged from small to large, and all the data are divided into four parts. The numbers in the first 25% position are taken as the upper quartile value, recorded as Q1, and the numbers in the 75% position are taken as the lower quartile value, recorded as Q3, and the interquartile range IQR = Q3-Q1; the outlier judgment basis is: greater than Q1+k*IQR or less than Q3-k*IQR, k=1.5; according to the collected data time sequence, find the first data greater than or equal to the upper quartile value, identify the data before it as feed data, delete all of them, search from the end of the data forward to the first data greater than or equal to the lower quartile value, identify it as retract data, delete all of it, and the remaining is the normal milling sensor data;

[0015] Hampel filter deburring method: First take a window of length k and calculate the median m of all data x in this window i and standard deviation δ i, set the threshold to 3, if xm i |>3δ i , just use m i The value of replace x;

[0016] Wavelet denoising method: select wavelet signals such as sym and db wavelet functions, perform p-layer wavelet decomposition on the input signal after hampel filtering, and obtain cd1, cd2, cd3...cdp, cap; secondly, the threshold function uses the hard threshold method, w is the original wavelet coefficient, λ is the given threshold, and finally cd1, cd2, cd3…cdp, cap are reconstructed into the signal after removing high-frequency clutter;

[0017] Step 4: Normalize and extract features of the preprocessed data;

[0018] Normalization calculation formula: x is each time series data, x min is the minimum value of each channel data, x max is the maximum value of each channel data, and y is the normalized data;

[0019] Feature extraction process: Feature extraction includes the extraction process of three types of features: time domain, frequency domain, and time-frequency domain; time domain features include peak value, average value, root mean square value, root square amplitude, kurtosis, skewness, peak index, pulse index, waveform index, margin index, kurtosis index, and skewness index; frequency domain features include center of gravity frequency, frequency variance, and mean square frequency; time-frequency domain features include wavelet energy characteristics of each sub-band after wavelet decomposition;

[0020] Step 5: Screen the extracted features;

[0021] Use principal component analysis (PCA) to calculate the characteristic sample X containing m variables and n data. n×m The co-matrix difference ∑ m×m , according to the co-array difference, find m eigenvalues ​​and sort them from small to large to get λ1,λ2...λ m and its corresponding m eigenvectors T1, T2...T m The original sample of m variables is reduced to p variables after principal component analysis. The number of principal components is related to the principal component contribution rate and the cumulative contribution rate. The contribution rate of the kth principal component is The greater the contribution rate of the principal component, the more information of the original data is preserved; the cumulative contribution rate of the first q principal components of the sample is If the contribution rate reaches more than 80%, the first q principal components are selected;

[0022] Step 6: Use the whale optimization algorithm to optimize the hyperparameters of the support vector machine model training to obtain the WOA-SVM regression model and classification model;

[0023] The whale optimization algorithm consists of three steps: surrounding prey - bubble net hunting - searching for prey;

[0024] S1 surrounds prey:

[0025] Mathematical model: D=|CX * (t)-X(t)|,X(t+1)=X * (t)-AD, t is the current iteration number, A and C are the coefficients, X * (t) represents the best whale position vector so far, X(t) represents the current whale position vector; A and C are obtained by the following formula: A = 2ar1-a, C = 2r2, Where r1 and r2 are random numbers between 0 and 1, t represents the current iteration number, T max is the maximum number of iterations;

[0026] S2 Bubble Net Predation: Humpback whales swim towards their prey in a spiral motion. The mathematical model is X(t+1)=X*(t)+D p e bl cos(2πl), where D p =|X * (t)-X(t)| represents the distance between the whale and the prey, b is a constant that defines the shape of the spiral, l is a random number in (-1,1), and X * (t) represents the best position vector so far; while the whale is swimming towards its prey in a spiral, it needs to shrink the encirclement. Assuming P i The probability of choosing the shrinking and surrounding mechanism and 1-P i The probability of choosing the spiral model to update the whale position, mathematical model: As the number of iterations t increases, the parameter A and the convergence factor a gradually decrease. If |A|<1, the whales gradually surround the current optimal solution, which is the local optimization stage in WOA.

[0027] S3 search prey:

[0028] Mathematical model D=|CX rand -X(t)|,X(t+1)=X rand -AD, where X rand A randomly selected whale position vector. When |A|≥1, a search agent is randomly selected to update the positions of other whales based on the randomly selected whale position, forcing the whale to deviate from the prey and find a more suitable prey. This can enhance the algorithm's exploration ability and enable the WOA algorithm to perform global search.

[0029] Step 7: Use the reduced dimensionality features as input features and use the pre-trained whale optimized support vector machine regression model to predict the wear value of the manufacturing equipment;

[0030] Step 8: Classify the tool wear status according to the manufacturing equipment wear value in advance, and divide it into corresponding different wear states according to the different manufacturing equipment wear values. Compare it with the tool wear status output by the WOA-SVM classification model to determine the tool wear status.

[0031] Compared with the prior art, the advantages of the present invention are:

[0032] 1. Step six of the present invention uses the whale optimization algorithm to optimize the penalty parameters of the support vector machine, effectively suppressing overfitting, and can adaptively adjust the gamma value of the support vector machine, effectively controlling the problem of excessive decision boundaries. The classification effect is best when the number of classification categories is greater than 120. Compared with SVM, better results can be achieved under the condition of unbalanced data sets.

[0033] 2. The tool state assessment method of the support vector machine based on the whale optimization algorithm of the present invention performs multi-dimensional comprehensive optimization on the characteristics of the collected signals in different time domains, frequency domains, and time-frequency domains. It is compatible with wavelet packet decomposition and Hilbert-Huang transform. It converges quickly on the PHM2010 milling cutter wear dataset and has strong global search capabilities. The trained model has an accuracy improvement of about 5% compared with the SVM model. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 It is a flow chart of the present invention.

[0035] Figure 2 This is the flowchart of the whale optimization algorithm. DETAILED DESCRIPTION

[0036] The specific embodiments of the present invention are described in further detail below with reference to the accompanying drawings.

[0037] Reference Figure 1 ,A tool condition evaluation method based on a support vector machine ,tool condition assessment method based on an improved whale ,optimization algorithm, includes the following steps:

[0038] Step 1: Install a three-axis cutting force sensor between the workpiece and the platform to measure the cutting force signals in the x, y, and z directions. Place three vibration sensors in the z direction of the spindle and in the x and y directions perpendicular to the spindle on the side of the workpiece near the bottom to measure the vibration signals in the x, y, and z directions respectively. Install an acoustic emission signal sensor on the side of the workpiece near the bottom to measure the sound signal.

[0039] Step 2: Collect three-dimensional cutting force signals, vibration signals, and acoustic emission signals between the manufacturing equipment and the workpiece under different wear conditions. The signals are evenly divided into eight samples at a sampling frequency of 50 kHz. Each tool wear condition has a data set consisting of three sampling points. The data set is then labeled in ascending order of wear as 1, 2, 3, ..., for a total of seven channels of data.

[0040] Step 3: Preprocess the collected signal, including effective data interception, Hampel filtering to remove burrs, and wavelet noise reduction to obtain a relatively pure signal;

[0041] Among them, the effective data interception method is as follows: the values ​​of each channel are arranged from small to large, and all the data are divided into 4 parts. The numbers in the first 25% position are taken as the upper quartile value, recorded as Q1, and the numbers in the 75% position are taken as the lower quartile value, recorded as Q3, and the interquartile range IQR = Q3-Q1; the outlier judgment basis is: greater than Q1+k*IQR or less than Q3-k*IQR, k=1.5; according to the collected data time sequence, find the first data greater than or equal to the upper quartile value, identify the data before it as feed data, delete all of them, search from the end of the data forward to the first data greater than or equal to the lower quartile value, identify it as retract data, delete all of it, and the remaining is the normal milling sensor data;

[0042] Hampel filter deburring method: First take a window of length k and calculate the median m of all data x in this window i and standard deviation δ i , set the threshold to 3, if |xm i |>3δ i Just use m i The value of replace x;

[0043] Wavelet denoising method: The wavelet signal sym wavelet function is selected to perform three-layer wavelet decomposition on the input signal after hampel filtering to obtain cd1, cd2, cd3, cd3, ca3; secondly, the threshold function uses the hard threshold method, w is the original wavelet coefficient, λ is the given threshold, and finally cd1, cd2, cd3, cd3, ca3 are reconstructed into the signal after removing high-frequency clutter;

[0044] Step 4: Normalize and extract features of the preprocessed data;

[0045] Normalization calculation formula: x is each time series data, x min is the minimum value of each channel data, x max is the maximum value of each channel data, and y is the normalized data;

[0046] The feature extraction process involves extracting three types of features: time domain, frequency domain, and time-frequency domain. Time domain features include peak value, average value, root mean square value, root square amplitude, kurtosis, skewness, peak index, pulse index, waveform index, margin index, kurtosis index, and skewness index. Frequency domain features include centroid frequency, frequency variance, and mean square frequency. Time-frequency domain features include the wavelet energy characteristics of each sub-band after wavelet decomposition.

[0047] Step 5: Screen the extracted features;

[0048] For the time domain characteristic index, principal component analysis (PCA) is used to calculate the co-array difference of the characteristic samples containing 12 variables and 100,000 data. Based on the co-array difference, 12 eigenvalues ​​are obtained and sorted from small to large to obtain λ1,λ2…λ 12 and its corresponding 12 eigenvectors T1, T2…T 12 ,After principal component analysis, the 12 variables in the original sample were reduced to 7 variables, and the cumulative contribution rate of the first 7 principal components of the sample reached more than 80%, so the first 7 principal components were selected;

[0049] Similarly, for the time domain and time-frequency domain feature indicators, 2 and 5 principal components are selected respectively.

[0050] Step 6: Use the whale optimization algorithm to optimize the hyperparameters of the support vector machine model training to obtain the WOA-SVM regression model and classification model;

[0051] The whale optimization algorithm is an algorithm based on the whale's hunting method, which is called the "bubble net foraging method". Figure 2 , mainly includes three steps: surrounding prey - bubble net hunting - searching for prey.

[0052] S1 surrounds prey:

[0053] Mathematical model: D=|CX * (t)-X(t)|,X(t+1)=X * (t)-AD, t is the current iteration number, A and C are the coefficients, X * (t) represents the best whale position vector so far, X(t) represents the current whale position vector; A and C are obtained by the following formula: A = 2ar1-a, C = 2r2, Where r1 and r2 are random numbers between 0 and 1, t represents the current iteration number, T max is the maximum number of iterations.

[0054] S2 Bubble Net Predation: Humpback whales swim towards their prey in a spiral motion. The mathematical model is X(t+1)=X*(t)+D p e blcos(2πl), where D p =|X * (t)-X(t)| represents the distance between the whale and the prey, b is a constant that defines the shape of the spiral, l is a random number in (-1,1), and X * (t) represents the best position vector so far; while the whale swims towards its prey in a spiral, it also needs to shrink the encirclement. Assuming P i The probability of choosing the shrinking and surrounding mechanism and 1-P i The probability of choosing the spiral model to update the whale position, mathematical model: As the number of iterations t increases, the parameter A and the convergence factor a gradually decrease. If |A|<1, the whales gradually surround the current optimal solution, which belongs to the local optimization stage in WOA.

[0055] S3 search prey:

[0056] Mathematical model D=|CX rand -X(t)|,X(t+1)=X rand -AD, where X rand is a randomly selected whale position vector. When |A|≥1, a search agent is randomly selected to update the positions of other whales according to the randomly selected whale position, forcing the whales to deviate from the prey and find a more suitable prey. This can enhance the exploration ability of the algorithm and enable the WOA algorithm to perform global search.

[0057] Step 7: Use the reduced dimension features as input features and use the pre-trained whale optimized support vector machine regression model to predict the output of tool wear value;

[0058] Step 8: The tool wear state is divided into three different wear states according to the different tool wear VB values: the initial wear VB value is 0-0.1mm, the normal wear VB value is 0.1-0.15mm, and the severe wear VB value is above 0.15mm. The tool wear state is compared with the tool wear state output by the WOA-SVM classification model to determine the tool wear state.

[0059] The support vector machine tool state assessment method of the whale optimization algorithm of the present invention is based on the mathematical model of the whale optimization algorithm of surrounding prey-bubble net predation-searching prey. By performing multi-dimensional comprehensive optimization on the characteristics of the collected signals in different time domains, frequency domains, and time-frequency domains, the method has fast convergence speed and strong global search capability, thereby improving the accuracy of mechanical fault assessment.

Claims

1. A method for evaluating the condition of aerospace manufacturing equipment, characterized in that: The following steps are involved: Step 1: Install a three-axis cutting force sensor between the workpiece and the platform to measure the cutting force signals in the x, y, and z directions. Place three vibration sensors in the z direction of the spindle and in the x and y directions perpendicular to the spindle on the side of the workpiece near the bottom to measure the vibration signals in the x, y, and z directions respectively. Install an acoustic emission signal sensor on the side of the workpiece near the bottom to measure the sound signal. Step 2: Collect three-dimensional cutting force signals, vibration signals, and acoustic emission signals between the manufacturing equipment and the workpiece under different wear conditions. The signals are evenly divided into multiple samples at a sampling frequency of 30 kHz to 60 kHz. Each manufacturing equipment wear state has a data set consisting of m sampling points. The data set is then labeled in ascending order of wear as 1, 2, 3, ..., n, for a total of n channels of data. Step 3: Preprocess the collected signal, including effective data interception, Hampel filtering to remove burrs, and wavelet noise reduction to obtain a relatively pure signal; Step 4: Normalize and extract features of the preprocessed data; Step 5: Screen the extracted features; Step 6: Use the whale optimization algorithm to optimize the hyperparameters of the support vector machine model training to obtain the WOA-SVM regression and classification models; The whale optimization algorithm consists of three steps: surrounding prey - bubble net hunting - searching for prey: S1 surrounds prey: Mathematical model: D=|CX * (t)-X(t)|,X(t+1)=X * (t)-AD, t is the current iteration number, A and C are the coefficients, X * (t) represents the best whale position vector so far, X(t) represents the current whale position vector; A and C are obtained by the following formula: A = 2ar1-a, C = 2r2, Where r1 and r2 are random numbers between 0 and 1, t represents the current iteration number, T max is the maximum number of iterations; S2 Bubble Net Predation: Humpback whales swim towards their prey in a spiral motion. The mathematical model is X(t+1)=X*(t)+D p e bl cos(2πl), where D p =|X * (t)-X(t)| represents the distance between the whale and the prey, b is a constant that defines the shape of the spiral, l is a random number in (-1, 1), and X * (t) represents the best position vector so far; while the whale swims towards its prey in a spiral, it also needs to shrink the encirclement. Assuming P i The probability of choosing the shrinking and surrounding mechanism and 1-P i The probability of choosing the spiral model to update the whale position, mathematical model: As the number of iterations t increases, the parameter A and the convergence factor a gradually decrease. If |A|<1, the whales gradually surround the current optimal solution, which is the local optimization stage in WOA. S3 search prey: Mathematical model D=|CX rand -X(t)|,X(t+1)=X rand -AD, where X rand A randomly selected whale position vector. When |A|≥1, a search agent is randomly selected to update the positions of other whales based on the randomly selected whale position, forcing the whale to deviate from the prey and find a more suitable prey. This can enhance the algorithm's exploration ability and enable the WOA algorithm to perform global search. Step 7: Use the reduced dimensionality features as input features and use the trained whale optimized support vector machine regression model to predict the wear value of the manufacturing equipment. Step 8: Classify the tool wear status according to the manufacturing equipment wear value in advance, and divide it into corresponding different wear states according to the different manufacturing equipment wear values. Compare it with the tool wear status output by the WOA-SVM classification model to determine the tool wear status.

2. The method for evaluating the state of aerospace manufacturing equipment according to claim 1, wherein: The effective data interception method in the step three is as follows: arrange the values ​​of each channel from small to large, divide all the data into four parts, take the numbers in the first 25% position as the upper quartile value, recorded as Q1, and the numbers in the 75% position as the lower quartile value, recorded as Q3, and the quartile distance IQR = Q3-Q1; the basis for judging abnormal values: greater than Q1+k*IQR or less than Q3-k*IQR, k=1.5; find the first data greater than or equal to the upper quartile value according to the collected data time sequence, identify the data before it as feed data, delete all of them, search from the end of the data to the first data greater than or equal to the lower quartile value, identify it as retract data, delete all of them, and the rest is the sensor data for normal milling.

3. The method for evaluating the state of aerospace manufacturing equipment according to claim 1, wherein: The hampel filter deburring method in step 3: first take a window of length k and calculate the median m of all data x in this window i and standard deviation δ i , set the threshold to 3, if |xm i |>3δ i , just use m i The value of substituted for x.

4. The method for evaluating the state of aerospace manufacturing equipment according to claim 1, wherein: In the step 3, the wavelet denoising method comprises: selecting a wavelet signal such as sym or db wavelet function, performing p-layer wavelet decomposition on the input signal after hampel filtering, and obtaining cd1, cd2, cd3 ... cdp, cap; Secondly, the threshold function uses the hard threshold method. w is the original wavelet coefficient, λ is the given threshold, and finally cd1, cd2, cd3…cdp, cap are reconstructed into the signal after removing high-frequency clutter.

5. The method for evaluating the state of aerospace manufacturing equipment according to claim 1, wherein: The step 4 is specifically as follows: Normalization calculation formula: x is each time series data, x min is the minimum value of each channel data, x max is the maximum value of each channel data, and y is the normalized data; Feature extraction process: Feature extraction includes the extraction process of three types of features: time domain, frequency domain, and time-frequency domain; Time domain features include peak value, average value, root mean square value, root square amplitude, kurtosis, skewness, peak index, pulse index, waveform index, margin index, kurtosis index, and skewness index; frequency domain features include center of gravity frequency, frequency variance, and mean square frequency; time-frequency domain features include wavelet energy characteristics of each sub-band after wavelet decomposition.

6. The method for evaluating the state of aerospace manufacturing equipment according to claim 1, wherein: The step five is specifically as follows: Use principal component analysis (PCA) to calculate the characteristic sample X containing m variables and n data. n×m The co-matrix difference ∑ m×m , according to the co-array difference, find m eigenvalues ​​and sort them from small to large to get λ1, λ2…λ m and its corresponding m eigenvectors T1, T2…T m , the original sample m variables are reduced to p variables after principal component analysis. The number of principal components selected is related to the principal component contribution rate and the cumulative contribution rate; the kth principal component contribution rate The greater the contribution rate of the principal component, the more information of the original data is preserved; the cumulative contribution rate of the first q principal components of the sample is When the contribution rate reaches more than 80%, the first q principal components are selected.

Images