Friction deposition forming defect detection method based on three-dimensional force time-frequency domain characteristics

Abstract

The application discloses a friction deposition forming defect detection method based on three-dimensional force time-frequency domain features, adopts a trained defect detection model to detect defects of samples to obtain a prediction result, wherein the defect detection model is trained by using samples in a data set; a data set is obtained by the following method: collecting original three-dimensional force signals; obtaining decomposed three-dimensional force signals and preprocessing the same; dividing three-dimensional force signals in a stable deposition stage to obtain a plurality of sample units; calculating features of each signal curve in each sample unit; screening strong correlation features and weak correlation features; screening N3 strong correlation features from the strong correlation features; obtaining a final feature set H according to information gains of the N3 strong correlation features and the weak correlation features; and constructing a data set. The friction deposition forming defect detection method reduces dimensionality of the number of force signal features, removes redundant information, improves the connection between features and defects, and improves the accuracy of the prediction result.

Description

Technical Field

[0001] This invention belongs to the field of additive manufacturing quality monitoring technology, specifically relating to a method for detecting defects in friction deposition forming based on three-dimensional force time-frequency domain characteristics. Background Technology

[0002] In recent years, friction deposition forming (FDF) has attracted much attention as a novel solid-state additive manufacturing method. During FDF, the deposited material undergoes plastic deformation under the action of extrusion and friction, forming a metallurgical bond with the substrate or the previous deposited layer. Since the deposition process does not involve the melting and solidification of metals, metallurgical defects such as porosity and cracks can be effectively avoided. However, if the parameters of the friction deposition process are not properly selected, internal defects such as weak bonds and grooves, as well as surface defects, can still occur, thus affecting the quality of the deposited part. Currently, commonly used non-destructive testing methods mainly include: ultrasonic testing (UT), X-ray computed tomography (X-CT), and real-time monitoring methods based on sensor signals (such as temperature and mechanical signal monitoring).

[0003] Ultrasonic testing can effectively identify pores and cracks inside deposited parts, and has the advantages of being non-destructive and convenient. However, because ultrasonic testing has high requirements for the material surface and is difficult to test deposited parts with complex geometries, it is difficult to use it for identifying friction deposition defects.

[0004] X-ray computed tomography (CT) can accurately display the three-dimensional defect distribution inside deposited parts and provide high spatial resolution. However, CT is expensive and the scanning and data analysis are time-consuming, making it difficult to achieve online real-time detection.

[0005] Among real-time process monitoring methods, mechanical signal monitoring is the most widely used. Force signals contain richer information about material flow and forming, and can be measured and acquired online in real time. In the field of friction stir welding (FSW), researchers have applied force signals to detect defect types. FSW is a solid-state joining technology that uses frictional heat generated between a high-speed rotating stirring head and the workpiece to create plastic flow and achieve material bonding. Compared with the FSW process, FDF not only involves solid-state bonding at the material interface but also includes the plasticization of the bar stock, material flow, and layer-by-layer deposition, making the force signals in the FDF process more complex, dynamic, and exhibiting distinct stage characteristics. To date, the force signals measured in FDF have a relatively single dimension and cannot fully reflect the information contained in the deposition process. In the FDF process, there are many features in the force signals; therefore, dimensionality reduction processing is needed during analysis to eliminate redundant features and improve analysis efficiency.

[0006] Current dimensionality reduction methods, such as Principal Component Analysis (PCA), are unsupervised linear dimensionality reduction methods. This method fails to fully consider the relationship between each feature and the classification target, and retains some information that is irrelevant to the discrimination task. The principal components obtained after dimensionality reduction are linear combinations of the original features, lacking intuitive physical interpretation and having a relatively vague physical meaning. Summary of the Invention

[0007] In view of the shortcomings of existing technologies, the purpose of this invention is to provide a defect detection model.

[0008] Another objective of this invention is to provide a method for detecting defects in triboelectric deposition forming based on three-dimensional force time-frequency domain characteristics. This method enables the prediction of the presence or absence of defects in the deposited part and the classification and prediction of defects in the deposited part.

[0009] Another object of the present invention is to provide the use of the above-mentioned defect detection model for real-time defect detection in triboelectric deposition forming.

[0010] This invention is achieved through the following technical solution.

[0011] A defect detection model, wherein each input sample includes: the value of a feature in the final feature set H corresponding to a sample unit; the method for obtaining the final feature set H includes:

[0012] S1, Prepare the original three-dimensional force signal. The original three-dimensional force signal is the signal of the force between the workpiece and the spindle in three mutually perpendicular directions during the same time period in the process of friction deposition forming.

[0013] S2. Variational mode decomposition is performed on the original three-dimensional force signal to obtain three sets of intrinsic mode functions (IMFs). Each IMF set includes K IMFs. The IMF with the lowest center frequency in each IMF set is selected as the selected IMF. The three selected IMFs are combined and used as the decomposed three-dimensional force signal. Gaussian smoothing is then applied to obtain three signal curves as the preprocessed three-dimensional force signal.

[0014] S3, the three-dimensional force signal is divided into multiple sample units according to a fixed window size, and the multiple sample units form a sample set D; the features of each signal curve in each sample unit in the sample set D are calculated, and M features are obtained for each sample unit;

[0015] S4. Take each type of feature of all sample units in the sample set D as a variable, and calculate the correlation coefficient between any two variables. Set a correlation threshold. If the correlation coefficient between a certain variable and the remaining M-1 variables is less than or equal to the correlation threshold, then the type of feature corresponding to that variable is regarded as a weakly correlated feature; otherwise, it is regarded as a strongly correlated feature, resulting in N strongly correlated features and N2 weakly correlated features.

[0016] Hierarchical clustering is used to divide N1 strongly correlated features into multiple clusters; a greedy algorithm is used to filter the strongly correlated features in each cluster, resulting in a total of N3 strongly correlated features after filtering; the N3 strongly correlated features obtained after filtering and N2 weakly correlated features form a feature set;

[0017] Calculate the information gain of each feature in the feature set, select features in the feature set whose information gain value is greater than or equal to 0.01, and form the final feature set H.

[0018] In the above technical solution, the characteristics of each signal curve include one or more of the following: maximum value (max), minimum value (min), mean value (mean), median value (med), peak-to-peak value (ran), average absolute value (arv), standard deviation (std), kurtosis (Kur), skewness (Ske), root mean square (Rms), root mean square amplitude (Srma), waveform factor (Wf), peak factor (Pf), impulse factor (If), margin factor (Cf), centroid frequency (FC), root mean square frequency (Rmsf), and frequency standard deviation (Rvf).

[0019] In the above technical solution, in step S4, the input for hierarchical clustering is the correlation coefficient of N1 strongly correlated features that are greater than the correlation threshold.

[0020] In the above technical solution, in step S4, the input of the greedy algorithm is the strongly correlated features within the selected cluster and the correlation coefficient greater than the correlation threshold corresponding to each strongly correlated feature within the cluster. The output of the greedy algorithm is at least one strongly correlated feature that is representative of the cluster.

[0021] In the above technical solution, the greedy algorithm is as follows: traverse each cluster, iterate each cluster according to the following cyclic steps until the correlation coefficients in the coefficient set of the cluster are 0, and form a strong correlation feature representative set of all selected strong correlation features during the iteration process. The strong correlation feature representative set obtained for all clusters is the N3 strong correlation features; wherein, the cyclic steps are: form a coefficient set of correlation coefficients of the strong correlation features in the cluster, select the strong correlation feature with the most correlation coefficients in the coefficient set as the selected strong correlation feature, delete the correlation coefficient corresponding to the selected strong correlation feature from the coefficient set, and delete the selected strong correlation feature from the cluster.

[0022] In the above technical solution, the defect detection model is constructed using the random forest algorithm.

[0023] In the above technical solution, the five-fold cross-validation method is used to train the defect detection model, and the grid search method is used to optimize the parameters of the defect detection model.

[0024] In the above technical solution, the output of the defect detection model is the prediction result. When the defect detection model is used for binary classification, the prediction result is no defect or defective. The true value of the sample is whether the deposited part corresponding to the sample unit is defective or defective.

[0025] When the defect detection model is used for three-class classification, the prediction result is good forming, minor defects, or serious defects. The true value of the sample is that the deposited part corresponding to the sample unit has no defects, the deposited part has minor defects, or the deposited part has serious defects.

[0026] A method for detecting defects in triboelectric deposition forming based on three-dimensional force time-frequency domain features is proposed. The method uses the defect detection model trained above to detect defects in the sample to be tested and obtains the prediction results corresponding to the sample to be tested.

[0027] The above-mentioned defect detection model is used for real-time defect detection in triboelectric deposition forming.

[0028] The present invention has the following advantages due to the adoption of the above technical solutions:

[0029] 1. The defect detection model constructed in this invention is based on the original three-dimensional force signal during the triboelectric deposition process of the deposited part, which realizes real-time monitoring and evaluation of the forming quality of the deposited part during the triboelectric deposition process. This not only improves the intelligence level of the triboelectric deposition process, but also provides solid theoretical and technical support for online parameter adjustment and defect control.

[0030] 2. The triboelectric deposition forming defect detection method of the present invention performs dimensionality reduction processing on the number of features of the force signal, ensuring that the 21 types of features after dimensionality reduction have clear physical meaning and effectively remove redundant information, ensuring sufficient preservation of discrimination information, further improving the connection between features and defects, providing a basis for the prediction of the defect detection model, and improving the accuracy of the prediction results of the defect detection model. Attached Figure Description

[0031] Figure 1 This is a flowchart of Example 1;

[0032] Figure 2 A schematic diagram of the triboelectric deposition process for forming the deposited part;

[0033] Figure 3 The images show the original three-dimensional force signal and the decomposed three-dimensional force signal.

[0034] Figure 4 The resulting clustering tree diagram output by hierarchical clustering;

[0035] Figure 5 The result is a graph showing the information gain values ​​for the features.

[0036] Figure 6 This is the prediction result of the present invention;

[0037] Figure 7 The image shows the confusion matrix and ROC curve corresponding to the prediction results of Example 6. Detailed Implementation

[0038] The following is a detailed description of the triboelectric deposition forming defect detection method based on three-dimensional force time-frequency domain characteristics of the present invention, with reference to the accompanying drawings.

[0039] In the following embodiments, the definition of a minor defect in the deposited part is: in a cross-section of the deposited layer perpendicular to the direction of the deposited part, the number of pixels in the defective portion of the deposited layer cross-section accounts for less than 20% of the total number of pixels in the deposited layer cross-section; the definition of a severe defect in the deposited part is: in a cross-section of the deposited layer perpendicular to the direction of the deposited part, the number of pixels in the defective portion of the deposited layer cross-section accounts for more than 20% of the total number of pixels in the deposited layer cross-section.

[0040] In the four three-dimensional force sensors, each three-dimensional force sensor collects the three-directional force signals between the workpiece and the spindle. The three-directional force signals collected by the four three-dimensional force sensors are superimposed to obtain the original three-dimensional force signal.

[0041] Example 1

[0042] like Figure 1 As shown, a method for obtaining a dataset includes:

[0043] S1, four force sensors are respectively installed at the four corners of the bottom of the friction deposition forming fixture (a workpiece is fixed on the top surface of the fixture; friction deposition forming is used to form a deposition layer on the workpiece). The spindle contacts the workpiece, and the spindle performs friction deposition forming on the workpiece to form a deposition layer, resulting in a deposited part, such as... Figure 2 As shown, during the friction deposition modeling process, four force sensors (with a sampling frequency set to 4000Hz) are used to collect the original three-dimensional force signals between the workpiece and the spindle within the same time period. These original three-dimensional force signals consist of three mutually perpendicular force signals between the workpiece and the spindle: axial force signal, horizontal force signal, and lateral force signal. Specifically, the directions of the forces in these three signals are mutually perpendicular. The axial force signal is the force signal along the spindle axis.

[0044] Original three-dimensional force signal such as Figure 3 As shown in (a), Figure 3 In (a), Z-Force is the axial original force signal, X-Force is the horizontal original force signal, and Y-Force is the lateral original force signal. Figure 3 In (a), the horizontal axis represents the acquisition time, and the vertical axis represents the force value, with the unit of the vertical axis being kilonewtons. The force values ​​of the horizontal and lateral raw force signals at different acquisition times correspond to... Figure 3 (a) shows the left vertical axis, where the axial original force signal corresponds to the force value at different acquisition times. Figure 3 The right longitudinal axis of (a) shows that the axial original force signal, horizontal original force signal and lateral original force signal exhibit periodic sinusoidal fluctuations, and there are obvious high-frequency abrupt noise signals and clutter.

[0045] S2 includes S2-1, S2-2, and S2-3:

[0046] S2-1, Variational Mode Decomposition (VMD) is performed on the original three-dimensional force signal (Zhang Ping, Li Yongqiang, Xing Hualiang. Optimization of Hybrid Energy Storage Capacity for Smoothing Wind Power Fluctuations Based on Variational Mode Decomposition [J]. Power Generation Technology, 1-10 [2025-03-19].), resulting in three sets of intrinsic mode functions. The axial original force signal, the horizontal original force signal, and the lateral original force signal each correspond to one set of intrinsic mode functions. Each set of intrinsic mode functions includes K intrinsic mode functions, and each intrinsic mode function is a characteristic component at different center frequencies.

[0047] In this embodiment, K=3 is set, and each intrinsic mode function set includes: a first intrinsic mode function, a second intrinsic mode function, and a third intrinsic mode function. The frequency bands of the center frequencies corresponding to the first, second, and third intrinsic mode functions are from high to low. The first and second intrinsic mode functions fluctuate frequently and contain random disturbances or noise components. The third intrinsic mode function has a high degree of smoothness, a small overall change amplitude, and is more consistent with the low-frequency components of the original three-dimensional force signal. The regularization penalty coefficient of variational mode decomposition is set to 2500 to optimize the smoothness of the intrinsic mode functions and improve the accuracy of the center frequencies.

[0048] S2-2, in each set of intrinsic mode functions, the intrinsic mode function with the lowest center frequency is selected as the selected intrinsic mode function (i.e., in this embodiment, the third intrinsic mode function of each set of intrinsic mode functions is selected as the selected intrinsic mode function). The selected intrinsic mode functions of the three sets of intrinsic mode functions are combined and used as the decomposed three-dimensional force signal. The decomposed three-dimensional force signal is as follows: Figure 3 As shown in (b).

[0049] S2-3, Gaussian smoothing is performed on the decomposed three-dimensional force signal to obtain the preprocessed three-dimensional force signal. The preprocessed three-dimensional force signal includes three signal curves, each of which represents the force in one direction. Gaussian smoothing is used to remove noise and clutter from the decomposed three-dimensional force signal.

[0050] S3 includes S3-1 and S3-2;

[0051] S3-1, since the process of friction deposition is divided into three stages, including: initial in-situ deposition stage, stable deposition stage and termination deposition stage, the three-dimensional force signals corresponding to the stable deposition stage are screened from the pre-processed three-dimensional force signals. The three-dimensional force signals of the stable deposition stage are divided according to a fixed window size to obtain multiple sample units. Each sample unit is three signal curves corresponding to the same time period. Multiple sample units form a sample set D.

[0052] In this embodiment, the fixed window size is 800 points, each sample unit includes 800 sample points, and the sample set D includes 3600 sample units, of which 1200 sample units correspond to deposited parts without defects, 1200 sample units correspond to deposited parts with slight defects, and 1200 sample units correspond to deposited parts with severe defects.

[0053] S3-2, calculate the features of each signal curve in each sample unit of the sample set D. The number of features for each signal curve is M / 3, and each sample unit has M types of features. The features of each signal curve include: time domain features and frequency domain features. The time domain features are: maximum value (max), minimum value (min), mean (mean), median (med), peak-to-peak value (ran), mean absolute value (arv), standard deviation (std), kurtosis (Kur), skewness (Ske), root mean square (Rms), root mean square amplitude (Srma), waveform factor (Wf), peak factor (Pf), impulse factor (If), and margin factor (Cf). The frequency domain features are: centroid frequency (FC), root mean square frequency (Rmsf), and frequency standard deviation (Rvf).

[0054] In this embodiment, since each signal curve has 18 features, the total number of features of the three signal curves in each sample unit is M = 3 * 18 = 54.

[0055] Step S3 is used to reflect the response characteristics of the three-dimensional force signal to the defect within a fixed time (fixed window).

[0056] S4 includes S4-1, S4-2, and S4-3, and the specific steps are as follows:

[0057] S4-1: Treat each feature of all sample units in the sample set D as a variable, and use the Pearson correlation coefficient (PCC) method to calculate the correlation coefficient between any two variables to obtain a correlation coefficient matrix; set a correlation threshold. If the correlation coefficient between a variable and the remaining M-1 variables is less than or equal to the correlation threshold, then the feature corresponding to that variable is considered a weakly correlated feature; otherwise, it is considered a strongly correlated feature (that is, if the correlation coefficient between a variable and any of the other variables is greater than the correlation threshold, then the feature corresponding to that variable is considered a strongly correlated feature), resulting in N1 strongly correlated features and N2 weakly correlated features.

[0058] The correlation coefficient matrix is ​​a symmetric matrix, and its diagonal elements represent the correlation coefficients of the variables themselves. The value of the diagonal elements is 1. Each element outside the diagonal in the correlation coefficient matrix represents the correlation coefficient between two variables.

[0059] In this embodiment, the correlation coefficient matrix has a dimension of 54×54; the correlation threshold is 0.9, N1=36, N2=18, and the correlation coefficients of the 45 pairs of strongly correlated features corresponding to the 36 strongly correlated features are shown in Table 1. In Table 1, “(x)” represents the horizontal original force signal corresponding to the feature, “(y)” represents the lateral original force signal corresponding to the feature, and “(z)” represents the axial original force signal corresponding to the feature.

[0060] Table 1

[0061]

[0062]

[0063] S4-2, hierarchical clustering (Cai Fapeng, Feng Ji, Yang Degang, et al. Hierarchical clustering algorithm based on natural neighborhood graph partitioning [J]. Computer Engineering and Science, 2025, 47(02):370-380.) is used to divide N1 strongly correlated features into multiple clusters; a greedy algorithm (Zhang Qian, Nie Yibing, Dong Cunjun. Research on route optimization based on greedy algorithm and dynamic programming [J]. Modern Business and Trade Industry, 2025, (07):236-238.) is used to filter the strongly correlated features in each cluster, and after filtering, a total of N3 strongly correlated features are obtained from all clusters; the N3 strongly correlated features obtained after filtering and N2 weakly correlated features form a feature set, and the specific steps are as follows:

[0064] S4-2-1, the input of hierarchical clustering is the correlation coefficient of N1 strongly correlated features that are greater than the correlation threshold. Hierarchical clustering divides the N1 strongly correlated features into multiple clusters. When dividing into multiple clusters, the formula for calculating the distance between two strongly correlated features is: distance = 1 - correlation coefficient. The clustering method in the hierarchical clustering algorithm (when calculating the distance between two clusters) adopts the average link method.

[0065] In this embodiment, the shearing threshold for hierarchical clustering is set to 0.1, and the multiple clusters output by hierarchical clustering are arranged as follows: Figure 4 The clustering tree shown intuitively demonstrates the clustering relationships and similarities among the strongly correlated features.

[0066] S4-2-2 employs a greedy algorithm to filter strongly correlated features within each cluster. The input to the greedy algorithm is the strongly correlated features within the cluster to be filtered, along with the correlation coefficient greater than a correlation threshold for each strongly correlated feature within the cluster. The output of the greedy algorithm is at least one strongly correlated feature representative of the cluster, forming a representative set of strongly correlated features. The specific greedy algorithm is as follows:

[0067] Iterate through each cluster, and for each cluster, follow the following iterative steps until the correlation coefficients in the coefficient set of the cluster are 0. During the iteration, all the selected strongly correlated features are combined into a representative set of strongly correlated features. The iterative steps are as follows: form a coefficient set of correlation coefficients of strongly correlated features in the cluster; each time, select the strongly correlated feature with the most correlation coefficients in the coefficient set as the selected strongly correlated feature; delete the correlation coefficient corresponding to the selected strongly correlated feature from the coefficient set and delete the selected strongly correlated feature from the cluster.

[0068] The set of strongly correlated features obtained from all clusters contains N3 strongly correlated features.

[0069] In this embodiment, N3 = 12, and the strongly correlated features represent N3 strongly correlated features in the set, such as... Figure 4 The strongly correlated features highlighted in bold are: mean (x), impulse factor (x), root mean square frequency (x), mean absolute value (x), mean (y), mean absolute value (y), impulse factor (y), root mean square frequency (y), mean (z), standard deviation (z), impulse factor (z), and root mean square frequency (z).

[0070] S4-2-3 makes the N3 strongly correlated features and N2 weakly correlated features in the strongly correlated feature representative set form a feature set.

[0071] S4-3, calculate the information gain (IG) of each type of feature in the feature set (Chen Xi. Research on the evaluation model of volleyball training effect based on information gain and random forest algorithm [J]. Journal of Kashgar University, 2024, 45(06): 79-84.), select features in the feature set with an information gain value greater than or equal to 0.01 and form the final feature set H, which includes N4 types of features.

[0072] In this embodiment, as Figure 5 The figure shows the information gain value of each type of feature in the final feature set H. Figure 5 The abbreviations for the features are preceded by x, y, or z, where x represents the feature corresponding to the horizontal original force signal, y represents the feature corresponding to the lateral original force signal, and z represents the feature corresponding to the axial original force signal. N4 = 21. The 21 types of features are: impulse factor (z), mean absolute value (x), maximum value (x), peak-to-peak value (x), peak-to-peak value (y), standard deviation (x), impulse factor (x), waveform factor (x), maximum value (y), standard deviation (y), centroid frequency (x), standard deviation (z), mean (z), kurtosis (x), mean (x), mean absolute value (y), kurtosis (y), kurtosis (z), minimum value (y), centroid frequency (y), and root mean square frequency (y).

[0073] Step S4-3 sorts the importance of each type of feature in the feature set according to the information gain value, and selects the features in the feature set that are more important in distinguishing defects.

[0074] S5, Construct the dataset, which consists of multiple samples. Each sample includes: the value of the feature in the final feature set H corresponding to a sample unit;

[0075] In this embodiment, the dataset includes 3600 samples.

[0076] Example 2

[0077] A training method for a defect detection model includes: training the defect detection model using a five-fold cross-validation method based on samples in a dataset to obtain a trained defect detection model. The true value of each sample is whether the deposited part corresponding to the sample unit is defect-free or defective. When the deposited part has a slight defect or a serious defect, it is considered that the deposited part is defective.

[0078] The defect detection model is constructed using the random forest algorithm. The input of the defect detection model is the sample in Example 1, and the output is the prediction result, which is either no defect or defective. During the training process, the grid search method (Zhao Yucheng, Li Yingjian, Shen Shimin, et al. Research on injector fault diagnosis based on grid search and voting classification model [J]. Machine Tool & Hydraulics, 2024, 52(05):213-220.) is used to optimize the parameters of the defect detection model.

[0079] In this embodiment, the random forest algorithm sets up 30 decision trees, each decision tree has a maximum of 19 branches, the maximum number of features allowed per decision tree is set to 7, and each decision tree leaf node requires at least 10 samples.

[0080] Example 3

[0081] A method for detecting defects in triboelectric deposition forming based on three-dimensional force time-frequency domain features includes: using the defect detection model trained in Example 2 to detect defects in the sample to be tested, and obtaining the prediction result corresponding to the sample to be tested.

[0082] The method for obtaining the sample to be tested is the same as the method for obtaining the sample in Example 1.

[0083] Example 4

[0084] A training method for a defect detection model is basically the same as in Example 2, except that the predicted result output by the defect detection model is either well-formed, slightly defective, or severely defective. The true value of each sample is the deposited part corresponding to that sample unit, indicating that the deposited part has no defects, slightly defective, or severely defective.

[0085] Example 5

[0086] A method for detecting defects in triboelectric deposition forming based on three-dimensional force time-frequency domain features includes: using the trained defect detection model of Example 4 to detect defects in the sample to be tested, and obtaining the prediction result corresponding to the sample to be tested.

[0087] The method for obtaining the sample to be tested is the same as the method for obtaining the sample in Example 1.

[0088] Example 6

[0089] A method for detecting defects in triboelectric deposition forming based on three-dimensional force time-frequency domain characteristics is basically the same as that in Example 5, except that each sample in this example includes the value of the feature in the final feature set H′ corresponding to a sample unit.

[0090] The method for obtaining the final feature set H′ is basically the same as the “method for obtaining a dataset” in Example 1. The only difference is that step S4 in Example 1 is replaced by the principal component analysis (PCA) algorithm. That is, the PCA algorithm is used to reduce the dimensionality of the M-class features, and the resulting multiple features are used as the final feature set H′.

[0091] like Figure 6 Figures (a)-(c) show the accuracy of the defect detection model in Example 4 under different parameters. The values ​​highlighted in the figure are the optimal parameters.

[0092] The confusion matrix corresponding to the prediction results of samples in the dataset in Example 4 is as follows: Figure 6 As shown in (e), the visualization of any corresponding decision tree is as follows: Figure 6 As shown in (f), the prediction accuracy of Example 4 is 98.8%; the confusion matrix corresponding to the prediction results of the samples in the dataset of Example 6 is as follows. Figure 7 As shown in (a), Figure 7 (b) shows the ROC curves for each category. The accuracy of the prediction results in Example 6 is 95.7%.

[0093] In summary, Example 4 effectively distinguished well-formed samples; and significantly improved overall accuracy in differentiating between minor and major defects. Example 4 can accurately determine the presence of forming defects in three-dimensional force signals within a short time, providing a theoretical basis and technical support for real-time parameter adjustment and quality control in continuous triboelectric deposition manufacturing processes based on force signals. The triboelectric deposition forming defect detection method in Example 6 is particularly inadequate in classifying minor defects, with 120 samples being misclassified.

[0094] Example 7-1

[0095] A method for detecting defects in triboelectric deposition forming based on three-dimensional force time-frequency domain characteristics is basically the same as that in Example 3, except that the defect detection model is constructed using a decision tree.

[0096] Example 7-2

[0097] A method for detecting defects in triboelectric deposition forming based on three-dimensional force time-frequency domain characteristics is basically the same as that in Example 5, except that the defect detection model is constructed using a decision tree.

[0098] Example 8-1

[0099] A method for detecting defects in triboelectric deposition forming based on three-dimensional force time-frequency domain characteristics is basically the same as that in Example 3, except that the defect detection model is constructed using a support vector machine (SVM).

[0100] Example 8-2

[0101] A method for detecting defects in triboelectric deposition forming based on three-dimensional force time-frequency domain characteristics is basically the same as that in Example 5, except that the defect detection model is constructed using a support vector machine (SVM).

[0102] Example 9-1

[0103] A method for detecting defects in triboelectric deposition forming based on three-dimensional force time-frequency domain characteristics is basically the same as that in Example 3, except that the defect detection model is constructed using the k-nearest neighbor algorithm.

[0104] Example 9-2

[0105] A method for detecting defects in triboelectric deposition forming based on three-dimensional force time-frequency domain characteristics is basically the same as that in Example 5, except that the defect detection model is constructed using the k-nearest neighbor algorithm.

[0106] The accuracy rates of the prediction results for Examples 7-1, 7-2, 8-1, 8-2, 9-1, 9-2, 3, and 5 are shown in Table 4. Figure 6 (d) shows Figure 6 In (d), DD represents "predicted result is no defect or defective", and DC represents "predicted result is well formed, slightly defective or seriously defective". In this case, the defect detection model of Example 3 has a precision of 98.5%, a recall of 98.8%, an F1 value of 98.65%, and an AUC (Area Under the Curve) of 0.9998, which is close to 1, which fully demonstrates that the defect detection model performs excellently in distinguishing between positive and negative samples.

[0107] Table 4

[0108]

[0109] In summary, the feature selection and dimensionality reduction methods in the triboelectric deposition forming defect detection method based on three-dimensional force time-frequency domain features of the present invention have a high response to defects and a significant impact on the model accuracy.

[0110] The present invention has been described above by way of example. It should be noted that any simple modifications, alterations or other equivalent substitutions that can be made by those skilled in the art without creative effort without departing from the core of the present invention fall within the protection scope of the present invention.

Claims

1. A defect detection model, wherein each input sample includes: The values ​​of features in the final feature set H corresponding to a sample unit; characterized in that the method for obtaining the final feature set H includes: S1, Prepare the original three-dimensional force signal. The original three-dimensional force signal is the signal of the force between the workpiece and the spindle in three mutually perpendicular directions during the same time period in the process of friction deposition forming. S2. Variational mode decomposition is performed on the original three-dimensional force signal to obtain three sets of intrinsic mode functions (IMFs). Each IMF set includes K IMFs. The IMF with the lowest center frequency in each IMF set is selected as the selected IMF. The three selected IMFs are combined and used as the decomposed three-dimensional force signal. Gaussian smoothing is then applied to obtain three signal curves as the preprocessed three-dimensional force signal. S3, the three-dimensional force signal is divided into multiple sample units according to a fixed window size, and the multiple sample units form a sample set D; the features of each signal curve in each sample unit in the sample set D are calculated, and M features are obtained for each sample unit; S4. Take each type of feature of all sample units in the sample set D as a variable, and calculate the correlation coefficient between any two variables. Set a correlation threshold. If the correlation coefficient between a certain variable and the remaining M-1 variables is less than or equal to the correlation threshold, then the type of feature corresponding to that variable is regarded as a weakly correlated feature; otherwise, it is regarded as a strongly correlated feature, resulting in N strongly correlated features and N2 weakly correlated features. Hierarchical clustering is used to divide N1 strongly correlated features into multiple clusters; a greedy algorithm is used to filter the strongly correlated features in each cluster, resulting in a total of N3 strongly correlated features after filtering; the N3 strongly correlated features obtained after filtering and N2 weakly correlated features form a feature set; Calculate the information gain of each feature in the feature set, select features in the feature set whose information gain value is greater than or equal to 0.01, and form the final feature set H.

2. The defect detection model according to claim 1, characterized in that, The characteristics of each signal curve include one or more of the following: maximum value, minimum value, mean value, median value, peak-to-peak value, mean absolute value, standard deviation, kurtosis, skewness, root mean square (RMS), RMS amplitude, waveform factor, peak factor, impulse factor, margin factor, centroid frequency, RMS frequency, and frequency standard deviation.

3. The defect detection model according to claim 1, characterized in that, In S4, the input for hierarchical clustering is the correlation coefficient of N1 strongly correlated features that are greater than the correlation threshold.

4. The defect detection model according to claim 1, characterized in that, In S4, the input of the greedy algorithm is the strongly correlated features within the selected cluster and the correlation coefficient greater than the correlation threshold corresponding to each strongly correlated feature within the cluster. The output of the greedy algorithm is at least one strongly correlated feature that is representative of the cluster.

5. The defect detection model according to claim 4, characterized in that, The greedy algorithm is as follows: Iterate through each cluster, and for each cluster, follow the following iterative steps until the correlation coefficients in the coefficient set of the cluster are 0. All the selected strongly correlated features during the iteration process are combined into a representative set of strongly correlated features. The representative set of strongly correlated features obtained from all clusters is the N3 strongly correlated features. The iterative steps are as follows: Form a coefficient set of the correlation coefficients of the strongly correlated features in the cluster. Each time, select the strongly correlated feature with the most correlation coefficients in the coefficient set as the selected strongly correlated feature. Delete the correlation coefficient corresponding to the selected strongly correlated feature from the coefficient set and delete the selected strongly correlated feature from the cluster.

6. The defect detection model according to claim 1, characterized in that, The defect detection model is constructed using the random forest algorithm.

7. The defect detection model according to claim 1, characterized in that, The defect detection model was trained using five-fold cross-validation, and the parameters of the defect detection model were optimized using a grid search method.

8. The defect detection model according to claim 1, characterized in that, The output of the defect detection model is the prediction result. When the defect detection model is used for binary classification, the prediction result is no defect or defective. The true value of the sample is whether the deposited part corresponding to the sample unit is defective or defective. When the defect detection model is used for three-class classification, the prediction result is good forming, minor defects, or serious defects. The true value of the sample is that the deposited part corresponding to the sample unit has no defects, the deposited part has minor defects, or the deposited part has serious defects.

9. A method for detecting defects in triboelectric deposition forming based on three-dimensional force time-frequency domain characteristics, characterized in that, The defect detection model trained according to any one of claims 1 to 8 is used to detect defects in the sample to be tested, and the prediction result corresponding to the sample to be tested is obtained.

10. The use of any one of the defect detection models in claims 1 to 8 for real-time defect detection in triboelectric deposition forming.

Images