A Feature-Adaptive Filtering-Based Method and System for Monitoring the Surface Quality of Milled Thin-Wall Parts

Abstract

This invention discloses a method and system for monitoring the surface quality of milled thin-walled parts based on feature-adaptive filtering, belonging to the field of surface quality monitoring technology for milled thin-walled parts. The method includes: sequentially calculating the feature information capacity, maximum information number, and distance correlation coefficient between each feature and the target vector to obtain a sequence ranking of features sorted in descending order based on feature information capacity, maximum information number, and distance correlation coefficient; calculating the sum of the three sorting indices for each feature vector to obtain a comprehensive feature index for each feature; sorting the features in ascending order based on the comprehensive feature index; selecting the top p feature vectors in the comprehensive ranking and concatenating them sequentially to obtain an adaptively filtered input feature matrix; and constructing a nonlinear real-time mapping relationship model between the input feature matrix and the target feature vector. This invention, based on adaptive filtering of single-channel cutting signal features, eliminates the adverse effects of unstable dynamic characteristics of the cutting system on monitoring accuracy, thereby improving monitoring accuracy.

Description

Technical Field

[0001] This invention relates to the field of surface quality monitoring technology for thin-walled parts milling, and particularly to a method and system for monitoring surface quality of thin-walled parts milling based on feature-adaptive filtering. Background Technology

[0002] Thin-walled components, as core components in lightweight design for aerospace equipment, are widely used in aircraft fuselage frames, engine blades, and door structures. However, during milling, thin-walled components are prone to chatter and tool deformation due to their weak rigidity, resulting in uneven surface roughness distribution and directly affecting their service performance. The identification and calculation of workpiece surface roughness mainly fall into two categories: offline prediction methods based on process parameters and workpiece material properties, and online monitoring methods based on real-time sensing information and artificial intelligence during the machining process. Regarding roughness prediction, existing technologies disclose a milling surface roughness prediction method that establishes a contact relationship model between the cutting tool teeth and the residual material height based on the cutting tool diameter, bottom corner radius, bottom edge inclination angle, and machining process parameters. Regarding monitoring, existing technologies disclose an online monitoring method for milling surface roughness that considers the real-time state of the cutting tool. This method transforms real-time tool wear information during the machining process into feature vectors and introduces them into a surface roughness identification model, thereby achieving the identification of the milled surface roughness of thin-walled components.

[0003] The core challenge in monitoring the surface roughness of milled thin-walled parts stems from the nonlinear dynamic characteristics caused by their low rigidity. First, the low stiffness leads to time-varying modal characteristics in the cutting system; the workpiece-tool contact stiffness can fluctuate by 2-3 times its initial value during machining, making traditional monitoring models with fixed feature extraction parameters ill-suited to the dynamically changing frequency response. Second, the coupling effect of chatter and tool deformation causes the cutting signal to exhibit non-stationary characteristics. For example, in the machining of titanium alloy housings, the chatter frequency randomly drifts within the range of 800-2500Hz, leading to a sharp increase in the misjudgment rate of conventional frequency domain analysis methods. To compensate for the effects of these unstable dynamic characteristics, existing data-driven models have to introduce high-dimensional signal features from multiple sensor sources, such as multi-sensor fusion of cutting force and noise signals, 128-dimensional time-frequency joint features, and tool wear features. This results in high monitoring costs and an exponential increase in computational complexity during model training and inference, failing to meet the real-time control requirements of five-axis machining centers. Summary of the Invention

[0004] To address the shortcomings of existing technologies, the present invention aims to provide a method and system for monitoring the surface quality of milled thin-walled parts based on feature-adaptive filtering. By adaptively filtering the features of a single-channel cutting signal, the adverse effects of the unstable dynamic characteristics of the cutting system on monitoring accuracy are eliminated, thereby improving the monitoring accuracy.

[0005] To achieve the above objectives, the present invention is implemented through the following technical solution: In a first aspect, embodiments of the present invention provide a method for monitoring the surface quality of milled thin-walled parts based on feature-adaptive filtering, including: Extract the time-domain, frequency-domain, and time-frequency-domain features of the processing vibration signal, construct the original feature matrix of the sensor data, and arrange each feature in the original feature matrix in sequence; Calculate the feature information capacity, maximum information number, and distance correlation coefficient between each feature and the target vector in turn, and obtain the sequence ranking of each feature after sorting it in descending order according to the feature information capacity, maximum information number, and distance correlation coefficient; For each feature vector, calculate the sum of the three sorting indices to obtain the feature comprehensive index of each feature; sort the features in ascending order based on the feature comprehensive index; select the top p feature vectors in the comprehensive ranking and concatenate them to obtain the adaptively filtered input feature matrix; Based on gated recurrent units and fully connected networks, a nonlinear real-time mapping model between the input feature matrix and the target feature vector is constructed to obtain the surface roughness of the workpiece in real time.

[0006] As a further implementation, when there are ties, the features with the same overall feature number will be sorted in ascending order based on their feature number.

[0007] As a further implementation, the calculation process for adaptive sorting is as follows: ; Where σ represents the index sequence result of sorting all features in ascending order of feature comprehensive index, argsort() ↑ This indicates the calculation of the index sequence arranged in ascending order. r j F Indicates the first j Each feature is ranked by its index based on its feature information capacity. r j M Indicates the first j Each feature is ranked by its ordinal number based on the maximum information content index. r j D Indicates the first j Each feature is ranked based on its distance correlation coefficient.

[0008] As a further implementation, all features are sorted from largest to smallest according to their Feature Information Capacity (FIC) index, resulting in an index sequence of all features arranged in descending order of FIC value. This yields the sorting number for each feature based on its FIC index. r i F : ; in, Indicates an indicator function, σ F ( i ) indicates the sorted order. i The index of a feature in the original feature set.

[0009] As a further implementation method, based on the maximum information number (MIC) of each feature... i Sort all features from largest to smallest to obtain an index sequence of all features in descending order of MIC value, thus obtaining the sorting number of each feature based on the maximum information number index. r i M : .

[0010] As a further implementation method, based on the distance correlation coefficient of each feature... DC i Sort all features from largest to smallest to obtain the result based on... DC The index column of all features is sorted in descending order of values, thus obtaining the sort number of each feature based on the distance correlation coefficient. r i D : .

[0011] As a further implementation, the input feature matrix is: ; in, p Here, is a hyperparameter related to monitoring accuracy, representing the dimension of the input feature matrix. f σ(p) express f σ(p) Indicates the order after comprehensive sorting. p The original feature corresponding to the feature index of the bit.

[0012] Secondly, embodiments of the present invention also provide a feature-adaptive filtering-based thin-walled part milling surface quality monitoring system, comprising: The feature extraction module is used to extract the time-domain features, frequency-domain features, and time-frequency-domain features of the processing vibration signal, construct the original feature matrix of the sensor data, and arrange each feature in the original feature matrix in sequence. The sequence ranking acquisition module is used to calculate the feature information capacity, maximum information number and distance correlation coefficient between each feature and the target vector in turn, and obtain the sequence ranking of each feature after sorting in descending order according to the feature information capacity, maximum information number and distance correlation coefficient. The input feature matrix acquisition module is used to calculate the sum of the three sorting indices for each feature vector to obtain the feature comprehensive index of each feature; the features are sorted in ascending order based on the feature comprehensive index; the top p feature vectors in the comprehensive ranking are selected and concatenated to obtain the adaptively filtered input feature matrix; The mapping relationship model construction module is used to construct a nonlinear real-time mapping relationship model between the input feature matrix and the target feature vector based on gated recurrent units and fully connected networks, so as to obtain the surface roughness of the workpiece in real time.

[0013] Thirdly, embodiments of the present invention also provide an electronic device, including a memory and a processor, wherein the memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to perform the feature-adaptive filtering method for monitoring the surface quality of milled thin-walled parts.

[0014] Fourthly, embodiments of the present invention also provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the feature-adaptive filtering method for monitoring the surface quality of milled thin-walled parts.

[0015] The beneficial effects of this invention are as follows: (1) The present invention calculates the feature information capacity between the feature and the target vector in sequence, and sorts the features in descending order to obtain the ranking of each feature in the sequence after sorting according to the feature information capacity; calculates the maximum information number between the feature and the target vector in sequence, and sorts the features in descending order to obtain the ranking of each feature in the sequence after sorting according to the maximum information number; calculates the distance correlation coefficient between the feature and the target vector in sequence, and sorts the features in descending order to obtain the ranking of each feature in the sequence after sorting according to the distance correlation coefficient; by calculating the sum of the above three sorting numbers for each feature vector, the feature comprehensive number of each feature can be obtained, and the features are sorted in ascending order based on the feature comprehensive number; thus, adaptive filtering of features is realized. The feature filtering process balances the multidimensional evaluation index independent of the feature distribution, and can autonomously adapt to the linear, nonlinear and other types of correlation between the feature vector and the target vector caused by unstable dynamic characteristics, effectively eliminating the influence of unstable dynamic characteristics of the cutting system on the feature vector distribution.

[0016] (2) The present invention can achieve accurate monitoring of milling surface roughness by using only the sensing information of one channel during the cutting process. The monitoring process does not require the construction of a fusion strategy for multiple channel sensing data, the integration of tool wear status information during the cutting process, or the input of process parameter data, which significantly reduces the amount of calculation in the monitoring process and effectively avoids the impact of changes in process conditions on the roughness monitoring accuracy. Attached Figure Description

[0017] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0018] Figure 1 This is a flowchart of a method for monitoring the surface quality of milled thin-walled parts according to one or more embodiments of the present invention.

[0019] Figure 2 This refers to the size of the feature information capacity and the sorting number based on the feature information capacity index according to one or more embodiments of the present invention; Figure 3 This refers to the maximum number of information items according to one or more embodiments of the present invention and the sorting number based on the maximum number of information items index; Figure 4 This refers to the magnitude of the distance correlation coefficient and the sorting number based on the distance correlation coefficient according to one or more embodiments of the present invention; Figure 5 This invention ranks features based on the combined feature sequence numbers of each feature according to one or more embodiments. Figure 6 This is a comparison diagram of the feature distribution before and after adaptive filtering according to one or more embodiments of the present invention; wherein, (a) represents the features obtained by adaptive filtering, and (b) represents the unfiltered features; Figure 7 This is a comparison diagram of the surface roughness monitoring results and the measured surface roughness results according to one or more embodiments of the present invention. Detailed Implementation

[0020] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0021] Example 1: This embodiment provides a feature-adaptive filtering method for monitoring the surface quality of milled thin-walled parts, including: Extract the time-domain, frequency-domain, and time-frequency-domain features of the processing vibration signal, construct the original feature matrix of the sensor data, and arrange each feature in the original feature matrix in sequence; Calculate the feature information capacity, maximum information number, and distance correlation coefficient between each feature and the target vector in turn, and obtain the sequence ranking of each feature after sorting it in descending order according to the feature information capacity, maximum information number, and distance correlation coefficient; For each feature vector, calculate the sum of the three sorting indices to obtain the feature comprehensive index of each feature; sort the features in ascending order based on the feature comprehensive index; select the top p feature vectors in the comprehensive ranking and concatenate them to obtain the adaptively filtered input feature matrix; Based on gated recurrent units and fully connected networks, a nonlinear real-time mapping model between the input feature matrix and the target feature vector is constructed to obtain the surface roughness of the workpiece in real time.

[0022] Specifically, it includes the following steps: S1. Accelerometers are used to collect cutting vibration signals during the milling of thin-walled parts. v The accelerometer can be fixed to the workpiece surface near the cutting area using double-sided tape without affecting the relative movement of the workpiece and the tool. By measuring the surface roughness value after actual machining using a digital microscope, the target vector required for training the recognition model can be obtained. t .

[0023] S2. Processing vibration signal v The time-domain features, frequency-domain features, and time-frequency-domain features are extracted respectively. The calculation formulas for each feature are existing technologies and will not be elaborated here.

[0024] This forms the original feature matrix of the sensor data. F =( V t , V f , V tf ),in, V t This represents the time-domain features extracted based on the processing vibration signal. V f This represents the frequency domain features extracted based on the processing vibration signal. V tf This represents the time-frequency domain features extracted based on the processing vibration signal.

[0025] S3. Arrange each feature of the sensing signal in the time domain, frequency domain, and time-frequency domain sequentially, and the original feature matrix can be expanded as follows: F =( f 1, f 2,…, f i ,… f n )∈R m×n ,in, f i Represents the first element in the original feature matrix of the sensor data. i 1 eigenvector m This represents the size of the cutting signal sample space. nThe dimension of the feature space represents the total number of features of all sensing signals.

[0026] S4. For all features F =( f 1, f 2,…, f i ,… f n ) Calculate the features sequentially f i With the target vector t of The capacity of feature information between FIC i :

[0027] In equation (1), σ () indicates that the Sigmoid function is calculated. ρ Spearman ( f i , t ) indicates the calculation of eigenvectors f i With the target vector t Spearman's rank correlation coefficient between them I* ( f i , t ) indicates the calculation of eigenvectors f i With the target vector t Inter-information exchange.

[0028] S5. Based on the feature information capacity index of each feature FIC Sort all features from largest to smallest to get the result by... FIC The index sequence of all features sorted in descending order of values:

[0029] In equation (2), argsort() ↓ This indicates that the index sequence is returned in descending order, such as... σ F ( j ) indicates the sorted order. j The index of a feature in the original feature set.

[0030] Based on this, each feature can be obtained. f i Sorting order based on feature information capacity index r i F :

[0031] In equation (3), This indicates an indicator function; its value is 1 if the condition is true, and 0 otherwise, thus returning the characteristic. f i Ranking within the sequence after sorting based on the feature information capacity index.

[0032] S6. For all features F =( f 1, f 2,…, f i ,… f n ) Calculate the features sequentially f i With the target vector t Maximum number of information between MIC i :

[0033] In equation (4), B ( m ) represents a function related to the sample size, in this embodiment B ( m )= m 0.6 .

[0034] S7. Based on the maximum information number of each feature MIC i Sort all features from largest to smallest to get the result by... MIC The index sequence of all features sorted in descending order of values:

[0035] Thus, each feature can be obtained. f i Based on the maximum information content index MIC i Sort number r i M :

[0036] S8. For all features F =( f 1, f 2,…, f i ,… f n ) Calculate the features sequentially f i With the target vector tDistance correlation coefficient DC i :

[0037] In equation (7), dCov is the distance covariance and dVar is the distance variance.

[0038] S9. Based on the distance correlation coefficient of each feature DC i Sort all features from largest to smallest to obtain the result based on... DC A sequence of indices for all features sorted in descending order of values;

[0039] Thus, each feature can be obtained. f i Based on distance correlation coefficient DC i Sort number r i D :

[0040] S10. By calculating the sum of the three sorting indices for each feature vector, the feature comprehensive index of each feature can be obtained. Based on the feature comprehensive index, the features are sorted in ascending order. When there are ties, the features with ties in the feature comprehensive index are sorted in ascending order, thus achieving adaptive sorting of cutting signal features. The calculation process can be expressed as follows:

[0041] In equation (10), σ represents the index sequence result of all features sorted in ascending order by feature comprehensive index, argsort() ↑ This indicates the calculation of the index sequence arranged in ascending order. r j F Indicates the first j Each feature is ranked by its index based on its feature information capacity. r j M Indicates the first j Each feature is ranked by its ordinal number based on the maximum information content index. r j D Indicates the first j Each feature is ranked based on its distance correlation coefficient.

[0042] S11. Select the top-ranked overall pThe feature vectors are concatenated sequentially, thus obtaining the input feature matrix after adaptive filtering:

[0043] In equation (11), p Here, is a hyperparameter related to monitoring accuracy, representing the dimension of the input feature matrix. f σ(p) express f σ(p) Indicates the order after comprehensive sorting. p The original feature corresponding to the feature index of the bit.

[0044] S12. Construct a surface roughness recognition network.

[0045] After obtaining the input feature matrix through feature adaptive filtering, a nonlinear real-time mapping relationship model between the input feature matrix and the target feature vector can be constructed based on gated recurrent units (GRUs) and fully connected networks.

[0046] Since feature-adaptive filtering eliminates the adverse effects of unstable dynamic characteristics of the cutting system on monitoring accuracy, accurate identification of the surface roughness of milled titanium alloy thin-walled parts can be achieved using only a 4-layer network structure. The roughness identification network structure includes 2 unidirectional GRU layers and 2 fully connected layers. The number of units in each of the 2 GRU layers is 128, the number of units in the first fully connected layer is 64, and the number of units in the last fully connected layer is 1.

[0047] The calculation process of the gated loop unit is expressed as follows:

[0048] In the above formula, W * , U * , b * ,(*∈{ h,r , z}) represents the weight matrix and bias vector. z i It's an update gate. r i To reset the door, This is the current state. This refers to the state at the previous moment. It is a candidate state.

[0049] The formula for calculating a fully connected layer is:

[0050] In equation (16), W This is the network weight matrix.b For bias vectors, X This represents the output vector of the previous layer.

[0051] S13. Train and deploy the recognition model and monitor online.

[0052] Based on S1~S11, adaptive filtering of cutting signal features is completed using sensing data from the cutting process of titanium alloy thin-walled parts. The surface roughness recognition model constructed in S12 is trained based on the input feature matrix obtained in S11 and the actual measured surface roughness label data. After the model training is completed, the feature adaptive filtering method and model are deployed on the edge device in the workshop. S1~S11 are repeated, and the feature matrix is ​​input into the trained recognition model, which can identify the surface roughness of the workpiece in real time during the processing.

[0053] Therefore, adaptive filtering of the original features is achieved through adaptive sorting of cutting signal features. Since the feature filtering process balances multidimensional evaluation indicators independent of feature distribution, the filtering method can autonomously adapt to various types of correlations between feature vectors and target vectors caused by unstable dynamic characteristics, such as linear and nonlinear relationships, and effectively eliminate the influence of unstable dynamic characteristics of the cutting system on feature vector distribution.

[0054] Example 2: To verify the feasibility of the surface quality monitoring method for thin-walled parts milling described in Example 1, a milling experiment of titanium alloy thin-walled parts was conducted in this example. A rectangular thin plate of titanium alloy was machined by using an indexable two-flute end mill through climb milling and side milling. One cut along the length of the rectangular thin plate was counted as one cut, and each cutter performed a total of 100 cuts.

[0055] During machining, the spindle speed was 8000 r / min, the feed rate was 1280 mm / min, the radial cutting width was 0.2 mm, and the axial cutting depth was 4 mm. A Dytran 3263A1 triaxial accelerometer was installed behind the workpiece's machining area to collect vibration signals in three mutually perpendicular directions coinciding with the machine coordinate system. The collected cutting vibration signals were then saved to the data storage system using a National Instrument PXIe-4464 data acquisition card.

[0056] After every 10 cuts, the machine is stopped, and the surface roughness of the workpiece is measured using a Keyence laser confocal microscope. To avoid the influence of random sampling, surface roughness data is collected sequentially at five equidistant points on the cut surface during the measurement process. The average roughness of each cut is calculated by averaging the five equidistant points, serving as the roughness data label. Subsequently, based on the measured results, the workpiece surface roughness data corresponding to each cut is obtained through nonlinear interpolation. This allows for the labeling of the cut sensing data for each cut, thus obtaining the target feature vector.

[0057] To improve computational efficiency, this embodiment only selects... x The directional vibration signal is used as the sole input channel data. According to S2, 12 time-domain features, 12 frequency-domain features, and 8 time-frequency-domain features are extracted for each channel. According to S3, the original feature matrix of the sensing data is expanded. F =( V t , V f , V tf )∈R 100*32 That is, the feature matrix size is 100*32, where 100 represents the 100 cuts performed. Since a total of 32 dimensions of features were extracted from the time domain, frequency domain, and time-frequency domain, the original feature matrix has a dimension of 32.

[0058] Based on S4, the feature information capacity index between each feature and the target vector is calculated sequentially. Then, based on S5, the rank of each feature in the sequence sorted according to the feature information capacity index is calculated, such as... Figure 2 As shown, the left vertical axis displays the feature information capacity calculated for each feature, and the right vertical axis displays the sequence number of each feature after sorting based on the feature information capacity.

[0059] Furthermore, according to S6, the maximum information number index between each feature and the target vector is calculated sequentially. Then, according to S7, all features can be sorted from largest to smallest according to their maximum information number, thus obtaining the ranking of all features in the sequence based on the maximum information number index. Figure 3 As shown.

[0060] Calculate the distance correlation coefficient for each feature based on S8 and S9, and then rank them according to the distance correlation coefficient index, such as... Figure 4 As shown. Based on S10, the feature comprehensive ranking for each feature can be obtained as follows: Figure 5As shown in the figure, the comprehensive ranking number corresponding to each feature is displayed, and the features are sorted according to the ranking of the comprehensive ranking number on the horizontal axis. This realizes the adaptive sorting of cutting signal features, and the input feature matrix can be obtained according to S11. Figure 6 The image shows a comparison of the feature distributions of the adaptively filtered input feature matrix and the unfiltered feature matrix. Figure 6 Figure (a) shows the distribution of the features obtained without adaptive filtering, that is, the top six features in the feature comprehensive ranking. Figure 6 Figure (b) shows the distribution of the top six features after feature aggregation ranking, representing unfiltered features. (By...) Figure 6 A comparison of (a) and (b) reveals that the feature distribution obtained after adaptive filtering has fewer outliers, indicating that it is easier to establish a mapping relationship between features and target vectors.

[0061] Furthermore, a surface roughness recognition network can be constructed according to S12, and the training and deployment of the recognition network can be completed according to S13.

[0062] In this embodiment, 80% of the cutting data from a single tool is set as the training set, and the remaining 20% ​​is set as the test set. The model parameters during the training process are set as follows: batch size is 16, learning rate is 0.001, number of training rounds is 200, and the time window length of the cutting data is set to 10. That is, the surface roughness Ra value of the current cutting data is monitored using the first 10 cutting data. The Adam algorithm is used to optimize the training process, thereby completing the training of the model.

[0063] Figure 7 The image shows a comparison between the surface roughness monitoring results and the measured surface roughness results. Although a total of 100 cuts were performed, the time window length was set to 10, so... Figure 7 The results shown only include surface roughness monitoring results after the 11th to 100th cuts, based on... Figure 7 The proposed method was found to be able to identify the surface roughness of milled thin-walled parts. When using this trained model for online monitoring, only the data collected from the first 10 cuts needs to be analyzed. x By inputting cutting vibration data in the specified direction into the model, the surface roughness value at the current cutting moment can be output. Based on this, online identification of the surface roughness of thin-walled titanium alloy parts can be achieved.

[0064] Example 3: This embodiment provides a feature-adaptive filtering-based surface quality monitoring system for milled thin-walled parts, including: The feature extraction module is used to extract the time-domain features, frequency-domain features, and time-frequency-domain features of the processing vibration signal, construct the original feature matrix of the sensor data, and arrange each feature in the original feature matrix in sequence. The sequence ranking acquisition module is used to calculate the feature information capacity, maximum information number and distance correlation coefficient between each feature and the target vector in turn, and obtain the sequence ranking of each feature after sorting in descending order according to the feature information capacity, maximum information number and distance correlation coefficient. The input feature matrix acquisition module is used to calculate the sum of the three sorting indices for each feature vector to obtain the feature comprehensive index of each feature; the features are sorted in ascending order based on the feature comprehensive index; the top p feature vectors in the comprehensive ranking are selected and concatenated to obtain the adaptively filtered input feature matrix; The mapping relationship model construction module is used to construct a nonlinear real-time mapping relationship model between the input feature matrix and the target feature vector based on gated recurrent units and fully connected networks, so as to obtain the surface roughness of the workpiece in real time.

[0065] Example 4: This embodiment provides an electronic device, including a memory and a processor. The memory stores a computer program, and the processor runs the computer program to enable the electronic device to perform the feature-adaptive filtering method for monitoring the surface quality of milled thin-walled parts as described in Embodiment 1.

[0066] Example 5: This embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the feature-adaptive filtering method for monitoring the surface quality of milled thin-walled parts as described in Embodiment 1.

[0067] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for monitoring the surface quality of milled thin-walled parts based on feature-adaptive filtering, characterized in that, include: Extract the time-domain, frequency-domain, and time-frequency-domain features of the processing vibration signal, construct the original feature matrix of the sensor data, and arrange each feature in the original feature matrix in sequence; Calculate the feature information capacity, maximum information number, and distance correlation coefficient between each feature and the target vector in turn, and obtain the sequence ranking of each feature after sorting it in descending order according to the feature information capacity, maximum information number, and distance correlation coefficient; For each feature vector, calculate the sum of the three sorting indices to obtain the feature composite index for each feature; The features are sorted in ascending order based on their comprehensive index; the top p feature vectors in the comprehensive ranking are selected and concatenated to obtain the adaptively filtered input feature matrix. Based on gated recurrent units and fully connected networks, a nonlinear real-time mapping relationship model between the input feature matrix and the target feature vector is constructed to obtain the surface roughness of the workpiece in real time. The calculation process for adaptive sorting is as follows: ； Where σ represents the index sequence result of all features sorted in ascending order by feature comprehensive index, argsort() ↑ This indicates the calculation of the index sequence arranged in ascending order. r j F Indicates the first j Each feature is ranked by its index based on its feature information capacity. r j M Indicates the first j Each feature is ranked by its ordinal number based on the maximum information content index. r j D Indicates the first j Each feature is ranked based on its sequence number according to the distance correlation coefficient index; Sort all features in descending order of their Feature Information Capacity (FIC) values ​​to obtain an index sequence of all features, thus yielding the sorting number for each feature based on its FIC value. r i F : ； in, Indicates an indicator function, σ F ( i ) indicates the sorted order. i The index of the feature in the original feature set; Based on the maximum information content (MIC) of each feature i Sort all features from largest to smallest to obtain an index sequence of all features in descending order of MIC value, thus obtaining the sorting number of each feature based on the maximum information number index. r i M : ； Based on the distance correlation coefficient of each feature DC i Sort all features from largest to smallest to obtain the result based on... DC The index column of all features is sorted in descending order of values, thus obtaining the sort number of each feature based on the distance correlation coefficient. r i D : 。 2. The method for monitoring the surface quality of milled thin-walled parts based on feature-adaptive filtering according to claim 1, characterized in that, When there are ties in ranking, the features with the same overall ranking will be sorted in ascending order of feature number.

3. The method for monitoring the surface quality of milled thin-walled parts based on feature-adaptive filtering according to claim 1, characterized in that, The input feature matrix is: ； in, p For hyperparameters related to monitoring accuracy, it represents the dimension of the input feature matrix; f σ(p) Indicates the order after comprehensive sorting. p The original feature corresponding to the feature index of the bit.

4. A feature-adaptive filtering-based surface quality monitoring system for milled thin-walled parts, characterized in that, The method for monitoring the surface quality of milled thin-walled parts based on feature-adaptive filtering as described in any one of claims 1-3 includes: The feature extraction module is used to extract the time-domain features, frequency-domain features, and time-frequency-domain features of the processing vibration signal, construct the original feature matrix of the sensor data, and arrange each feature in the original feature matrix in sequence. The sequence ranking acquisition module is used to calculate the feature information capacity, maximum information number and distance correlation coefficient between each feature and the target vector in turn, and obtain the sequence ranking of each feature after sorting in descending order according to the feature information capacity, maximum information number and distance correlation coefficient. The input feature matrix acquisition module is used to calculate the sum of the three sorting indices for each feature vector to obtain the feature comprehensive index of each feature; the features are sorted in ascending order based on the feature comprehensive index; the top p feature vectors in the comprehensive ranking are selected and concatenated to obtain the adaptively filtered input feature matrix; The mapping relationship model construction module is used to construct a nonlinear real-time mapping relationship model between the input feature matrix and the target feature vector based on gated recurrent units and fully connected networks, so as to obtain the surface roughness of the workpiece in real time.

5. An electronic device, characterized in that, The device includes a memory and a processor, the memory being used to store a computer program, and the processor running the computer program to enable the electronic device to perform the feature-adaptive filtering method for monitoring the surface quality of milled thin-walled parts according to any one of claims 1-3.

6. A computer-readable storage medium, characterized in that, It stores a computer program that, when executed by a processor, implements the feature-adaptive filtering method for monitoring the surface quality of milled thin-walled parts as described in any one of claims 1-3.

Images