Milling chatter online learning monitoring method based on sample secondary screening

Abstract

The milling chatter online learning monitoring method based on sample secondary screening belongs to the technical field of milling chatter monitoring, and solves the problem of poor learning real-time caused by sample redundancy in online incremental learning of an existing online monitoring model. The sample secondary screening strategy based on distance and density is introduced, the influence of sample redundancy on online learning of the model is avoided, the real-time of online learning of the model is improved, and high-precision and high-real-time online learning monitoring of milling chatter is realized. The present application is mainly used for online real-time monitoring of the processing state of thin-walled parts in the milling process.

Description

Technical Field

[0001] This invention belongs to the field of milling chatter monitoring technology. Background Technology

[0002] During milling, chatter is easily generated due to the low rigidity of the cutting tool or workpiece. Chatter reduces the surface quality of the workpiece, affects machining efficiency, and shortens the service life of machine tools and cutting tools, severely impacting the operational performance of enterprises and thus attracting significant attention. Traditional offline chatter detection methods require manual machine shutdown for inspection, which not only reduces production efficiency but also fails to meet the requirements of online intelligent control. Therefore, researching online monitoring methods for milling chatter and achieving online chatter monitoring is of great significance. Simultaneously, with the rise of intelligent manufacturing technology, there is an urgent need for an online monitoring method that can achieve high-precision, high-real-time monitoring of milling chatter, thereby providing a foundation for subsequent control of milling chatter.

[0003] Currently, researchers have conducted some research on chatter monitoring in milling. In the patent "An Online Monitoring Method for Milling Chatter (CN201810089272.1)," the cutting force signal is decomposed using wavelet transform. By extracting the correlation coefficients of the cutting force signals Fx and Fy and the variance characteristics of the high-frequency components, a multivariate regression model is established to achieve online monitoring of chatter during the milling process. In the patent "A Monitoring Method for Chatter in Machining Process (CN201910832660.9)," chatter is monitored by calculating the fractal dimension and comparing it with a monitoring threshold. In the patent "An Online Monitoring Method and Device for Machine Tool Cutting Chatter (CN202011251745.7)," the vibration signal is first decomposed and reconstructed using wavelet packets. Then, singular value decomposition is performed on the reconstructed signal, and singular value entropy features are extracted. Finally, the singular value entropy feature curve is used to determine whether chatter has occurred. However, the above milling chatter monitoring models are all static models, which do not have the ability to learn online during the online monitoring process. As the cutting progresses, the machining process will continuously change (tool wear, removal of workpiece material, etc.), and the monitoring accuracy of the static model will continue to decline. To solve this problem, the ISK-Means & OS-ELM chatter online learning monitoring method was proposed in the literature "Wang Zhixue, YueCaixu, Liu Xianli, et al. On-line Evolutionary Identification Technology for Milling Chatter of Thin Walled Parts Based on the Incremental-Sparse K-Means and the Online Sequential Extreme Learning Machine. The International Journal of Advanced Manufacturing Technology, 2023, 5-6(128): 2001-2011". This method realizes online learning monitoring of milling chatter. The specific algorithm flow is as follows: Figure 1 As shown, this model uses samples with consistent monitoring results from ISK-Means and OS-ELM (achieving a single sample selection) for online learning of the chatter monitoring model, thus solving the problem of spurious samples affecting the online learning process and improving the accuracy of online monitoring of milling chatter. However, this model uses all samples selected in the first selection for online learning, without considering the role of each sample in model updates. Therefore, it suffers from poor real-time performance in online learning due to sample redundancy, limiting its practical application. Summary of the Invention

[0004] The purpose of this invention is to solve the problem of poor real-time performance of existing online chatter monitoring models due to redundancy of online incremental learning samples. This invention provides an online learning monitoring method for milling chatter based on secondary sample screening.

[0005] A milling chatter online learning monitoring method based on secondary sample screening includes the following steps:

[0006] Step S1: Train the OS-ELM model and ISK-Means model offline to obtain the pre-trained OS-ELM model and ISK-Means model, and then proceed to step S2.

[0007] Step S2: Preprocess the cutting force signal and acceleration signal collected during the i-th sampling period online at preset time intervals during the cutting process to form a corresponding real-time sample x. i Simultaneously, steps S3 and S4 are executed; wherein, the sample is a vector composed of the multi-scale arrangement entropy features of the acceleration signal and the energy entropy features of the cutting force signal;

[0008] Step S3: Apply the pre-trained OS-ELM model to sample x. i The corresponding cutting state is identified online, and the identification result is output. It is determined whether the machining process has ended. If the result is yes, the identification ends; if the result is no, i = i + 1, the initial value of i is 1, and step S2 is executed. The identification result is either chattering state or stable state.

[0009] Step S4: Transfer the online collected sample x i After saving to the online sample set, the sample collection time is counted to determine whether there are any sample collection times that reach the threshold T0. If the result is yes, step S5 is executed; if the result is no, step S4 is returned.

[0010] Step S5: Input all samples in the online sample set into the pre-trained ISK-Means model, output the clustering results, update the cluster center of each processing state according to the clustering results, label the processing state category of each sample in the online sample set according to the clustering results, and the clustering result is flutter state or stable state. At the same time, clear the online sample set and reset the sample collection time to zero, and execute step S6.

[0011] Step S6: For the same input sample, compare the output results of the pre-trained OS-ELM model and the pre-trained ISK-Means model, and use the samples with consistent results as the incremental learning samples to form an incremental sample set, thereby achieving a first screening of samples, and then proceed to step S7.

[0012] Step S7: Determine whether there is a number of incremental learning samples for one or more processing states in the incremental sample set that exceeds the threshold N. If the result is yes, then perform secondary sample screening based on distance and density for the incremental learning samples for one or more processing states, and retain N incremental learning samples accordingly to update the incremental sample set. Use the updated incremental sample set to train the pre-trained OS-ELM model online, update the model parameters, and execute step S3.

[0013] If the result is negative, use each of the proposed incremental learning samples in the current incremental sample set as the input to the pre-trained OS-ELM model, and use the processing state corresponding to each proposed incremental learning sample as the label to train the pre-trained OS-ELM model online, update the model parameters, and execute step S3.

[0014] Preferably, the implementation methods for secondary sample selection based on distance and density include:

[0015] S71. For the incremental learning samples to be used in various processing states where the number of samples exceeds a threshold N, calculate the distance d between each incremental learning sample to the cluster center corresponding to its state under each processing state. ci The cluster center corresponding to the given state is the latest cluster center of the corresponding state obtained by the pre-trained ISK-Means model; d ci Let be the distance between the i-th incremental learning sample and its corresponding cluster center;

[0016] S72. Calculate the density ρ of each sample in the incremental learning samples for the same processing state. i , ρ i Let be the density of the i-th sample to be used for incremental learning;

[0017] S73, based on density ρ i and distance d ci Construct weights w i ;w i Let be the weight of the i-th incremental learning sample to be used;

[0018] S74, Based on the weight w i Calculate the probability p(x) of each proposed incremental learning sample being selected. i );p(x i ) represents the probability that the i-th incremental learning sample will be selected;

[0019] S75. Based on the probability p(x) of all proposed incremental learning samples being selected under the same processing state. i According to the roulette wheel method, N incremental learning samples of the same processing state are extracted N times to obtain N incremental learning samples.

[0020] Preferably,

[0021] Where τ() is a piecewise function, d ij Let d be the Euclidean distance between the i-th intended incremental learning sample and the j-th intended incremental learning sample. c Let N be the cutoff distance, and N′ be the distance x after the first sample selection. i The total number of incremental learning samples to be used in the current state.

[0022] Preferably,

[0023] e is the base of the natural logarithm, d ci For d ci The result after normalization, d max This represents the distance from the cluster center to which the sample to be incrementally learned is located under the same processing conditions.

[0024] Preferably, In the formula, N′ represents the value of x after the first sample selection. i The total number of incremental learning samples to be used in the current state.

[0025] Preferably, the implementation methods for obtaining the pre-trained OS-ELM model and ISK-Means model include:

[0026] The cutting force and acceleration signals collected at each sampling time point during the historical cutting process are preprocessed to form a corresponding sample. The processing status category is labeled according to the surface quality of the workpiece corresponding to the sample to obtain an offline sample set.

[0027] Each sample in the offline sample set is used as the input to the OS-ELM model, and the processing state corresponding to the sample is used as the label. The OS-ELM model is trained offline to obtain the pre-trained OS-ELM model.

[0028] Meanwhile, the processing state corresponding to each sample in the offline sample set is used as a label, and the ISK-Means model is used to perform cluster training on the offline sample set. Based on the label corresponding to the sample and the clustering results, the cluster center of the category to which each processing state belongs is determined, and the pre-trained ISK-Means model is obtained.

[0029] Preferably, the implementation method for preprocessing the cutting force signal and acceleration signal collected in each sampling period includes:

[0030] Energy entropy features are extracted from the cutting force signal collected during the current sampling period to obtain the energy entropy features of the cutting force signal; at the same time, multi-scale arrangement entropy features are extracted from the acceleration signal collected during the same sampling period to obtain the multi-scale arrangement entropy features of the acceleration signal, thus completing the signal preprocessing.

[0031] Preferably, the acceleration signal is the acceleration signal of the milling cutter or the acceleration signal of the workpiece.

[0032] Advantages of this invention:

[0033] (1) A secondary screening strategy based on distance and density was introduced to reasonably screen the samples for online learning of the milling chatter monitoring model, which solved the problem of poor real-time performance caused by sample redundancy during online learning of the milling chatter monitoring model and improved the efficiency and real-time performance of online learning of the model.

[0034] (2) An online learning model was used to monitor milling chatter, which solved the problem that the accuracy of the static monitoring model for milling chatter continued to decline due to dynamic changes in the machining process (tool wear, removal of workpiece material, etc.), and improved the dynamic monitoring accuracy of milling chatter. Attached Figure Description

[0035] Figure 1 This is a flowchart of an online learning and monitoring algorithm for milling chatter in existing technologies;

[0036] Figure 2 This is a schematic diagram showing the installation positions of the rotation force gauge and the acceleration sensor;

[0037] Figure 3 This is a schematic diagram of the online learning and monitoring method for milling chatter based on secondary sample screening as described in this invention. Detailed Implementation

[0038] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0039] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0040] Example 1:

[0041] See Figure 3This embodiment describes the online learning and monitoring method for milling chatter based on secondary sample screening. Step S1: Offline training is performed on the OS-ELM model and the ISK-Means model to obtain the pre-trained OS-ELM model and the ISK-Means model. Step S2 is then executed.

[0042] In practical applications, cutting force and acceleration signals during the cutting process are acquired online using a force gauge and an accelerometer. These signals are then divided into sub-signals with the same number of sampling points at regular time intervals (t), i.e., signals obtained from each sampling period. Corresponding features are then extracted from each sub-signal and applied to chatter detection. The acceleration signal can be either the milling cutter's acceleration signal or the workpiece's acceleration signal.

[0043] Specifically, the implementation methods for obtaining the pre-trained OS-ELM model and ISK-Means model include:

[0044] The cutting force and acceleration signals collected at each sampling time point during the historical cutting process are preprocessed to form a corresponding sample. The processing status category is labeled according to the surface quality of the workpiece corresponding to the sample to obtain an offline sample set.

[0045] Each sample in the offline sample set is used as the input to the OS-ELM model, and the processing state corresponding to the sample is used as the label. The OS-ELM model is trained offline to obtain the pre-trained OS-ELM model.

[0046] Simultaneously, the processing state corresponding to each sample in the offline sample set is used as a label. The ISK-Means model is then used for cluster training on the offline sample set. Based on the label corresponding to the sample and the clustering results, the cluster center of each processing state is determined, resulting in the pre-trained ISK-Means model. Specifically, the preprocessing methods for the cutting force and acceleration signals collected at each sampling time period include:

[0047] Energy entropy features are extracted from the cutting force signal collected during the current sampling period to obtain the energy entropy features of the cutting force signal; at the same time, multi-scale arrangement entropy features are extracted from the acceleration signal collected during the same sampling period to obtain the multi-scale arrangement entropy features of the acceleration signal, thus completing the signal preprocessing.

[0048] Multi-scale permutation entropy feature extraction of acceleration signals can be achieved using existing techniques, specifically:

[0049] First, the acquired acceleration signal is decomposed using EMD, and the signal is reconstructed based on the energy proportion of each mode. Then, the multi-scale permutation entropy features of the reconstructed signal are extracted and applied to flutter monitoring. When using multi-scale permutation entropy for flutter monitoring, coarse-grained processing of the reconstructed signal is required. Assume the original sequence of the signal reconstructed based on energy proportion is X = {x1, x2, ..., x...} n}, then the coarse-grained sequence can be calculated using equation (1).

[0050]

[0051] In the formula: Let $\frac{j}{j}$ be the j-th element in the coarse-grained sequence, and $\lambda$ be the scale factor, which can be selected based on the number of tool teeth. Then, the permutation entropy value of the coarse-grained sequence is calculated according to the single-scale permutation entropy principle and applied to chatter monitoring. After coarse-graining, the interference of random signals during machining is reduced, which is more conducive to monitoring the machining status.

[0052] Energy entropy feature extraction of force signals can be achieved using existing technologies, specifically:

[0053] The acquired force signal is decomposed into I-level wavelet packet decomposition. At the I-level, the force signal is decomposed into m (m=2) I ) sub-signals, and then extract the energy E1, E2, ..., E of the m sub-signals on the I-th layer. m Since the sub-signals of wavelet packet decomposition are orthogonal, the total energy E of the signal can be calculated according to equation (2):

[0054] E = E1 + E2 + ... + E m (2);

[0055] The normalized energy ratio of each sub-signal can then be calculated according to equation (3):

[0056]

[0057] In the formula, k = 1, 2, ..., m.

[0058] The energy entropy of the force signal at the macroscopic scale can be calculated according to equation (4):

[0059]

[0060] Step S2: Preprocess the cutting force signal and acceleration signal collected during the i-th sampling period online at preset time intervals during the cutting process to form a corresponding real-time sample x. i Simultaneously, steps S3 and S4 are executed; wherein, the sample is a vector composed of the multi-scale arrangement entropy features of the acceleration signal and the energy entropy features of the cutting force signal;

[0061] Step S3: Apply the pre-trained OS-ELM model to sample x. i The corresponding cutting state is identified online, and the identification result is output. It is determined whether the machining process has ended. If the result is yes, the identification ends; if the result is no, i = i + 1, the initial value of i is 1, and step S2 is executed. The identification result is either chattering state or stable state.

[0062] Step S4: Transfer the online collected sample x i After saving to the online sample set, the sample collection time is counted to determine whether there are any sample collection times that reach the threshold T0. If the result is yes, step S5 is executed; if the result is no, step S4 is returned.

[0063] Step S5: Input all samples in the online sample set into the pre-trained ISK-Means model, output the clustering results, update the cluster center of each processing state according to the clustering results, label the processing state category of each sample in the online sample set according to the clustering results, and the clustering result is flutter state or stable state. At the same time, clear the online sample set and reset the sample collection time to zero, and execute step S6.

[0064] Step S6: For the same input sample, compare the output results of the pre-trained OS-ELM model and the pre-trained ISK-Means model, and use the samples with consistent results as the incremental learning samples to form an incremental sample set, thereby achieving a first screening of samples, and then proceed to step S7.

[0065] Step S7: Determine whether there is a number of incremental learning samples for one or more processing states in the incremental sample set that exceeds the threshold N. If the result is yes, then perform secondary sample screening based on distance and density for the incremental learning samples for one or more processing states, and retain N incremental learning samples accordingly to update the incremental sample set. Use the updated incremental sample set to train the pre-trained OS-ELM model online, update the model parameters, and execute step S3.

[0066] If the result is negative, use each of the proposed incremental learning samples in the current incremental sample set as the input to the pre-trained OS-ELM model, and use the processing state corresponding to each proposed incremental learning sample as the label to train the pre-trained OS-ELM model online, update the model parameters, and execute step S3.

[0067] The online learning and monitoring method for milling chatter based on secondary sample screening described in this embodiment involves two sample screenings for each model update during the entire monitoring process. The first screening is to avoid the impact of spurious samples on the accuracy of online OS-ELM model updates, and the second screening is to avoid the impact of sample redundancy on the real-time performance of online OS-ELM model updates. This improves the real-time performance of online model learning and achieves high-precision and high-real-time online learning and monitoring of milling chatter.

[0068] Furthermore, methods for implementing secondary sample selection based on distance and density include:

[0069] S71. For the incremental learning samples to be used in various processing states where the number of samples exceeds a threshold N, calculate the distance d between each incremental learning sample to the cluster center corresponding to its state under each processing state. ci The cluster center corresponding to the given state is the latest cluster center of the corresponding state obtained by the pre-trained ISK-Means model; d ci Let be the distance between the i-th incremental learning sample and its corresponding cluster center;

[0070] S72. Calculate the density p of each sample in the incremental learning samples for the same processing state. i p i Let be the density of the i-th sample to be used for incremental learning; specifically,

[0071]

[0072] Where τ() is a piecewise function, d ij Let d be the Euclidean distance between the i-th and j-th proposed incremental learning samples. c Let N be the cutoff distance, and N′ be the distance x after the first sample selection. i The total number of incremental learning samples to be used in the current state. S73, based on density p i and distance d ci Construct weights w i ;w i Let be the weight of the i-th incremental learning sample to be used;

[0073] Specifically,

[0074] in,

[0075] e is the base of the natural logarithm, d′ ci For d ci The result after normalization, d maxThis represents the distance from the cluster center to which the sample to be incrementally learned is located under the same processing conditions.

[0076] S74, Based on the weight w i Calculate the probability p(x) of each proposed incremental learning sample being selected. i );p(x i ) represents the probability that the i-th incremental learning sample will be selected; specifically,

[0077]

[0078] In the formula, N′ represents the value of x after the first sample selection. i The total number of incremental learning samples to be used in the current state;

[0079] S75. Based on the probability p(x) of all proposed incremental learning samples being selected under the same processing state. i According to the roulette wheel method, N incremental learning samples of the same processing state are extracted N times to obtain N incremental learning samples.

[0080] This preferred embodiment introduces a secondary sample screening strategy based on distance and density, which avoids the impact of sample redundancy on the online evolutionary learning of the milling chatter model, thereby improving the real-time performance of the model's online evolution and realizing high-precision, high-real-time online evolutionary monitoring of milling chatter.

[0081] As an example, a three-axis vertical CNC milling machine will be used to illustrate the implementation of the present invention in detail.

[0082] The first step is the acquisition of force and acceleration signals during the processing.

[0083] Mount the rotary force gauge onto the spindle of the CNC milling machine, and then clamp the milling cutter using the rotary force gauge's tool holder. Attach the accelerometer to the opposite side of the workpiece's surface to be machined using coupling adhesive, such as... Figure 2 As shown. After the sensors were set up, cutting was performed and cutting force and acceleration signals were collected. During the machining process, the sampling frequency of both force and acceleration signals was set to 10kHz. Then, the signal was divided into sub-signals with 512 sampling points at time intervals of 0.0512s. Finally, the corresponding features were extracted from each sub-signal and applied to chatter monitoring.

[0084] The second step is to extract features based on the signal.

[0085] Step 21: Extraction of multi-scale permutation entropy features from acceleration signals

[0086] The accelerometer signal obtained in the first step is decomposed into EMD, and the top three modes with the highest energy proportion are selected to reconstruct the signal. Then, the reconstructed signal is coarsened according to formula (1), and the multi-scale arrangement entropy feature of the reconstructed signal is extracted. Here, the scale factor λ = 4 (the same as the number of tool teeth).

[0087] Step 22: Extraction of energy entropy features from force signals

[0088] As an example, the Dmey wavelet basis can be used to decompose the force signal obtained in the first step into a 3-layer wavelet packet. Then, the total energy of each frequency band in the third layer is calculated according to formula (2), the normalized energy ratio is calculated according to formula (3), and finally the energy entropy feature is calculated according to formula (4).

[0089] Step 23 involves combining the multi-scale arrangement entropy features of the acceleration signal and the energy entropy features of the force signal into a vector, which is then applied to the training and monitoring of the flutter monitoring model.

[0090] The third step is to construct an online learning monitoring model for milling chatter based on secondary sample screening.

[0091] I. Offline Training of ISK-Means & OS-ELM Models

[0092] The ISK-Means model was configured with two cluster centers: a stable state and a fluttering state. The OS-ELM model used the sigmoidal function as the activation function, and the number of hidden layer neurons was set to 20. A variable-cut deep machining experiment was conducted, and signals were acquired. Then, following the first and second steps, the features of the offline samples were obtained. The sample categories were categorized based on the surface quality of the workpiece. After this processing, 300 samples were obtained from the stable state and 210 samples from the fluttering state.

[0093] The acquired samples were divided into a test set and a training set in a 7:3 ratio. The training set was used to train both models offline, and the test set was used to test the trained models. The accuracy threshold T for the test set was set to 90%. After this setup, the offline test accuracy of the ISK-Means model was 90.2%, and the offline test accuracy of the OS-ELM model was 95%, both exceeding the set threshold. Therefore, they can be applied to online monitoring and online learning of milling chatter.

[0094] II. Online Monitoring and Learning of Milling Chatter Monitoring Model Based on Secondary Sample Screening

[0095] Combination Figure 3 The detailed explanation of the online monitoring and learning process of the milling chatter monitoring model based on secondary sample screening is as follows:

[0096] Step 1: During online monitoring, the processing status is monitored in real time using the OS-ELM model, and the current processing status is output. Simultaneously, newly acquired samples are collected. When the running cycle reaches the threshold of 20.48 seconds (when the number of newly added samples reaches 400), the ISK-Means model is used to cluster the newly added samples, obtaining new cluster centers and the category of each sample.

[0097] Step 2 involves selecting samples whose monitoring results are consistent between the OS-ELM model and the ISK-Means model as the incremental learning samples to be used, thus completing the first screening of the samples. After this screening, the number of samples in the stable processing state and the number of samples in the flutter processing state in the sample set to be used for incremental learning are 235.

[0098] Step 3: Determine if the number of incremental learning samples for each category exceeds the set threshold. Here, the secondary screening threshold is set to 150. After the first screening, the number of samples in the stable processing state within the incremental learning sample set is 235, which exceeds the set threshold. Therefore, a secondary sample screening based on distance and density is required. The number of samples in the flutter processing state is 109, which is less than the set threshold, so no secondary sample screening is needed.

[0099] Step 4: Calculate the distance from each steady-state sample in the incremental learning sample set to its own category center.

[0100] Step 5: Calculate the density of each steady-state sample point using formula (5). When calculating the sample density, a cutoff distance d needs to be set. c Here, the distances from the steady-state samples calculated in step 4 to the cluster centers are arranged in ascending order, and the value at the 2nd percentile is taken as d. c The value of d c =0.013.

[0101] Step 6: Calculate the weight of each steady-state sample according to formula (6).

[0102] Step 7: Calculate the probability of each steady-state sample being selected according to formula (7).

[0103] Step 8: Extract incremental learning samples from the stable state according to the roulette wheel rule to obtain 150 stable state samples, thus achieving the second screening of samples in the stable state.

[0104] Step 9: Combine the 150 samples obtained after secondary screening in the steady state and the 109 samples obtained after primary screening in the fluttering state to form the final incremental learning samples, totaling 259, and then apply them to the online learning of the OL-ELM model.

[0105] Step 10: Continue to monitor the processing status online using the updated ISK-Means & OL-ELM & and repeat the above steps to achieve online monitoring and evolutionary learning of the model.

[0106] While the invention has been described herein with reference to specific embodiments, it should be understood that these embodiments are merely examples of the principles and applications of the invention. Therefore, it should be understood that many modifications can be made to the exemplary embodiments, and other arrangements can be designed without departing from the spirit and scope of the invention as defined by the appended claims. It should be understood that different dependent claims and features described herein can be combined in ways different from those described in the original claims. It is also understood that features described in conjunction with individual embodiments can be used in other described embodiments.

Claims

1. A method for online learning and monitoring of milling chatter based on secondary sample screening, characterized in that, The method includes the following steps: Step S1: Train the OS-ELM model and ISK-Means model offline to obtain the pre-trained OS-ELM model and ISK-Means model, and then proceed to step S2. Step S2: Preprocess the cutting force signal and acceleration signal collected during the i-th sampling period online at preset time intervals during the cutting process to form a corresponding real-time sample x. i Simultaneously, steps S3 and S4 are executed; wherein, the sample is a vector composed of the multi-scale arrangement entropy features of the acceleration signal and the energy entropy features of the cutting force signal; Step S3: Apply the pre-trained OS-ELM model to sample x. i The corresponding cutting state is identified online, and the identification result is output. It is determined whether the machining process has ended. If the result is yes, the identification ends; if the result is no, i = i + 1, the initial value of i is 1, and step S2 is executed. The identification result is either chattering state or stable state. Step S4: Transfer the online collected sample x i After saving to the online sample set, the sample collection time is counted to determine whether there are any sample collection times that reach the threshold T0. If the result is yes, step S5 is executed; if the result is no, step S4 is returned. Step S5: Input all samples in the online sample set into the pre-trained ISK-Means model, output the clustering results, update the cluster center of each processing state according to the clustering results, label the processing state category of each sample in the online sample set according to the clustering results, and the clustering result is flutter state or stable state. At the same time, clear the online sample set and reset the sample collection time to zero, and execute step S6. Step S6: For the same input sample, compare the output results of the pre-trained OS-ELM model and the pre-trained ISK-Means model, and use the samples with consistent results as the incremental learning samples to form an incremental sample set, thereby achieving a first screening of samples, and then proceed to step S7. Step S7: Determine whether there is a number of incremental learning samples for one or more processing states in the incremental sample set that exceeds the threshold N. If the result is yes, then perform secondary sample screening based on distance and density for the incremental learning samples for one or more processing states, and retain N incremental learning samples accordingly to update the incremental sample set. Use the updated incremental sample set to train the pre-trained OS-ELM model online, update the model parameters, and execute step S3. If the result is negative, use each of the proposed incremental learning samples in the current incremental sample set as the input to the pre-trained OS-ELM model, and use the processing state corresponding to each proposed incremental learning sample as the label to train the pre-trained OS-ELM model online, update the model parameters, and execute step S3.

2. The online learning monitoring method for milling chatter based on secondary sample screening according to claim 1, characterized in that, Methods for implementing secondary sample selection based on distance and density include: S71. For the incremental learning samples to be used in various processing states where the number of samples exceeds a threshold N, calculate the distance d between each incremental learning sample to the cluster center corresponding to its state under each processing state. ci The cluster center corresponding to the given state is the latest cluster center of the corresponding state obtained by the pre-trained ISK-Means model; d ci Let be the distance between the i-th incremental learning sample and its corresponding cluster center; S72. Calculate the density ρ of each sample in the incremental learning samples for the same processing state. i , ρ i Let be the density of the i-th sample to be used for incremental learning; S73, based on density ρ i and distance d ci Construct weights w i ;w i Let be the weight of the i-th incremental learning sample to be used; S74, Based on the weight w i Calculate the probability p(x) of each proposed incremental learning sample being selected. i );p(x i ) represents the probability that the i-th incremental learning sample will be selected; S75. Based on the probability p(x) of all proposed incremental learning samples being selected under the same processing state. i According to the roulette wheel method, N incremental learning samples of the same processing state are extracted N times to obtain N incremental learning samples.

3. The online learning monitoring method for milling chatter based on secondary sample screening according to claim 2, characterized in that, Where τ() is a piecewise function, d ij Let d be the Euclidean distance between the i-th intended incremental learning sample and the j-th intended incremental learning sample. c Let N be the cutoff distance, and N′ be the distance x after the first sample selection. i The total number of incremental learning samples to be used in the current state.

4. The online learning monitoring method for milling chatter based on secondary sample screening according to claim 2, characterized in that, e is the base of the natural logarithm, d′ ci For d ci The result after normalization, d max This represents the distance from the cluster center to which the sample to be incrementally learned is located under the same processing conditions.

5. The online learning monitoring method for milling chatter based on secondary sample screening according to claim 2, characterized in that, In the formula, N′ represents the value of x after the first sample selection. i The total number of incremental learning samples to be used in the current state.

6. The online learning monitoring method for milling chatter based on secondary sample screening according to claim 1, characterized in that, The methods for obtaining the pre-trained OS-ELM model and ISK-Means model include: The cutting force and acceleration signals collected at each sampling time point during the historical cutting process are preprocessed to form a corresponding sample. The processing status category is labeled according to the surface quality of the workpiece corresponding to the sample to obtain an offline sample set. Each sample in the offline sample set is used as the input to the OS-ELM model, and the processing state corresponding to the sample is used as the label. The OS-ELM model is trained offline to obtain the pre-trained OS-ELM model. Meanwhile, the processing state corresponding to each sample in the offline sample set is used as a label, and the ISK-Means model is used to perform cluster training on the offline sample set. Based on the label corresponding to the sample and the clustering results, the cluster center of the category to which each processing state belongs is determined, and the pre-trained ISK-Means model is obtained.

7. The online learning monitoring method for milling chatter based on secondary sample screening according to claim 1 or 6, characterized in that, The methods for preprocessing the cutting force and acceleration signals collected at each sampling time period include: Energy entropy features are extracted from the cutting force signal collected during the current sampling period to obtain the energy entropy features of the cutting force signal; at the same time, multi-scale arrangement entropy features are extracted from the acceleration signal collected during the same sampling period to obtain the multi-scale arrangement entropy features of the acceleration signal, thus completing the signal preprocessing.

8. The online learning monitoring method for milling chatter based on secondary sample screening according to claim 1 or 6, characterized in that, The acceleration signal can be the acceleration signal of the milling cutter or the acceleration signal of the workpiece.

Images