Cross-domain cutter breakage monitoring method and system based on transfer learning

Abstract

The present invention provides a cross-domain cutter breakage monitoring method and system based on transfer learning, which minimizes the distribution difference between samples from different domains and the dispersion of intra-class samples while preserving the prior distribution structure attributes of the training data. Furthermore, when there are no broken samples in the target domain, the source domain knowledge is effectively transferred to the target domain to train a cross-domain monitoring model. Furthermore, the present invention also uses an improved multi-scale symbol dynamic entropy to break away from dependence on machine tool and cutter parameters, thereby alleviating the drawback of losing useful information in the traditional mean-based coarse-graining process.

Description

[Technical Field]

[0001] (CROSS-REFERENCE TO RELATED APPLICATIONS) This invention claims priority to a Chinese patent application filed with the State Intellectual Property Office of the People's Republic of China on August 8, 2023, bearing application number 202310994189.X and entitled "Cross-domain cutter breakage monitoring method and system based on transfer learning," the entire contents of which are incorporated by reference into and constitute part of the present invention for all purposes.

[0002] The present invention relates to the technical field of cutter breakage monitoring, and to a cross-domain cutter breakage monitoring method and system based on transfer learning. [Background technology]

[0003] The discussion in this section merely provides background information related to the present invention and does not necessarily constitute prior art.

[0004] Milling is one of the major manufacturing technologies in modern industry, and milling cutters are a critical component of the milling system. During milling, factors such as severe wear on the cutting edge, improper machining conditions, and uneven finishing allowances can all cause the cyclic impact load on the milling cutter to increase rapidly, ultimately leading to cutter breakage. Failure to timely detect and replace a broken cutter during milling can result in damage to the workpiece, machine tool failure, and even operator safety. Therefore, the research and application of cutter breakage monitoring technology is extremely important to prevent catastrophic accidents during the milling process.

[0005] Current cutter breakage detection methods are primarily directed and indirect. Direct detection relies on machine vision to directly measure changes in the geometry of the cutting edge. However, due to the influence of cutting fluid and lighting conditions, it requires stopping the machine for detection and is therefore not applicable online. Data-driven indirect detection methods, which use physical signals closely related to the cutter, such as cutting force, vibration, and current, are practical for online detection. Conventional indirect cutter breakage monitoring methods collect signals and then extract features, such as the number of signal abnormal peak periods, frequency domain fusion features, and average frequency, from them. These features are then input into a machine learning classification model to build a cutter breakage monitoring model. However, the calculations of these methods depend on relevant machine tool and cutter parameters, such as spindle speed and cutter tooth count.

[0006] Traditional machine learning classification models, including support vector machines (SVN), k-nearest neighbors (KNN), and artificial neural networks (ANN), can effectively detect cutter breakage under fixed cutting conditions if the feature distributions of the training and test sets are consistent. However, changes in cutting conditions can alter signal characteristics and further change the feature distribution. Therefore, traditional cutter breakage monitoring methods face challenges in identifying samples from the new distribution, necessitating model retraining. However, model retraining requires the inclusion of a large number of markup samples from all classes. Unfortunately, due to strict safety restrictions implemented during production, it is difficult to obtain a sufficient number of breakage samples under various operating conditions. Therefore, the inability to obtain a sufficient number of breakage samples under changing operating conditions (i.e., the target domain) is a major obstacle limiting the application of cutter breakage monitoring systems in actual industrial production.

[0007] For the same machining type, the operating principles of the machining process are similar, so the predicted new data and historical training data contain similar feature information. Feature-based transfer learning aims to train a cross-domain monitoring model by applying extensive knowledge from one domain (i.e., the source domain) to another related but different new domain (i.e., the target domain). This means that the training data and test data in cutter breakage monitoring can have different distributions. Therefore, transfer learning can be used to solve the problem of difficult-to-obtain target domain samples. However, traditional transfer learning methods, including transfer component analysis (TCA), joint distribution adaptation (JDA), and transfer joint matching (TJM), all assume that the target domain contains a large amount of unmarked data for all classes. When training a cutter breakage monitoring model in actual production, only normal samples from the target domain can usually be collected. Therefore, the above-mentioned methods cannot be effectively applied to cross-domain cutter breakage monitoring. Furthermore, the application environment of these methods is bearing fault diagnosis, which has stable operating conditions, and they do not take into account the issue of overlapping edge samples between classes due to the complex milling process. Summary of the Invention

[0008] To solve the above problems, the present invention proposes a cross-domain cutter damage monitoring method and system based on transfer learning, which minimizes the distribution difference between samples from different domains and the dispersion of samples within classes while preserving the prior distribution structure attributes of the training data. Furthermore, if there are no damaged samples in the target domain, the knowledge of the source domain is effectively transferred to the target domain to train a cross-domain monitoring model. Furthermore, the present invention also uses an improved multi-scale symbol dynamic entropy to break away from dependence on machine tool and cutter parameters, thereby alleviating the drawback of losing useful information in the traditional mean-based coarse-graining process.

[0009] According to some embodiments, the present invention adopts the following technical solutions. acquiring a vibration signal associated with the milling process; Extracting features from the feed direction vibration signal by improved multi-scale symbol dynamic entropy, and constructing a training sample set for a transfer learning model including samples of all classes in the source domain and normal samples in the target domain, and a test set for a supervised model including samples of all classes in the target domain; a multi-scale symbolic dynamic entropy refinement process converting a time series of vibration signals in the feed direction into a symbolic time series, and dividing the symbolic time series into non-overlapping segments by scale; Computing coarse-grained time series of different types of data within each segment; Calculating the symbolic dynamic entropy of different types of coarse-grained time series at each scale; amalgamating each symbol dynamic entropy to form a final multi-scale symbol dynamic entropy; Using a transfer learning model to obtain a mapping matrix and further calculate a generic feature representation to apply the source domain knowledge to the target domain; training a cross-domain cutter breakage monitoring model based on a generic feature representation of a training sample set; and calculating a generalized feature representation of the test set using the transfer learning model and inputting the generalized feature representation into the cross-domain cutter breakage monitoring model to realize cross-domain cutter breakage monitoring.

[0010] In an alternative embodiment, the segments are equal length segments.

[0011] In alternative embodiments, the different types of coarse-grained time series include mean-based coarse-grained time series, root-mean-square-based coarse-grained time series, and maximum-based coarse-grained time series.

[0012] As an optional embodiment, the specific process of calculating the symbolic dynamic entropy of different types of coarse-grained time series at each scale is to reconstruct the time series into a series of embedding vectors based on phase space embedding theory, where the embedding vectors are related to embedding dimensions and time lags; Calculating symbol dynamic entropy from the state mode probabilities and state transition probabilities of the embedding vector; Includes.

[0013] In an optional embodiment, the process of constructing the transfer learning model includes constructing constraint terms based on locality-preserving projection, constructing maximum average discrepancy constraint terms based on a non-parametric distance estimation method, calculating corresponding constraint terms according to the class centers of samples of each class in the source domain and the target domain, calculating regularization constraint terms, assigning corresponding weights based on all the constraint terms, respectively, and taking the corresponding minimum sum as the objective function to form a transfer learning model.

[0014] In an optional embodiment, in the absence of corrupted samples of the target domain, the performance of the transfer learning model is evaluated and parameter optimization is also performed on the transfer learning model parameters using Bayesian optimization, and in the optimization process, the Bayesian optimization objective function is the minimum ratio of the maximum average mismatch distance as the numerator and the data center-to-class center distance as the denominator.

[0015] In an optional embodiment, the generic feature representation of the training sample set output by the transfer learning model is input into a cross-domain cutter breakage monitoring model to train the cross-domain cutter breakage monitoring model, and the cross-domain cutter breakage monitoring model is a decision tree classifier.

[0016] a signal acquisition module configured to acquire vibration signals associated with the milling process; A feature extraction module configured to extract features from the feed direction vibration signal by improved multi-scale symbol dynamic entropy, and construct a training sample set for a transfer learning model including samples of all classes in the source domain and normal samples in the target domain, and a test set for a supervision model including samples of all classes in the target domain, a multi-scale symbolic dynamic entropy refinement process converting a time series of vibration signals in the feed direction into a symbolic time series, and dividing the symbolic time series into non-overlapping segments by scale; Computing coarse-grained time series of different types of data within each segment; Calculating the symbolic dynamic entropy of different types of coarse-grained time series at each scale; a feature extraction module including: fusing each symbol dynamic entropy to form a final multi-scale symbol dynamic entropy; a transfer learning module configured to use a transfer learning model to obtain a mapping matrix and further compute a generic feature representation to apply source domain knowledge to a target domain; a damage monitoring module configured to train a cross-domain cutter breakage monitoring model based on the generic feature representation of the training sample set, calculate the generic feature representation of the test set using a transfer learning model, and input the calculated generic feature representation into the cross-domain cutter breakage monitoring model to realize cross-domain cutter breakage monitoring; A cross-domain cutter breakage monitoring system based on transfer learning.

[0017] A computer-readable storage medium having a plurality of commands stored thereon, said commands being suitable for being loaded by a processor of a terminal device to perform the steps of said method.

[0018] A terminal device comprising a processor and a computer-readable storage medium, the processor for implementing each command and the computer-readable storage medium for storing a plurality of commands, the commands being suitable for being loaded by the processor to perform the steps of the method.

[0019] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0020] The transfer learning model proposed in this invention constructs constraint terms based on locality-preserving projection, constructs maximum average discrepancy constraint terms based on non-parametric distance estimation methods, calculates corresponding constraint terms according to the class centers of samples of each class in the source and target domains, and calculates regularization constraint terms, thereby minimizing the distribution difference between samples from different domains and the dispersion of samples within classes while preserving the prior distribution structure attributes of the training data. If there are no corrupted samples in the target domain, the knowledge of the source domain can be effectively transferred to the target domain to train a cross-domain supervision model.

[0021] The objective function for evaluating the transfer learning model proposed in this invention can realize the optimization of the transfer learning model parameters without the corrupted samples of the target domain, which overcomes the drawback that the parameter optimization in the traditional method relies on identifying the faulty samples of the target domain by the supervision model, and is more suitable for practical application.

[0022] The improved multi-scale symbolic dynamic entropy proposed in this invention, instead of the symbolic dynamic entropy of the coarse-grained time series based on the average value, calculates the coarse-grained time series based on the average value, RMS and maximum value respectively, and finally fuses them into one feature vector, breaking away from the dependence on the machine tool and cutter parameters and alleviating the drawback of losing useful information in the traditional coarse-grained process based only on the average value.

[0023] In order to make the above-mentioned objects, features and advantages of the present invention more clearly comprehensible, a particularly preferred embodiment of the present invention will be described in detail below with reference to the accompanying drawings.

[0024] The drawings in the specification that form a part of this invention are intended to provide a further understanding of the invention, and the illustrative embodiments of the invention and the descriptions thereof are intended to interpret the invention and are not intended to unduly limit the invention. [Brief explanation of the drawings]

[0025] [Figure 1] 1 is an overview of a cross-domain cutter breakage monitoring frame. [Figure 2] FIG. 1 is a schematic diagram of an experimental cutter. [Figure 3] The error bar graphs of the proposed improved multi-scale symbol dynamic entropy for different cutter conditions are shown. [Figure 4] This is a 2D visualization of the original features and the generic feature representation output by the transfer learning model. [Figure 5] 10 is a column chart of the accuracy of the submitted method when monitoring cross-domains. DETAILED DESCRIPTION OF THE INVENTION

[0026] In the following the invention will be further explained with reference to the figures and examples.

[0027] It should be pointed out that the following detailed description is all exemplary and is intended to further explain the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art.

[0028] It should be noted that the terminology used herein is for the purpose of describing specific embodiments only and is not intended to limit the exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular is intended to include the plural, and it should also be understood that when the terms "comprise" and / or "include" are used herein, they indicate the presence of features, steps, operations, devices, assemblies, and / or combinations thereof.

[0029] Example 1 This embodiment provides a cross-domain cutter breakage monitoring method based on transfer learning, which, as shown in Figure 1, includes the following steps S1, S2, S3, S4, and S5.

[0030] In step S1, the vibration signals related to the milling process are sensed by an acceleration sensor (the sensor selected in this embodiment has the model number Dytran 3263A2 and a sensitivity of 100mv / g) attached to the non-rotating part of the spindle of the machine tool, and collected by a data acquisition system (DH5922D), and the collected signals are transmitted to a computer via a gigabit network cable.

[0031] In step S2, features are extracted from the vibration signal in the feed direction by the proposed improved multi-scale symbol dynamic entropy (IMSDE), and a training sample set for the transfer learning model containing samples of all classes in the source domain and normal samples in the target domain, and a test set for the supervision model containing samples of all classes in the target domain, are constructed.

[0032] In step S2, the calculation of the improved multi-scale symbol dynamic entropy includes the following steps S2.1 and S2.2.

[0033] In step S2.1, we convert a given time series G into a symbolic time series S={s1, s2, …, s N}, and then divide the symbolic time series into non-overlapping equal-length segments at scale τ. Several statistics of the data within each segment are calculated to construct different types of coarse-grained time series, as follows: Coarse-grained time series based on mean values ​​are

number

number

number

[0034] In step S2.2, we calculate the symbolic dynamic entropy (SDE) of different types of coarse-grained time series at each scale. Taking the original symbolic time series (i.e., τ = 1) as an example, the calculation of symbolic dynamic entropy includes the following sub-steps: step S2.2.1, step S2.2.2, and step S2.2.3.

[0035] In step S2.2.1, a sequence S = {s1, s2, ..., s N} into a set of embedding vectors as follows:

number

[0036] In step S2.2.2, the symbol dynamic entropy is calculated from the state mode probability and state transition probability of the embedding vector, and the formula is:

number

number

number

[0037] In step S2.2.3, calculate the improved multi-scale symbolic dynamic entropy as follows:

number

[0038] In the present invention, the experimental cutter is shown in Figure 2. In Test 1, the corresponding cutting parameters are spindle speeds of 600, 1200 and 1800 rpm, feed rate of 60 mm / min, axial cutting depth of 2 mm, and radial cutting depth of 2 mm. In Test 2, the cutting parameters are spindle speeds of 600, 1200 and 1800 rpm, feed rate of 120 mm / min, axial cutting depth of 2 mm, and radial cutting depth of 2 mm.

[0039] In the present invention, the calculation parameters of IMSDE were set as follows: embedding dimension m=2, time lag λ=1, number of symbols ε=25.

[0040] The IMSDE features extracted by step S2 are shown in Figure 3. To better illustrate the high-scale advantages of this method, each signal sample in the figure contains 4000 sample points and the scale factor is τ = 10. As can be seen from Figure 3, this method can extract features with good segmentation performance, and as the scale increases, the effective information in the signal is not lost.

[0041] In step S3, after obtaining the original IMSDE feature set, a transfer learning model is used to obtain a mapping matrix to calculate a generic feature representation. The rich knowledge of one domain (i.e., the source domain) is then applied to a new, related but different domain (i.e., the target domain). The source domain and the target domain are denoted by Ds={X s ,Y s} and Dt={X t ,Y t} is defined as

number

number

[0042] The total objective function of the transfer learning model is

number

[0043] In step S3, constructing the transfer learning model mainly includes the following six sub-steps: step S3.1, step S3.2, step S3.3, step S3.4, step S3.5, and step S3.6.

[0044] In step S3.1, construct the constraint term L LPP based on Locality Preserving Projection (LPP). Given the training sample set

Number

Number

[0045] Distance d ij is calculated based on the cosine distance as follows:

number

number

number

number

number

[0046] In step S3.2, the constraint term L is calculated based on the non-parametric distance estimation method - Maximum Mean Discrepancy (MMD). MMD The feature representation after dimensionality reduction based on kernel LPP is z(x i )=A T φ(X) T φ(xi) and L MMD is defined as follows:

number

number

number

number

number

number

number

[0047] In step S3.3, the constraint term L is defined by defining the class centers. CSD The main role of this constraint is to increase the intra-class aggregation degree of samples in each class and alleviate the overlap problem of edge samples between classes in complex milling. The class centers of samples of each class in each domain are as follows:

number

number

number

number

number

[0048] L based on a training sample set containing samples of all classes in the source domain and normal samples in the target domain CSD is defined as follows:

number

number

number

[0049] In step S3.4, we define the regularization constraint term L R can control the complexity of the model by limiting the model parameters and reduce the risk of overfitting, and is defined as follows:

number

[0050] In step S3.5, the overall objective function is set and solved. Based on the above substeps, the overall objective function is set as follows:

number

number

number

[0051] Based on constrained optimization theory, Lagrange multipliers

number

number

number

number

[0052] In step S3.6, the performance of the transfer learning model is evaluated in the absence of corrupted samples in the target domain using an objective function based on the submitted data centers and class centers, and parameter optimization is performed on the transfer learning model parameters using Bayesian optimization.

[0053] Pseudo target domain sample dataset

number

number

number

number

number

number

number

number

[0054] Data Center-Class Center Distance (D DCD ) is calculated as follows:

number

number

[0055] In this embodiment, the kernel mapping is realized by a Gaussian kernel function, the parameter of the heat kernel function is σ=1, and other transfer learning model parameters that need to be optimized by Bayes and the proposed objective function and their selection ranges are as follows:

[0056] The bandwidth δ of the Gaussian kernel function can be selected within the range [10 -3 ,10], the selection range of the feature dimension l after dimensional reduction is [2,4], the selection range of the number of nearest neighbors k in the KNN classifier is [2,6], and L LPP The trade-off factor μ of the term is selected in the range [10 -3 ,10 3 ] and L MMD The trade-off factor ρ is selected within the range [10 -3 ,10 3 ] and L R The range of selection of the trade-off factor γ is [10 -3 ,10 2 ].

[0057] In this example, six types of monitoring tasks are defined: A → B, B → A, A → C, C → A, B → C, and C → B. Here, A → B indicates that A is the source domain, B is the target domain, and a transfer learning model is constructed using samples of all classes in A and normal samples in B.

[0058] In this example, Figure 4(a) shows the 2D t-SNE visualization of the original IMSDE in Test 1, and Figure 4(b) shows the generalized feature representation output by the transfer learning model for monitoring task A → B in Test 1. To meet the requirements of online monitoring, each signal sample here contains 1,000 sample points. As can be seen from Figure 4, the proposed transfer learning model can effectively reduce the distribution difference between different domains and further realize cross-domain cutter breakage monitoring.

[0059] In step S4, a cross-domain cutter breakage monitoring model is trained based on the generic feature representation of the training sample set. In this embodiment, the cross-domain monitoring model employs a decision tree classifier. The generic feature representation of the training sample set of the transfer learning model is input into the decision tree classifier to train the cross-domain cutter breakage monitoring model.

[0060] In step S5, the transfer learning model calculates a generic feature representation of the test set and inputs it into the cross-domain monitoring model to realize cross-domain cutter breakage monitoring. Figure 5 shows the classification accuracy of six cross-domain monitoring tasks in Test1 and Test2, with 100 samples for each class. As can be seen from Figure 5, the cross-domain cutter breakage monitoring method based on transfer learning proposed in this embodiment can effectively realize cutter breakage monitoring when there are no broken samples in the target domain.

[0061] The values, model number selections, etc. in the above examples are merely exemplary and may be adjusted or substituted in other examples.

[0062] The present invention further provides the following product examples.

[0063] A cross-domain cutter breakage monitoring system based on transfer learning, comprising: a signal acquisition module configured to acquire vibration signals associated with the milling process; A feature extraction module configured to extract features from the feed direction vibration signal by improved multi-scale symbol dynamic entropy, and construct a training sample set for a transfer learning model including samples of all classes in the source domain and normal samples in the target domain, and a test set for a supervision model including samples of all classes in the target domain, a multi-scale symbolic dynamic entropy refinement process converting a time series of vibration signals in the feed direction into a symbolic time series, and dividing the symbolic time series into non-overlapping segments by scale; Computing coarse-grained time series of different types of data within each segment; Calculating the symbolic dynamic entropy of different types of coarse-grained time series at each scale; a feature extraction module including: fusing each symbol dynamic entropy to form a final multi-scale symbol dynamic entropy; a transfer learning module configured to use a transfer learning model to obtain a mapping matrix and further compute a generic feature representation to apply source domain knowledge to a target domain; a damage monitoring module configured to train a cross-domain cutter breakage monitoring model based on the generic feature representation of the training sample set, calculate the generic feature representation of the test set using a transfer learning model, and input the calculated generic feature representation into the cross-domain cutter breakage monitoring model to realize cross-domain cutter breakage monitoring; Equipped with.

[0064] A computer-readable storage medium having a plurality of commands stored thereon, said commands being suitable for being loaded by a processor of a terminal device to perform the steps of said method.

[0065] A terminal device comprising a processor and a computer-readable storage medium, the processor for implementing each command, the computer-readable storage medium for storing a plurality of commands, the commands being suitable for being loaded by the processor to perform the steps of the method.

[0066] Those skilled in the art should understand that embodiments of the present invention may be provided as a method, a system, or a computer program product. Therefore, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware. The present invention may also take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, magnetic disk memory, CD-ROM, optical memory, etc.) containing computer-usable program code.

[0067] The present invention has been described with reference to flowcharts and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the present invention. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and combinations of flows and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions may be provided to a processor of a common computer, a special purpose computer, an embedded processor, or other programmable data processing apparatus to produce a machine, whereby the instructions, executed by the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in one or more flows of the flowcharts and / or one or more blocks of the block diagrams.

[0068] These computer program instructions may be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to operate in a particular manner, whereby the instructions stored in the computer-readable memory create an article of manufacture that includes instruction means for implementing the functions specified in one or more flows of the flowcharts and / or one or more blocks of the block diagrams.

[0069] These computer program instructions may be loaded into a computer or other programmable data processing apparatus to cause the computer or other programmable data processing apparatus to perform a series of operational steps to generate a computer-executed process, whereby the instructions executed on the computer or other programmable data processing apparatus provide steps for implementing the functions specified in one or more flows of the flowcharts and / or one or more blocks of the block diagrams.

[0070] The above description is only a preferred embodiment of the present invention, and is not intended to limit the present invention. Those skilled in the art can make various modifications and changes to the present invention. Any modifications, equivalent replacements, improvements, etc. made without departing from the spirit and principle of the present invention shall fall within the protection scope of the present invention.

[0071] Although the above describes specific embodiments of the present invention with reference to the drawings, it does not limit the protection scope of the present invention, and those skilled in the art should understand that various modifications or variations obtained by those skilled in the art based on the technical solutions of the present invention without requiring creative efforts still fall within the protection scope of the present invention.

Claims

1. acquiring a vibration signal associated with the milling process; Extracting features from the feed direction vibration signal by improved multi-scale symbol dynamic entropy, and constructing a training sample set for a transfer learning model including samples of all classes in the source domain and normal samples in the target domain, and a test set for a supervised model including samples of all classes in the target domain; a multi-scale symbolic dynamic entropy refinement process converting a time series of vibration signals in the feed direction into a symbolic time series, and dividing the symbolic time series into non-overlapping segments by scale; calculating different types of coarse-grained time series of data within each segment, wherein the different types of coarse-grained time series include a mean-based coarse-grained time series, a root-mean-square-based coarse-grained time series, and a maximum-based coarse-grained time series; Calculating the symbolic dynamic entropy of different types of coarse-grained time series at each scale; amalgamating each symbol dynamic entropy to form a final multi-scale symbol dynamic entropy; obtaining a mapping matrix using a transfer learning model, and further computing a generic feature representation to apply source domain knowledge to a target domain; The total objective function of the transfer learning model is [Number 53] where L ＬＰＰ is a constraint term constructed based on locality preserving projection (LPP), and L ＭＭＤ is a constraint term constructed based on the non-parametric distance estimation method - maximum mean discrepancy (MMD), and L ＣＳＤ is the constraint term proposed by defining the class center, and L Ｒ is the regularization constraint term, and the trade-off parameters μ, ρ, and γ are all greater than 0. If there are no corrupted samples in the target domain, evaluate the performance of the transfer learning model and also use Bayesian optimization to perform parameter optimization on the transfer learning model parameters, in the optimization process, use the minimum ratio of the maximum average discrepancy distance as the numerator and the data center-class center distance as the denominator as the Bayesian optimization objective function; training a cross-domain cutter breakage monitoring model based on a generic feature representation of a training sample set; and calculating a general feature representation of the test set using the transfer learning model and inputting the general feature representation into the cross-domain cutter breakage monitoring model to realize cross-domain cutter breakage monitoring.

2. The cross-domain cutter breakage monitoring method based on transfer learning according to claim 1 , wherein the segments are equal-length segments.

3. The specific process of calculating the symbolic dynamic entropy of different types of coarse-grained time series at each scale is to reconstruct the time series into a series of embedding vectors based on phase space embedding theory, where the embedding vectors are related to the embedding dimension and the time lag; Calculating symbol dynamic entropy from the state mode probabilities and state transition probabilities of the embedding vector; The cross-domain cutter breakage monitoring method based on transfer learning according to claim 1, comprising:

4. 2. The cross-domain cutter breakage monitoring method based on transfer learning according to claim 1, wherein the process of constructing the transfer learning model includes: constructing constraint terms based on locality-preserving projection; constructing maximum average discrepancy constraint terms based on a non-parametric distance estimation method; calculating corresponding constraint terms according to the class centers of samples of each class in the source domain and the target domain; calculating regularization constraint terms; assigning corresponding weights based on all the constraint terms; and using the corresponding minimum sum as the objective function to form the transfer learning model.

5. 2. The cross-domain cutter breakage monitoring method based on transfer learning according to claim 1, characterized in that the generalized feature representation of the training sample set output by the transfer learning model is input into the cross-domain cutter breakage monitoring model to train the cross-domain cutter breakage monitoring model, and the cross-domain cutter breakage monitoring model is a decision tree classifier.

6. a signal acquisition module configured to acquire vibration signals associated with the milling process; A feature extraction module configured to extract features from the feed direction vibration signal by improved multi-scale symbol dynamic entropy, and construct a training sample set for a transfer learning model including samples of all classes in the source domain and normal samples in the target domain, and a test set for a supervision model including samples of all classes in the target domain, a multi-scale symbolic dynamic entropy refinement process converting a time series of vibration signals in the feed direction into a symbolic time series, and dividing the symbolic time series into non-overlapping segments by scale; calculating different types of coarse-grained time series of data within each segment, wherein the different types of coarse-grained time series include a mean-based coarse-grained time series, a root-mean-square-based coarse-grained time series, and a maximum-based coarse-grained time series; Calculating the symbolic dynamic entropy of different types of coarse-grained time series at each scale; a feature extraction module including: fusing each symbol dynamic entropy to form a final multi-scale symbol dynamic entropy; a transfer learning module configured to obtain a mapping matrix using a transfer learning model, and further to compute a generic feature representation and apply source domain knowledge to a target domain, The total objective function of the transfer learning model is [Number 54] where L ＬＰＰ is a constraint term constructed based on locality preserving projection (LPP), and L ＭＭＤ is a constraint term constructed based on the non-parametric distance estimation method - maximum mean discrepancy (MMD), and L ＣＳＤ is the constraint term proposed by defining the class center, and L Ｒ is the regularization constraint term, and the trade-off parameters μ, ρ, and γ are all greater than 0. a transfer learning module that evaluates the performance of the transfer learning model when there are no corrupted samples in the target domain, and also uses Bayesian optimization to perform parameter optimization on the transfer learning model parameters, and in the optimization process, the Bayesian optimization objective function is the minimum ratio of the maximum average mismatch distance as the numerator and the data center-class center distance as the denominator; a damage monitoring module configured to train a cross-domain cutter breakage monitoring model based on the generic feature representation of the training sample set, calculate the generic feature representation of the test set using a transfer learning model, and input the calculated generic feature representation into the cross-domain cutter breakage monitoring model to realize cross-domain cutter breakage monitoring; A cross-domain cutter breakage monitoring system based on transfer learning, comprising:

7. A computer-readable storage medium having stored thereon a plurality of commands, said commands being suitable for being loaded by a processor of a terminal device to carry out the steps of the method according to any one of claims 1 to 5.

8. 6. A terminal device comprising a processor and a computer-readable storage medium, the processor being adapted to implement each command and the computer-readable storage medium being adapted to store a plurality of commands, the commands being suitable to be loaded by the processor to perform the steps of the method according to any one of claims 1 to 5.

Images