Multi-working-condition tool wear prediction method based on reciprocal Gaussian sampling element learning
By employing the reciprocal Gaussian sampling element learning method, the problem of accurately measuring tool wear under different working conditions is solved, enabling rapid adaptation and efficient prediction, thereby improving the accuracy and efficiency of tool wear monitoring.
Patent Information
- Application Number
- CN202511562972.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-10-30
- Publication Date
- 2026-01-23
AI Technical Summary
The degree of tool wear is difficult to measure accurately and uniformly in real time, especially since the wear rate and turning point vary significantly under different working conditions, making it difficult to achieve precision control.
A meta-learning method based on reciprocal Gaussian sampling is adopted. Through feature selection, meta-learner and reciprocal Gaussian sampling mechanism, monitoring signals are acquired and feature selection and parameter initialization are performed. Support set and query set are divided and learner parameters are updated to achieve accurate prediction of tool wear.
It achieves rapid adaptation and prediction capability under new working conditions, improves the practicality and application prospects of tool wear prediction, enhances the sampling probability and data quality of samples deviating from typical conditions, and improves learning efficiency.
Smart Images
Figure CN121374282A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application relates to the technical field of numerical control machining process monitoring, in particular to a multi-working-condition tool wear prediction method based on inverse Gaussian sampling meta-learning. BACKGROUND
[0002] The field of machining is constantly developing towards ultra-high precision machining, which requires the precision of all parts of the machining equipment to be controllable during the machining process. The tool, as a part directly involved in the work of the machining equipment, the degree of wear at the contact with the workpiece directly affects the precision of the workpiece. However, the degree of wear of the tool is always changing as the machining process proceeds. More unfortunately, the degree of wear of the tool is generally difficult to measure directly. In addition, the stage evolution characteristics of tool wear under different working conditions are different, and the wear rates and turning points of each stage are obviously different, resulting in difficulty in real-time and accurate measurement and unified monitoring of tool wear. SUMMARY
[0003] The purpose of the application is to provide a multi-working-condition tool wear prediction method based on inverse Gaussian sampling meta-learning, which realizes accurate measurement of tool wear through feature screening, meta-learner and inverse Gaussian sampling mechanism. In order to achieve the above purpose, the application provides the following scheme: In a first aspect, the application provides a multi-working-condition tool wear prediction method based on inverse Gaussian sampling meta-learning, comprising: S1: acquiring monitoring signals in a tool cutting process under different historical working conditions and actual wear values corresponding to the monitoring signals; S2: for any historical working condition, performing feature screening on the monitoring signals under the historical working condition to obtain screened features; S3: initializing parameters of a task learner, parameters of a meta-learner and parameters of an inverse Gaussian sampling mechanism; S4: based on the initialized inverse Gaussian sampling mechanism, dividing the screened features into a support set and a query set; S5: on the support set, based on wear prediction values output by the initialized task learner and actual wear values, obtaining a total support loss of the historical working condition; S6: updating the parameters of the task learner based on the total support loss of the historical working condition; S7: on the query set, based on wear prediction values output by the updated task learner and actual wear values, obtaining a total query loss of the historical working condition; S8: updating the parameters of the meta-learner and the parameters of the inverse Gaussian sampling mechanism according to the sum of the total query losses of all historical working conditions; S9: repeating S3 to S8 until a set number of times is reached, and obtaining the final parameters of the meta-learner. S10: predicting the tool wear value based on the monitoring signal collected in real time under the current working condition and the final parameters of the meta-learner.
[0004] According to the specific embodiments provided in the application, the application has the following technical effects: The application provides a tool wear prediction method under multiple working conditions based on inverse Gaussian sampling meta-learning, acquires monitoring signals under different historical working conditions to provide high-quality original data basis for meta-learning, improves the quality of data and the efficiency of subsequent learning by performing feature screening on the monitoring signals, intelligently screens by using an inverse Gaussian sampling mechanism, enhances the sampling probability of samples deviating from the typical state, and introduces the inverse Gaussian sampling mechanism into the optimization process of meta-learning to realize self-learning and self-adaptation of the sampling strategy; the application realizes rapid adaptation and few-sample prediction ability under new working conditions through the meta-learning framework, and has good practicability and popularization prospect for tool wear prediction under multiple variable working conditions. BRIEF DESCRIPTION OF DRAWINGS
[0005] In order to more clearly illustrate the technical solutions in the embodiments of the application or the prior art, the drawings needed in the embodiments will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the application, and for those skilled in the art, other drawings can be obtained from these drawings without creative labor.
[0006] Figure 1 A flowchart of a tool wear prediction method under multiple working conditions based on inverse Gaussian sampling meta-learning provided by an embodiment of the application is shown in the figure. Figure 2 As shown in the figure, the tool wear prediction method under multiple working conditions based on inverse Gaussian sampling meta-learning provided by an embodiment of the application includes the following steps. Figure 1 A detailed flowchart of S2 of the tool wear prediction method under multiple working conditions based on inverse Gaussian sampling meta-learning is shown in the figure. Figure 3 A feature screening index descending order histogram provided by an embodiment of the application is shown in the figure. Figure 4 A wear physical model fitting effect diagram provided by an embodiment of the application is shown in the figure. Figure 5 A comparison diagram of tool wear prediction results obtained by different methods provided by an embodiment of the application is shown in the figure. Figure 6 A structural diagram of a computer device provided by an embodiment of the application is shown in the figure. DETAILED DESCRIPTION
[0007] With reference to the drawings and embodiments of the present application, the technical solutions in the embodiments of the present application will be described clearly and completely. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of the present application.
[0008] In order to make the purpose, characteristics and advantages of the present application more obvious and easy to understand, the present application will be further described in detail below with reference to the drawings and specific embodiments.
[0009] In one exemplary embodiment, as shown in Figure 1 A method for multi-working-condition tool wear prediction based on inverse Gaussian sampling meta-learning is provided, which is executed by a computer device, specifically, can be executed by a terminal or a server, or can be executed by a terminal and a server together. The present application includes the following S1 to S10. Wherein: S1: Obtain monitoring signals in the tool cutting process under different historical working conditions, and actual wear values corresponding to the monitoring signals.
[0010] S2: For any historical working condition, the monitoring signals under the historical working condition are subjected to feature screening to obtain screened features.
[0011] S3: Initialize the parameters of the task learner, the parameters of the meta-learner and the parameters of the inverse Gaussian sampling mechanism. The task learner is constructed based on the encoder architecture of the Transformer. The encoder includes an embedding layer, a position encoding layer, a multi-head attention layer and a feedforward network connected in turn. The output of the position encoding layer is connected in residual with the output of the multi-head attention layer, and the output of the multi-head attention layer is connected in residual with the output of the feedforward network. The encoder further passes through a global attention layer and a linear layer to realize the output of the task learner.
[0012] Wherein, the embedding layer of the encoder is realized by a full connection layer, which can be represented as: . is the output of the embedding layer, is the weight matrix of the full connection layer, is the bias term, is the input matrix, is the activation function.
[0013] The position encoding layer realizes the data sequence by using different frequency sine and cosine functions, that is, .
[0014] .
[0015] Wherein, is a position index, is a dimension index, is a dimension of an input of an encoder, is a position index.
[0016] S4: Based on the initialized inverse Gaussian sampling mechanism, the filtered features are divided into a support set and a query set.
[0017] S5: On the support set, based on the wear prediction value output by the initialized task learner and the actual wear value, the total support loss of the historical working condition is obtained.
[0018] S6: Update the parameters of the task learner based on the total support loss of the historical working condition.
[0019] S7: On the query set, based on the wear prediction value output by the updated task learner and the actual wear value, the total query loss of the historical working condition is obtained.
[0020] S8: According to the sum of the total query losses of all historical working conditions, update the parameters of the meta-learner and the parameters of the inverse Gaussian sampling mechanism.
[0021] S9: Repeat S3 to S8 until a set number of times to obtain the final parameters of the meta-learner.
[0022] S10: Based on the real-time monitoring signals collected under the current working condition and the final parameters of the meta-learner, the tool wear value is predicted.
[0023] By implementing the above S1 to S10, the application realizes the rapid adaptation and few-sample prediction ability under new working conditions.
[0024] In another exemplary embodiment of the application, the monitoring signals include a plurality of monitoring sub-signals; as Figure 2 S2 specifically includes the following S201 to S209: S201, determine the actual wear value of each monitoring sub-signal, and determine a plurality of wear increments according to the actual wear value.
[0025] S202, extract the features of each monitoring sub-signal, and determine a plurality of feature increments according to the features extracted from each monitoring sub-signal, and determine the feature monotonicity (Monotonicity, Mon) based on the plurality of feature increments.
[0026] S203, a plurality of distribution models are used to fit a plurality of wear increments respectively to obtain the distribution parameters of each distribution model after fitting the wear increments. The wear increments are fitted into normal distribution, lognormal distribution, gamma distribution, three-parameter Weibull distribution and other distribution models by using nonlinear least squares method.
[0027] S204, fitting the distribution parameters of the wear increment based on each distribution model, and determining the optimal distribution model by using information criterion.
[0028] S205, fitting the multiple feature increments by using the optimal distribution model, to obtain the distribution parameters of the feature increments fitted by the optimal distribution model.
[0029] S206, determining the JS divergence (Jensen-Shannon, JS) based on the distribution parameters of the wear increment fitted by the optimal distribution model and the distribution parameters of the feature increments fitted by the optimal distribution model.
[0030] S207, determining the maximum information coefficient (Maximum Information Coefficient, MIC) based on the features extracted from each monitoring sub-signal and the actual wear value of each monitoring sub-signal.
[0031] S208, determining the feature selection index (Feature Selection Index, FSI) based on the JS divergence, the maximum information coefficient, and the feature monotonicity.
[0032] S209, screening the features extracted from each monitoring sub-signal based on the feature selection index, to obtain the screened features of each monitoring sub-signal.
[0033] The JS divergence is used to measure the distribution similarity between the actual wear value and the features, the maximum information coefficient is used to determine the correlation degree between the actual wear value and the features, the feature monotonicity is used to measure the direction of feature change, and the feature selection index is used to screen the time-frequency domain features similar to the actual wear value and its distribution in multiple dimensions.
[0034] In another exemplary embodiment of the present application, S202 specifically includes: For any monitoring sub-signal, the time domain features and the time-frequency domain features of the monitoring sub-signal are calculated.
[0035] The time domain features and the time-frequency domain features are normalized to obtain the time-frequency domain feature matrix of the monitoring sub-signal.
[0036] According to the time-frequency domain feature matrix of each monitoring sub-signal, multiple feature increments are determined.
[0037] The above S201 to S209 are the processing steps of one monitoring signal of any historical working condition. In addition, any historical working condition of the present application includes multiple monitoring signals, including main shaft vibration signals, main shaft current signals, cutting sound signals, cutting forces, etc.
[0038] Specifically, the specific process of constructing the time-frequency domain feature matrix comprises: installing an acceleration sensor on a spindle of a numerical control machine tool, collecting vibration signals of a indexable carbide milling insert in a cutting process by using the acceleration sensor; threading a current sensor through a U-phase cable of the numerical control machine tool, and collecting a current signal in the cutting process; installing a sound sensor beside a steel workpiece for collecting a sound signal in the cutting process; and installing a dynamometer between the workpiece and a workbench for collecting a cutting force in the cutting process. Each time of milling adopts a reverse milling mode to shorten the processing time, and the above signals are collected once and the wear of the insert is measured once each time of milling until the insert reaches a failure standard. The application repeatedly collects signals under six different historical working conditions.
[0039] Under historical working conditions 1, 2, 4 and 5, the mean value, standard deviation, square root amplitude, root mean square value and maximum value of the spindle vibration signal, the spindle current signal, the cutting sound signal, the cutting force and other monitoring signals are calculated.
[0040] Then each feature is normalized according to Min-Max, and the normalized features are spliced according to rows and columns to obtain a time-frequency domain feature matrix. The time-frequency domain feature matrix can be represented as follows: .
[0041] Wherein, M is a time-frequency domain feature matrix, is the pth time domain feature of the cth monitoring sub-signal at t time points, is the qth frequency domain feature of the cth monitoring sub-signal at t time points.
[0042] In another exemplary embodiment of the application, in S201, the monotonicity calculation formula of the feature is: .
[0043] Wherein, ε( ) is a unit step function, ΔF is a feature increment, is the number of features in a monitoring sub-signal.
[0044] In S204, the expression of the information criterion is: .
[0045] .
[0046] Wherein, represents the number of parameters of the distribution model, represents the number of wear increments, represents the maximum likelihood estimate value of the distribution model, is the Akaike information criterion, Bayesian information criterion.
[0047] In S206, the JS divergence is calculated. The calculation formula is: .
[0048] wherein, is the distribution parameter of the optimal distribution model fitted after the wear increment, is the distribution parameter of the optimal distribution model fitted after the feature increment, is the Kullback-Leibler divergence between distributions, M is an intermediate distribution model, and the calculation formula is: .
[0049] .
[0050] wherein, is the optimal distribution model, is the probability of the occurrence of the zth event in the optimal distribution model, is the probability of the occurrence of the zth event in the intermediate distribution model.
[0051] In S207, the calculation formula of the maximum mutual information coefficient is: .
[0052] wherein, Y is the actual wear value, F is a feature in the time-frequency domain feature matrix, the value space of Y is divided into intervals, the value space of F is divided into intervals, ≈0.6, n is the sample number, is a function of the sample number, I(Y;F) is the maximum mutual information coefficient between the actual wear value and the feature, . is the joint distribution of the actual wear value and the feature, is the distribution of the actual wear value, is the distribution of the feature.
[0053] In S208, the following formula is used to determine the feature screening index: .
[0054] wherein, is the feature screening index, is the JS divergence, is the maximum mutual information coefficient, is the feature monotonicity, a weight for JS divergence, a weight for maximum mutual information coefficient, a weight for feature monotonicity, u v w =1.
[0055] In another exemplary embodiment of the present application, S4 specifically comprises the following S401 to S404: S401, taking a screened feature of a monitoring sub-signal as a sample, and constructing a sample set based on screened features of each monitoring sub-signal under the historical working condition.
[0056] S402, calculating a sampling weight of each sample by using an initialized inverse Gaussian sampling mechanism according to an actual wear value corresponding to each sample.
[0057] S403, normalizing the sampling weight of each sample to obtain a probability distribution of each sample.
[0058] S404, randomly sampling to generate a support set based on the probability distribution of each sample, and taking samples in the sample set that are not selected into the support set as a query set. The expression of the inverse Gaussian sampling mechanism is: .
[0059] wherein, is an actual wear value corresponding to the sample, is a sampling weight of the sample, is a first sampling parameter, is a second sampling parameter.
[0060] In another exemplary embodiment of the present application, S5 specifically comprises the following S501 to S503: S501, obtaining a data loss of the historical working condition based on a wear prediction value output by the initialized task learner and an actual wear value on the support set.
[0061] S502, obtaining a physical loss of the historical working condition based on a wear prediction value output by the initialized task learner and a physical wear value obtained by a wear physical model; wherein the wear physical model is constructed based on cutting parameters of the historical working condition.
[0062] Specifically, according to a three-stage wear law of the tool, the cutting parameters of each working condition are considered to construct a wear physical model of the three-stage tool. The three stages are an initial wear stage, a normal wear stage and an acute wear stage of the tool. The wear physical model is: .
[0063] wherein, is a wear initial value, is a back engagement, is a radial depth of cut, is a cutting speed, is a feed per tooth, is a first correction coefficient related to the cutting material and the cutting environment, is a second correction coefficient related to the cutting material and the cutting environment, is a first depth of cut index, is a second depth of cut index, is a speed index, is a feed index, is a first cutting duration index, is a second cutting duration index. The depth of cut index, the speed index, and the feed index are pre-set parameters. The wear initial value is random or specified, and the present application specifies it according to prior knowledge.
[0064] S503, based on the data loss of the historical working condition and the physical loss of the historical working condition, obtaining a total support loss of the historical working condition.
[0065] In another exemplary embodiment of the present application, the total support loss of the i-th historical working condition is calculated using the following formula: .
[0066] .
[0067] .
[0068] wherein, is the i-th historical working condition, is the data loss of the i-th historical working condition, is a task learner, is a parameter of the task learner, is a parameter of the inverse Gaussian sampling mechanism, is a probability distribution of the i-th sample on the support set, is a wear prediction value of the task learner on the i-th sample on the support set, is an actual wear value of the i-th sample, is the physical loss of the i-th historical working condition, is a physical wear value of the i-th sample, is the i-th historical working condition, is the i-th sample, is the i-th sample, is the i-th historical working condition, is the i-th sample, is the i-th sample, is the i-th sample, is the i-th sample, Total support loss for each historical operating condition.
[0069] In another exemplary embodiment of this application, the parameter update process in the task learner in S6 is as follows: .
[0070] in, The learning rate of the task learner. The updated parameters for the task learner. These are the parameters of the task learner. The gradient of the parameters of the task learner.
[0071] In another exemplary embodiment of this application, S7 specifically includes: Based on the updated wear prediction and actual wear values output by the task learner, the total query loss for the historical operating conditions is obtained on the query set.
[0072] Based on the updated wear prediction and actual wear values output by the updated task learner, the updated data loss for the historical operating conditions is obtained on the query set.
[0073] Based on the updated wear prediction value output by the task learner and the physical wear value obtained from the wear physics model, the updated physical loss of the historical working conditions is obtained.
[0074] Based on the data loss and physical loss after the historical operating conditions are updated, the total query loss of the historical operating conditions is obtained.
[0075] The following formula is used to calculate the first... Total query loss for each historical scenario: .
[0076] .
[0077] .
[0078] in, For the first A historical working condition, For the first Data loss after updating historical operating conditions For the updated task learner, The updated parameters for the task learner. The parameters are for the reciprocal Gaussian sampling mechanism. To the support set The probability distribution of each sample For the updated task learner on the support set a wear prediction value of the sample, a wear value of the first historical working condition, an actual wear value of the sample, a wear value of the first historical working condition, a total query loss of the first historical working condition. a physical wear value of the sample, a wear value of the first historical working condition, a total query loss of the first historical working condition.
[0079] In another exemplary embodiment of the present application, the updating process of the parameters of the meta-learner is as follows: .
[0080] The updating process of the parameters of the inverse Gaussian sampling mechanism is as follows: .
[0081] wherein β is the learning rate of the meta-learner, and p(T) represents the distribution of all working conditions.
[0082] In another exemplary embodiment of the present application, S1 to S9 are offline training steps of the meta-learner, and the final parameters of the meta-learner are obtained by training the meta-learner. In the training of the meta-learner, the normalized inverse Gaussian sampling mechanism is used as the weight to be respectively incorporated into the physical loss and the data loss of the task learner, so as to improve the utilization rate of the boundary neighborhood data and improve the training efficiency. In one meta-iteration in the training, the data loss and the physical loss of any historical working condition are calculated and used to update the task learner by gradient descent, and then the sum of the data loss and the physical loss of all historical working conditions is used to update the meta-learner. Moreover, the parameters of the inverse Gaussian sampling mechanism are also updated by using the sum of the data loss and the physical loss of all historical working conditions, so as to dynamically adjust the attention degree to the data boundary area. In S10, a small amount of data of the tool under the current working condition is collected as a sample in the cutting process, and the learner is adapted to the current working condition by several steps of gradient update, so as to realize the real-time and accurate monitoring of the tool wear under the working condition.
[0083] In the present application, the feature screening index contains information such as the direction of feature change, the correlation strength with wear, and the similarity degree with wear distribution. The features are screened by the feature screening index to maximize the effectiveness of the training data and improve the training efficiency. The introduction of the wear physical model constrains the solution space of the task learner. The meta-learning method is used to train the meta-learner to learn the common features of different historical working conditions. In fact, the goal of meta-learning is to initialize a set of parameters, so that the distance from the meta-learner to all historical working conditions is minimized, and the meta-learner can quickly fine-tune to adapt to the current working condition.
[0084] The present application can realize the extraction of wear-sensitive features from the monitoring signals of the cutting process through multi-dimensional feature screening. Information criterion is used to determine the distribution characteristics of tool wear, which can ensure that the distribution model of tool wear is the best choice. The designed feature screening index is used to identify multi-dimensional sensitive features such as wear and its distribution, which can improve the reliability of data and the training efficiency of meta-learner.
[0085] The present application independently designs a task learner based on the decoder architecture of the Transformer, which can well map the wear-sensitive features of each historical working condition. The position information is fully considered and the meta-learner representation is enhanced by position encoding of the screened feature sequence. The multi-head attention mechanism can make the task learner pay attention to the information in different position and different representation subspace at the same time.
[0086] The present application constructs a reciprocal Gaussian sampling meta-learner that fuses three-stage wear physical information, which can well learn the common features of different historical working conditions. A three-stage wear physical model considering cutting parameters is introduced into the loss function of the learner, which reduces the dependence of the meta-learner on monitoring data and the sensitivity to noise. The reciprocal Gaussian sampling mechanism introduces the loss function, which enhances the attention of the meta-learner to the boundary neighborhood data. The parameters of the reciprocal Gaussian sampling mechanism are updated by the total query loss to adaptively adjust the attention to the boundary neighborhood data. The reciprocal Gaussian sampling meta-learner is used to train the meta-learner, which significantly improves the training efficiency of the meta-learner and quickly adapts to multiple new working conditions in testing.
[0087] In another exemplary embodiment of the present application, as shown in Figure 3 The features are arranged in descending order of FSI, and the top 8 FSI corresponding features are selected as the data under the corresponding working condition.
[0088] As shown in Figure 4 The fitting effect diagram of the wear physical model under six historical working conditions is shown. The wear physical model has very high consistency with the actual wear value, which illustrates the applicability and superiority of the wear physical model.
[0089] The final parameters of the meta-learner are saved for fine-tuning testing under historical working condition 3 and historical working condition 6 using a small amount of samples.
[0090] As shown in Figure 5The figure shows the comparison of tool wear prediction results obtained by different methods. Comparison 1 uses a classic LSTM network, and the filtered features are used as input. Comparison 2 is a Transformer, and the input is the unfiltered features. Comparison 3 is an advanced tool wear prediction method, and the input is the filtered features. To verify the performance of the present application, the hyperparameters of all comparison methods are adjusted to the best settings of the corresponding meta-learner as much as possible, except for the number of iterations and the present application.
[0091] Intuitively, compared with other comparison methods, the predicted tool wear of the present application is almost consistent with the actual wear value. On historical working condition 3 and historical working condition 6, the prediction absolute error greater than 0.002mm is only 9.13% and 5.59% respectively, and the maximum absolute error is only 0.028mm and 0.040mm respectively. This shows that the present application has very excellent performance in the task of tool wear prediction under new working conditions.
[0092] The prediction results of comparison 1 are only slightly worse than those of comparison 2, which proves that the meta-learner trained by the distribution features selected by the multi-dimensional feature selection method according to the data and its distribution characteristics has a certain effectiveness.
[0093] Given the leading advantage of comparison 2 in the multi-modal task, its results in the tool wear prediction task are also good, which also proves the correctness of using only the decoder architecture of the Transformer as the task learner of the present application.
[0094] For comparison 3, the effect is slightly worse than that of other comparison methods due to the low number of training iterations, which also proves that the present application has high training efficiency in the task of tool wear prediction under the same low number of iterations.
[0095] In an exemplary embodiment, a computer device, which can be a server or a terminal, is provided, and its internal structure diagram can be as shown in Figure 6As shown in the figure. The computer device includes a processor, a memory, an input / output interface (Input / Output, referred to as I / O) and a communication interface. Among them, the processor, the memory and the input / output interface are connected through the system bus, and the communication interface is connected to the system bus through the input / output interface. Among them, the processor of the computer device is used to provide computing and control capability. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store monitoring signals in the tool cutting process under different working conditions. The input / output interface of the computer device is used to exchange information between the processor and external devices. The communication interface of the computer device is used to communicate with the terminal outside through network connection. The computer program is executed by the processor to realize a tool wear prediction method based on inverse Gaussian sampling meta-learning under multiple working conditions.
[0096] Those skilled in the art can understand that, Figure 6 The skilled in the art can understand that,
[0097] The technical features of the above embodiments can be combined arbitrarily, in order to make the description simple, not all possible combinations of the technical features in the above embodiments are described, however, as long as the combination of the technical features does not exist contradictory, it should be considered as the scope of the present application.
[0098] The principles and implementation modes of the present application are described by applying specific examples in this paper, the above embodiment is only used to help understand the method and its core idea of the present application; at the same time, for those skilled in the art, according to the idea of the present application, the specific implementation mode and application range will be changed. In summary, the content of the present application should not be understood as a limitation.
Claims
1. A multi-working-condition tool wear prediction method based on inverse Gaussian sampling meta-learning, characterized in that, The method comprises: S1: acquiring monitoring signals in a cutting process of a tool under different historical working conditions and actual wear values corresponding to the monitoring signals; S2: for any historical working condition, performing feature screening on the monitoring signals under the historical working condition to obtain screened features; S3: initializing parameters of a task learner, parameters of a meta learner, and parameters of a reverse Gaussian sampling mechanism; S4: based on the initialized reverse Gaussian sampling mechanism, dividing the screened features into a support set and a query set; S5: based on wear prediction values output by the initialized task learner and actual wear values, obtaining a total support loss of the historical working condition on the support set; S6: updating the parameters of the task learner based on the total support loss of the historical working condition; S7: based on wear prediction values output by the updated task learner and actual wear values, obtaining a total query loss of the historical working condition on the query set; S8: updating the parameters of the meta learner and the parameters of the reverse Gaussian sampling mechanism according to a sum of the total query losses of all historical working conditions; S9: repeatedly performing S3 to S8 until a set number of times is reached, to obtain final parameters of the meta learner; S10: predicting a wear value of the tool based on real-time acquired monitoring signals under a current working condition and the final parameters of the meta learner.
2. The inverse Gaussian sampling based meta-learning multi-condition tool wear prediction method according to claim 1, characterized in that, The monitoring signals comprise a plurality of monitoring sub-signals; The feature screening on the monitoring signals under the historical working condition comprises: determining actual wear values of each monitoring sub-signal, determining a plurality of wear increments according to the actual wear values; extracting features of each monitoring sub-signal, determining a plurality of feature increments according to the extracted features of each monitoring sub-signal, and determining feature monotonicity based on the plurality of feature increments; fitting the plurality of wear increments by using a plurality of distribution models to obtain distribution parameters of each distribution model after fitting the wear increments; determining an optimal distribution model by using an information criterion based on the distribution parameters of each distribution model after fitting the wear increments; fitting the plurality of feature increments by using the optimal distribution model to obtain distribution parameters of the optimal distribution model after fitting the feature increments; determining a JS divergence based on the distribution parameters of the optimal distribution model after fitting the wear increments and the distribution parameters of the optimal distribution model after fitting the feature increments; determining a maximum mutual information coefficient based on the extracted features of each monitoring sub-signal and the actual wear values of each monitoring sub-signal; determining a feature screening index based on the JS divergence, the maximum mutual information coefficient, and the feature monotonicity; screening the features extracted from each monitoring sub-signal based on the feature screening index to obtain screened features of each monitoring sub-signal.
3. The inverse Gaussian sampling based meta-learning multi-condition tool wear prediction method according to claim 2, characterized in that, An expression of the information criterion is: ； ； wherein, denotes the number of parameters of the distribution model, denotes the number of wear increments, denotes the maximum likelihood estimate of the distribution model, is the Akaike information criterion, is the Bayesian information criterion.
4. The inverse Gaussian sampling based meta-learning multi-condition tool wear prediction method according to claim 2, characterized in that, The feature screening index is determined by using the following formula: ； wherein, is a feature screening index, is a JS divergence, is a maximum mutual information coefficient, is a feature monotonicity, is a weight of the JS divergence, is a weight of the maximum mutual information coefficient, is a weight of the feature monotonicity.
5. The inverse Gaussian sampling based meta-learning multi-condition tool wear prediction method according to claim 2, characterized in that, The screened features are divided into the support set and the query set based on the initialized reverse Gaussian sampling mechanism, and the division specifically comprises: taking the screened features of one monitoring sub-signal as one sample, and constructing a sample set based on the screened features of each monitoring sub-signal under the historical working condition; calculating sampling weights of the samples by using the initialized reverse Gaussian sampling mechanism according to the actual wear values corresponding to the samples; and The sampling weights of each sample are normalized to obtain a probability distribution of each sample; A support set is randomly sampled based on the probability distribution of each sample, and samples not selected into the support set from the sample set are taken as a query set.
6. The inverse Gaussian sampling based meta-learning multi-condition tool wear prediction method according to claim 5, characterized in that, The expression of the reciprocal Gaussian sampling mechanism is: ； wherein, is an actual wear value corresponding to the sample, is a sampling weight of the sample, is a first sampling parameter, is a second sampling parameter.
7. The inverse Gaussian sampling based meta-learning multi-condition tool wear prediction method according to claim 5, characterized in that, On the support set, a total support loss of the historical working condition is obtained based on the wear prediction value output by the initialized task learner and the actual wear value, specifically including: On the support set, a data loss of the historical working condition is obtained based on the wear prediction value output by the initialized task learner and the actual wear value. A physical loss of the historical working condition is obtained based on the wear prediction value output by the initialized task learner and a physical wear value obtained from a wear physical model, wherein the wear physical model is constructed based on the cutting parameters of the historical working condition. The total support loss of the historical working condition is obtained based on the data loss of the historical working condition and the physical loss of the historical working condition.
8. The inverse Gaussian sampling based meta-learning multi-condition tool wear prediction method according to claim 7, characterized in that, The total support loss for the nth historical operating condition is calculated using the following equation: Total support loss = (1 - support level) * (1 - support level) * (1 - support ； ； ； wherein, is the data loss of the th historical working condition, is the data loss of the th historical working condition, is the task learner, is the parameter of the task learner, is the parameter of the inverse Gaussian sampling mechanism, is the probability distribution of the th sample on the support set, is the wear-out prediction value of the th sample on the support set by the task learner, is the actual wear-out value of the th sample, is the physical data loss of the th historical working condition, is the physical wear-out value of the th sample, is the total support loss of the th historical working condition.
9. The inverse Gaussian sampling based meta-learning multi-condition tool wear prediction method according to claim 7, characterized in that, The total query loss for the nth historical working condition is calculated using the following formula: ； ； ； wherein, is the data loss of the th historical working condition, is the data loss of the th historical working condition after updating, is the updated task learner, is the updated parameter of the task learner, is the parameter of the inverse Gaussian sampling mechanism, is the probability distribution of the th sample on the support set, is the wear prediction value of the th sample on the support set by the updated task learner, is the actual wear value of the th sample, is the physical data loss of the th historical working condition after updating, is the physical wear value of the th sample, is the total query loss of the th historical working condition.