A cutting load prediction method based on deep transfer learning

Abstract

This invention relates to a cutting load prediction method based on deep transfer learning, focusing on composite material machining processes. The method includes: data sources of sensor data and cutting parameters, time-frequency domain feature extraction, a cutting load prediction model based on transfer learning, and finally, cutting load prediction. The method first collects data such as sound, vibration, and spindle current during composite material machining, as well as cutting parameters such as spindle speed, depth of cut, and feed rate. Then, time-frequency domain features are extracted from the sensor data to obtain feature data related to the cutting load. Next, transfer learning is used to train a cutting load prediction model for machining under different tools and cutting parameters, ultimately achieving cutting load prediction for machining processes with different tools and cutting parameters.

Description

Technical Field

[0001] This invention relates to the research field of intelligent manufacturing, and in particular to a cutting load prediction method based on deep transfer learning, specifically involving sensor data analysis and cutting load prediction during composite material processing. Background Technology

[0002] In the machining of composite material products, the problem of uneven cutting load caused by factors such as dynamic wear of tools, uneven distribution of allowance, and material anisotropy is mainly addressed by relying on the judgment of cutting load based on human experience and manually changing parameters such as spindle speed and feed rate to achieve the highest possible machining efficiency and quality. However, this method has weak data collection and quantitative analysis capabilities, insufficient precision in adapting quality and efficiency control, and lacks timeliness.

[0003] Cutting load prediction technology based on real-time sensing of machining condition data analyzes and predicts cutting loads from multi-source heterogeneous measured data during machining, and then dynamically adjusts spindle speed and feed rate. This provides a theoretical basis for adaptive control of cutting loads, effectively improving machining efficiency and ensuring machining quality. However, different tools have different effects on cutting loads under different cutting parameters and wear levels. Therefore, a universal model is needed to predict cutting loads under different tool conditions. Summary of the Invention

[0004] The technical problem solved by this invention is to overcome the shortcomings of the prior art and propose a cutting load prediction method based on transfer learning to achieve online prediction of cutting load, thereby providing a basis for adaptive load control and further improving machining efficiency and quality.

[0005] The solution of this invention is: a cutting load prediction method based on transfer learning, which includes the following steps:

[0006] S1. Construct a cutting load prediction model. The input of the cutting load prediction model is a feature vector composed of sensor feature data and cutting parameters during the composite material cutting process, and the output is the cutting load.

[0007] S2. Construct the source domain sensor feature matrix X_Tool from the M feature vectors corresponding to the reference tool and reference cutting parameters with known cutting load. s The target domain sensor feature matrix X_Tool is constructed by using the M feature vectors corresponding to the tool under test with known cutting load and the cutting parameters under test. t ;

[0008] S3. Utilizing the source domain sensor feature matrix X_Tool sEach feature vector is used as a training sample, and the cutting load corresponding to the training sample is used as the label data to pre-train the cutting load prediction model, which is suitable for the cutting parameter load model of reference tool and reference cutting parameters.

[0009] S4. Utilizing the target domain sensor feature matrix X_Tool t Each feature vector is used as a training sample, and a transfer learning training is performed on the cutting parameter load model applicable to the reference tool and reference cutting parameters to obtain a cutting load prediction model applicable to the tool under test and the cutting parameters under test.

[0010] S5. Input the sensor feature data and cutting parameters collected online during the machining process of the tool under test and the cutting parameters under test into the trained cutting load prediction model, and output the predicted cutting load online.

[0011] Preferably, the source domain sensor feature matrix X_Tool s and target domain sensor feature matrix X_Tool t The matrix form is as follows:

[0012]

[0013] Where, x ij ,i∈[1,M],j∈[1,N] are elements of the feature matrix X_Tool, i represents the row of the feature matrix; j represents the column of the feature matrix; M represents the number of eigenvectors of the feature matrix; N represents the number of columns of the feature matrix; N=h×r+1, h is the number of features measured by each sensor, and r represents the total number of sensors.

[0014] Preferably, the source domain sensor feature matrix X_Tool s and target domain sensor feature matrix X_Tool t Obtained through the following method:

[0015] S1.1. Perform time-frequency domain analysis on the raw data collected by the sensor during the cutting process to obtain sensor characteristic data;

[0016] S1.2 Calculate and sort the monotonicity of the sensor feature data;

[0017] S1.3 Select the h sensor feature data with the highest monotonicity ranking from the sensor feature data;

[0018] S1.4 Combine the sensor feature data with the cutting parameter feed rate f to form a feature matrix.

[0019] Preferably, the monotonicity Mon of the j-th feature in the feature matrix X_Tool j :

[0020]

[0021] Where x :j Let positive(diff(x) represent the j-th feature. ij ()) represents the number of times the next eigenvalue of two adjacent eigenvalues ​​is greater than the previous eigenvalue in the M eigenvalues ​​of the j-th feature, starting from the first one. Similarly, negative(diff(x)) represents the number of times the next eigenvalue of two adjacent eigenvalues ​​is greater than the previous eigenvalue. ij The number of times that the next feature value of the M features in the j-th feature is smaller than the previous feature value.

[0022] Preferably, the raw data collected by the sensor during the cutting process includes sound pressure data, vibration amplitude data in three directions on the spindle and spindle box, and three-phase current data of the spindle controller.

[0023] Preferably, the cutting load prediction model is based on a long short-term memory network and includes an input module, a feature extraction module, a prediction module, and an output module.

[0024] Preferably, the pre-training process is as follows: The labeled source domain sensor feature matrix X_Tool s Using the feature vectors as training samples and the first loss function L1 as the optimization objective, the cutting load prediction model is pre-trained to obtain a cutting parameter load model suitable for source domain data.

[0025] Preferably, the first loss function L1 is:

[0026]

[0027] Among them, CL i and , respectively, represent the true value and the predicted value of the cutting load, and M is the number of eigenvectors in the input feature matrix.

[0028] Preferably, the transfer learning training process is as follows:

[0029] S4.1 Freeze the prediction module in the pre-trained cutting load prediction model;

[0030] S4.2, The labeled source domain sensor feature matrix X_Tool s The partial feature vectors of the target domain sensor feature matrix and the partial feature vectors of the unlabeled target domain sensor feature matrix are used as training samples. The total loss function L1+L2 is used as the optimization objective. In the training step S4.1 frozen cutting load prediction model, the frozen cutting load prediction model is fine-tuned. L1 is the first loss function and L2 is the second loss function.

[0031] S4.3, Freeze the feature extraction module in the cutting load prediction model after S4.2 fine-tuning, and extract the labeled target domain feature matrix X_Tool t Using a subset of feature vectors as training samples and the first loss function L1 as the optimization objective, the frozen cutting load prediction model in this step is trained to obtain the final cutting load prediction model.

[0032] Preferably, the second loss function L2:

[0033]

[0034]

[0035] in, This is an empirical estimate of MMD, calculated using a Gaussian kernel function, where σ is the kernel bandwidth and M is the number of eigenvectors in the input feature matrix. and Let a and b be the a-th and b-th eigenvectors in the source domain eigenma matrix. and Let a and b be the a-th and b-th eigenvectors of the target domain feature matrix. The advantages of this invention compared to existing technologies are:

[0036] (1) This invention addresses the problem of unbalanced load in composite material processing, which requires manual qualitative judgment of the load. It analyzes and predicts the cutting load by collecting sensor data under different cutting parameters during processing online. It realizes online analysis and prediction of cutting load based on measured data. Compared with the current cutting load calculation method that is independent of cutting parameters, it can provide a reliable quantitative method for adaptive control of cutting load.

[0037] (2) This invention addresses the problem of uneven cutting load caused by factors such as uneven allowance distribution and material anisotropy in composite materials. It uses deep transfer learning to train a cutting load prediction model under different cutting parameters. It can make full use of measured data to fit the load prediction model and obtain a more accurate prediction model than the mechanism model. It provides an effective solution for dynamic cutting load prediction under different cutting parameters and different tool wear conditions of different cutting tools. Attached Figure Description

[0038] Figure 1 This invention provides a cutting load prediction method based on deep transfer learning.

[0039] Figure 2 This is the cutting load prediction process according to an embodiment of the present invention;

[0040] Figure 3 This is an LSTM-based cutting load prediction model for an embodiment of the present invention;

[0041] Figure 4 This invention provides an embodiment based on a multi-stage transfer learning training strategy.

[0042] Figure 5 This is a step-by-step transfer training of the cutting load prediction model in an embodiment of the present invention. Detailed Implementation

[0043] The present invention will be further described below with reference to the embodiments.

[0044] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention are within the scope of protection of the present invention.

[0045] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0046] This invention provides a cutting load prediction method based on transfer learning. First, time-frequency domain analysis is performed on sensor data such as current and vibration collected during composite material machining. Since tool wear, depth of cut, and feed rate are positively correlated with cutting load, the sensor feature data are sorted by monotonicity, and feature data with high monotonicity are selected to form a feature matrix along with cutting parameter data. Second, the input / output interface size of the model data is analyzed using the feature matrix, and a cutting load prediction model framework based on deep transfer learning is constructed. Finally, a model training process and mechanism are established, and a model training loss function is constructed to achieve composite material cutting load prediction based on deep transfer learning.

[0047] like Figure 1 As shown, the specific steps of the cutting load prediction method based on transfer learning provided by this invention are as follows:

[0048] (1) Construct a cutting load prediction model. The input of the cutting load prediction model is a feature vector composed of sensor feature data and cutting parameters during the composite material cutting process, and the output is the cutting load.

[0049] The raw data collected by the sensor during the cutting process includes sound pressure data (Pa), vibration amplitude data in three directions on the spindle and spindle box (V1x, V1y, V1z, V2x, V2y and V2z), and three-phase current data of the spindle controller (Cu, Cv and Cw).

[0050] The time-frequency domain characteristics of the aforementioned sensors include: mean, skewness, kurtosis, peak-to-peak value, root mean square, energy, crest factor, impulse factor, shape factor, marginal factor, standard deviation, mean spectral kurtosis, standard deviation spectral kurtosis, skewness spectral kurtosis, kurtosis kurtosis, and wavelet energy.

[0051] Under the same cutting parameters and material system, the cutting load gradually increases with the machining process, exhibiting time-series characteristics. LSTM has the ability to capture long-term dependencies in time series, making it well-suited for solving time-series prediction problems. Therefore, this invention constructs a cutting load prediction model based on Long Short-Term Memory (LSTM) networks. The cutting load prediction model, based on LSTM networks, includes an input module, a feature extraction module, a prediction module, and an output module, as follows: Figure 3 As shown.

[0052] The model comprises nine layers: an input module containing source and target domain feature data, a feature extraction module consisting of two LSTM layers, a prediction module consisting of two feature matrix transformation layers and four LSTM layers, and an output layer of Dense. The model's structure and hyperparameter settings are shown in Table 1. The first layer receives the input data, layers 2 and 3 extract features from the input data, layers 4 and 5 adjust the extracted features to the desired output data shape, layers 6 to 8 train the correlation between the feature data and the target value, and finally, the target data is output through the ninth layer. The LSTM in the model uses the ReLU activation function, the input data shape is (5, h×r+1), and the output data shape is (1, 1). Furthermore, Adam is chosen as the optimization function to minimize the loss function at each stage.

[0053] Table 1. Structure and parameters of the tool degradation state prediction model

[0054]

[0055]

[0056] S2. Construct the source domain sensor feature matrix X_Tool from the M feature vectors corresponding to the reference tool and reference cutting parameters with known cutting load. s The target domain sensor feature matrix X_Tool is constructed by using the M feature vectors corresponding to the tool under test with known cutting load and the cutting parameters under test. t ;

[0057] The source domain sensor feature matrix X_Tool s and target domain sensor feature matrix X_Tool t The matrix form is as follows:

[0058]

[0059] Where, x ij ,i∈[1,M],j∈[1,N] are elements of the feature matrix X_Tool, i represents the row of the feature matrix; j represents the column of the feature matrix; M represents the number of eigenvectors of the feature matrix; N represents the number of columns of the feature matrix; N=h×r+1, h is the number of features measured by each sensor, and r represents the total number of sensors.

[0060] like Figure 2 As shown, the source domain sensor feature matrix X_Tool s and target domain sensor feature matrix X_Tool t Obtained through the following method:

[0061] S1.1. Perform time-frequency domain analysis on the raw data collected by the sensor during the cutting process to obtain sensor characteristic data;

[0062] S1.2 Calculate and sort the monotonicity of the sensor feature data;

[0063] Monotonicity Mon of the j-th feature in the feature matrix X_Tool j :

[0064]

[0065] Where x :j Let positive(diff(x) represent the j-th feature. ij ()) represents the number of times the next eigenvalue of two adjacent eigenvalues ​​is greater than the previous eigenvalue in the M eigenvalues ​​of the j-th feature, starting from the first one. Similarly, negative(diff(x)) represents the number of times the next eigenvalue of two adjacent eigenvalues ​​is greater than the previous eigenvalue. ij The number of times that the next feature value of the M features in the j-th feature is smaller than the previous feature value.

[0066] S1.3 Select the h sensor feature data with the highest monotonicity ranking from the sensor feature data;

[0067] S1.4. Combine the sensor feature data with the cutting parameter feed speed f to form a feature matrix, which serves as the input data for the prediction model.

[0068] This invention employs a multi-stage model training mechanism, such as... Figure 4As shown, the cutting parameter load model training process is divided into a pre-training stage and a fine-tuning stage, which are steps S3 and S4 below.

[0069] The model is first pre-trained using a subset of data from the feature matrix X_Tool to enable initial prediction capabilities, with an L1 loss function. Then, using a loss function L2 based on Maximum Mean Discrepancy (MMD), transfer learning is performed on the sensor feature matrices for different tool machining operations to obtain the cutting load prediction model training diagram, as shown below. Figure 5 As shown.

[0070] S3. Utilizing the source domain sensor feature matrix X_Tool s Each feature vector is used as a training sample, and the cutting load corresponding to the training sample is used as the label data to pre-train the cutting load prediction model. This model is applicable to the cutting parameter load model with reference tool and reference cutting parameters. This step is the pre-training stage.

[0071] The pre-training process is as follows: The labeled source domain sensor feature matrix X_Tool s Using the feature vectors as training samples and the first loss function L1 as the optimization objective, the cutting load prediction model is pre-trained to obtain a cutting parameter load model suitable for source domain data.

[0072] The first loss function L1 is:

[0073]

[0074] Among them, CL i and , respectively, represent the true value and the predicted value of the cutting load, and M is the number of eigenvectors in the input feature matrix.

[0075] S4. Utilizing the target domain sensor feature matrix X_Tool t Using some feature vectors as training samples, a transfer learning training is performed on the cutting parameter load model applicable to the reference tool and reference cutting parameters to obtain a cutting load prediction model applicable to the tool under test and the cutting parameters under test; this step is the fine-tuning stage.

[0076] The transfer learning training process is as follows:

[0077] S4.1 Freeze the prediction module in the pre-trained cutting load prediction model;

[0078] S4.2, The labeled source domain sensor feature matrix X_Tool sThe partial feature vectors of the target domain sensor feature matrix and the partial feature vectors of the unlabeled target domain sensor feature matrix are used as training samples. The total loss function L1+L2 is used as the optimization objective. In the training step S4.1 frozen cutting load prediction model, the frozen cutting load prediction model is fine-tuned. L1 is the first loss function and L2 is the second loss function.

[0079] S4.3, Freeze the feature extraction module in the cutting load prediction model after S4.2 fine-tuning, and extract the labeled target domain feature matrix X_Tool t Using a subset of feature vectors as training samples and the first loss function L1 as the optimization objective, the frozen cutting load prediction model in this step is trained to obtain the final cutting load prediction model.

[0080] Loss L2 utilizes "MMD" and "L" z The calculated values ​​are "MMD" and "L". z "These represent the maximum mean difference and zero label of the feature extraction module outputs for the two-domain data under dual input conditions."

[0081] Construct the maximum mean difference (MMD) between the source domain feature matrix and the target domain feature matrix. Let X be the feature matrix of the source domain in the feature matrix of sensor data under different cutting parameters. s and the feature matrix X of the target domain t , respectively, follow p and q distributions, as shown in equation (3), which defines the MMD between the two distribution functions.

[0082]

[0083] Where f belongs to F, which is a set of smooth functions defined in the Reproducing Kernel Hilbert Space (RKHS), denoted as H. Here, := signifies the definition, and sup(·) is the maximum upper bound of the input. In transfer learning, MMD serves as a metric for measuring the difference between the source and target domains. Based on the above MMD, a transfer learning loss function is constructed:

[0084] Constructing the L2 transfer learning loss function:

[0085]

[0086] MMD is calculated as follows, where Lz is the zero-value label.

[0087] The second loss function L2:

[0088]

[0089]

[0090] in, It is an empirical estimate of MMD, calculated using a Gaussian kernel function, where σ is the kernel bandwidth and M is the number of eigenvectors in the input feature matrix. and Let a and b be the a-th and b-th eigenvectors in the source domain eigenma matrix. and Let a and b be the a-th and b-th eigenvectors of the target domain feature matrix.

[0091] S5. Input the sensor feature data and cutting parameters collected online during the machining process of the tool under test and the cutting parameters under test into the trained cutting load prediction model, and output the predicted cutting load online.

[0092] Based on the above LSTM-based cutting load model and multi-stage transfer training, the following cutting load prediction model is obtained:

[0093]

[0094] Among them, China F CL This represents the cutting load prediction model. X_Tool t-ls ,X_Tool t-ls+1 ,...,X_Tool t-1 This represents the input data from ls sampling points before time t, used to predict the cutting load at time t.

[0095] In summary, the cutting load prediction method based on deep transfer learning proposed in this invention first collects sensor data collected during the tool wear process under different tool specifications and cutting parameters to construct a feature matrix; then, it constructs an LSTM-based cutting load prediction model, with the feature matrix as input and the cutting load as output; next, it constructs a loss function for model training, including a pre-training loss function L1 and a fine-tuning loss function L2 based on MMD; finally, it uses partial feature matrix data under a certain parameter or tool to pre-train the cutting load prediction model, and then uses the remaining feature data of that parameter or tool to transfer train the model with feature data under other parameters or tools, obtaining a cutting load prediction model that can predict the cutting load under different tools and cutting parameters. Finally, it uses data collected online during the machining process to input into the model, and can output the predicted cutting load online, thus solving the problem of online cutting load prediction under different tools or cutting parameters.

[0096] Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make possible changes and modifications to the technical solutions of the present invention by utilizing the methods and techniques disclosed above without departing from the spirit and scope of the present invention. Therefore, any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solutions of the present invention shall fall within the protection scope of the technical solutions of the present invention.

Claims

1. A cutting load prediction method based on transfer learning, characterized in that... Includes the following steps: S1. Construct a cutting load prediction model. The input of the cutting load prediction model is a feature vector composed of sensor feature data and cutting parameters during the composite material cutting process, and the output is the cutting load. S2. Construct the source domain sensor feature matrix from the M feature vectors corresponding to the reference tool and reference cutting parameters with known cutting load. X_Tool s The target domain sensor feature matrix is ​​constructed by using M feature vectors corresponding to the tool under test with known cutting load and the cutting parameters under test. X_Tool t ; S3, Utilizing the feature matrix of source domain sensors X_ Tool s Each feature vector is used as a training sample, and the cutting load corresponding to the training sample is used as the label data to pre-train the cutting load prediction model, which is suitable for the cutting load model of reference tool and reference cutting parameters. S4. Utilizing the target domain sensor feature matrix X_Tool t Using some feature vectors as training samples, a transfer learning training is performed on the cutting parameter load model applicable to the reference tool and reference cutting parameters to obtain a cutting load prediction model applicable to the tool under test and the cutting parameters under test. S5. Input the sensor feature data and cutting parameters acquired online during the machining process of the tool under test and the cutting parameters under test into the trained cutting load prediction model, and output the predicted cutting load online; The source domain sensor feature matrix X_Tool s and target domain sensor feature matrix X_Tool t The matrix form is as follows: in, Characteristic matrix elements, Represents the rows of the characteristic matrix; Represents the columns of the characteristic matrix; M This indicates the number of eigenvectors in the characteristic matrix. N Indicates the number of columns in the characteristic matrix; N=h× r+1, h is the number of features measured by each sensor, and r represents the total number of sensors; Source domain sensor feature matrix X_Tool s and target domain sensor feature matrix X_Tool t Obtained through the following method: S1.

1. Perform time-frequency domain analysis on the raw data collected by the sensor during the cutting process to obtain sensor characteristic data; S1.2 Calculate and sort the monotonicity of the sensor feature data; S1.3 Select the h sensor feature data with the highest monotonicity ranking from the sensor feature data; S1.4, Combine sensor characteristic data with cutting parameters and feed rate. f Combined to form a feature matrix; The cutting load prediction model is based on a long short-term memory network and includes an input module, a feature extraction module, a prediction module, and an output module. The pre-training process is as follows: The labeled source domain sensor feature matrix... X_ Tool s Using the feature vectors as training samples and the first loss function L1 as the optimization objective, the cutting load prediction model is pre-trained to obtain a cutting load model suitable for source domain data.

2. The cutting load prediction method based on transfer learning according to claim 1, characterized by the following features matrix X_Tool Monotonicity of the j-th feature : in Represents the j-th feature, This represents the number of times, starting from the first of the M feature values ​​in the j-th feature, the subsequent feature value is greater than the preceding feature value. Similarly, The number of times the next feature value of the M features in the j-th feature is smaller than the previous feature value.

3. The cutting load prediction method based on transfer learning according to claim 1, characterized in that... The raw data collected by the sensor during the cutting process includes sound pressure data, vibration amplitude data in three directions on the spindle and spindle box, and three-phase current data of the spindle controller.

4. The cutting load prediction method based on transfer learning according to claim 1, characterized in that... The first loss function L1 is: in, and , respectively, represent the true value and the predicted value of the cutting load, and M is the number of eigenvectors in the input feature matrix.

5. The cutting load prediction method based on transfer learning according to claim 1, characterized in that... The transfer learning training process is as follows: S4.1 Freeze the prediction module in the pre-trained cutting load prediction model; S4.2, Label the source domain sensor feature matrix X_ Tool s The partial feature vectors of the target domain sensor feature matrix and the partial feature vectors of the unlabeled target domain feature matrix are used as training samples. The total loss function L1+L2 is used as the optimization objective. In the training step S4.1, the frozen cutting load prediction model is fine-tuned, with L1 being the first loss function. L 2. Second loss function; S4.3, Freeze the feature extraction module in the cutting load prediction model after S4.2 fine-tuning, and extract the labeled target domain feature matrix. X_Tool t Using a subset of feature vectors as training samples and the first loss function L1 as the optimization objective, the frozen cutting load prediction model in this step is trained to obtain the final cutting load prediction model.

6. The cutting load prediction method based on transfer learning according to claim 5, characterized in that... The second loss function L 2: in, yes MMD The empirical estimate is calculated using a Gaussian kernel function, where σ is the kernel bandwidth. M The number of eigenvectors in the input feature matrix. and Let a and b be the a-th and b-th eigenvectors in the source domain eigenma matrix. and Let a and b be the a-th and b-th eigenvectors of the target domain feature matrix.

Images