Tool wear prediction method for data enhancement based on RealNVP model

Abstract

The invention provides a tool wear prediction method for data enhancement based on a RealNVP model, and aims to solve the problem that a traditional model is insufficient in accuracy under rare wear conditions. According to the method, the sparse data in the training set is enhanced, so that the prediction accuracy and robustness of the model under the condition of small data volume and abrasion are remarkably improved. The method specifically comprises the steps of data collection, selection and preprocessing; configuring a RealNVP model comprising a deep neural network coupling layer, wherein the RealNVP model is used for executing complex data transformation; enhancing data in a potential space through the model, and generating a new sample with an actual operation value to complete the expansion of a data set; and training a deep learning model using the extended data set. Performance evaluation is carried out on a model of a data set before and after being enhanced by using RealNVP, and it is proved that the model provided by the invention has remarkable advantages in wear prediction accuracy and robustness.

Description

Technical Field

[0001] The invention belongs to the technical field of tool wear prediction, and in particular relates to a tool wear prediction method based on a RealNVP model for data enhancement. Background Art

[0002] The combination of information and communication technology and advanced manufacturing has ushered in a new era of intelligent manufacturing, which is a key trend in the future of the industry. Predictive health management (PHM) is crucial in this transformation and is widely used in manufacturing systems to maintain the integrity, precision and efficiency of cutting tools, especially those that are susceptible to thermomechanical coupling in metal cutting. Past studies have shown that abnormal conditions of tools cause about 10-40% of machine downtime, seriously affecting product quality and easily causing damage to machine tools. Therefore, implementing real-time tool condition monitoring (TCM) is the key to maintaining stable production, minimizing quality issues and preventing equipment damage due to tool wear or failure, while also avoiding profit losses due to premature tool replacement.

[0003] Most current tool wear monitoring methods use data-driven methods, which are divided into direct and indirect methods. The specific classification depends on the type of signal data collected. The direct method uses tool images to collect status information, and the results are accurate, but this method complicates image acquisition due to the complexity of tool movement, chip generation, and coolant, increasing the cost of imaging equipment. The indirect method infers tool wear through signals collected by sensors during the machining process. Although this method is not as accurate as the direct method, it can predict tool wear in real time during the production process, ensuring safer and more reliable production activities.

[0004] During the training process of existing tool wear prediction models based on sensor data, there is often a situation where the training set has too much data in the stable wear stage, but too little data in the initial wear stage and the rapid wear stage. This results in the trained model having low accuracy in predicting the wear values ​​in the initial wear stage and the rapid wear stage of the tool, affecting the overall prediction effect of the tool prediction model. Summary of the invention

[0005] The present invention aims to solve the problem that the existing tool wear prediction model has insufficient accuracy under rare wear conditions, and provides a tool wear prediction method based on the RealNVP model for data enhancement.

[0006] The technical solution of the present invention is a tool wear prediction method based on RealNVP model for data enhancement, the method comprising:

[0007] Step 1: Collect sample data and preprocess,

[0008] During the cutting process, the tool cutting force, tool vibration signal, and sound signals in the X, Y, and Z directions are collected; after each processing is completed, the tool wear status is observed through an electron microscope, the wear of the three-edge milling cutter in X, Y, and Z is quantified, and the average wear value is calculated to evaluate the overall wear value; and the data collected from multiple sensors is combined with the tool wear results observed by the electron microscope to construct a comprehensive data set containing the corresponding relationship between sensor signals and tool wear;

[0009] De-noising and normalizing the comprehensive dataset;

[0010] Step 2: Build a sample enhancement model,

[0011] A RealNVP model with ten coupled layers was constructed. Each coupled layer consisted of a deep neural network with two hidden layers and 512 hidden units in each hidden layer.

[0012] Each coupling layer uses a non-volume preserving transformation, the transformation method is:

[0013] y 1:d =x 1:d

[0014] y a+1:D =x a+1:D ☉exp(s(x 1：d ))+t(x 1:d )

[0015] Among them, x 1：d and x d+1：D are the first and second halves of the input vector x, y 1：d ,y d+1：D Represents the first and second halves of the transformed output vector, s and t are the first half x 1：d The function generates scale and translation parameters, D represents the total number of dimensions of the data, d represents the dimension of the current operation, and ⊙ represents element-level multiplication, that is, X t，d+1：D Each element in is related to exp(s(X t，d )) is multiplied; this structure ensures that part of the input x 1：d Directly mapped to output y 1：d , and the other part x d+1：D After transformation; this design keeps the transformation reversible while ensuring that part of the input information is directly passed to the output, increasing the flexibility and expressiveness of the model;

[0016] Among them, y 1：d ,y d+1：D Represents the first and second halves of the transformed output vector, x 1：d and xd+1：D They are the first and second halves of the input vector x, respectively. s() and t() are the first half of x. 1：d Functions that generate scale and translation transforms;

[0017] Step 3: Sample data enhancement and training set construction;

[0018] The model constructed in step 2 generates new samples in the latent space based on the existing samples, and then uses the inverse transformation f -1 () maps the new sample from the latent space back to the original feature space. The inverse transformation formula is:

[0019]

[0020] Among them, z is a new sample in the latent space sampled from the latent Gaussian distribution N(0,I), and f -1 is the inverse mapping function of RealNVP;

[0021] Step 4: Use all the samples obtained to train a deep convolutional neural network-multi-channel bidirectional long short-term memory model, and use the trained deep convolutional neural network-multi-channel bidirectional long short-term memory model to predict tool wear.

[0022] Furthermore, in step 2, in each coupling layer of the model, the s() and t() functions are dynamically updated according to the forward and backward propagation of the data; specifically, when the data flows through the coupling layer, the output of the s() and t() functions is calculated in real time according to the input data of the layer to ensure that the functions can adapt to the characteristics of the current data. During the training process, the gradient of the loss function relative to the parameters of the s() and t() functions is calculated, and these gradients are used to update the parameters to optimize the model performance; this update mechanism enables the s() and t() functions to effectively adapt to changes in data characteristics and enhance the data processing capabilities of the model;

[0023] The update method is:

[0024] s′(x)=s(x)·(1+α·σ(W·x+b))

[0025] t′(x)=t(x)+β·σ(W·x+b)

[0026] Among them, σ is the activation function, α and β are the adjustment rates, W and b are the weights and biases, which are learned during the model training process. s(x) and t(x) represent the scale transformation and translation transformation functions before updating, and s′(x) and t′(x) represent the scale transformation and translation transformation functions after updating.

[0027] Furthermore, in step 3, the newly generated samples are screened to select data with tool wear values ​​in the range of 150 to 300 for subsequent training.

[0028] Furthermore, in step 3, when the new sample data is inverted by the model function f -1 After conversion and input into the model, the s() and t() functions are updated according to the characteristics of the generated data and their deviation from the characteristics of the original data; the update method is:

[0029] s′ new (x) = s old (x)+λ s ·(μ real -μ gen )

[0030] t′ new (x) = t old (x)+λ t ·(σ real -σ gen )

[0031] Among them, μ real and σ real are the mean and standard deviation of the real data, μ gen and σ gen is the mean and standard deviation of the generated data; s and λ t is the learning rate parameter, which is used to adjust the amplitude of the scale and translation function update, s old (x), t old (x) scale transformation and translation transformation function before updating, s′ new (x), t′ new (x) represents the updated scale transformation and translation transformation function;

[0032] Such updates rely specifically on assessing the model's adaptability to new situations, such as how the model performs under new data distributions and whether it meets predetermined performance criteria. This update mechanism allows the model to not only process data seen during training, but also adapt to completely new types of data, thereby improving its versatility and flexibility.

[0033] Step 2 relies on regular model training and data batches, while the update of step 3 relies on new data generation and the satisfaction of specific conditions, such as the performance of the model when processing unconventional data or newly generated data. That is, the update of step 2 is more regular and periodic, and the update of step 3 may be more conditional, relying on the adaptation and performance of the model in the new situation.

[0034] Furthermore, the loss function of the deep convolutional neural network-multi-channel bidirectional long short-term memory model in step 4 adopts the mean square error, and the formula is: where y i is the true label, is the predicted value, n is the total number of samples in the data set, and the mean absolute percentage error MAPE is used as the evaluation indicator. The formula is:

[0035] The method of the present invention performs data enhancement on the stage with less data in the training set, and by expanding the sample size of this stage, the accuracy and reliability of the model in predicting tool wear are improved. This method not only enhances the prediction ability of the model within the normal working range, but also greatly improves the robustness of the model when facing rare wear patterns. The method of the present invention performs data enhancement on the rapid wear stage in the training set based on the RealNVP model, so that the sample size of this stage is expanded to improve the accuracy and reliability of the model in predicting tool wear. In this way, rapid wear data with high variability can be generated to more accurately reflect the wear that the tool may suffer in this stage. This method not only enhances the prediction ability of the model within the normal working range, but also greatly improves the robustness of the model when facing rare wear patterns. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 A diagram showing the execution steps of a tool wear prediction method for data enhancement based on the RealNVP model provided by the present invention;

[0037] Figure 2 This is a basic principle diagram of the RealNVP network in the present invention. DETAILED DESCRIPTION

[0038] The present invention is further described below in conjunction with the accompanying drawings.

[0039] Reference Figure 1 , Figure 2 ,The present invention proposes a method for training a tool wear prediction model based on the RealNVP data enhancement model, such as Figure 1 As shown, the method for data enhancement of the tool wear prediction model mainly includes the following steps:

[0040] 1. Data selection and preprocessing:

[0041] In this tool experiment, the following machining parameters were used: spindle speed of 10400RPM, feed rate of 1555mm / min, cutting depth of 0.125mm, radial cutting depth of 0.2mm. Kistler brand sensors were used to monitor key data on the machine tool work area and tool, as follows: Kistler 8152 sensor was used to monitor cutting force, Kistler 8636C sensor was used to monitor vibration signals, and Kistler 9265B sensor was used to monitor acoustic emission signals, covering the three directions of X, Y, and Z.

[0042] After each machining cycle, the wear state of the tool is accurately observed and recorded in detail through an IECIA MZ12 electron microscope. The wear of the X, Y, and Z three-edge milling cutters is quantified, and the average wear value is calculated to evaluate the overall wear degree. The data collected from multiple sensors is combined with the tool wear results observed by the electron microscope to construct a comprehensive data set containing the corresponding relationship between sensor signals and tool wear for subsequent analysis and wear prediction.

[0043] All data undergo preprocessing steps such as denoising and standardization before constructing the dataset to ensure the accuracy of analysis and prediction.

[0044] First, normalization and other preprocessing are performed on all samples to ensure data quality and meet the requirements of subsequent model training. After completing these preprocessing, samples with wear values ​​exceeding 150 are screened out and defined as key data sets. These preprocessing steps include removing outliers, supplementing missing data, and data standardization. The methods for data standardization are:

[0045]

[0046] Where X is the original data, X min and X max are the minimum and maximum values ​​of the data respectively.

[0047] Configure the RealNVP model:

[0048] The present invention develops a RealNVP model consisting of ten coupled layers, each of which is constructed by a deep neural network, each of which contains two hidden layers, and each hidden layer has 512 hidden units. Figure 2 The model architecture shown, these coupling layers are specially designed to perform non-volume preserving transformations, which optimize the data characteristics of high wear samples by learning complex nonlinear scaling s and translation transformations t. The non-volume preserving transformation of RealNVP is implemented in the coupling layer, and the specific form is as follows:

[0049] y 1：d =x1:d

[0050] y d+1:D =x d+1:D ☉exp(s(x 1:d )+t(x 1:d )

[0051] Among them, x 1：d and x d+1：D are the first and second halves of the input vector x, respectively, and s and t are the first half of x. 1：d This structure ensures that part of the input x 1：d Directly mapped to output y 1：d , and the other part x d+1：D This design keeps the transformation reversible while ensuring that part of the input information is directly transferred to the output, increasing the flexibility and expressiveness of the model.

[0052] At the same time, in order to improve the adaptability and prediction accuracy of the model, a dynamic adjustment mechanism based on real-time data feedback is introduced. This mechanism dynamically adjusts the parameters of the s and t functions by monitoring model performance and input data changes:

[0053] Adjustment factor=σ(W·x+b)

[0054] s′(x)=s(x)·(1+α·Adjustment factor)

[0055] t′(x)=t(x)+β·Adjustment factor

[0056] Among them, σ is the activation function, α and β are the adjustment rates, W and b are the weights and biases, which are learned during the model training process. Through this mechanism, the s and t functions in the coupling layer can self-adjust according to the changes in input data and model performance feedback.

[0057] 3. Data enhancement and training set construction:

[0058] The present invention maps and enhances the high-wear data in the latent space through a well-configured RealNVP model. The model uses the Adam optimizer and optimizes the parameters in 50 training cycles. The loss function used is the negative of the data log-likelihood, ensuring that the data distribution in the latent space is consistent with the original data distribution. In the data enhancement stage, the model generates 200 new data samples and uses the inverse transformation f -1These samples are mapped from the latent space back to the original feature space. These newly generated data samples are screened and only samples with wear values ​​between 150 and 300 are selected for further analysis. The formula for the inverse transform is:

[0059]

[0060] where z is sampled from the latent Gaussian distribution N(0, I), and f -1 is the inverse mapping function of RealNVP.

[0061] At the same time, in order to ensure the quality of generated data, the model introduces a data-based real-time feedback mechanism to dynamically adjust the scale s and translation t function parameters in the RealNVP model. The adjustment of these parameters is based on the statistical difference between the generated data and the actual data. The specific adjustment method is as follows:

[0062] s' new (x) = s old (x)+λ s ·(μ real -μ gen )

[0063] t' new (x) = t old (x)+λ t ·(σ real -σ gen )

[0064] Among them, μ real and σ real are the mean and standard deviation of the real data, μ gen and σ gen are the mean and standard deviation of the generated data. s and λ t is the learning rate parameter, which is used to adjust the amplitude of the scale and translation function updates. This mechanism ensures that the generated data is not only statistically similar to the real data, but also adapts to and reflects the complex wear phenomena that may occur under real operating conditions.

[0065] 4. Deep learning model training:

[0066] The DCNN-McBiLSTM model is trained using the expanded training set. In this model structure, the convolutional neural network layer is responsible for extracting spatial features, while the bidirectional long short-term memory network captures the dynamic changes in the time series data, enabling the model to accurately predict complex tool wear patterns. During the model training process, the dropout regularization technique is also used to avoid overfitting, and the batch normalization technique is used to speed up the convergence of the network. The loss function of the deep learning model uses the mean square error (MSE), which is where y i is the true label, is the predicted value, and the mean absolute percentage error MAPE is used as the evaluation index, the formula is: Where: n is the total number of samples in the data set, and y i is the true label, is the predicted value.

[0067] 5. Performance evaluation and analysis:

[0068] When evaluated on an independent test set, the prediction model trained using the training data augmented with the RealNVP model showed superior performance. Compared with the model trained on the original data alone, the prediction accuracy was significantly improved, with the mean squared error (MSE) reduced by 28.6% from 44.05μ 2 Reduced to 31.44μ 2 , while the mean absolute percentage error (MAPE) also decreased by 35%, from 4.6% to 2.99%. This result fully demonstrates the robustness and accuracy of the model in the normal operating range and less wear modes.

Claims

1. A tool wear prediction method based on RealNVP model with data enhancement, the method comprising: Step 1: Collect sample data and preprocess, During the cutting process, the tool cutting force, tool vibration signal, and sound signals in the X, Y, and Z directions are collected; after each processing is completed, the tool wear status is observed through an electron microscope, the wear of the three-edge milling cutter in X, Y, and Z is quantified, and the average wear value is calculated to evaluate the overall wear value; and the data collected from multiple sensors is combined with the tool wear results observed by the electron microscope to construct a comprehensive data set containing the corresponding relationship between sensor signals and tool wear; De-noising and normalizing the comprehensive dataset; Step 2: Build a sample enhancement model, A RealNVP model with ten coupled layers was constructed. Each coupled layer consisted of a deep neural network with two hidden layers and 512 hidden units in each hidden layer. Each coupling layer uses a non-volume preserving transformation, the transformation method is: and 1：d =x 1：d y d+1：D =x d+1：D ⊙exp(s(x 1：d ))+t(x 1：d ) Among them, x 1：d and x d+1:D are the first and second halves of the input vector x, y 1：d ,y d+1：D Represents the first and second halves of the transformed output vector, s and t are the first half x 1：d The function generates scale and translation parameters, D represents the total number of dimensions of the data, d represents the dimension of the current operation, and ⊙ represents element-level multiplication, that is, X y，d+1：D Each element in is equal to exp(s(X t，d )) is multiplied; this structure ensures that part of the input x 1：d Directly mapped to output y 1：d , and the other part x d+1：D After transformation; Among them, y 1：d ,y d+1：D Represents the first and second halves of the transformed output vector, x 1：d and x d+1：D They are the first and second halves of the input vector x, respectively. s() and t() are the first half of x. 1：d Functions that generate scale and translation transforms; Step 3: Sample data enhancement and training set construction; The model constructed in step 2 generates new samples in the latent space based on the existing samples, and then uses the inverse transformation f -1 () maps the new sample from the latent space back to the original feature space. The inverse transformation formula is: Among them, z is a new sample in the latent space sampled from the latent Gaussian distribution N(0, I), and f -1 is the inverse mapping function of RealNVP; Step 4: Use all the samples obtained to train a deep convolutional neural network-multi-channel bidirectional long short-term memory model, and use the trained deep convolutional neural network-multi-channel bidirectional long short-term memory model to predict tool wear.

2. A tool wear prediction method based on RealNVP model for data enhancement as claimed in claim 1, characterized in that: In step 2, in each coupling layer of the model, the s() and t() functions are dynamically updated according to the forward and backward propagation of the data; specifically, when the data flows through the coupling layer, the output of the s() and t() functions is calculated in real time according to the input data of the layer to ensure that the functions can adapt to the characteristics of the current data. During the training process, the gradient of the loss function relative to the parameters of the s() and t() functions is calculated, and these gradients are used to update the parameters to optimize the model performance; this update mechanism enables the s() and t() functions to effectively adapt to changes in data characteristics and enhance the data processing capabilities of the model; The update method is: s′(x)=s(x)·(1+a·σ(W·x+b)) t′(x)=t(x)+β·σ(W·x+b) Among them, σ is the activation function, a and β are the adjustment rates, W and b are the weights and biases, which are learned during the model training process. s(x) and t(x) represent the scale transformation and translation transformation functions before updating, and s′(x) and t′(x) represent the scale transformation and translation transformation functions after updating.

3. A tool wear prediction method based on RealNVP model for data enhancement as claimed in claim 1, characterized in that: In the step 3, the newly generated samples are screened to select data with tool wear values ​​in the range of 150 to 300 for subsequent training.

4. The tool wear prediction method based on RealNVP model for data enhancement as claimed in claim 1, characterized in that: In step 3, when the new sample data is inverted by the model function f -1 After conversion and input into the model, the s() and t() functions are updated according to the characteristics of the generated data and their deviation from the characteristics of the original data; the update method is: s′ new (x)=s old (x)+a s ·(m real -m gen ) t′ new (x)=t old (x)+λ t ·(s real -s gen ) Among them, μ real and σ real are the mean and standard deviation of the real data, μ gen and σ gen is the mean and standard deviation of the generated data; s and λ t is the learning rate parameter, which is used to adjust the amplitude of the scale and translation function update, s old (x), t old (x) scale transformation and translation transformation function before updating, s′ new (x), t′ new (x) represents the updated scale transformation and translation transformation function.

5. The tool wear prediction method based on RealNVP model for data enhancement as claimed in claim 1, characterized in that: The loss function of the deep convolutional neural network-multi-channel bidirectional long short-term memory model in step 4 adopts the mean square error, and the formula is: where y i is the true label, is the predicted value, n is the total number of samples in the data set, and the mean absolute percentage error MAPE is used as the evaluation indicator. The formula is:

Images