A method for predicting the wear trend of gear hobbing tools based on DRSN and APF

Abstract

This invention belongs to the field of intelligent manufacturing auxiliary technology, and particularly relates to a method for predicting the wear trend of hobbing tools based on DRSN and APF, including the following steps: S1, constructing the framework of a hobbing tool wear trend prediction model; S2, acquiring multiple sets of historical machining data, each set of historical machining data including a set of hobbing tool vibration data throughout the entire machining life cycle of the workpiece, and the surface roughness value of the machined workpiece; processing the surface roughness value of the workpiece as a label for the wear trend value of the hobbing tool, and preprocessing the vibration signal data of the hobbing tool to obtain training data; then using the training data to train the trend prediction model; S3, acquiring real-time machining vibration data of the hobbing tool, preprocessing it, and inputting it into the hobbing tool wear trend prediction model to obtain the predicted wear trend value of the hobbing tool. This method can conveniently and effectively predict hobbing tool wear, thereby making the operational decisions of the hobbing machine tool's operating status more timely and effective.

Description

Technical Field

[0001] This invention belongs to the field of intelligent manufacturing auxiliary technology, and in particular relates to a method for predicting the wear trend of gear hobbing tools based on DRSN and APF. Background Technology

[0002] During high-speed gear hobbing, the cutting force and frictional heat generated by the interaction between the hob (i.e., "hob tool") and the workpiece cause wear on the hob surface, which in turn affects the surface quality and dimensional accuracy of the workpiece. Therefore, studying how to accurately identify and effectively predict the wear state of the hob is of great significance for enterprises to improve product quality, increase production efficiency and reduce costs.

[0003] After identifying the wear condition of the hob, if it is determined to be in a severely worn stage, a decision needs to be made regarding the machine tool's operating status. Directly stopping the machine to replace or sharpen the hob would increase production costs and waste tool resources. However, if, by adjusting the machining parameters, the hob can continue machining while maintaining stable machining quality until its wear trend reaches a threshold, then it can be scrapped or sharpened, making full use of the hob's remaining service life. Therefore, real-time prediction of hob wear trends or wear amounts can provide decision-makers with a basis for judgment, which is beneficial for improving machine tool production efficiency.

[0004] Currently, in the field of tool health management and maintenance, the use of deep learning models based on sensor data to identify and predict tool wear conditions is widely applied. However, in actual machining processes, constantly starting and stopping the machine tool to measure the wear of the hob surface can cause temperature imbalances in the machine tool system, leading to an increase in factors affecting hob wear and hindering the exploration of hob wear trend prediction based on vibration characteristics. Furthermore, the complex structure of the hob makes it difficult to directly measure the wear over its entire lifespan. In summary, direct online measurement of hob surface wear has limitations, affecting research on hob wear prediction. However, without measuring hob surface wear, it is difficult to train the corresponding wear prediction model. Therefore, deep learning-based hob wear prediction is very difficult to conduct in this field.

[0005] Therefore, how to predict hob wear in a relatively convenient and effective way, so as to make the operational decisions of hob machine tool operation status more timely and effective, has become an urgent problem to be solved. Summary of the Invention

[0006] To address the shortcomings of the existing technology, this invention provides a method for predicting the wear trend of hobbing tools based on DRSN and APF, which can predict hob wear relatively conveniently and effectively, thereby making the operational decisions of hobbing machine tools more timely and effective.

[0007] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0008] A method for predicting the wear trend of gear hobbing tools based on DRSN and APF includes the following steps:

[0009] S1. A framework for constructing a hob wear trend prediction model, used to predict the hob wear trend value based on the hob vibration signal data;

[0010] S2. Acquire multiple sets of historical processing data. Each set of historical processing data includes a set of hob vibration data for the entire life cycle of workpiece processing, as well as the surface roughness value of the processed workpiece. Process the surface roughness value of the workpiece as a label for the wear trend value of the hob, and preprocess the vibration signal data of the hob to obtain training data. Then use the training data to train the trend prediction model.

[0011] S3. After acquiring and preprocessing the real-time machining vibration data of the hob, input it into the hob wear trend prediction model trained in S2 to obtain the predicted wear trend value of the hob; wherein, the real-time machining vibration data includes hob vibration data of multiple sets of workpiece machining life cycle.

[0012] S4. Set up an observation equation to represent the relationship between the wear trend value and the wear trend state of the hob, wherein the wear trend state includes multiple wear parameters; and set up a state transition equation to represent the relationship between the wear trend states of the hob at adjacent times.

[0013] S5. Using the predicted wear trend value obtained in S3 as the observation value, and combining it with the observation equation and state transition equation set in S4, the wear state of the hob is estimated using the auxiliary particle filter algorithm to obtain the estimated values ​​of each state parameter in the wear state.

[0014] S6. Using the estimated values ​​of each state parameter obtained in S5, perform multi-step predictions for each state parameter using the preset ARIMA model to obtain the predicted values ​​of each state parameter for the next multiple steps.

[0015] S7. Using the predicted values ​​of each state parameter for multiple future steps, combined with the observation equation constructed in S4, the hob wear trend value for multiple future steps is obtained.

[0016] S8. The predicted wear trend value obtained in S3 and the future multi-step hob wear trend value obtained in S7 are used as hob wear prediction data to assist in subsequent operation decisions regarding the hob machine tool's operating status.

[0017] Compared with the prior art, the present invention has the following beneficial effects:

[0018] 1. Due to the complexity of the hob structure, it is difficult to directly measure the surface wear of the hob online, making it challenging to find a specific value to characterize the degree of hob wear. This invention shifts the focus away from the hob itself and adopts a different approach: indirectly reflecting the hob wear trend through the surface roughness of the workpiece (hob teeth) machined by the hob. Because different hob wear conditions naturally result in different workpiece surface roughness, and hob wear directly affects workpiece surface roughness, the workpiece surface roughness can indirectly reflect the degree of hob wear.

[0019] Based on this technical approach, this invention constructs corresponding training data to train a hob wear trend prediction model. Thus, during subsequent actual machining, only vibration data from the hob machining process needs to be collected to obtain the predicted workpiece surface roughness, i.e., the hob wear trend value, through the hob wear trend prediction model. This solves the problems of not being able to directly measure the hob surface wear online and finding specific values ​​that characterize the degree of hob wear.

[0020] 2. Based on obtaining a certain number of real-time hob wear trend values, this method also combines an auxiliary particle filter algorithm and an ARIMA model to estimate the hob wear trend values ​​for future multiple steps. This allows for the prediction of hob wear levels over a future period. The real-time wear trend values ​​can be used to assist in subsequent operational decisions regarding the hob machine's operating status. For example, if the hob wear trend begins to increase rapidly after a certain period, the hob operating status can be changed by adjusting the machining process parameters to fully utilize the tool's lifespan, thus ensuring machining quality while maximizing the remaining tool life. Furthermore, if adjusting the machining process parameters does not improve the workpiece quality and the tool wear trend continues to increase, the machine can be stopped to prevent dangerous situations.

[0021] In summary, this method can predict hob wear relatively conveniently and effectively, thereby enabling more timely and effective operational decisions regarding the hob machine's operating status.

[0022] Preferably, the hob wear trend prediction model includes a backbone network unit, a residual shrinkage processing unit, and a feature fusion unit connected in sequence; the backbone network unit includes a convolutional layer, a normalization layer, and a ReLU activation layer connected in sequence; the residual shrinkage processing unit includes multiple cascaded residual shrinkage modules; the feature fusion unit includes an adaptive average pooling layer and a fully connected layer; the preprocessed real-time machining vibration data is input to the backbone network unit and passes through the backbone network unit, the residual shrinkage processing unit, and the feature fusion unit in sequence to obtain the predicted wear trend value of the hob.

[0023] Using such a network architecture, the hob wear trend prediction model can learn features from signals containing strong background noise, automatically calculate thresholds to remove features irrelevant to the current task, and focus on important features, thereby ensuring the effectiveness of the prediction results.

[0024] Preferably, when constructing the framework of the hob wear trend prediction model, the residual shrinkage module selected is a residual shrinkage module with different thresholds between channels.

[0025] Since each channel has an independent threshold, the residual shrinkage module with different thresholds per channel outperforms the residual shrinkage module with shared thresholds per channel. Therefore, the framework for constructing the hob wear trend prediction model is selected using the residual shrinkage module with different thresholds per channel.

[0026] Preferably, in S2, the process of processing the surface roughness value of the workpiece as a label for the wear trend value of the hob includes: smoothing the obtained surface roughness value of the workpiece using a smoothing prior method to obtain the hob wear trend curve, defining the value on the curve as the wear trend value of the hob, and then using these values ​​on the curve as labels for the wear trend values ​​of the vibration signal data in the corresponding historical processing data group.

[0027] This solves the problem of errors in the measured roughness and unclear roughness change trend caused by human factors and measuring equipment, thus ensuring the effectiveness of subsequent operations.

[0028] Preferably, in S4, the state transition equation is:

[0029]

[0030] In the formula, x t The hob wear trend state at time t is determined by four state parameters a. t b t c t d t The vector [a] t ,b t ,c t ,d t ] T Indicates; ω t-1 This represents the process error, which conforms to a Gaussian distribution with a mean of 0 and a variance of 1e-6.

[0031] Preferably, in S4, the observation equation is:

[0032]

[0033] In the formula, z t v represents the hob wear trend value predicted by the hob wear trend prediction model at time t;t The observation error is represented by a Gaussian distribution with a mean of 0 and a variance of 25e-6; T is the time length.

[0034] Thus, after establishing the state transition equation and the observation equation, based on the obtained observation values, the estimated values ​​of the four state parameters {a} at different times can be obtained. t ,b t ,c t ,d t Then, by substituting these values ​​into the observation equation, we can obtain the estimated values ​​of the observations at time t+1.

[0035] Preferably, in S6, the values ​​of each parameter in the preset ARIMA model are set to p=1, d=1, and q=1.

[0036] Experiments have verified that this parameter setting is the optimal parameter combination. Attached Figure Description

[0037] To make the objectives, technical solutions, and advantages of the invention clearer, the invention will now be described in further detail with reference to the accompanying drawings, wherein:

[0038] Figure 1 This is a flowchart of the present invention;

[0039] Figure 2 This is a flowchart illustrating the workflow of the hob wear trend prediction model, auxiliary particle filter algorithm, and ARIMA model in Example 1.

[0040] Figure 3 This is a schematic diagram of the hob wear trend prediction model in Example 1;

[0041] Figure 4 This is a schematic diagram of the basic structure of the two residual shrinkage modules in Example 1;

[0042] Figure 5 This is a schematic diagram of the Bayesian estimation process in Example 1;

[0043] Figure 6 This is a schematic diagram of the gear hobbing experimental platform in Example 2;

[0044] Figure 7 The following are time-domain waveforms and their spectra of some vibration signals collected during the processing in Example 2;

[0045] Figure 8 This is a schematic diagram of the workpiece surface roughness measuring instrument in Example 2;

[0046] Figure 9 The surface roughness of the workpiece in Example 2 and the extracted hob wear trend curve;

[0047] Figure 10 This is a schematic diagram showing the partitioning of the experimental dataset for predicting the wear trend of the hobbing cutter in Example 2;

[0048] Figure 11 This is a hob wear trend prediction diagram from the hob wear trend prediction model in Example 2;

[0049] Figure 12 This is a schematic diagram illustrating the prediction performance of different models on the test set in Example 2;

[0050] Figure 13 This is a schematic diagram illustrating the state parameter prediction results using the ARIMA model in Example 2.

[0051] Figure 14 This is a schematic diagram of wear trend prediction based on the APF and ARIMA methods in Example 2;

[0052] Figure 15 This is a schematic diagram illustrating the multi-step prediction effect of different models on the wear trend of the hob in Example 2. Detailed Implementation

[0053] The following detailed explanation illustrates the specific implementation methods:

[0054] Example 1

[0055] like Figure 1 , Figure 2 As shown, this embodiment discloses a method for predicting the wear trend of gear hobbing tools based on DRSN and APF, including the following steps:

[0056] S1. A framework for constructing a hob wear trend prediction model is established to predict the hob wear trend value based on the hob vibration signal data.

[0057] In practical implementation, the framework structure of the hob wear trend prediction model is as follows: Figure 3 As shown in the diagram. Here, the framework of the hob wear trend prediction model refers to the hob wear trend prediction model before it has been trained. The hob wear trend prediction model includes a backbone network unit, a residual shrinking processing unit, and a feature fusion unit connected in sequence. The backbone network unit includes a convolutional layer, a normalization layer, and a ReLU activation layer connected in sequence. The residual shrinking processing unit includes multiple cascaded residual shrinking modules. The feature fusion unit includes an adaptive average pooling layer and a fully connected layer. The preprocessed real-time machining vibration data is input to the backbone network unit and passes through the backbone network unit, the residual shrinking processing unit, and the feature fusion unit in sequence to obtain the predicted wear trend value of the hob.

[0058] The forward propagation process of the hob wear trend prediction model is as follows: The input signal tensor, with a batch size of 32, 1 channel, and a signal length of 4096, is first shortened by a one-dimensional convolutional layer. Then, features are extracted through sequentially stacked residual shrinking modules. The network continuously increases the number of channels while reducing the input tensor size. Finally, the features are input to an adaptive average pooling layer to extract the average value of each channel, completing the feature extraction. The features are flattened into a one-dimensional vector and then input into a fully connected layer. The final output is the hob wear trend value.

[0059] Using such a network architecture, the hob wear trend prediction model can learn features from signals containing strong background noise, automatically calculate thresholds to remove features irrelevant to the current task, and focus on important features, thereby ensuring the effectiveness of the prediction results.

[0060] In the hobbing cutter wear trend prediction model, the residual shrinkage module is a key core component. It employs a Deep Residual Shrinkage Network (DRSN) architecture, integrating residual network structure and attention mechanisms, and incorporating filtering concepts from signal processing. It effectively filters out noise or redundant information by utilizing an adaptive soft threshold function. In this module, the residual network structure provides effective information transmission and learning capabilities, while the attention mechanism helps focus on important signal features. This combination makes the residual shrinkage module more efficient and accurate when processing complex signals. There are two types of residual shrinkage modules: channel-specific threshold type and channel-shared threshold type. The basic structures of the two types of residual shrinkage modules are as follows: Figure 4 (a) and Figure 4 As shown in (b), C represents the number of channels of the input feature and W represents the length of the input feature.

[0061] exist Figure 4 In the diagram, BN represents batch normalization, and ReLU and Sigmoid represent non-linear activation functions. As can be seen from the figure, the threshold of the different threshold-type residual shrinkage modules for each channel is calculated based on the relationship between the feature map and the weight vector α. c The product of these factors yields the threshold τ. c The threshold τ is a vector, while the threshold of the inter-channel shared threshold residual shrinkage module is calculated as the product of the average value of the feature map channels and the weight α, resulting in a positive threshold τ. Since each channel has an independent threshold, the inter-channel different threshold residual shrinkage module outperforms the inter-channel shared threshold residual shrinkage module. Therefore, this method uses inter-channel different threshold residual shrinkage modules to construct the framework of the hob wear trend prediction model.

[0062] S2. Acquire multiple sets of historical processing data. Each set of historical processing data includes a set of hob vibration data for the entire life cycle of workpiece processing, as well as the surface roughness value of the processed workpiece. Process the surface roughness value of the workpiece as a label for the wear trend value of the hob, and preprocess the vibration signal data of the hob to obtain training data. Then use the training data to train the trend prediction model.

[0063] The process of processing the workpiece surface roughness value as a label for the hob's wear trend value includes: smoothing the obtained workpiece surface roughness value using a smoothing prior method to obtain the hob wear trend curve, defining the values ​​on this curve as the hob wear trend value, and then using these values ​​on the curve as labels for the wear trend values ​​of the vibration signal data in the corresponding historical machining data set. This solves the problem of errors in the measured roughness and unclear roughness change trends caused by human factors and measuring equipment, thus ensuring the effectiveness of subsequent operations.

[0064] In practice, the training data can be divided into a training set, a validation set, and a test set. The training set is input into the model for training, and the validation set is used to verify the model's training effect. The optimal model parameters corresponding to the validation set are then saved.

[0065] S3. After acquiring and preprocessing the real-time machining vibration data of the hob, input it into the hob wear trend prediction model trained in S2 to obtain the predicted wear trend value of the hob; wherein, the real-time machining vibration data includes hob vibration data of multiple sets of workpiece machining life cycle.

[0066] S4. Set up an observation equation to represent the relationship between the wear trend value and the wear trend state of the hob, wherein the wear trend state includes multiple wear parameters; and set up a state transition equation to represent the relationship between the wear trend states of the hob at adjacent times.

[0067] In practical implementation, the state transition equation is as follows:

[0068]

[0069] In the formula, x t The hob wear trend state at time t is determined by four state parameters a. t b t c t d t The vector [a] t ,b t ,c t ,d t ] T Indicates; ω t-1This represents the process error, which conforms to a Gaussian distribution with a mean of 0 and a variance of 1e-6.

[0070] The observation equation is:

[0071]

[0072] In the formula, z t v represents the hob wear trend value predicted by the hob wear trend prediction model at time t; t The observation error is represented by a Gaussian distribution with a mean of 0 and a variance of 25e-6; T is the time length.

[0073] Thus, after establishing the state transition equation and the observation equation, based on the obtained observation values, the estimated values ​​of the four state parameters {a} at different times can be obtained. t ,b t ,c t ,d t Then, by substituting these values ​​into the observation equation, we can obtain the estimated values ​​of the observations at time t+1.

[0074] Because sensor signals are affected by various random factors during acquisition, the signal-to-noise ratio of hob vibration signals is often low. Therefore, it is difficult to directly obtain the specific relationship between the hob wear trend state and the hob wear trend value. In addition, it is currently impossible to determine the transformation relationship of the hob wear trend state at different times, thus making it impossible to obtain the analytical expressions of the state transition function and the observation function. This method, based on relevant data and drawing on the wear degradation laws of commonly used cutting tools such as milling cutters and turning tools, sets up the aforementioned observation equation and state transition equation.

[0075] S5. Using the predicted wear trend value obtained in S3 as the observation value, and combining it with the observation equation and state transition equation set in S4, the wear state of the hob is estimated using the auxiliary particle filter algorithm to obtain the estimated values ​​of each state parameter in the wear state.

[0076] Auxiliary Particle Filter (APF) is an improved algorithm based on the standard particle filter, but with a different particle update method. The standard particle filter utilizes Monte Carlo sampling to perform an iterative Bayesian estimation process, which can estimate the parameters or states of nonlinear non-Gaussian systems. Its main idea is to generate a set of weighted random samples, called particles, based on the prior probability distribution of the state. Then, based on the obtained observations, the weights of the particles are continuously adjusted to correct the initial prior probability, making it approximate the posterior probability density of the true state. When the sample size is large, the state values ​​simulated by Monte Carlo can be used to approximate the optimal estimate.

[0077] According to Bayesian theory, when the observation value z at time t+1 is obtained based on the hob wear trend prediction model... t+1 Then, it is necessary to calculate the posterior probability density p(x) of the hob wear trend state. t+1 |z t+1 Based on Bayesian theory, the posterior probability density p(x) is obtained. t+1 |z t+1 This requires two steps: iteratively performing prediction and updating. The specific process is as follows: Figure 5 As shown, where x t This represents the prior value of the hob wear trend at time t; z t x' represents the observation value at time t; t Represents the state x at time t. t The posterior estimate is obtained by estimating the four state parameters separately using an auxiliary particle filter algorithm, for each state x. t Corresponding to a set of parameters {a t ,b t .c t ,d t The auxiliary particle filter algorithm used in this method to update and predict the state essentially involves updating and predicting four state parameters.

[0078] Standard particle filtering does not utilize current observation information during resampling, resulting in significant blindness and a tendency to degenerate. In contrast, auxiliary particle filtering algorithms utilize current observation information to sample particles from a joint importance density distribution, increasing particle diversity and effectiveness. The particle weights are flatter, effectively addressing particle degeneration and thus providing more accurate estimation results.

[0079] S6. Using the estimated values ​​of each state parameter obtained in S5, perform multi-step predictions for each state parameter using a preset ARIMA model to obtain the predicted values ​​for each state parameter in the future. In specific implementation, the values ​​of each parameter in the preset ARIMA model are set to p=1, d=1, and q=1.

[0080] The Autoregressive Differential Moving Average (ARMA) model is a method for predicting nonlinear stationary time series. The main idea of ​​this model is to learn the patterns that change over time from historical data and use the learned patterns to predict future trends. This method has a simple structure and calculation, high short-term prediction accuracy, and is suitable for nonlinear prediction tasks. Therefore, this method uses the ARIMA model to achieve the task of predicting the future multi-step trend of hobbing tool wear.

[0081] If the hob wear trend value at time t is z t Not only with the hob wear trend value z at the previous p moments t-1 ,zt-2 ,…,z t-p It is related to, and also to, the disturbance ε at the previous q times. t-1 ,ε t-2 ,...,ε t-q If a dependency exists, the current wear trend value of the hob can be described by an autoregressive moving average (ARMA) (p, q) model, i.e.:

[0082]

[0083] The ARMA(p,q) model is based on a stationary sequence. However, the tool wear sequence has a clear increasing trend and is a non-stationary sequence. The ARIMA model introduces the difference term I(d) into the ARMA model, performs d-order difference operations on the hob wear trend sequence to stabilize it, and then establishes the ARMA model.

[0084] The AIC (A-Information Criterion) criterion is a commonly used method for selecting parameters in ARIMA models. The AIC criterion function is defined as follows:

[0085] AIC(p,d,q)=(nd)logσ 2 +2(p+q+1)logn;

[0086] Where n is the size of the input hob wear trend sequence, p, d, and q are the undetermined parameters of the ARIMA(p,d,q) model, and σ 2 This is the average sum of the fitted residuals.

[0087] For a given combination of parameters P of p, d, and q, if:

[0088]

[0089] Then p0, d0, and q0 are chosen as the optimal parameter combination for the ARIMA(p, d, q) model. Eight parameter combinations P are set as follows: (1, 1, 1), (1, 1, 2), (1, 2, 1), (1, 2, 2), (2, 1, 1), (2, 1, 2), (2, 2, 1), (2, 2, 2). Experiments have verified that the optimal parameter combination is p = 1, d = 1, and q = 1.

[0090] At the initial moment, 500 particles are generated for each parameter based on the mean and variance in Table 1. The average value of the particles for each parameter is calculated and then substituted into the observation equation to obtain the estimated value of the observation at the initial moment.

[0091] Table 1. Mean and variance of generated particles for each parameter at the initial time.

[0092]

[0093] After establishing the state transition equation and observation equation, estimates of the four state parameters {a} at different times can be obtained based on the acquired observations. t ,b t ,c t ,d t Then, substituting these values ​​into the observation equation, we can obtain the tracking value of the observed value at time t+1. After obtaining the estimated values ​​of the four parameters from the beginning to time t, we use the ARIMA model to predict these four parameters for multiple future steps. Substituting the predicted values ​​into the observation equation, we can obtain the predicted value z of the hobbing wear trend for multiple future steps. msp .

[0094] S7. Using the predicted values ​​of each state parameter for multiple future steps, combined with the observation equation constructed in S4, the hob wear trend value for multiple future steps is obtained.

[0095] S8. The predicted wear trend value obtained in S3 and the future multi-step hob wear trend value obtained in S7 are used as hob wear prediction data to assist in subsequent operation decisions regarding the hob machine tool's operating status.

[0096] Due to the complexity of hob structures, it is difficult to directly measure the surface wear of hobs online, making it challenging to find specific values ​​that characterize the degree of hob wear. This invention shifts the focus away from the hob itself and adopts a different approach: indirectly reflecting the hob wear trend by analyzing the surface roughness of the workpiece (hob teeth) machined by the hob. Since different hob wear conditions naturally result in different workpiece surface roughness, and hob wear directly affects workpiece surface roughness, the workpiece surface roughness can indirectly reflect the degree of hob wear. Based on this approach, this invention constructs corresponding training data to train a hob wear trend prediction model. Thus, during subsequent actual machining, only vibration data from the hob machining process needs to be collected to obtain the predicted workpiece surface roughness, i.e., the hob wear trend value, through the hob wear trend prediction model. This solves the problem of not being able to directly measure hob surface wear online and finding specific values ​​that characterize the degree of hob wear.

[0097] In addition, based on obtaining a certain number of real-time hob wear trend values, this method also combines an auxiliary particle filter algorithm and an ARIMA model to estimate the hob wear trend values ​​for future multiple steps. This allows for the prediction of hob wear levels over a future period. The real-time wear trend values ​​can be used to assist in subsequent operational decisions regarding the hob machine's operating status. For example, if the hob wear trend begins to increase rapidly after a certain period, the hob operating status can be changed by adjusting the machining process parameters to fully utilize the tool's remaining life while ensuring machining quality. Furthermore, if adjusting the machining process parameters does not improve the workpiece quality and the tool wear trend continues to increase, the machine can be stopped to prevent dangerous situations.

[0098] Example 2

[0099] To better illustrate the effectiveness of this method, the following verification experiment is provided.

[0100] Current research primarily focuses on commonly used cutting tools such as milling cutters and turning tools, and most studies are based on laboratory platforms and publicly available datasets such as PHM2010 and NASA for the identification and prediction of tool wear conditions. There is a lack of validation for practical engineering applications, and research on the identification and prediction of wear conditions for hobbing tools during actual machining processes is limited. Therefore, this validation experiment uses hob vibration signals throughout its entire lifespan to experimentally verify the proposed method. The method is compared with other models and its effectiveness is analyzed from both qualitative and quantitative perspectives.

[0101] Gear hobbing experimental platform and data acquisition

[0102] This verification experiment mainly describes the process of collecting hob vibration data, including the description of the hobbing machine tool platform scenario and the construction of the hob wear trend prediction dataset.

[0103] Gear hobbing machine tool platform scenario description

[0104] The hob vibration signal originates from the vibration generated by the hob base during gear machining on a high-speed dry-cut automatic gear hobbing machine (model YE120CNC). The experimental platform machining scenario is as follows: Figure 6 As shown. (a) Gear machining area; (b) Panoramic view of the experimental equipment; (c) Hob; (d) Gear blank. From... Figure 6 As shown in (a), two vibration acceleration sensors are arranged on the gear hobbing machine tool, installed on the hob base and the machine tool column, respectively. The main process parameters and data processing hardware and software parameters involved in the experiment are shown in Tables 2 and 3.

[0105] Table 2 Main process parameters during gear hobbing machine machining

[0106]

[0107] Table 3 shows the computer hardware and software parameters involved in processing vibration data.

[0108]

[0109] (1) Vibration data acquisition

[0110] During data acquisition, the vibration signal of the hobbing cutter base collected by the piezoelectric accelerometer is first input into the data acquisition card via the sensor cable. After data conversion, it is stored in the portable hard drive. The parameters of the data acquisition equipment involved in the experiment are shown in Table 4.

[0111] Table 4. Parameters of the data acquisition equipment involved in the experiment.

[0112]

[0113] The time-domain waveforms and spectra of some vibration signals collected during gear hobbing are as follows: Figure 7 As shown, there is no significant difference in the signal waveform at different time stages, but there are obvious differences in the spectrum.

[0114] (2) Surface roughness measurement of workpiece

[0115] Surface roughness is a quantitative indicator that indirectly reflects the surface quality of a workpiece. Its influencing factors include cutting force, cutting speed, tool wear, and cutting fluid. In continuous, stable machining operations using a cutting tool, cutting speed and cutting force can be considered constant. Furthermore, cutting fluid is not used in high-speed dry-cut automatic gear hobbing machines. Therefore, the main influencing factor on workpiece surface roughness is related to tool wear. Higher tool wear results in a duller cutting edge, leading to increased chips and friction during machining, thus increasing workpiece surface roughness.

[0116] To mitigate the impact of machining temperature on hob wear, this experiment began with the machining of the 61st workpiece, measuring the surface roughness of the workpieces in cycles of 18 workpieces. A roughness measuring instrument such as... Figure 8 As shown, a total of 200 workpiece surface roughness values ​​were measured, and the results are as follows. Figure 9 As shown by the black curve in the middle.

[0117] To address the issues of measurement roughness errors and unclear roughness variation trends caused by human factors and measuring equipment, a smoothing prior method is used to smooth the roughness. The results are as follows: Figure 9As shown by the red curve, this curve is defined as the "hob wear trend curve," and the values ​​on this curve are called "hob wear trend values." The hob wear stages are divided based on the hob wear trend values, and these values ​​are used as labels for predicting wear trends.

[0118] Construction of hob wear trend prediction dataset

[0119] The extracted hob wear trend curve is used as the wear trend prediction label, and the corresponding vibration signals are uniformly and non-overlappingly segmented into segments using a fixed-length window (window size set to 4096). To enable the model to learn the wear trend throughout the entire lifespan of the hob on the same training or test set, a set of samples consisting of vibration data at the same location corresponding to all wear trend values ​​is called a data set. A schematic diagram of the hob wear trend prediction experimental dataset partitioning is shown below. Figure 10 As shown in Table 5, 80 sets of data samples were selected as the training set, 20 sets as the validation set, and 1 set as the test set. The results of constructing the hob wear trend prediction dataset are shown in Table 5.

[0120] Table 5 Results of constructing the hobbing cutter wear trend prediction dataset

[0121]

[0122] Predicting hob wear trends

[0123] Performance evaluation indicators

[0124] To compare the predictive performance of different models, the following four quantitative indicators were selected to evaluate the model prediction results. These four indicators are used to measure the deviation between the predicted value and the true value to judge the accuracy of the model.

[0125] Mean Square Error (MSE) is calculated by averaging the squares of the differences between the predicted and actual values, as shown in the following formula. MSE represents the difference between the predicted and actual values, indicating that the model fits well.

[0126]

[0127] The root mean square error (RMSE) is calculated by taking the square root of the square of the differences between the predicted and actual values. As shown in the formula below, it measures the average magnitude of the prediction error. A smaller RMSE indicates a smaller difference between the predicted and actual values.

[0128]

[0129] Mean Absolute Error (MAE) is calculated by averaging the absolute values ​​of the differences between the predicted and actual values. A smaller MAE indicates that the model's predicted values ​​are less different from the actual values, and that the model fits the data better.

[0130]

[0131] R 2 The R-value, known as the coefficient of determination, is a statistical indicator used to evaluate the goodness of fit of a regression prediction model. A higher R-value indicates a better fit. 2 The value indicates that the model fits the data well and its predictions can explain the variability of the dependent variable well, that is, the model fits well.

[0132]

[0133] Hob wear trend prediction results and comparative analysis

[0134] The training set was input into the hob wear trend prediction model for training. The parameters used during model training are shown in Table 6.

[0135] Table 6. Parameters used when training the pre-model of hob wear trend.

[0136]

[0137] The wear trend prediction effect of the hob wear trend prediction model was observed using a test set, and the results are as follows: Figure 11 As shown in the figure, the hob wear trend prediction model can accurately predict the wear trend value, with a small error between the predicted and actual values. Furthermore, the predicted trend is consistent with the experimentally measured wear trend curve, effectively reflecting the deterioration trend of hob wear. However, the hob wear trend prediction model is not entirely consistent with the actual hob wear trend value. This is mainly because the collected dataset may contain signals from other sources, interfering with the hob wear difference. Additionally, the sample size in the dataset is relatively small, limiting its ability to train the neural network model. Therefore, future research requires further supplementation of the dataset and source tracing of the collected signals to enable the neural network model to learn features from sufficiently large and clean signals.

[0138] To better demonstrate the good predictive performance of the hob wear trend prediction model, it was compared with ResNet18 residual neural network, Time Series Transformer, Bidirectional Long Short-Time Neural Network (BiLSTM), Time Series Convolutional Neural Network (TCN), Support Vector Regression (SVR), and Random Forest Regression (RF) models. The prediction results are as follows: Figure 12 As shown. With Figure 11A comparison of the hob wear trend prediction results of the hob wear trend prediction model shows that the proposed method outperforms the six comparative models. In addition, among the six models, the residual neural network has a better prediction effect than the other five models. The prediction effect in the initial wear and normal wear stages is generally consistent with the actual hob wear trend value. However, in the later severe wear stage, the error between the actual value and the predicted value is large. The prediction effect of the time series convolutional neural network is the worst, with a significant error between the actual value and the predicted value.

[0139] The performance evaluation results of each model are shown in Table 7. As can be seen from the table, the hobbing trend prediction model proposed in this invention exhibits the best prediction performance, with R... 2 The value was improved by 0.198 compared to the suboptimal residual neural network model. The TCN model had the worst prediction performance. 2 The value is only -7.7921, indicating that the TCN model has very poor applicability to this test set. Looking at the other three quantitative indicators, the RMSE, MAE, and MAPE values ​​of the hob wear trend prediction model proposed in this invention are 0.018, 0.0131, and 0.0213 respectively, and compared with the prediction results of the other six models, the hob wear trend prediction model has the smallest prediction error. Therefore, the hob wear trend prediction model used in this invention has excellent predictive performance and can accurately reflect changes in wear trends.

[0140] Table 7 Performance Indicators of Wear Trend Tracking for Different Models

[0141]

[0142] Multi-step prediction effect and comparative analysis of hob wear trend

[0143] When hob wear reaches a severe stage, predicting the future wear trend of the hob in advance plays a crucial role in subsequent decision-making. To verify the hob wear trend prediction method based on APF (Auxiliary Particle Filtering) and ARIMA proposed in this invention, 190 wear trend values ​​predicted by the hob wear trend prediction model were used as observations in the auxiliary particle filtering method. These observations were sequentially input into the auxiliary particle filtering algorithm to estimate the hob wear trend state, resulting in the change curves of the four parameter estimates corresponding to the state, as shown below. Figure 13 As shown by the black curve in the middle, by inputting all the estimated values ​​of the four parameters into the observation equation in the auxiliary particle filter, we obtain the following: Figure 14 The black curve in the middle shows 190 tracking values. After obtaining the parameter change curves of the state, the ARIMA model is used to predict the four parameters for the next 10 steps, and the prediction results are as follows. Figure 13As shown by the red curve, the predicted values ​​can well reflect the trend changes based on the existing 190 parameter values.

[0144] Substituting the four state parameter values ​​for the next 10 steps into the observation equation sequentially, the resulting 10 observation values ​​are the predicted hob wear trend values ​​obtained based on the APF and ARIMA methods, such as... Figure 14 As shown by the green curve. Figure 14 The black curve represents the predicted hob wear trend value obtained by inputting the vibration signal into the hob wear trend prediction model. This value is used as the observation value in the auxiliary particle filter method. The black curve represents the tracking value obtained after state estimation using the auxiliary particle filter method. Figure 14 As can be seen, within the range of 1-190 feeds, the error between the tracking value obtained using the auxiliary particle filter method and the predicted value obtained by the hob wear trend prediction model is small, indicating that the method proposed in this invention has a good tracking ability for hob wear trend values. Within the range of 190-200 feeds, the proposed hob wear prediction method based on APF and ARIMA models shows good short-term prediction performance. Although not a perfect match, the predicted values ​​are basically consistent with the actual hob wear trend values, and can well characterize the deterioration trend of hob wear.

[0145] To better verify the predictive performance of the proposed method, three time series prediction models—Holt-Winters Seasonal Exponential Smoothing (HW), Holt's Linear Exponential Smoothing (Holt), and the Gray Model (GM)—were used to estimate the four state parameters for the next 5, 10, 15, and 20 steps, respectively. The parameter values ​​were then substituted into the observation equation to achieve multi-step prediction of the hob wear trend. Furthermore, without using the auxiliary particle filtering method, based on the prediction results of the hob wear trend prediction model, the ARIMA, HW, Holt, and GM models were directly used to predict the hob wear trend for the next multiple steps. The results are as follows. Figure 15 As shown. From Figure 15 (b) Figure 15 (c) Figure 15 As can be seen in (d), the method based on APF and ARIMA proposed in this invention has good prediction performance and can well reflect the changes in wear trend. Especially in the prediction of the next 10 steps, the difference between the changing trend and the actual hob wear trend value is small, which shows that the method proposed in this invention has good predictive performance. Figure 15 (c) and Figure 15As can be seen from (d), directly using the ARIMA model for prediction does not yield good results, while the method combining APF and ARIMA achieves the best prediction performance. This demonstrates that the method proposed in this invention has significantly improved the prediction performance. Figure 15 As can be seen in (a), the prediction results based on the method proposed in this invention and the Holt method are similar, which is mainly related to the prediction step size set, but overall they are consistent with the trend of hob wear.

[0146] Performance evaluation metrics were used to compare the prediction performance of different models from a quantitative analysis perspective, and the results are shown in Table 8. Compared with other methods, this method shows superior prediction performance in the next 10, 15, and 20 steps, with the minimum errors in MSE, MAPE, and RMSE. 2 The value is the largest. In addition, the combined model based on APF and GM has the worst prediction effect. This is because the GM model mainly deals with the fact that the generated sequence has an exponential change law and can only describe the monotonic change process, so it is not suitable to use the GM model for prediction.

[0147] Table 8 Comparison of results for multi-step prediction of hobbing wear trend based on different time series prediction methods

[0148]

[0149]

[0150] This verification experiment validates the performance of the hob wear trend prediction model from the perspective of practical engineering applications. First, hob vibration signals were collected from an actual high-speed dry-cut automatic gear hobbing machine, and an experimental dataset was constructed based on these signals. Second, the vibration signals were input into the hob wear trend prediction model to achieve real-time prediction of the hob wear trend value. The model achieved an RMSE of only 0.018 and an R² of 0.9762 on the test set. Finally, the APF and ARIMA models were used to predict the future hob wear trend for 5, 10, 15, and 20 steps. The overall prediction results were good, except for a slightly weaker performance in the 5-step prediction; the prediction results for the other hob wear trends were good, demonstrating that the method proposed in this invention has excellent predictive performance.

[0151] To address the difficulty in directly measuring the wear of hob surfaces, which hinders the construction of effective wear prediction labels and impedes the development of deep learning-based hob wear prediction models, this method uses workpiece surface roughness to indirectly reflect the wear trend of the hob. The hob wear trend curve, obtained by smoothing the extracted workpiece surface roughness, is used as a prediction label. Based on this label and vibration signals, a framework for a hob wear trend prediction model is constructed. Inputting the obtained vibration signals into the trained model allows for real-time prediction of the hob wear trend value. Furthermore, based on a certain number of hob wear trend values, the prediction results from the hob wear trend prediction model are used as observations, and auxiliary particle filtering and an ARIMA model are employed to predict the hob wear trend values ​​for future multiple steps. Experimental results show that the proposed method has good prediction performance. In predicting the hob wear trend for the next 10, 15, and 20 steps, RMSE, MAE, MAPE, and R... 2 The values ​​are all optimal. In the next 10 prediction steps, the RMSE between the predicted value and the actual value of the proposed method is only 0.0302. In the next 5 prediction steps, the prediction effect of the proposed method is slightly worse than that of the Holt model, but the RMSE value between the two differs by only 0.0075. Therefore, the proposed hob wear trend prediction method based on the hob wear trend prediction model and APF is suitable for the hob wear trend prediction task.

[0152] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit the technical solutions. Those skilled in the art should understand that any modifications or equivalent substitutions to the technical solutions of the present invention without departing from the spirit and scope of the present invention should be covered within the scope of the claims of the present invention.

Claims

1. A method for predicting the wear trend of gear hobbing tools based on DRSN and APF, characterized in that, Includes the following steps: S1. A framework for constructing a hob wear trend prediction model, used to predict the hob wear trend value based on the hob vibration signal data; S2. Acquire multiple sets of historical processing data. Each set of historical processing data includes a set of hob vibration data for the entire life cycle of workpiece processing, as well as the surface roughness value of the processed workpiece. Process the surface roughness value of the workpiece as a label for the wear trend value of the hob, and preprocess the vibration signal data of the hob to obtain training data. Then use the training data to train the trend prediction model; S3. After acquiring and preprocessing the real-time machining vibration data of the hob, input it into the hob wear trend prediction model trained in S2 to obtain the predicted wear trend value of the hob; wherein, the real-time machining vibration data includes hob vibration data of multiple sets of workpiece machining life cycle. S4. Set up an observation equation to represent the relationship between the wear trend value and the wear trend state of the hob, wherein the wear trend state includes multiple wear parameters; and set up a state transition equation to represent the relationship between the wear trend states of the hob at adjacent times. S5. Using the predicted wear trend value obtained in S3 as the observation value, and combining it with the observation equation and state transition equation set in S4, the wear state of the hob is estimated using the auxiliary particle filter algorithm to obtain the estimated values ​​of each state parameter in the wear state. S6. Using the estimated values ​​of each state parameter obtained in S5, perform multi-step predictions for each state parameter using the preset ARIMA model to obtain the predicted values ​​of each state parameter for the next multiple steps. S7. Using the predicted values ​​of each state parameter for multiple future steps, combined with the observation equation constructed in S4, the hob wear trend value for multiple future steps is obtained. S8. The predicted wear trend value obtained in S3 and the future multi-step hob wear trend value obtained in S7 are used as hob wear prediction data to assist in subsequent operation decisions based on the hob machine tool's operating status. The hob wear trend prediction model includes a backbone network unit, a residual shrinkage processing unit, and a feature fusion unit connected in sequence. The backbone network unit includes a convolutional layer, a normalization layer, and a ReLU activation layer connected in sequence. The residual shrinkage processing unit includes multiple cascaded residual shrinkage modules. The feature fusion unit includes an adaptive average pooling layer and a fully connected layer. The preprocessed real-time machining vibration data is input to the backbone network unit and passes through the backbone network unit, the residual shrinkage processing unit, and the feature fusion unit in sequence to obtain the predicted wear trend value of the hob.

2. The method for predicting the wear trend of gear hobbing tools based on DRSN and APF as described in claim 1, characterized in that: When constructing the framework of the hob wear trend prediction model, the residual shrinkage module selected is a residual shrinkage module with different thresholds between channels.

3. The method for predicting the wear trend of gear hobbing tools based on DRSN and APF as described in claim 1, characterized in that: In S2, the process of processing the surface roughness value of the workpiece as a label for the wear trend value of the hob includes: smoothing the obtained surface roughness value of the workpiece using a smoothing prior method to obtain the hob wear trend curve, defining the value on the curve as the wear trend value of the hob, and then using these values ​​on the curve as labels for the wear trend values ​​of the vibration signal data in the corresponding historical processing data group.

4. The method for predicting the wear trend of gear hobbing tools based on DRSN and APF as described in claim 1, characterized in that: In S4, the state transition equation is: In the formula, x t The hob wear trend state at time t is determined by four state parameters a. t b t c t d t The vector [a] t ,b t ,c t ,d t ] T Indicates; ω t-1 This represents the process error, which conforms to a Gaussian distribution with a mean of 0 and a variance of 1e-6.

5. The method for predicting the wear trend of gear hobbing tools based on DRSN and APF as described in claim 4, characterized in that: In S4, the observation equation is: In the formula, z t ν represents the hob wear trend value predicted by the hob wear trend prediction model at time t; t The observation error is represented by a Gaussian distribution with a mean of 0 and a variance of 25e-6; T is the time length.

6. The method for predicting the wear trend of gear hobbing tools based on DRSN and APF as described in claim 1, characterized in that: In S6, the preset ARIMA model has the parameters set to p=1, d=1, and q=1.

Images