A feed rate adaptive control system and method that accounts for tool wear

Abstract

The application discloses a kind of feed speed adaptive control system and method considering tool wear, including multi-source signal feature screening module, tool wear monitoring module and feed speed self-adaptive control module;Multi-source signal feature screening module carries out the feature selection of multi-source sensing signal in processing based on domain adversarial neural network, and the feature that is sensitive to tool wear change and insensitive to feed speed change is screened out;Tool wear monitoring module accurately monitors the wear of grinding wheel under different feed speeds;Feed speed self-adaptive control module is formulated based on fuzzy control theory for different tool wear monitoring results under different feed speeds, and the feed speed adaptive control strategy for different tool wear stages is formulated, and the feed speed in part processing is adaptively regulated.The application has important practical significance for improving the intelligentization of part processing process, improving processing efficiency, prolonging tool life and ensuring the processing quality of parts.

Description

Technical Field

[0001] This invention relates to the field of intelligent manufacturing technology, and specifically to a feed rate adaptive control system and method that takes into account tool wear. Background Technology

[0002] With the rapid development of high-end manufacturing technologies, key industrial sectors such as aerospace, energy equipment, and precision instruments are facing increasingly stringent requirements for the processing of complex components. These components generally have typical characteristics such as complex structures, high dimensional accuracy requirements, and diverse materials. In particular, when high-strength alloys and composite materials are involved, their inherent low thermal conductivity and high hardness lead to significant fluctuations in cutting forces and abnormal temperature rises during processing, placing higher demands on existing processing technologies.

[0003] Currently, most CNC machine tools still employ fixed-parameter machining modes, where parameters such as feed rate and spindle speed are typically preset based on process experiments or empirical formulas. Domestic and international scholars have conducted extensive research on optimizing fixed-parameter machining, including establishing cutting parameter databases and developing process optimization algorithms. However, in actual machining processes, factors such as the tool-workpiece contact state and machine tool dynamic characteristics constantly change, making it difficult to maintain the preset fixed parameters at their optimal state. This is especially true during long-term continuous machining, where the time-varying nature of the machining system can lead to decreased machining quality and reduced efficiency. Therefore, monitoring the machining process and adjusting machining parameters in real time is crucial for achieving intelligent machining, improving part machining efficiency, and ensuring part machining quality.

[0004] In recent years, significant progress has been made in adaptive control technology for machining parameters based on process monitoring. Researchers dynamically adjust parameters such as feed rate through real-time feedback of signals such as cutting force and vibration. German scholars have developed a feed rate adjustment system based on cutting force control, while a Japanese team has proposed an adaptive control method using vibration signals. However, most of these studies focus on maintaining a stable cutting state, neglecting the changing patterns of tool wear, a crucial factor. In fact, tool wear significantly alters the geometry and cutting performance of the tool, and its evolution exhibits distinct stages. Tools at different wear stages display different cutting characteristics, requiring control systems to identify wear states and implement differentiated control strategies. However, existing adaptive control methods have failed to establish a correlation mechanism with tool wear states, leading to the application of the same control strategy at different tool wear stages. This limits further improvements in machining efficiency and affects the stability of machining quality. Summary of the Invention

[0005] To address the problems of existing machining parameter control technologies, such as simplistic control strategies, inability to adapt to tool wear variations, and unstable machining quality, this invention proposes an adaptive feed rate control system and method that considers tool wear. It establishes separate feed rate control strategies for the initial and stable wear stages of the tool, monitors tool wear in real time during part machining, determines the tool's wear stage, and selects the corresponding feed rate control strategy to control the feed rate during machining. This achieves adaptive feed rate control that takes tool wear into account during machining. This invention has significant practical implications for improving the intelligence of part machining processes, increasing machining efficiency, extending tool life, and ensuring machining quality.

[0006] To achieve the above-mentioned technical objectives, the technical solution adopted by the present invention is as follows:

[0007] In a first aspect, the present invention discloses a feed rate adaptive control system that takes into account tool wear, the system comprising a multi-source signal feature screening module, a tool wear monitoring module, and a feed rate adaptive control module;

[0008] The multi-source signal feature screening module constructs a feature screening model based on a domain adversarial neural network. The feature screening model includes a feature extractor, a tool wear monitor, a working condition classifier, and a gradient flipping layer.

[0009] The feature extractor extracts feature sets from sample data of source and target operating conditions. The sample data includes all source operating condition monitoring signal samples with tool wear labels, several target operating condition monitoring signal samples with tool wear labels, and target operating condition monitoring signal samples without tool wear labels. The tool wear monitor predicts the corresponding tool wear label based on the feature set output by the feature extractor. The operating condition classifier is used to distinguish the source operating conditions of the feature set output by the feature extractor. The gradient flip layer updates the parameters in the tool wear monitor and the operating condition classifier using the gradient descent method, so that the parameters in the feature extractor are updated in a gradient ascending manner. Through adversarial training with the operating condition classifier, it outputs fused features that are strongly correlated with tool wear and weakly correlated with feed rate, so that the feature distribution of the source and target operating conditions tends to be consistent.

[0010] The tool wear monitoring module is trained using the feature set selected by the feature filtering model, and is used to monitor the wear of the grinding wheel at different feed rates, and output the tool wear monitoring results at different feed rates.

[0011] The feed rate adaptive control module combines the tool wear monitoring results under different feed rates and formulates an adaptive control strategy for feed rate based on fuzzy control theory for different tool wear stages, so as to adaptively regulate the feed rate during the part machining process.

[0012] Furthermore, the monitoring signals include various part machining process signals such as cutting power, cutting force, and vibration.

[0013] Furthermore, the tool wear monitor uses gradient descent to update its parameters, with the loss function being:

[0014] ;

[0015] In the formula, and These are the network parameters for the feature extractor and the tool wear predictor, respectively. and These represent the number of samples with tool wear labels in the target operating condition and the source operating condition, respectively. and These are monitoring signal characteristics of the target working condition and the source working condition, respectively, along with tool wear label data pairs. This represents all monitoring signal characteristics of the target operating condition – tool wear label data pairs. and These are the predicted tool wear values ​​for the target working condition and the source working condition samples, respectively. This represents all monitoring signal characteristics of the source operating condition – tool wear label data pairs.

[0016] Furthermore, the working condition classifier updates its parameters using a gradient descent algorithm, with the loss function being:

[0017] ;

[0018] In the formula, and Here, N represents the network parameters for the feature extractor and the working condition classifier, respectively, and N is the number of samples. For monitoring signal characteristics of source or target operating conditions - operating condition label data pairs, These are the predicted operating conditions output by the operating condition classifier.

[0019] Furthermore, the gradient flipping layer includes a monitoring signal reconstruction module, a parameter initialization module, and a model parameter update module;

[0020] The monitoring signal reconstruction module is used to convert multiple input monitoring signals into... The feature matrices are then stacked into tensors and used as inputs to the feature extractor and the working condition classifier.

[0021] The parameter initialization module is used to initialize model parameters. , and ,in, , and These are the network parameters for the feature extractor, the working condition classifier, and the tool wear predictor, respectively.

[0022] The model parameter update module calculates the classification loss based on the loss function of the working condition classifier and the tool wear prediction loss based on the loss function of the tool wear predictor. It then iteratively updates the network parameters using the following formula until the preset maximum number of iterations is reached, and finally outputs the updated network parameters:

[0023]

[0024] In the formula, , and These are the updated network parameters. For learning rate, For domain adversarial hyperparameters.

[0025] Furthermore, the tool wear monitoring module includes a Transformer encoder, a gated loop unit, and a fully connected layer connected in sequence;

[0026] The Transformer encoder processes the grinding wheel wear monitoring signals at different feed rates and uses its output as the input to a gated loop unit. The output of the gated loop unit is then passed through a fully connected layer to perform regression prediction on tool wear, outputting the tool wear monitoring results at different feed rates. The calculation formula of the tool wear monitoring module is as follows:

[0027]

[0028] In the formula, and For the weights of the feedforward neural network, These are the input query vector, key, and weight, respectively. and The bias of the feedforward neural network is defined by Layerorm, where M is the layer normalization function. ulti This represents a multi-head attention mechanism, where ffn represents a feedforward neural network, ReLU is the activation function, and Sublayer is the processing function for both the attention mechanism and the feedforward neural network. The state was hidden in the previous moment. and The weight matrix is ​​used to calculate the candidate hidden states, and tanh is the activation function. This is the output of the Transformer encoder; Z represents the tool wear prediction result. t To update the gate, R t To reset the door.

[0029] Furthermore, the feed rate adaptive control module includes a force measuring instrument, a target cutting force determination module, a two-dimensional fuzzy controller, and a control module;

[0030] The force gauge is used to monitor the real-time cutting force during the machining process;

[0031] The target cutting force determination module analyzes the experimental data to determine the target cutting force during the machining process;

[0032] The two-dimensional fuzzy controller is constructed based on fuzzy control theory. Its inputs are the difference between the real-time cutting force and the target cutting force, as well as the rate of change of the real-time cutting force, and its output is the adjustment amount of the feed rate.

[0033] The control module adjusts the feed rate according to the adjustment amount output by the two-dimensional fuzzy controller;

[0034] Furthermore, the two-dimensional fuzzy controller, in conjunction with the tool wear characteristics during the machining process, divides the entire machining process into an initial wear stage and a stable wear stage. Then, based on experience, a knowledge base is constructed. For the uncertain rules in the knowledge base, the influence on the feed rate control is evaluated through orthogonal experiments to determine the optimal fuzzy rule for each stage. Specifically, in the initial tool wear stage, the tool is not fully broken in, and the change in cutting force is greater than a preset threshold. In the stable tool wear stage, the cutting performance of the tool tends to stabilize, and the change in cutting force is less than the preset threshold.

[0035] Secondly, the present invention discloses a feed rate adaptive control method considering tool wear, characterized in that the method is implemented based on the system described above; the method includes the following steps:

[0036] Collect sample data of source conditions and target conditions. The sample data includes all source condition monitoring signal samples with tool wear labels, several target condition monitoring signal samples with tool wear labels, and target condition monitoring signal samples without tool wear labels.

[0037] A feature selection model is constructed based on a domain adversarial neural network. The feature selection model includes a feature extractor, a tool wear monitor, a working condition classifier, and a gradient flipping layer. The feature selection model is trained using sample data. Specifically, the feature extractor extracts feature sets from sample data of source and target working conditions. The tool wear monitor predicts the corresponding tool wear label based on the feature sets output by the feature extractor. The working condition classifier distinguishes the source working conditions of the feature sets output by the feature extractor. The gradient flipping layer updates the parameters in the tool wear monitor and the working condition classifier using gradient descent, so that the parameters in the feature extractor are updated in a gradient ascending manner. Through adversarial training with the working condition classifier, it outputs fused features that are strongly correlated with tool wear and weakly correlated with feed rate, making the feature distributions of the source and target working conditions tend to be consistent.

[0038] A tool wear monitoring module was constructed. After training with the feature set output by the feature selection model, the grinding wheel wear under different feed rates was monitored, and the tool wear monitoring results under different feed rates were output.

[0039] Based on the tool wear monitoring results at different feed rates, an adaptive control strategy for feed rate is formulated according to fuzzy control theory for different tool wear stages, so as to adaptively regulate the feed rate during the part machining process.

[0040] Furthermore, the process of adaptively controlling the feed rate during part machining includes the following steps:

[0041] For the initial wear stage and the stable wear stage of the tool, respectively, empirical fuzzy control rules are formulated;

[0042] For the uncertain rules in the knowledge base of the two wear stages, orthogonal experiments are set up. The fuzzy rules of the two stages are determined by the improvement of machining efficiency, the change of steady-state maximum error and steady-state error. The fuzzy control rules of feed rate for the initial wear stage and steady wear stage of the tool are formulated.

[0043] Set the initial feed rate and perform part machining, collect monitoring signals during the machining process in real time, and use the tool wear monitoring model to output the tool wear monitoring results during the machining process in real time.

[0044] Based on the real-time output of tool wear monitoring results, a tool wear curve is plotted to determine the current wear stage of the tool. The corresponding fuzzy control rule for the feed rate is selected to control the feed rate during the machining process until the machining is completed. During the machining process, when the tool is detected to have passed the stable wear stage, a brand new tool is replaced to continue machining.

[0045] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0046] First, the feed rate adaptive control system and method of the present invention, which takes into account tool wear, can realize real-time monitoring and adaptive adjustment of feed rate during part machining, and has good applicability and high controllability.

[0047] Secondly, the feed rate adaptive control system and method of the present invention, which takes into account tool wear, considers the tool wear condition when performing adaptive control of feed rate. For the initial wear stage and the stable wear stage of the tool, respectively, the feed rate control strategy is formulated, which helps to give full play to the machining performance of the machine tool, improve machining efficiency, ensure the machining quality of parts and extend the tool life.

[0048] Third, the feed rate adaptive control system and method of the present invention, which takes into account tool wear, controls the feed rate during the machining process through fuzzy control, making the control process more scientific and reasonable. Attached Figure Description

[0049] Figure 1 This is a schematic diagram illustrating the structural principle of the feed rate adaptive control system considering tool wear according to the present invention.

[0050] Figure 2 This is a schematic diagram of the feature selection model structure based on DANN;

[0051] Figure 3 A comparison chart of the prediction performance of the constructed feature selection model;

[0052] Figure 4 A schematic diagram of the structure of the constructed Transformer-GRU tool wear model;

[0053] Figure 5 A schematic diagram of the structure of the constructed two-dimensional fuzzy control system;

[0054] Figure 6 A fuzzy subset of input and output variables;

[0055] Figure 7 This is a schematic diagram illustrating the changing trend of a fuzzy system.

[0056] Figure 8 This is a fuzzy control response diagram for the initial tool wear stage during the machining process of a certain part.

[0057] Figure 9 This is a fuzzy control response diagram for the stable wear stage of the tool during the machining process of a certain part.

[0058] Figure 10 The graph shows a comparison of the feed rate control effects during the machining process of a certain part. (a) corresponds to the initial wear stage, and (b) corresponds to the stable wear stage.

[0059] Figure 11 This is a comparison chart of tool life during the machining process of a certain part. Detailed Implementation

[0060] The embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0061] like Figure 1 As shown, the present invention discloses a feed rate adaptive control system that takes into account tool wear. The system includes a multi-source signal feature screening module, a tool wear monitoring module, and a feed rate adaptive control module.

[0062] The multi-source signal feature screening module constructs a feature screening model based on a domain adversarial neural network. The feature screening model includes a feature extractor, a tool wear monitor, a working condition classifier, and a gradient flipping layer.

[0063] The feature extractor extracts feature sets from sample data of source and target operating conditions. The sample data includes all source operating condition monitoring signal samples with tool wear labels, several target operating condition monitoring signal samples with tool wear labels, and target operating condition monitoring signal samples without tool wear labels. The tool wear monitor predicts the corresponding tool wear label based on the feature set output by the feature extractor. The operating condition classifier is used to distinguish the source operating conditions of the feature set output by the feature extractor. The gradient flip layer updates the parameters in the tool wear monitor and the operating condition classifier using the gradient descent method, so that the parameters in the feature extractor are updated in a gradient ascending manner. Through adversarial training with the operating condition classifier, it outputs fused features that are strongly correlated with tool wear and weakly correlated with feed rate, so that the feature distribution of the source and target operating conditions tends to be consistent.

[0064] The tool wear monitoring module is trained using the feature set selected by the feature filtering model, and is used to monitor the wear of the grinding wheel at different feed rates, and output the tool wear monitoring results at different feed rates.

[0065] The feed rate adaptive control module combines the tool wear monitoring results under different feed rates and formulates an adaptive control strategy for feed rate based on fuzzy control theory for different tool wear stages, so as to adaptively regulate the feed rate during the part machining process.

[0066] like Figure 2 As shown, the multi-source signal feature selection module constructs a feature selection model based on DANN. The feature selection model includes a feature extractor, a tool wear monitor, a working condition classifier, and a gradient flipping layer.

[0067] The feature extractor of the feature selection model extracts features from the source and target operating condition data. The input consists of all source operating condition monitoring signal samples with tool wear labels, a small number of target operating condition monitoring signal samples with tool wear labels, and target operating condition monitoring signal samples without tool wear labels.

[0068] The tool wear monitor's function is to accurately predict the tool wear label based on the feature set output by the feature extractor. The training process for the tool wear monitor uses gradient descent, and the loss function is:

[0069]

[0070] In the formula, For tool wear predictor parameters, and These represent the number of samples with tool wear labels in the target operating condition and the source operating condition, respectively. and These are monitoring signal characteristics of the target working condition and the source working condition, respectively, along with tool wear label data pairs. This represents all monitoring signal characteristics of the target operating condition – tool wear label data pairs. and Predicted tool wear values ​​for target and source operating conditions, respectively. This represents all monitoring signal characteristics of the source operating condition – tool wear label data pairs.

[0071] The purpose of the working condition classifier is to accurately distinguish the source working conditions of the input features. Its parameters are updated using the gradient descent algorithm, and the loss function is:

[0072]

[0073] In the formula, , These are the parameters for the feature extractor and the operating condition classifier, respectively, where N is the number of samples. For monitoring signal characteristics of source or target operating conditions - operating condition label data pairs, These are the predicted operating conditions output by the operating condition classifier.

[0074] The gradient flip layer updates the parameters in the feature extractor using a gradient-ascending method, increasing the classification error of the condition classifier. Through adversarial training with the condition classifier, the feature extractor outputs fused features that are strongly correlated with tool wear and weakly correlated with condition information (feed rate), making the feature distributions of the source and target conditions more consistent. That is:

[0075]

[0076] In the formula, The output of the feature extractor The marginal distribution of the source operating condition data, The marginal distribution of the target operating condition data.

[0077] The feature selection algorithm based on the DANN feature selection model takes the signal feature array, grinding wheel wear value, and working condition label as input, and outputs the model parameters. , and The specific algorithm is as follows:

[0078] Step 1: Reconstruct monitoring signal characteristics: Transform the characteristics of multi-source sensor monitoring signals into... The feature matrices are then stacked into a tensor form and used as the input to the network.

[0079] Step 2: Initialize model parameters , ,

[0080] Step 3: Calculate the classification loss based on the loss function of the working condition classifier, and calculate the tool wear prediction loss based on the loss function of the tool wear predictor. Iteratively update the model parameters using the following formula:

[0081]

[0082] Repeat the iterations until the maximum number of iterations is reached.

[0083] like Figure 3 As shown, to verify the effectiveness of the proposed feature selection model, typical CNN and SVR (Support Vector Regression) models were constructed as comparison models. These models have no condition classifier and their other structures are the same as DANN. Mean Absolute Error (MAE) was used to measure the performance of the tool wear prediction model. As can be seen from the figure, under known conditions (training and test sets), DANN, CNN, and SVR have similar prediction accuracy. In tool wear prediction under unknown conditions (validation set), DANN performs best, reducing the prediction error by 53.85% and 47.72% compared to CNN and SVR, respectively.

[0084] like Figure 4 As shown, the tool wear monitoring model is obtained by combining a Transformer encoder and a GRU. The model uses the output of the Transformer encoder as the input of the GRU, and the output of the GRU passes through a linear layer to achieve regression prediction of tool wear.

[0085] Specifically, the process for predicting tool wear using the Transformer-GRU tool wear monitoring model is as follows:

[0086] Step 1: Data Standardization and Dataset Partitioning: Extracting Wear Sequences Here, c represents the operating condition, and N represents the number of wear labels for the corresponding operating condition. To reduce the impact of data distribution on network training, the wear label data is normalized, and then the normalized operating condition data is divided into training and testing sets as needed.

[0087] Step 2: Sliding time window processing: Perform sliding window processing on the training set and the test set to obtain the input sequences of the training set and the test set respectively;

[0088] Step 3: Training the model: Input the standardized training set into the Transformer-GRU model, set the root mean square error (RMSE) as the loss function of the model, and use the adaptive motion estimation algorithm (Adam) to optimize the loss;

[0089] Step 4: Input the test set into the trained model and calculate the tool wear prediction result according to the following formula. ,

[0090]

[0091] In the formula, and For the weights of the feedforward neural network, These are the input query vector, key, and weight, respectively. and The bias of the feedforward neural network is defined by Layerorm, where M is the layer normalization function. ulti This represents a multi-head attention mechanism, where ffn represents a feedforward neural network, ReLU is the activation function, and Sublayer is the processing function for both the attention mechanism and the feedforward neural network. The state was hidden in the previous moment. and The weight matrix is ​​used to calculate the candidate hidden states, and tanh is the activation function. This is the output of the Transformer encoder; Z represents the tool wear prediction result. t To update the gate, R t To reset the door.

[0092] The feed rate adaptive control module includes a force gauge, a target cutting force determination module, a two-dimensional fuzzy controller, and a control module. The force gauge monitors the cutting force during machining; the target cutting force determination module experimentally determines the target cutting force; the two-dimensional fuzzy controller is constructed based on fuzzy control theory, and its structure is as follows: Figure 5As shown, this is used for real-time control of the feed rate during machining. Its inputs are the difference between the cutting force and the target force, and the rate of change of the cutting force; the output is the feed rate adjustment. The fuzzy subsets of the input and output variables are as follows: Figure 6 As shown.

[0093] The two-dimensional fuzzy controller constructs fuzzy rule knowledge bases for the initial wear stage and the stable wear stage of tool wear respectively. When constructing the knowledge base of the two-dimensional fuzzy controller, the knowledge base is first constructed based on experience. Then, for the uncertain rules in the knowledge base, the influence on the feed rate control is evaluated through orthogonal experiments to determine the optimal fuzzy rule for each stage.

[0094] Specifically, the feed rate control strategies for the initial and stable wear stages of the tool are as follows: In the initial wear stage, the tool is not fully broken in, resulting in a rapid wear rate and significant changes in cutting force. During this stage, the cutting force should be rapidly adjusted to quickly transition to the stable wear stage. Therefore, a rapid transition strategy is adopted: the feed rate is adaptively adjusted based on real-time changes in cutting force to quickly transition to the stable wear stage, reduce steady-state error, and improve machining efficiency. Simultaneously, entry and exit protection is implemented to prevent sudden changes in cutting force. In the stable wear stage, the tool's cutting performance tends to stabilize, the wear rate slows down, and the cutting force fluctuates within a certain range. During this stage, the cutting force should be controlled to fluctuate within a small range around the target force to reduce steady-state error, improve efficiency, and extend tool life as much as possible. Therefore, a constant force cutting strategy is adopted: by adjusting the feed rate in real-time, the fluctuation of cutting force is further reduced, and the cutting force is controlled within 5% above and below the target force.

[0095] Taking the machining process of a certain part as an example, the initial feed rate is 200 r / min. A feed rate adaptive control method considering tool wear includes the following steps:

[0096] Step 1: Build a hardware acquisition system that includes signals from the part machining process such as power, force, and vibration, and a tool wear measurement system;

[0097] Step 2: Conduct machining experiments at different feed rates to obtain machining process signals and corresponding tool wear change curves at different feed rates. Simultaneously, based on the cutting force changes and part machining conditions during the machining process at different feed rates, determine the target cutting force for subsequent feed rate control. The principle for determining the target cutting force is: the minimum cutting force in the stable wear stage of the experiment corresponding to the experiment that improves part machining efficiency, meets part machining quality requirements, and does not exhibit machining abnormalities such as chipping or tool breakage is the target cutting force in the initial wear stage, and the average value of the cutting force in the stable wear stage is the target cutting force in the stable wear stage.

[0098] Machining experiments were conducted on this part using different feed rates. It was found that a feed rate of 300 r / min significantly improved machining efficiency, ensured the part surface quality met requirements, and prevented machining anomalies such as tool breakage or chipping. Therefore, the minimum cutting force during the stable tool wear stage of the machining process corresponding to a feed rate of 300 r / min was set to the target cutting force during the initial tool wear stage (48 N), and the average cutting force during the stable wear stage was set to the target force during the stable wear stage (66 N). The upper limit for feed rate adjustment was set to 300 r / min.

[0099] Step 3: Use a feature filtering model based on DANN to extract features from the machining process signals, extract a fusion feature set that is strongly correlated with tool wear and weakly correlated with working condition information (feed speed), and train and test the constructed Transformer-GRU tool wear monitoring model to ensure the tool wear monitoring accuracy of the Transformer-GRU model under different feed speeds.

[0100] Step 4: For the initial wear stage and the stable wear stage of the tool, according to... Figure 7 Based on the changing trends of the fuzzy system shown, empirical fuzzy control rules as shown in Table 1 are formulated.

[0101] Table 1. Fuzzy control rules for the initial wear stage based on experience.

[0102]

[0103] Table 2. Fuzzy control rules for the empirically-based stable wear stage

[0104]

[0105] Step 5: For the uncertain rules in the knowledge base of the two wear stages, set up orthogonal experiments, and determine the specific fuzzy rules by improving the machining efficiency, the change of the steady-state maximum error and the steady-state error, so as to formulate the fuzzy control rules for the feed rate in the initial wear stage and the steady wear stage of the tool.

[0106] The AG of the fuzzy control rules for the initial wear stage shown in Table 1 is an uncertain rule. These seven rules are used as seven factors, and each factor has three levels. The efficiency increase and the maximum steady-state error are used as evaluation indicators. An L18(3^7) orthogonal experiment was set up. The experimental results are shown in Tables 3 and 4.

[0107] Table 3. Range analysis of efficiency increase during the initial wear stage

[0108]

[0109] Table 4 Range Analysis of Steady-State Maximum Error

[0110]

[0111] As can be seen from Tables 3 and 4, the optimal results for rules A through F are consistent based on the efficiency increase and the maximum steady-state error. Since rule G has a relatively small impact on the efficiency increase, the optimal level of rule G is determined based on the maximum steady-state error. The optimal rule combination for the initial wear stage is A3B3C3D2E2F3G2. The corresponding overall fuzzy rules for the initial wear stage are shown in Table 5.

[0112] Table 5. Overall Fuzzy Rule Table for the Initial Wear Stage

[0113]

[0114] The AE of the fuzzy control rules in the initial wear stage shown in Table 2 is an uncertain rule. These five rules are used as five factors, and each factor has four levels. The efficiency increase and steady-state error are used as evaluation indicators. An L16 (4^5) orthogonal experiment is set up. The experimental results are shown in Tables 6 and 7.

[0115] Table 6. Range analysis of efficiency increase during the steady wear stage

[0116]

[0117] Table 7 Range Analysis of Steady-State Error

[0118]

[0119] As can be seen from Tables 6 and 7, the optimal results for rules A and D based on efficiency increase and steady-state error are consistent. Rule B has a relatively small impact on steady-state error, and rule E has a relatively small impact on efficiency increase. Rule C has a large impact on both efficiency increase and steady-state error, but the optimal results based on the two indicators are inconsistent. Considering the strategy of constraining the cutting of fluctuations to reduce steady-state error during the stable wear stage, rule C is selected based on the optimal result of steady-state error. Therefore, the optimal rule combination for the stable wear stage is A1B4C4D3E1, and the overall rules for the corresponding stable wear stage are shown in Table 8.

[0120] Table 8. Overall Fuzzy Rule Table for Stable Wear Stage

[0121]

[0122] Step 6: Set the initial feed rate and perform part machining. Collect signals during the machining process in real time and use the constructed Transformer-GRU tool wear monitoring model to monitor tool wear during the machining process in real time.

[0123] Step 7: Based on the real-time monitoring of tool wear results, plot the tool wear curve, determine the current wear stage of the tool, and select the corresponding fuzzy control rule to control the feed rate during the machining process until the machining is completed. During the machining process, if the tool is detected to have passed the stable wear stage, a brand new tool needs to be replaced to continue machining.

[0124] Specifically, in step seven, the fuzzy control response of the feed rate during the initial wear stage is as follows: Figure 8 As shown in the figure, after intervention control, the feed rate was adjusted from 200 mm / min to 300 mm / min within 0.3 seconds, improving the air cutting efficiency. Milling times of 4.2 s and 11.4 s correspond to the entry and exit processes, respectively; at these times, the feed rate was rapidly reduced to prevent excessive cutting force. During the cutting process from 5 to 11.4 s and the air cutting process after 12.4 s, the feed rate was maintained at 300 mm / min to quickly transition through the initial tool wear stage. Machining efficiency was improved by 34.7%, and the maximum steady-state error was 3.75 N.

[0125] Specifically, in step seven, the fuzzy control response of the feed rate during the stable wear stage is as follows: Figure 9 As shown in the figure, during the air cutting process, the feed rate is rapidly adjusted to 300 mm / min. During the entry and exit processes, the feed rate is rapidly reduced to prevent excessive cutting force. The overall cutting process shows an accelerating trend, and the machining efficiency is improved by 33.1% compared to uncontrolled cutting, with a steady-state error of 2.43 N.

[0126] Two brand-new cutting tools were used to machine a certain part. The machining process was compared using the method proposed in this invention and a traditional constant cutting force control method with a single control strategy. The results are as follows: Figure 10 As shown in the figure, the method proposed in this invention can rapidly increase the cutting force during the initial wear stage, thereby enabling the tool to transition from the initial wear stage to the stable wear stage more quickly. During the stable wear stage, the method proposed in this invention can mitigate fluctuations in the cutting force. When using the proposed method for adaptive control of the feed rate, the steady-state error of the cutting force is only 2.43 N, a decrease of 44.65% compared to 4.39 N under a single control strategy. Simultaneously, the maximum cutting force deviation is only 5.69 N, a decrease of 35.49% compared to 8.82 N under a single control strategy.

[0127] Figure 11The figure shows the number of parts machined using the method proposed in this invention and the conventional constant cutting force control method with a single control strategy. As can be seen from the figure, when machining using the method proposed in this invention and the constant cutting force control method with a single control strategy, the tool can machine 15 parts and 13 parts respectively, indicating that the method proposed in this invention extends tool life by 15.38% compared to the constant cutting force control method with a single control strategy. Machining 15 parts using the control method proposed in this invention takes 2400 seconds, while the constant cutting force control method with a single control strategy machined 12 parts in the same time, indicating that the proposed method improves machining efficiency by 25% compared to the constant cutting force control method with a single control strategy.

[0128] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0129] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A feed rate adaptive control system considering tool wear, characterized in that, The system includes a feed rate adaptive control module, a multi-source signal feature filtering module, and a tool wear monitoring module; The feed rate adaptive control module includes a force measuring instrument, a target cutting force determination module, a two-dimensional fuzzy controller, and a control module; The force gauge is used to monitor the real-time cutting force during the machining process; The target cutting force determination module is used to conduct machining experiments at different feed rates, obtain machining process signals and corresponding tool wear change curves at different feed rates, and determine the target cutting force for subsequent feed rate control based on the cutting force changes and part machining conditions at different feed rates. The principle for determining the target cutting force is as follows: the minimum value of the cutting force in the stable wear stage of the tool corresponding to the experiment in which the part machining quality meets the requirements and no chipping or tool breakage occurs is taken as the target cutting force in the initial wear stage of the tool, and the average value of the cutting force in the stable wear stage is taken as the target cutting force in the stable wear stage. The two-dimensional fuzzy controller is constructed based on fuzzy control theory. Its inputs are the difference between the real-time cutting force and the target cutting force, as well as the rate of change of the real-time cutting force, and its output is the adjustment amount of the feed rate. The two-dimensional fuzzy controller, combined with the tool wear characteristics during machining, divides the entire machining process into an initial wear stage and a stable wear stage. Then, based on experience, a knowledge base is constructed. For uncertain rules in the knowledge base, orthogonal experiments are used to evaluate the impact on feed rate control, determining the optimal fuzzy rule for each stage. Specifically, the feed rate control strategies for the initial and stable wear stages are as follows: In the initial wear stage, a rapid transition strategy is adopted, adaptively adjusting the feed rate according to real-time changes in cutting force to quickly transition to the stable wear stage, reducing steady-state error and improving machining efficiency, while implementing entry and exit protection to prevent sudden changes in cutting force; In the stable wear stage, a constant force cutting strategy is adopted, further reducing cutting force fluctuations by adjusting the feed rate in real-time, controlling the cutting force within 5% above and below the target force. The control module adjusts the feed rate according to the adjustment amount output by the two-dimensional fuzzy controller; The multi-source signal feature screening module constructs a feature screening model based on a domain adversarial neural network. The feature screening model includes a feature extractor, a tool wear monitor, a working condition classifier, and a gradient flipping layer. The feature extractor extracts feature sets from sample data of source and target operating conditions. The sample data includes all source operating condition monitoring signal samples with tool wear labels, several target operating condition monitoring signal samples with tool wear labels, and target operating condition monitoring signal samples without tool wear labels. The tool wear monitor predicts the corresponding tool wear label based on the feature set output by the feature extractor. The operating condition classifier is used to distinguish the source operating conditions of the feature set output by the feature extractor. The gradient flip layer updates the parameters in the tool wear monitor and the operating condition classifier using the gradient descent method, so that the parameters in the feature extractor are updated in a gradient ascending manner. Through adversarial training with the operating condition classifier, it outputs fused features that are strongly correlated with tool wear and weakly correlated with feed rate, so that the feature distribution of the source and target operating conditions tends to be consistent. The tool wear monitoring module is trained using a feature set selected by a feature filtering model. It is used to monitor tool wear at different feed rates and output the tool wear monitoring results at different feed rates.

2. The feed rate adaptive control system considering tool wear according to claim 1, characterized in that, The monitoring signals include various part machining process signals such as cutting power, cutting force, and vibration.

3. The feed rate adaptive control system considering tool wear according to claim 1, characterized in that, The tool wear monitor uses the gradient descent method for parameter updates, with the loss function being: ； In the formula, and These are the network parameters for the feature extractor and the tool wear predictor, respectively. and These represent the number of samples with tool wear labels in the target operating condition and the source operating condition, respectively. and These are monitoring signal characteristics of the target working condition and the source working condition, respectively, along with tool wear label data pairs. This represents all monitoring signal characteristics of the target operating condition – tool wear label data pairs. and These are the predicted tool wear values ​​for the target working condition and the source working condition samples, respectively. This represents all monitoring signal characteristics of the source operating condition – tool wear label data pairs.

4. The feed rate adaptive control system considering tool wear according to claim 1, characterized in that, The operating condition classifier updates its parameters using a gradient descent algorithm, with the loss function being: ； In the formula, and Here, N represents the network parameters for the feature extractor and the working condition classifier, respectively, and N is the number of samples. For monitoring signal characteristics of source or target operating conditions - operating condition label data pairs, These are the predicted operating conditions output by the operating condition classifier. This represents all monitoring signal characteristics of the target operating condition – tool wear label data pairs. This represents all monitoring signal characteristics of the source operating condition – tool wear label data pairs.

5. The feed rate adaptive control system considering tool wear according to claim 1, characterized in that, The gradient flipping layer includes a monitoring signal reconstruction module, a parameter initialization module, and a model parameter update module; The monitoring signal reconstruction module is used to convert multiple input monitoring signals into... The feature matrix is ​​then stacked into a tensor form, which serves as the input to the feature extractor and the condition classifier. The parameter initialization module is used to initialize model parameters. , and ,in, , and These are the network parameters for the feature extractor, the working condition classifier, and the tool wear predictor, respectively. The model parameter update module calculates the classification loss based on the loss function of the working condition classifier and the tool wear prediction loss based on the loss function of the tool wear predictor. It then iteratively updates the network parameters using the following formula until the preset maximum number of iterations is reached, and finally outputs the updated network parameters: In the formula, , and These are the updated network parameters. For learning rate, For domain adversarial hyperparameters.

6. The feed rate adaptive control system considering tool wear according to claim 1, characterized in that, The tool wear monitoring module includes a Transformer encoder, a gated loop unit, and a fully connected layer connected in sequence. The Transformer encoder processes the grinding wheel wear monitoring signals at different feed rates and uses its output as the input to a gated loop unit. The output of the gated loop unit is then passed through a fully connected layer to perform regression prediction on tool wear, outputting the tool wear monitoring results at different feed rates. The calculation formula of the tool wear monitoring module is as follows: In the formula, and For the weights of the feedforward neural network, These are the input query vector, key, and weight, respectively. and The bias of the feedforward neural network is defined by Layerorm, where M is the layer normalization function. ulti This represents a multi-head attention mechanism, where ffn represents a feedforward neural network, ReLU is the activation function, and Sublayer is the processing function for both the attention mechanism and the feedforward neural network. The state was hidden in the previous moment. and The weight matrix is ​​used to calculate the candidate hidden states, and tanh is the activation function. This is the output of the Transformer encoder; Z represents the tool wear prediction result. t To update the gate, R t To reset the door.

7. A method for adaptive control of feed rate considering tool wear, characterized in that, The method is implemented based on the system described in any one of claims 1-6; the method includes the following steps: Based on the tool wear characteristics during machining, the entire machining process is divided into an initial wear stage and a stable wear stage. A knowledge base is then constructed based on experience. For uncertain rules in the knowledge base, orthogonal experiments are used to evaluate their impact on feed rate control, determining the optimal fuzzy rule for each stage. Specifically, the feed rate control strategies for the initial and stable wear stages are as follows: In the initial wear stage, a rapid transition strategy is adopted, adaptively adjusting the feed rate according to real-time changes in cutting force to quickly transition to the stable wear stage, reducing steady-state error and improving machining efficiency, while implementing entry and exit protection to prevent sudden changes in cutting force; In the stable wear stage, a constant force cutting strategy is adopted, further reducing cutting force fluctuations by adjusting the feed rate in real-time, controlling the cutting force within 5% above and below the target force. Machining experiments were conducted at different feed rates to obtain machining process signals and corresponding tool wear change curves at different feed rates. Simultaneously, based on the cutting force changes and part machining conditions during the machining process at different feed rates, the target cutting force for subsequent feed rate control was determined. The principle for determining the target cutting force is as follows: the minimum value of the cutting force in the stable wear stage of the tool corresponding to the experiment where the part machining quality meets the requirements and no chipping or tool breakage occurs is taken as the target cutting force in the initial wear stage, and the average value of the cutting force in the stable wear stage is taken as the target cutting force in the stable wear stage. Set the initial feed rate and perform part machining, collect monitoring signals during the machining process in real time, and use the tool wear monitoring model to output the tool wear monitoring results during the machining process in real time. Based on the real-time output of tool wear monitoring results, a tool wear curve is plotted to determine the current wear stage of the tool. The corresponding fuzzy control rule for the feed rate is selected, and the feed rate adjustment amount is output based on the difference between the real-time cutting force and the target cutting force, as well as the rate of change of the real-time cutting force. The feed rate is adjusted based on the output adjustment amount until the machining is completed. During the machining process, when the tool is detected to have passed the stable wear stage, a brand new tool is replaced to continue machining.

8. The feed rate adaptive control method considering tool wear according to claim 7, characterized in that, The method further includes: Collect sample data of source conditions and target conditions. The sample data includes all source condition monitoring signal samples with tool wear labels, several target condition monitoring signal samples with tool wear labels, and target condition monitoring signal samples without tool wear labels. A feature selection model is constructed based on a domain adversarial neural network. The model includes a feature extractor, a tool wear monitor, a working condition classifier, and a gradient flipping layer. The model is trained using sample data. Specifically, the feature extractor extracts feature sets from sample data of source and target working conditions. The tool wear monitor predicts corresponding tool wear labels based on the feature sets output by the feature extractor. The working condition classifier distinguishes the source working conditions from which the feature sets output by the feature extractor originate. The gradient flipping layer updates the parameters in the tool wear monitor and the working condition classifier using gradient descent, causing the parameters in the feature extractor to update in a gradient ascending manner. Through adversarial training with the working condition classifier, it outputs fused features that are strongly correlated with tool wear and weakly correlated with feed rate, making the feature distributions of the source and target working conditions more consistent. A tool wear monitoring module was constructed. After training with the feature set output by the feature selection model, the grinding wheel wear under different feed rates was monitored, and the tool wear monitoring results under different feed rates were output.

Images