Method for predicting and dynamically compensating tool wear in numerical control machining based on machine learning

Abstract

The application discloses a kind of based on machine learning's numerical control processing tool wear prediction and dynamic compensation method, belong to industrial control technical field.The method includes: acquisition multi-source sensor signal data, after pre-processing, according to mean and standard deviation, adaptive symbol mapping is carried out, and symbol sequence data is obtained;Self-encoder is constructed to extract features to symbol sequence, and obtain multi-dimensional fusion features;Physical information neural network is constructed, data-driven loss and physical constraint loss based on Usui wear rate model are fused, tool current wear and remaining service life are predicted;Future wear trend is represented by symbol transition probability matrix;With current wear state and future wear trend as state space, dynamic compensation strategy is generated by reinforcement learning model decision;Compensation strategy is converted into numerical control instruction execution, and closed-loop feedback and model updating are realized by digital twin body.The application improves tool wear prediction accuracy and cross-condition generalization ability, realizes prospective dynamic compensation.

Description

Technical Field

[0001] This invention relates to the field of industrial control technology, specifically to a method for predicting and dynamically compensating for wear of CNC machining tools based on machine learning. Background Technology

[0002] In CNC machining, tool wear is an unavoidable and highly complex dynamic process. Tool wear directly affects machining quality, machining efficiency, and tool life. Accurate tool wear prediction and timely compensation control are of great significance for ensuring machining accuracy and reducing production costs.

[0003] In existing technologies, tool wear prediction and compensation methods suffer from insufficient generalization ability. Traditional methods often rely on empirical or machine learning models built from single-condition data. When machining conditions change—such as changing workpiece materials, adjusting cutting parameters, or replacing different types of tools—the model's prediction accuracy drops sharply, making it difficult to adapt to the variable machining conditions in actual production. Feature extraction is manual and lacks sufficient information. Existing methods often rely on manual extraction of statistical features in the time and frequency domains. The quality of feature engineering is highly dependent on professional knowledge, and single-dimensional features cannot fully characterize the complex dynamic behavior of tool wear, easily overlooking the correlation and complementarity between multi-source sensor signals. Traditional compensation strategies are mostly fixed threshold strategies—only triggering tool replacement or regrinding when wear reaches a certain preset threshold. This "post-event response" strategy cannot proactively adjust according to the wear evolution trend, easily leading to untimely compensation (affecting machining quality) or overcompensation (wasting tool life). Summary of the Invention

[0004] The purpose of this invention is to provide a machine learning-based method for predicting and dynamically compensating CNC machining tool wear, in order to solve the problem of insufficient generalization ability of existing tool wear prediction and compensation methods.

[0005] The technical solution of the present invention to solve the above-mentioned technical problems is as follows: Machine learning-based methods for predicting and dynamically compensating for CNC machining tool wear include: Collect signal data from multiple sources during CNC machining and preprocess the data to obtain standardized multi-source sensor data. A machine learning-based multidimensional feature extraction model is constructed to extract features from the standardized multi-source sensor data and obtain multidimensional fused features. A wear prediction model is constructed to predict the current wear state and future wear trend of the tool based on the multi-dimensional fusion features; Based on the current wear status and the future wear trend, a wear compensation strategy is dynamically generated through a decision model. The wear compensation strategy is converted into CNC instructions and executed.

[0006] A further embodiment involves acquiring signal data from multiple sensors during CNC machining and preprocessing it to obtain standardized multi-source sensor data, including: Acquire cutting force signals, vibration signals, and temperature signals; The collected signals are filtered, denoised, and normalized to obtain standardized data; Calculate the statistical characteristics of the standardized data, including the mean and standard deviation; Based on the statistical characteristics, the symbol intervals are divided, and the standardized data is converted into symbol sequences.

[0007] A further embodiment, wherein dividing the symbol interval based on the statistical characteristics and converting the standardized data into a symbol sequence includes: The rule for dividing the symbol interval is as follows: Cutting force sign range: Vibration symbol range: Temperature symbol range: Where F represents the cutting force signal, V represents the vibration signal, and T represents the temperature signal; , , These are the average values ​​of the cutting force signal, vibration signal, and temperature signal, respectively. , , These represent the standard deviations of the cutting force signal, vibration signal, and temperature signal, respectively.

[0008] In a further embodiment, the multidimensional feature extraction model is an autoencoder, which consists of an encoder and a decoder; The encoder maps the input symbol sequence x to a low-dimensional latent space representation z: in, This represents the encoder network function defined by parameter φ; The decoder reconstructs the low-dimensional latent space representation z into a sequence of reconstructed symbols. : in, This represents the decoder network function defined by the parameter ψ; The multidimensional fusion feature is the low-dimensional latent space representation z; When training the autoencoder, minimize the input symbol sequence x and the reconstructed symbol sequence. Reconstruction error loss function between: Where N is the sample size. For the i-th input symbol sequence, For the i-th reconstructed symbol sequence, This represents the square of the Euclidean distance.

[0009] In a further embodiment, the wear prediction model is a physical information neural network, and the output of the physical information neural network includes the current wear amount. and remaining useful life (RUL); The loss function of the physical information neural network Data-driven loss terms and physical constraint loss term Weighted composition: Where λ is the weighting coefficient of the physical constraint term; The data-driven loss term Represented as: in, Let be the true value of tool wear for the i-th sample. This represents the true remaining useful life of the i-th sample. The physical constraint loss term Based on the Usui wear rate model: Where W represents tool wear, t represents machining time, and a, b, and c are material correlation coefficients. Where is the cutting speed, T is the cutting temperature, and V is the vibration amplitude; The physical constraint loss term Represented as: in, Let be the rate of change of the predicted tool wear value for the i-th sample over time.

[0010] In a further embodiment, the future wear trend is represented by a symbolic transition probability matrix; Constructing the symbol transition probability matrix includes: The number of transitions between symbols at adjacent time points was counted according to three categories: cutting force symbols, vibration symbols, and temperature symbols. , where k∈{F,V,T} represent the three categories of cutting force, vibration and temperature respectively, i represents the symbol of the current time, and j represents the symbol of the next time; Calculate the probability of transitioning from symbol i to symbol j: in, The total number of times symbol i appears in category k; Based on the symbol combination at the current moment and the symbol transition probability matrix, the symbol combination at future moments is predicted, and the wear state corresponding to the symbol combination at the future moment is taken as the future wear trend.

[0011] In a further embodiment, the decision-making model is a reinforcement learning model, wherein the reinforcement learning model uses the current wear state and the future wear trend to construct the state space. The motion space is constituted by the adjustment amount of the cutting parameters. The reward function is composed of a comprehensive index of machining quality and tool life. ; The objective function of the reinforcement learning model is to maximize the cumulative discount reward. Where θ represents the policy network parameters. Let θ be the policy function defined by parameter θ, γ be the discount factor and 0 < γ ≤ 1, and t be the time step index. Indicating in strategy The expected value of the following; The adjustment amounts of the cutting parameters include the spindle speed adjustment amount Δn, the feed rate adjustment amount Δf, and the depth of cut adjustment amount. At least one of them.

[0012] A further embodiment, the reward function in, For processing quality evaluation function, For tool life evaluation function, Let α be the processing efficiency evaluation function, and let α, β, and η be the weight coefficients of the corresponding evaluation function, where α > 0, β > 0, and η > 0.

[0013] Further plans also include: Construct a digital twin of the tool, which is used to map the state of the physical tool in real time; The processing data after implementing the wear compensation strategy is fed back to the digital twin to update at least one of the multidimensional feature extraction model, the wear prediction model, and the decision model.

[0014] A machine learning-based CNC machining tool wear prediction and dynamic compensation system includes: The data acquisition module is used to acquire signal data from multiple sources of sensors during CNC machining and to preprocess the data to obtain standardized multi-source sensor data. The feature extraction module is used to construct a multi-dimensional feature extraction model based on machine learning, and to extract features from the standardized multi-source sensor data to obtain multi-dimensional fused features. The wear prediction module is used to construct a wear prediction model and predict the current wear state and future wear trend of the tool based on the multi-dimensional fusion features. The compensation decision module is used to dynamically generate a wear compensation strategy based on the current wear state and the future wear trend through a decision model; The instruction execution module is used to convert the wear compensation strategy into CNC instructions and execute them.

[0015] The present invention has the following beneficial effects: This invention automatically extracts spatiotemporal fusion features from the original signal through multi-source sensor signal fusion and autoencoder deep feature extraction, overcoming the deficiency of insufficient information in manual feature extraction; through an adaptive symbol mapping mechanism, the symbolization process can adapt to different machining conditions, solving the problem of decreased prediction accuracy of traditional models across working conditions, and realizing high-precision and stable prediction of tool wear state across working conditions.

[0016] This invention integrates a physical mechanism model of tool wear as a prior constraint into the training process of a physical information neural network. By penalizing prediction results that violate the physical laws of wear through a physical constraint loss term, the model output can both fit the observed data and follow the basic laws of cutting mechanics. This avoids the problem of prediction distortion under extreme working conditions in pure data-driven methods and significantly improves the interpretability and engineering credibility of the model.

[0017] This invention mines the temporal patterns of wear evolution through symbolic transition probability matrix, enabling early perception of wear trends; it uses a reinforcement learning decision model to combine the current wear state with future trends as the basis for decision-making, achieving forward-looking optimization of compensation strategies; and it constructs a closed-loop feedback architecture through a digital twin, enabling the system to continuously evolve using actual processing results, realizing intelligent compensation from passive response to proactive foresight. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the CNC machining tool wear prediction and dynamic compensation method based on machine learning in an embodiment of the present invention.

[0019] Figure 2 This is a schematic diagram of the adaptive symbol mapping process in an embodiment of the present invention.

[0020] Figure 3 This is a schematic diagram of the physical information neural network in an embodiment of the present invention.

[0021] Figure 4 This is a schematic diagram illustrating the construction and trend prediction process of the symbol transition probability matrix in an embodiment of the present invention.

[0022] Figure 5 This is a schematic diagram of the closed-loop feedback architecture of the digital twin in an embodiment of the present invention. Detailed Implementation

[0023] The principles and features of the present invention are described below with reference to the accompanying drawings. The examples given are only for explaining the present invention and are not intended to limit the scope of the present invention.

[0024] Example 1 This embodiment provides a method for predicting and dynamically compensating for CNC machining tool wear based on machine learning.

[0025] like Figure 1 As shown, the method in this embodiment includes the following steps: Step S1: Collect signal data from multiple sources during CNC machining and preprocess the data to obtain standardized multi-source sensor data.

[0026] In this embodiment, multiple high-precision sensors are installed at key locations on the CNC machining equipment. A cutting force sensor is installed at the tool clamping area to monitor force changes in real time during the cutting process; a vibration sensor is fixed near the tool spindle to capture tool vibration signals; and a temperature sensor is placed near the tool to measure the temperature in the cutting area. Each sensor is connected to the CNC system via a data acquisition card, acquiring data in real time at a sampling frequency of 100Hz.

[0027] The sensor collects signal data including cutting force signal F, vibration signal V, and temperature signal T.

[0028] Preprocessing includes two sub-steps: filtering and denoising, and normalization.

[0029] Filtering and noise reduction: A low-pass filter is used to filter the cutting force signal F, vibration signal V, and temperature signal T respectively to remove high-frequency noise. The cutoff frequency of the low-pass filter is set to 10Hz.

[0030] Normalization: The min-max normalization method is used to normalize each signal data to the range [0,1]. , , in, , , These are the normalized cutting force signal, vibration signal, and temperature signal, respectively. , , These are the minimum values ​​of the corresponding signals; , , These represent the maximum values ​​of the corresponding signals.

[0031] Standardized data are obtained after normalization.

[0032] like Figure 2 As shown, this embodiment also includes the step of converting standardized data into a symbol sequence.

[0033] Calculate the mean and standard deviation of standardized data: The sign interval is divided based on the mean and standard deviation, as described above. The boundaries of the sign interval are dynamically adjusted according to real-time data, exhibiting adaptability.

[0034] At each sampling moment, the cutting force signal, vibration signal, and temperature signal are converted into corresponding symbol values ​​(0-8). The three symbol values ​​are combined to form a symbol sequence unit. Symbol sequence units from multiple consecutive moments are arranged in chronological order to form symbol sequence data.

[0035] For example, if at a certain moment the cutting force signal falls in the interval of symbol "1", the vibration signal falls in the interval of symbol "4", and the temperature signal falls in the interval of symbol "7", then the symbol sequence unit at that moment is "147".

[0036] Step S2: Construct a multi-dimensional feature extraction model based on machine learning to extract features from symbol sequence data and obtain multi-dimensional fused features.

[0037] In this embodiment, the multidimensional feature extraction model uses an autoencoder.

[0038] encoder Map the input symbol sequence x to a low-dimensional latent space representation z: decoder Reconstruct the low-dimensional latent space representation z into a sequence of reconstructed symbols. : The encoder consists of three fully connected layers with 128, 64, and 32 neurons in each layer, respectively; the decoder also consists of three fully connected layers with 32, 64, and 128 neurons in each layer. The encoder's output—the low-dimensional latent space representation z—is the multidimensional fused feature.

[0039] When training the autoencoder, minimize the reconstruction error loss function: Where N is the sample size. For the i-th input symbol sequence, Let i be the sequence of reconstructed symbols.

[0040] Training uses the backpropagation algorithm and the Adam optimizer to update the network parameters φ and ψ.

[0041] Step S3: Construct a wear prediction model to predict the current wear status and future wear trend of the tool based on multi-dimensional fusion features.

[0042] like Figure 3 As shown, the wear prediction model in this embodiment uses a physical information neural network (PINN).

[0043] The input to the physical information neural network is the multidimensional fusion feature z output from step S2, and the output includes the current wear level. And remaining useful life (RUL).

[0044] The loss function of a physical information neural network is a weighted sum of a data-driven loss term and a physical constraint loss term: Wherein, λ is the weighting coefficient of the physical constraint term, and in this embodiment, λ=0.1.

[0045] Data-driven loss term: The physical constraint loss term is based on the Usui wear rate model: The training process is divided into two stages: first, the network is pre-trained on a historical dataset to minimize the data-driven loss term; then, fine-tuning is performed by simultaneously minimizing the data-driven loss term and the physical constraint loss term.

[0046] Future wear trends are characterized by a sign transition probability matrix. For example... Figure 4 As shown, the method for constructing the symbol transition probability matrix is ​​as follows: The number of transitions between adjacent time points is counted for each of the three symbol classes. , k∈{F,V,T}).

[0047] Calculate the transition probability: Construct the cutting force sign transition probability matrix Vibration symbol transition probability matrix and temperature sign transition probability matrix .

[0048] The symbol transition probability matrix is ​​updated in real time—every time a new symbol sequence data is collected, the corresponding number of transitions and transition probabilities are updated.

[0049] Based on the current symbol combination and symbol transition probability matrix, the symbol combination at future times is predicted, thereby determining the future wear trend. The correspondence between wear states and symbol combinations is shown in Table 1.

[0050] Table 1 Correspondence between Wear Conditions and Symbol Combinations

[0051] Step S4: Based on the current wear status and future wear trend, dynamically generate a wear compensation strategy through a decision model.

[0052] In this embodiment, the decision-making model adopts a reinforcement learning model.

[0053] state space It consists of the current wear condition and the future wear trend.

[0054] Action space The cutting parameter adjustment amounts include spindle speed adjustment Δn, feed rate adjustment Δf, and depth of cut adjustment. .

[0055] Reward function: in, For processing quality evaluation function, For tool life evaluation function, Let α = 0.4, β = 0.3, and η = 0.3 be the processing efficiency evaluation function.

[0056] The objective function of a reinforcement learning model is to maximize the cumulative discount reward. Wherein, γ=0.95.

[0057] This embodiment uses the Proximal Policy Optimization (PPO) algorithm to train the reinforcement learning model. Both the policy network and the value network employ three fully connected neural networks, with 256 neurons in each hidden layer.

[0058] After training is complete, the real-time status will be displayed. Input the policy network, and the policy network outputs the optimal action. A wear compensation strategy is generated based on the actions.

[0059] Wear compensation strategies include: If the current wear condition is slight wear and the future wear trend is also slight wear: keep the current cutting parameters unchanged; If the current wear condition is slight wear and the future wear trend is moderate wear: reduce the cutting speed and increase the feed rate; If the current wear condition is moderate and the future wear trend is severe wear: Replace the tool; If the current wear condition is severe wear: Replace the tool; If the current wear condition is abnormal: stop processing.

[0060] Step S5: Convert the wear compensation strategy into CNC instructions and execute them.

[0061] The strategy for adjusting cutting parameters is to update the cutting parameters in real time through the parameter setting module of the CNC system; the strategy for changing tools is to be executed through the tool management module of the CNC system; and the strategy for stopping machining is to be executed through the emergency stop module of the CNC system.

[0062] After the CNC command is executed, the new machining data is collected by the sensor and fed back to the system for model updating.

[0063] like Figure 5 As shown, this embodiment also includes: constructing a digital twin of the tool to map the physical tool's state in real time. The digital twin receives signals from physical sensors to update its own state parameters and receives the simulation results of the wear compensation strategy in a virtual environment. The actual machining data after executing the compensation strategy is fed back to the digital twin, triggering online updates of the multi-dimensional feature extraction model, wear prediction model, and decision model.

[0064] Example 2 The difference between this embodiment and Embodiment 1 is that the multidimensional feature extraction model adopts a CNN-BiGRU hybrid model.

[0065] The CNN part extracts local spatial features of symbol sequences and consists of two convolutional layers and two pooling layers stacked together. The first convolutional layer contains 32 convolutional kernels (3×3), and the second convolutional layer contains 64 convolutional kernels (3×3). Each layer is followed by a ReLU activation function and a max pooling layer (2×2).

[0066] The BiGRU part extracts the temporal dependencies of symbol sequences and consists of two layers of bidirectional gated recurrent units stacked together, with 128 neurons in each hidden layer. The feature map output by the CNN is expanded step by step and then input into the BiGRU. The hidden state of the last time step of the BiGRU is used as the multidimensional fusion feature z.

[0067] The training method is the same as in Example 1, which is end-to-end training.

[0068] The remaining steps are the same as in Example 1.

[0069] Example 3 The difference between this embodiment and Embodiment 1 is that the wear prediction model is adaptively updated across operating conditions through transfer learning.

[0070] First, a physical information neural network is pre-trained on the source domain dataset to obtain the pre-trained model parameters.

[0071] When the processing conditions change, a small amount of target domain data is collected to fine-tune the pre-trained model. During fine-tuning, the parameters of the first two layers of the model are frozen, and only the parameters of the last two layers are updated.

[0072] The fine-tuning loss function is: The remaining steps are the same as in Example 1.

[0073] Example 4 The difference between this embodiment and Embodiment 1 is that the reinforcement learning model uses the soft actor-critic (SAC) algorithm.

[0074] The objective function of the SAC algorithm is: Where τ is the temperature parameter, For strategy In state Entropy below.

[0075] The SAC algorithm maximizes both cumulative reward and policy entropy, encouraging exploration and avoiding getting trapped in local optima.

[0076] The remaining steps are the same as in Example 1.

[0077] Technical effect verification Experiments were conducted using a CNC lathe to machine 45# steel workpieces and a carbide cutting tool. The method of this invention (Example 1) and a traditional method (a fixed threshold compensation method based on an empirical model) were compared.

[0078] Experimental results show that: (1) The root mean square error (RMSE) of the wear amount prediction method of the present invention is 0.012 mm, while the RMSE of the traditional method is 0.035 mm, which improves the prediction accuracy by about 65.7%.

[0079] (2) When the cutting speed is adjusted from 150m / min to 200m / min, the prediction accuracy of the method of the present invention decreases by 3.2%, and the prediction accuracy of the traditional method decreases by 28.6%. The method of the present invention has a stronger ability to generalize across working conditions.

[0080] (3) The method of the present invention extends the average tool life by 22.3% and improves the surface roughness of the machined surface. The surface roughness is stabilized below 0.8μm; the average tool life extension of traditional methods is only 8.1%, and the surface roughness is improved. The fluctuation range is between 0.6 μm and 1.5 μm.

[0081] Summary of working principles: Cutting force, vibration, and temperature signals during the cutting process are acquired in real time using multi-source sensors. After filtering, denoising, and normalization, the signals are mapped from continuous physical quantities to discrete symbols using the real-time calculated mean and standard deviation as boundaries, forming a symbol sequence. This mapping mechanism makes the symbolization process insensitive to absolute changes in signal amplitude but sensitive to relative distribution shifts, thus maintaining consistency in symbol representation under different machining conditions.

[0082] The symbol sequence is first input into an autoencoder, which compresses and extracts a low-dimensional latent space representation as a multi-dimensional fusion feature. This feature is then input into a physical information neural network. The network's loss function consists of two terms: a data-driven loss term measures the deviation between the predicted and actual values, and a physical constraint loss term applies physical constraints to the prediction results based on the Usui wear rate model. The two terms are weighted and jointly optimized to ensure that the network output both fits the data and follows the physical mechanism. The network output represents the current wear level, while the symbol transition probability matrix, based on historical symbol sequences, statistically analyzes the transition patterns between symbols at adjacent time points. This predicts future symbol combinations and their corresponding wear states, enabling early perception of future wear trends.

[0083] The current wear state and future wear trend form the state space, which is then input into the reinforcement learning model. The policy network maps the state to the optimal action—the adjustment of cutting parameters. The reward function integrates three indicators: machining quality, tool life, and machining efficiency, guiding the policy network to converge towards the optimization direction of extending tool life and ensuring machining quality.

[0084] After being parsed, the motion commands are converted into cutting parameter commands that the CNC system can execute. The actual machining effect after execution is transmitted back to the digital twin via sensors. The deviation is compared with the simulation prediction results of the digital twin, triggering online parameter updates of the intelligent model, thus forming a continuously evolving closed-loop system.

[0085] The core of this invention lies in expanding tool wear prediction and compensation from a simple data fitting problem to a time-series perception, forward-looking decision-making, and closed-loop evolution problem under the constraints of physical laws. Through the organic combination of symbolic representation, physical information neural networks, and reinforcement learning, it achieves improved prediction accuracy, enhanced cross-working-condition generalization ability, and forward-looking optimization of compensation strategies.

[0086] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting and dynamically compensating for CNC machining tool wear based on machine learning, characterized in that, Includes the following steps: S1. Collect signal data from multiple sources during CNC machining and preprocess the data to obtain standardized multi-source sensor data. S2. Construct a multi-dimensional feature extraction model based on machine learning to extract features from the standardized multi-source sensor data and obtain multi-dimensional fused features; S3. Construct a wear prediction model to predict the current wear state and future wear trend of the tool based on the multi-dimensional fusion features; S4. Based on the current wear status and the future wear trend, a wear compensation strategy is dynamically generated through a decision model; S5. Convert the wear compensation strategy into CNC instructions and execute them.

2. The method for predicting and dynamically compensating CNC machining tool wear based on machine learning according to claim 1, characterized in that, The process of acquiring signal data from multiple sources during CNC machining and preprocessing it to obtain standardized multi-source sensor data includes: Acquire cutting force signals, vibration signals, and temperature signals; The collected signals are filtered, denoised, and normalized to obtain standardized data; Calculate the statistical characteristics of the standardized data, including the mean and standard deviation; Based on the statistical characteristics, the symbol intervals are divided, and the standardized data is converted into symbol sequences.

3. The method for predicting and dynamically compensating CNC machining tool wear based on machine learning according to claim 1, characterized in that, Based on the aforementioned statistical characteristics, the data is divided into symbol intervals, and the standardized data is converted into symbol sequences, including: The rule for dividing the symbol interval is as follows: Cutting force sign range: Vibration symbol range: Temperature symbol range: Where F represents the cutting force signal, V represents the vibration signal, and T represents the temperature signal; , , These are the average values ​​of the cutting force signal, vibration signal, and temperature signal, respectively. , , These represent the standard deviations of the cutting force signal, vibration signal, and temperature signal, respectively.

4. The method for predicting and dynamically compensating CNC machining tool wear based on machine learning according to claim 1, characterized in that, The multidimensional feature extraction model is an autoencoder, which consists of an encoder and a decoder. The encoder maps the input symbol sequence x to a low-dimensional latent space representation z: in, This represents the encoder network function defined by parameter φ; The decoder reconstructs the low-dimensional latent space representation z into a sequence of reconstructed symbols. : in, This represents the decoder network function defined by the parameter ψ; The multidimensional fusion feature is the low-dimensional latent space representation z; When training the autoencoder, minimize the input symbol sequence x and the reconstructed symbol sequence. Reconstruction error loss function between: Where N is the sample size. For the i-th input symbol sequence, For the i-th reconstructed symbol sequence, This represents the square of the Euclidean distance.

5. The method for predicting and dynamically compensating CNC machining tool wear based on machine learning according to claim 1, characterized in that, The wear prediction model is a physical information neural network, and the output of the physical information neural network includes the current wear amount. and remaining useful life (RUL); The loss function of the physical information neural network Data-driven loss terms and physical constraint loss term Weighted composition: Where λ is the weighting coefficient of the physical constraint term; The data-driven loss term Represented as: in, Let be the true value of tool wear for the i-th sample. This represents the true remaining useful life of the i-th sample. The physical constraint loss term Based on the Usui wear rate model: Where W represents tool wear, t represents machining time, and a, b, and c are material correlation coefficients. Where is the cutting speed, T is the cutting temperature, and V is the vibration amplitude; The physical constraint loss term Represented as: in, Let be the rate of change of the predicted tool wear value for the i-th sample over time.

6. The method for predicting and dynamically compensating CNC machining tool wear based on machine learning according to claim 1, characterized in that, The future wear trend is represented by a symbolic transition probability matrix; Constructing the symbol transition probability matrix includes: The number of transitions between symbols at adjacent time points was counted according to three categories: cutting force symbols, vibration symbols, and temperature symbols. , where k∈{F,V,T} represent the three categories of cutting force, vibration and temperature respectively, i represents the symbol of the current time, and j represents the symbol of the next time; Calculate the probability of transitioning from symbol i to symbol j: in, The total number of times symbol i appears in category k; Based on the symbol combination at the current moment and the symbol transition probability matrix, the symbol combination at future moments is predicted, and the wear state corresponding to the symbol combination at the future moment is taken as the future wear trend.

7. The method for predicting and dynamically compensating CNC machining tool wear based on machine learning according to claim 1, characterized in that, The decision-making model is a reinforcement learning model, which uses the current wear state and the future wear trend to form the state space. The motion space is constituted by the adjustment amount of the cutting parameters. The reward function is composed of a comprehensive index of machining quality and tool life. ; The objective function of the reinforcement learning model is to maximize the cumulative discount reward. Where θ represents the policy network parameters. Let θ be the policy function defined by parameter θ, γ be the discount factor and 0 < γ ≤ 1, and t be the time step index. Indicating in strategy The expected value of the following; The adjustment amounts of the cutting parameters include the spindle speed adjustment amount Δn, the feed rate adjustment amount Δf, and the depth of cut adjustment amount. At least one of them.

8. The method for predicting and dynamically compensating CNC machining tool wear based on machine learning according to claim 7, characterized in that, The reward function: in, For processing quality evaluation function, For tool life evaluation function, Let α be the processing efficiency evaluation function, and let α, β, and η be the weight coefficients of the corresponding evaluation function, where α > 0, β > 0, and η > 0.

9. The method for predicting and dynamically compensating CNC machining tool wear based on machine learning according to claim 5, characterized in that, Also includes: Construct a digital twin of the tool, which is used to map the state of the physical tool in real time; The processing data after implementing the wear compensation strategy is fed back to the digital twin to update at least one of the multidimensional feature extraction model, the wear prediction model, and the decision model.

Images