Method and system for online compensation of wear and life prediction of a gear forming tool

Abstract

The present application relates to a kind of gear forming tool wear online compensation and life prediction method and system, by constructing the multi-dimensional digital twin mechanism model of gear forming tool, real-time acquisition multi-source sensing signal, extract dominant feature and with implicit feature and mechanism parameter coding vector fusion;Again based on the fusion feature construction graph neural network classification model, realize the online accurate discrimination of tool wear grade;Using reinforcement learning algorithm trains optimal compensation strategy, real-time adjustment processing parameter, forms " prediction-compensation-feedback " closed loop control;Establish the residual life prediction method of multi-model integration, and realize intelligent tool change decision optimization based on prediction result.The present application realizes the integrated management of gear forming tool wear feature extraction, condition monitoring, dynamic compensation and life prediction, improves the intelligent level of gear machining process and tool life.

Description

Technical Field

[0001] This invention belongs to the field of intelligent manufacturing and digital twin technology, specifically relating to a method and system for online compensation and life prediction of gear forming tool wear, which is applied to the prediction of gear forming tool life, condition detection and process control in high-precision gear machining. Background Technology

[0002] With the development of high-end equipment manufacturing and precision transmission systems, the application of gear forming machining in aerospace, automotive power systems, industrial robots and other fields is constantly expanding. As a key actuator in the gear machining process, the condition monitoring of gear forming tools is particularly important.

[0003] Currently, the main methods used for condition monitoring of gear forming tools are offline detection during shutdown, online indirect monitoring, and data-driven methods. However, all three methods have certain limitations in actual gear machining environments. Offline detection requires stopping the machine for sampling, which has a long cycle, significantly impacts production cycle time, and is difficult to reflect the dynamic evolution of wear. Online indirect monitoring is sensitive to changes in operating conditions, and the signal is easily affected by noise, load fluctuations, and thermal drift. Fixed thresholds are prone to failure, and early fine wear characteristics are difficult to reliably separate. Pure data-driven models have weak mechanistic constraints, are easily affected by distribution drift, have limited generalization ability, and are difficult to directly integrate with online compensation closed-loop control.

[0004] In recent years, with the rapid development of digital twin technology, its application to gear forming tool machining processes to achieve accurate modeling and intelligent prediction of the tool's operating status has become a research hotspot. However, current technologies still suffer from the following technical problems: insufficient wear feature extraction, weak accuracy and generalization ability of online monitoring, lack of active compensation closed-loop control mechanisms, and low life prediction accuracy that is disconnected from decision-making. These problems severely restrict the consistency between the virtual and real worlds, the adaptability across working conditions, and the online practicality of digital twin models. This is the deficiency of existing technologies.

[0005] In view of this, it is very necessary to provide a method and system for online compensation of wear and life prediction of gear forming tools in order to solve the above-mentioned defects in the prior art. Summary of the Invention

[0006] The purpose of this invention is to address the technical problems of insufficient wear feature extraction, weak accuracy and generalization ability of online monitoring, lack of active compensation closed-loop control mechanism, low life prediction accuracy and disconnection from decision-making in the existing technology. This invention provides a method and system for online wear compensation and life prediction of gear forming tools. By deeply coupling mechanism parameters and data models, an integrated management system covering implicit feature extraction, online status monitoring, dynamic compensation control and remaining life prediction is constructed, thereby solving the aforementioned technical problems in the existing technology.

[0007] To achieve the above objectives, the present invention provides the following technical solution: In a first aspect, the present invention provides a method for online wear compensation and life prediction of gear forming tools, specifically including the following steps: Step S1, the step of extracting the hidden features of wear on gear forming tools, in which: Utilizing the parameter space of the constructed digital twin mechanism model, multi-source sensor signals are simultaneously acquired. These signals are preprocessed, and dominant features are extracted through multi-dimensional feature extraction. Dimensionality reduction is then applied to the dominant features to obtain core features. A variational autoencoder is constructed to encode the core features into latent features. The parameters of the digital twin mechanism model are mapped to mechanism parameter encoding vectors and fused with the core and latent features to obtain fused features. A physical consistency loss is defined and combined with the variational autoencoder loss to form a total loss function for joint optimization. Step S2, the online monitoring and wear level determination of gear forming tool condition, in which: Online monitoring of the gear forming tool status is performed based on the fused features obtained in step S1; multiple levels are defined and labeled according to the wear amount on the tool's flank face; a feature relationship graph is constructed based on the feature correlation of the fused features; a graph neural network classifier is constructed to determine the wear level, and the global feature vector is aggregated through graph convolution to output the wear level probability distribution; lightweight model deployment is achieved through knowledge distillation and model quantization; and the wear status of the gear forming tool is determined by real-time data acquisition through online monitoring of the gear forming tool status. Step S3, the dynamic compensation strategy for gear forming tools, in which: A dynamic compensation mechanism is established based on the tool wear state identified in step S2. The machining parameter adjustment amount is defined as the control variable. A learning simulation environment is constructed using a digital twin mechanism model. A multi-objective reward function is designed. A deep deterministic strategy gradient algorithm is used to train the intelligent agent to make the optimal action decision based on the current state. Parameter adjustment instructions are output in real time and feedback is collected to form a "prediction-compensation-feedback" closed loop. Safety constraints and protection mechanisms are set. Step S4, the prediction of remaining tool life and intelligent tool change decision for gear forming, in which: Establish the correlation between wear and remaining life; construct prediction models based on long short-term memory networks, Transformer, and Hidden Markov-Bayes networks respectively; use the prediction results of multi-model integration and evaluate the uncertainty; set safety margins based on the prediction results to determine the timing of single-tool tool changes; establish a multi-objective optimization model for multi-tool scenarios, and use dynamic programming to solve the optimal tool change plan.

[0008] Furthermore, the extraction of the hidden wear features of the gear forming tool mentioned in step S1 specifically includes the following steps: Step S101, the step of multi-source sensor signal acquisition and explicit feature extraction, in which: Utilizing the parameter space of the established digital twin mechanism model, multi-source sensor signals are simultaneously acquired during the gear forming process, and a multi-source sensor signal set is constructed:

[0009] in, for A collection of multi-source sensor signals at any given time. This is the cutting force signal. It is a vibration acceleration signal. For acoustic emission signals, For temperature signals, Main spindle power signal; During the acquisition of multi-source sensor signals, the corresponding machining parameters are recorded simultaneously; the machining parameters include cutting speed, feed per tooth, and depth of cut. Furthermore, the acquired multi-source sensor signals are preprocessed, and a Hampel filter is used to remove outliers from the multi-source sensor signals. The Hampel filter can denoise and preserve fidelity, ensuring the accuracy and stability of subsequent wear feature extraction and modeling. The specific expression is as follows:

[0010] in, The signal before filtering. This is the filtered signal. This represents the absolute deviation of the median. This is the threshold coefficient; This represents the median of all samples within the current sliding window. Furthermore, multi-dimensional features are extracted from the preprocessed signal. These multi-dimensional features include time-domain statistical features, frequency-domain energy features, and time-frequency domain features based on wavelet packet decomposition. The time-domain features include mean, standard deviation, root mean square value, kurtosis, and peak value. The frequency-domain features are obtained by obtaining total power, center frequency, and bandwidth through fast Fourier transform. The time-frequency domain features are extracted by extracting the energy of each frequency band using wavelet packet decomposition. This achieves a comprehensive perception of the tool wear state from statistics, energy, to transient characteristics, providing input features with high discriminative power, low redundancy, and strong robustness for subsequent modeling and prediction. Furthermore, all features extracted from the multi-dimensional feature extraction are concatenated to form an explicit feature matrix. Principal component analysis is then used to perform dimensionality reduction on the explicit feature matrix, compressing high-dimensional, redundant, and correlated explicit features into low-dimensional, decorrelational core features. This provides efficient, robust, and low-noise input for subsequent latent feature extraction and modeling. The specific expression is as follows:

[0011] in, The principal component analysis transformation matrix is ​​used to retain principal components with a cumulative variance contribution rate of 90%-95%. It is an explicit characteristic matrix. The core feature matrix after dimensionality reduction; After removing outliers using Hampel filtering in step S101, the time-domain, frequency-domain, and wavelet packet time-frequency-domain features are extracted and concatenated into an explicit feature matrix. Principal component analysis is used to reduce the dimensionality and retain 90%–95% of the cumulative variance to obtain the core matrix. In step S101, the original high-dimensional features are compressed to low dimensions with an information loss of ≤10%, which reduces the input dimension and computational load of the subsequent VAE network while ensuring the integrity of early wear information.

[0012] Step S102, the step of extracting latent features and mapping mechanistic parameters based on variational autoencoders, in which: A variational autoencoder model is constructed to extract latent wear features. The encoder maps core features to the latent space, compressing shallow core features into low-dimensional, probabilistic, and information-dense latent features, revealing the deep evolutionary laws of tool wear, and endowing the model with uncertainty modeling and robust generalization capabilities; the specific expression is as follows:

[0013]

[0014]

[0015] in, Using latent variable vectors as latent features, For encoder parameters, and The mapping function of the neural network outputs the mean and standard deviation of the latent variables respectively. For element-wise multiplication, It is standard normally distributed random noise; Let be the mean vector of the latent space distribution. is the standard deviation vector of the latent space distribution; This is the core feature matrix.

[0016] Furthermore, the decoder reconstructs the original features and defines the variational autoencoder loss function; this makes the latent features z both compact and complete, as well as regular and sampleable, laying the foundation for subsequent high-precision, interpretable, and uncertain tool state modeling. The specific expression is as follows:

[0017] in, The reconstruction loss measures the error between the reconstructed features and the original features. The KL divergence loss is used, and the latent variable distribution is constrained to approximate a standard normal distribution. This is the balance coefficient; It is the total loss function of the variational autoencoder, which is the objective function that needs to be minimized when training a VAE.

[0018] Furthermore, by mapping the parameter space of the digital twin mechanism model to the feature space, the set of parameter spaces of the digital twin mechanism model is represented as:

[0019] in, For geometric parameter vectors, For physical property parameter vectors, For behavioral feature parameter vectors, For the degeneracy law function, The parameter space for the twinning mechanism model; Furthermore, the mechanistic parameters are encoded using a multilayer perceptron, with the specific expression as follows:

[0020] in, Encode the mechanism parameters into a vector. For multilayer perceptrons used for encoding mechanistic parameters, The parameter space for the twinning mechanism model; Furthermore, the core features, latent features, and mechanistic parameter encoding vectors are fused together:

[0021] in, The feature vectors are fused, incorporating explicit physical quantities, deep implicit patterns, and mechanistic knowledge. As the core feature, It is a latent feature. Encode the mechanism parameters into a vector; Step S102 outputs the mean and standard deviation of the latent variables through a variational autoencoder, and obtains the latent variables through the reparameter trick; then maps the digital twin mechanism parameter space into a mechanism parameter encoding vector through a multilayer perceptron; finally, the vectors are spliced ​​together to form a fusion feature vector; the "explicit-latent-mechanistic" fusion feature vector obtained in step S102 achieves a compact expression of the deep wear pattern and has a probability distribution form, providing a low-dimensional input that conforms to physical laws for subsequent graph neural network classification and Bayesian update.

[0022] Step S103, the step of physical consistency loss constraint and joint optimization, in which: The simulation output of tool wear state generated by the digital twin mechanism model is used as a supervision signal, and the obtained wear state is mapped to the latent space through an encoder to obtain the corresponding latent variables. The physical consistency loss is defined as follows:

[0023] in, To measure the difference between latent variables extracted through data-driven methods and latent variables from mechanism simulation, we consider the physical consistency constraint error. Latent variables extracted for data-driven models The latent variables mapped to the latent space of the mechanistic model simulation output serve as monitoring signals to ensure that the output of the data-driven model conforms to physical laws. Furthermore, a total loss function is established for joint optimization:

[0024] in, This is the physical consistency weighting coefficient, used to balance the variational autoencoder loss and the physical consistency constraint; The loss function of the variational autoencoder is... For physical consistency constraint error, This is the total loss function; Furthermore, the gradient descent method is used to simultaneously optimize the parameters of the variational autoencoder and the mechanistic parameter mapping network, so that the extracted latent features can accurately represent the data distribution and conform to physical laws, thereby enhancing the interpretability and generalization ability of the features. After training, the latent variables obtained by the encoder for each sample are the latent features of gear forming tool wear, which are used for subsequent state monitoring and life prediction.

[0025] The online monitoring of gear forming tool condition and wear level determination step S2 further includes the following steps: Step S201, the step of defining wear level and constructing monitoring feature input, in which: The fusion features extracted in step S1 are used as input to the state monitoring model to construct the monitoring feature input vector, as shown in the following expression:

[0026] in, The input for state monitoring features includes the encoding of core features, latent features, and mechanistic parameters; Features of fusion; Furthermore, based on the wear amount VB on the tool's flank face, a wear level classification standard is established, with the specific expression as follows:

[0027] in, The total number of wear levels. The wear threshold for each level is determined based on the tool type and machining requirements; Furthermore, samples in the historical dataset are labeled with the corresponding wear level categories based on the offline measured VB values, and a labeled training dataset is constructed for model training. Step S202, the step of constructing a graph structure based on feature correlation, in which: Using each dimension of the fused feature vector as a graph node, a feature relationship graph is constructed based on the correlation between features, and the Pearson correlation coefficient matrix between features is calculated. The specific expression is as follows:

[0028] in, Features and characteristics The correlation coefficient between them For covariance, and These are the standard deviations of the two features; Furthermore, the adjacency matrix of the graph is constructed based on the correlation coefficient matrix, as shown in the following expression:

[0029] in, It is an adjacency matrix; The correlation threshold is used to retain only connections between feature pairs with strong correlations. Features and characteristics The correlation coefficient between them; Furthermore, the adjacency matrix is ​​normalized, as shown in the following expression:

[0030] in, For the normalized adjacency matrix, For degree matrix, Represents a node The degree; Step S203, the step of constructing the graph neural network classification model, in which: A graph neural network classifier is constructed to determine wear level. The node feature update formula for the graph convolutional layer is as follows:

[0031] in, For the first The node feature matrix of the layer These are the initial node features. For the first The learnable weight matrix of the layer, For activation functions; For the normalized adjacency matrix, Furthermore, after multiple layers of graph convolution, graph pooling is used to aggregate the graph representation into a global feature vector. The probability distribution of wear level is then output through fully connected layers and softmax layers, as shown in the following expression:

[0032] in, For the predicted wear level probability distribution, The total number of convolutional layers in the graph. For graph pooling operations, It is a multilayer perceptron; For the first The node feature matrix of the layer; Softmax is the Softmax layer, a commonly used classification layer used to map input features to a probability distribution; Furthermore, a classification loss function is defined for model training, with the following specific expression:

[0033] in, The number of training samples, For the sample True label one-hot encoding, To predict probabilities, the cross-entropy loss function is used to optimize the model parameters; The classification loss function; Step S204, the steps of online monitoring implementation and lightweight model deployment, in which: During the online monitoring phase, at fixed time intervals Slide to acquire sensor data, with each sampling window having a length of [length missing]. The data within the window is processed according to step S1 to extract features. After obtaining the fused features, they are input into the graph neural network classifier, which outputs the wear level prediction result at the current moment. Furthermore, the confidence level of the prediction result is calculated, with the specific expression as follows:

[0034] in, To predict confidence levels, when the confidence level is below a set threshold... In such cases, a manual review mechanism may be triggered or the monitoring frequency may be increased. This represents the probability distribution of the predicted wear level. Furthermore, to meet the real-time requirements of industrial sites, knowledge distillation technology is used to compress the graph neural network into a lightweight student model. The student model is trained using soft labels from the teacher model, as shown in the following expression:

[0035] in, For distillation losses, KL divergence; Softmax is the Softmax classification layer; and The output logits of the student model and the teacher model are respectively. For temperature parameters, Furthermore, the lightweight model is quantized by converting floating-point parameters into INT8 or INT16 format to reduce model memory usage and computational load. It is then deployed to industrial control computers or edge computing devices to ensure that the single inference time meets the real-time monitoring requirements and to achieve online and accurate identification of tool wear status.

[0036] The dynamic compensation strategy for gear forming tools described in step S3 specifically includes the following steps: Step S301, the design and state-action space definition of the compensation control system, in which: Based on the wear state of the gear forming tool identified in step S2, a dynamic compensation control system is constructed, and the control variable is defined as the adjustment amount of the machining parameter, with the specific expression as follows:

[0037] in, for Action vector at time step, This is the amount of adjustment for the cutting speed. Adjustment amount for feed per tooth. This is the depth of cut adjustment amount. This is the adjustment amount for the radius compensation value of the gear forming tool; Furthermore, the adjustment range of each control variable is set to construct the motion space, as shown in the following expression:

[0038] in, for Action vector at any given moment; For the action space, and These are the minimum and maximum adjustment ranges for each parameter, determined based on machine tool performance and process requirements; Furthermore, the state space of the reinforcement learning algorithm is defined, integrating the wear state of the gear forming tool, machining parameters, and feature information. The specific expression is as follows:

[0039] in, For state vectors, The current wear level, This refers to the amount of wear. This is the current processing parameter vector. To fuse feature vectors, This represents the increment of processing error. Step S302, the step of constructing a reinforcement learning simulation environment based on a mechanism model, in which: A simulation environment for reinforcement learning training is constructed using the digital twin mechanism model in S1. The mechanism model calculates the processing result at the next moment based on the current state and action, as shown in the following expression:

[0040] in, For the predicted cutting force, Due to machining error, This represents the increase in wear. For mechanism simulation model, for The state vector at time t, for Action vector at any given moment; Furthermore, a multi-objective reward function is designed, comprehensively considering machining quality, gear forming tool life, and machining efficiency. The specific expression is as follows:

[0041] in, For instant rewards, These are the weighting coefficients. , For wear amount error, As a processing efficiency indicator, To constrain violations and penalties; Furthermore, the constraint penalty term is defined, with the following specific expression:

[0042] in, To constrain penalties, The penalty coefficient is... , , These are the safety thresholds for machining error, cutting force, and wear, respectively. , , These represent the machining error, cutting force, and wear amount at time t, respectively. This indicates that under other circumstances, the constraint penalty term is 0; Step S303, the deep deterministic policy gradient algorithm and policy network training steps, in which: The intelligent agent is trained using a deep deterministic policy gradient algorithm, and an Actor-Critic network architecture is constructed. The Actor network outputs a deterministic policy, the specific expression of which is as follows:

[0043] in, The action taken by the intelligent agent at time step t; For policy networks, For network parameters, For state vectors, To explore noise, and to enhance exploration capabilities during the training phase; Furthermore, the Critic network evaluates the value of state-action pairs, with the following specific expression:

[0044] in, For value networks, For network parameters, For state vectors, The action taken by the intelligent agent at time step t; As a discount factor, and For target network parameters; The value of state-action pairs evaluated for the Critic network; For immediate reward at time step t; The next state output by the target Actor network The optimal action under the given conditions; For the target Critic network; Expressing expectations; Furthermore, the Actor network parameters are updated via policy gradients, as shown in the following expression:

[0045] in, Let be the objective function of the policy network, and optimize the policy by maximizing the cumulative reward. These are the parameters of the policy network (Actor network); The value of state-action pairs evaluated for the Critic network; These are the parameters of the Critic network; The action output by the policy network; Output for Critic network The gradient of action a; Output for policy network The gradient with respect to the parameter θ; Expressing expectations; The key gradient expression is denoted as the gradient of the policy network objective function J(θ) with respect to the parameter θ.

[0046] Furthermore, the Critic network parameters are updated using the temporal difference error, as shown in the following expression:

[0047] in, This is the loss function of the Critic network, used to measure the difference between the predicted value and the target value; Expressing expectations; The value of state-action pairs evaluated for the Critic network; To optimize the value network for the target Q value, mean squared error loss is used. Furthermore, the specific expression for y is as follows:

[0048] in, R is the target Q value; R is the immediate reward. It is a discount factor; The next state; It is the action output by the target policy in state s′; It is the value of the target action in the next state s′ as evaluated by the target Critic network. and For target network parameters; Furthermore, the target network parameters are updated using a soft update method, with the specific expression as follows:

[0049]

[0050] in, and For target network parameters; This is the soft update coefficient, which is usually taken as a small value to ensure training stability; and These are the current network parameters; Step S304, the steps for implementing closed-loop control and establishing safety constraint mechanisms, in which: After training, the strategy network is deployed to the actual machining system. Based on the state monitoring results of step S2, the current state of the gear forming tool is predicted, and the strategy network outputs parameter adjustment instructions, the specific expression of which is as follows:

[0051] in, This is the optimal parameter adjustment amount. To train the converged policy network parameters; To train the converged policy network; It is a state vector; Furthermore, the adjustment command is sent to the CNC system for execution to update the actual machining parameters. The specific expression is as follows:

[0052] in, For the updated processing parameters, These are the original processing parameters. This is the optimal parameter adjustment amount; Furthermore, the machining results are collected and fed back in real time, including information such as cutting force, machining error, and tool wear, and the state vector is updated accordingly. This forms a closed-loop control chain of "prediction-compensation-feedback"; Furthermore, a security constraint mechanism is set to limit the adjustment amount of the policy network output, as shown in the following expression:

[0053] in, This is the adjustment amount after safety constraints. This is a truncation function. This is the optimal parameter adjustment amount; To adjust to the minimum value, This is the maximum adjustment value; Furthermore, when the cutting force exceeds the safety threshold, the wear reaches the limit value, or the machining error exceeds the tolerance range, the protection mechanism is triggered, switching to preset safety parameters or stopping machining to ensure the safety and stability of the machining process.

[0054] The gear forming tool remaining life prediction and intelligent tool change decision-making steps described in step S4 specifically include the following steps: Step S401, the step of defining the life of the gear forming tool and constructing the timing sequence, in which: Define a criterion for tool life termination: when the tool flank wear reaches a threshold or a sudden failure occurs, the tool life is determined to have ended. The specific expression is as follows:

[0055] in, This is a marker of the end of life. The wear threshold, For sudden failure events, Let t be the amount of wear at time t. In other cases; Furthermore, the remaining tool life is defined, with the following specific expression:

[0056] in, for The remaining service life at any given moment. The moment when the tool reaches the end of its lifespan, t is time t; Furthermore, a tool wear time series is constructed as input to the life prediction model, with the specific expression as follows:

[0057]

[0058] in, For time-series input sequences, For the length of the history window, for The state vector at time t, To accumulate processing load, Let t be the amount of wear at time t. To fuse feature vectors, For processing parameters; Step S402, the steps of the time-series lifetime prediction model, in which: A lifetime prediction model based on a long short-term memory network is constructed, and a recurrent neural network is used to capture the temporal evolution of wear. The specific expression is as follows:

[0059]

[0060] in, This represents the hidden state of the LSTM network. The remaining lifetime predicted by the LSTM model. Let be the input feature vector at time step t. To be at time step t The hidden state of 1, It is a multilayer perceptron; Furthermore, a Transformer-based self-attention lifetime prediction model is constructed, which captures long-range dependencies in time series through a multi-head attention mechanism. The specific expression is as follows:

[0061]

[0062] in, For the output of the self-attention mechanism, For query, key-value matrix, The normalization function ensures that the sum of the attention weights is 1. The dot product of the query matrix Q and the key matrix K represents the similarity between each query vector and each key vector; The dimension of the key vector. The remaining lifetime predicted by the Transformer model. For multi-head attention operation, The input sequence is a time series; Furthermore, the lifetime prediction loss function is defined, with the following specific expression:

[0063] in, The lifetime prediction loss function; The mean absolute error is used as the optimization objective to determine the number of training samples. The remaining lifetime of the i-th sample as predicted by the model; This represents the actual remaining useful life of the i-th sample. Step S403, the probabilistic lifetime prediction and multi-model fusion step, in which: A hidden Markov model is established to describe the stochastic evolution of the wear state of gear forming tools. The state transition probability and observation probability are defined, and their specific expressions are as follows:

[0064]

[0065] in, for The hidden wear and tear condition at all times, Let be the state transition probability. To observe the feature vector, For the observed probability distribution; Let be the state transition probability, representing the probability of transitioning from state i to state j at time step t; observe the probability distribution. This indicates that an observation is generated in state i. The probability of.

[0066] Furthermore, Bayesian filtering is used to update the posterior probability of the state online, with the specific expression as follows:

[0067] in, The posterior probability of the state is based on all historical observations; For all observation data from time step 1 to time step t, To observe the probability distribution, Based on all historical observations status The posterior probability, These are state parameters; Furthermore, based on the probability distribution of the remaining lifetime estimated from the posterior state probability, the lifetime prediction value of the Bayesian model is obtained. ; Furthermore, a multi-model ensemble method is used to fuse the results of the three prediction models, as shown in the following expression:

[0068] in, To integrate the prediction results, The weights for each model are determined based on the performance on the validation set. This represents the prediction result of the i-th prediction model at time step t; Furthermore, the prediction uncertainty is calculated, and the specific expression is as follows:

[0069] in, The standard deviation of the remaining life forecast is used to quantify the uncertainty of the forecast. To integrate the prediction results, These are the weight coefficients for each model. This represents the prediction result of the i-th prediction model at time step t; Step S404, the intelligent tool change decision optimization step, in which: Based on the remaining tool life prediction results, a safety margin is set to determine the tool change timing for a single tool. Tool change is recommended when the following conditions are met, as shown in the following expression:

[0070] in, The processing time required to complete the next workpiece For safety margin, To integrate the prediction results, The standard deviation of the remaining life prediction; Furthermore, for multi-tool collaborative machining scenarios, a multi-objective optimization model for tool change decisions is established, with the objective function being, in specific terms as follows:

[0071] in, The total number of cutting tools, To cover the cost of downtime for tool replacement, Total number of tool changes Wasting money on cutting tools For knives Remaining tool life at the time of tool change; Furthermore, constraints are set, with the specific expression as follows:

[0072] in, For knives The time required to complete the task should be considered to ensure that the tool life meets the task requirements. For knives The time required to complete the task should be considered to ensure that the tool life meets the task requirements. This represents the total number of cutting tools.

[0073] Furthermore, a dynamic programming algorithm is used to solve for the optimal tool change plan, the specific expression of which is as follows:

[0074] in, For the state value function, Let this be the tool change decision vector. For immediate costs; For immediate costs; To minimize the possible tool change decision u(t); The value function for subsequent states; Furthermore, it generates an optimal tool change schedule, outputs suggested tool change plans, and guides on-site operators to perform tool change operations, thereby achieving optimal utilization of tool life and minimizing machining costs.

[0075] Secondly, the present invention also provides a system for online compensation and life prediction of wear of gear forming tools, including a gear forming tool wear latent feature extraction module, a gear forming tool status online monitoring and wear level discrimination module, a gear forming tool dynamic compensation strategy module, and a gear forming tool remaining life prediction and intelligent tool replacement decision module. The gear forming tool wear latent feature extraction module utilizes the parameter space of the pre-constructed digital twin mechanism model to simultaneously acquire multi-source sensor signals. It preprocesses the multi-source sensor signals, performs multi-dimensional feature extraction to extract explicit features, and performs dimensionality reduction on the explicit features to obtain core features. A variational autoencoder is constructed to encode the core features into latent features. The digital twin mechanism model parameters are mapped to mechanism parameter encoding vectors and fused with the core and latent features to obtain fused features. A physical consistency loss is defined and combined with the variational autoencoder loss to form a total loss function for joint optimization. The online monitoring and wear level determination module for gear forming tools constructs an online monitoring model for gear forming tool status based on fused features obtained from the wear latent feature extraction module. It defines multiple levels based on the wear amount on the tool's flank face and labels the data. A feature relationship graph is constructed based on the feature correlation of the fused features. A graph neural network classifier is built to determine the wear level, and global feature vectors are aggregated through graph convolution to output the wear level probability distribution. Lightweight deployment of the model is achieved through knowledge distillation and model quantization. The constructed online monitoring model for gear forming tools collects data in real time and determines the tool's wear state. The gear forming tool dynamic compensation strategy module constructs a dynamic compensation mechanism based on the tool wear state identified by the online monitoring and wear level discrimination module. It defines the machining parameter adjustment amount as a control variable, utilizes a digital twin mechanism model to construct a learning simulation environment, designs a multi-objective reward function, trains an intelligent agent using a deep deterministic strategy gradient algorithm to make optimal action decisions based on the current state, outputs parameter adjustment commands in real time and collects feedback, forming a "prediction-compensation-feedback" closed loop, and sets safety constraints and protection mechanisms. The gear forming tool remaining life prediction and intelligent tool change decision module establishes a correlation model between wear and remaining life; constructs prediction models based on long short-term memory networks, Transformer, and Hidden Markov-Bayes; integrates and fuses the prediction results of multiple models and evaluates the uncertainty; sets a safety margin based on the prediction results to determine the tool change timing for a single tool; establishes a multi-objective optimization model for multi-tool scenarios, and uses dynamic programming to solve for the optimal tool change plan.

[0076] The beneficial effects of this invention are as follows: The aforementioned method for online wear compensation and life prediction of gear forming tools achieves deep extraction and enhanced interpretability of latent wear features by constructing a variational autoencoder model and introducing mechanistic parameter mapping and physical consistency loss constraints. It also achieves high-precision real-time online discrimination and lightweight deployment of wear levels by constructing a graph neural network classification model based on feature relationship graphs and employing knowledge distillation and model quantization techniques. Furthermore, it utilizes deep reinforcement learning algorithms combined with a mechanistic model simulation environment to construct a dynamic compensation strategy, enabling adaptive adjustment of machining parameters and "prediction-compensation-feedback" closed-loop control. By integrating long short-term memory networks, Transformer, and Bayesian models to establish a multi-model ensemble prediction method and combining it with multi-objective optimization algorithms, it achieves accurate prediction of remaining life and intelligent tool change decision optimization. Finally, by deeply coupling the mechanistic model and data model throughout the entire process of feature extraction, state monitoring, dynamic control, and life prediction, it maintains the consistency between the virtual and real worlds of the digital twin system and enhances engineering interpretability. This method effectively solves the technical problems of insufficient wear feature extraction, weak online monitoring accuracy and generalization ability, lack of active compensation closed-loop control mechanism, low life prediction accuracy, and disconnection from decision-making in existing technologies. Attached Figure Description

[0077] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.

[0078] Figure 1 This is a schematic flowchart of a method for online wear compensation and life prediction of gear forming tools according to the present invention.

[0079] Figure 2 This is a flowchart of the steps for extracting the latent wear features of gear forming tools based on a multidimensional digital twin mechanism model, according to the present invention.

[0080] Figure 3 This is a flowchart of the steps for online monitoring of the condition and wear level determination of gear forming tools according to the present invention.

[0081] Figure 4 This is a flowchart of the steps of the reinforcement learning-based dynamic compensation strategy for gear forming tools of the present invention.

[0082] Figure 5 This is a flowchart of the steps for predicting the remaining life of the gear forming tool and making intelligent tool changing decisions according to the present invention. Detailed Implementation

[0083] Example 1 To facilitate understanding of the present invention, the following is a specific machining example of machining a small-module external gear in the output stage of an RV reducer, where the gear forming tool is a hob.

[0084] like Figure 1 As shown, a method for online wear compensation and life prediction of gear forming tools includes the following steps: Step S1: Extracting latent features of hob wear based on a multidimensional digital twin mechanism model, obtaining the latent features of hob wear, such as... Figure 2 As shown, based on the established mechanism-data fusion digital twin dynamic model of the hobbing cutter, the mechanism parameter space is extracted and multi-source sensor signals are fused.

[0085] First, multi-source sensor signals are simultaneously acquired during the gear hobbing process:

[0086] in, This is the cutting force signal. It is a vibration signal. For acoustic emission signals, For temperature signals, It is a power signal; Furthermore, a Hampel filter is used to remove outliers:

[0087] in, This is the filtered signal. This represents the absolute deviation of the median. The threshold coefficient; the threshold coefficient The usual value is 3; Furthermore, time-domain, frequency-domain, and time-frequency-domain features are extracted, and a 3-level wavelet packet decomposition is performed using db4 wavelets to construct an explicit feature matrix. ; Furthermore, principal component analysis is used for dimensionality reduction, retaining 95% of the cumulative variance:

[0088] in, Principal component analysis transformation matrix, It is an explicit characteristic matrix. These are the explicit features after dimensionality reduction; Furthermore, a variational autoencoder is constructed to extract latent features, and the encoder maps to the latent space:

[0089]

[0090]

[0091] in, Using latent variable vectors as latent features, For encoder parameters, and The mapping function of the neural network outputs the mean and standard deviation of the latent variables respectively. For element-wise multiplication, It is standard normally distributed random noise; Furthermore, we define the variational autoencoder loss function:

[0092] in, The reconstruction loss measures the error between the reconstructed features and the original features. The KL divergence loss is used, and the latent variable distribution is constrained to approximate a standard normal distribution. The balance coefficient; Take 0.5; Furthermore, the parameter space is extracted from the digital twin mechanism model of the hobbing cutter. ,in For geometric parameter vectors, For physical property parameter vectors, For behavioral feature parameter vectors, The degeneracy law function; the geometric parameter vector The physical characteristic parameter vector includes the number of teeth, outer diameter, module, pressure angle, and helix angle of the hob. Including elastic modulus, hardness, coefficient of friction, and critical temperature; the behavioral characteristic parameter vector Including the wear coefficient and thermal deformation coefficient simulated by ANSYS, the degradation law function These are parameters for the empirical wear model; Furthermore, the mechanism parameters are encoded through the multilayer perceptron:

[0093] in, Encode the mechanism parameters into a vector; For mapping functions; Furthermore, the dominant features, latent features, and mechanistic parameter codes are fused together:

[0094] in, The feature vectors are fused, incorporating explicit physical quantities, deep implicit patterns, and mechanistic knowledge. Furthermore, a physical consistency loss is introduced:

[0095] in, To measure the difference between latent variables extracted through data-driven methods and latent variables from mechanism simulation, we consider the physical consistency constraint error. These are the implicit variables corresponding to the wear state in the simulation; Furthermore, establish the total loss function:

[0096] in The physical consistency weighting coefficient is used to balance the variational autoencoder loss and the physical consistency constraint; the physical consistency weighting coefficient Take 0.3; Furthermore, the gradient descent method is used to simultaneously optimize the parameters of the variational autoencoder and the mechanistic parameter mapping network, so that the extracted latent features can accurately represent the data distribution and conform to physical laws, thereby enhancing the interpretability and generalization ability of the features. After training, the latent variable z obtained by the encoder for each sample is the latent feature of gear forming tool wear, which is used for subsequent state monitoring and life prediction.

[0097] Step S2, Online Monitoring and Wear Level Determination of Hob Status: This step enables real-time determination of the hob wear level, such as... Figure 3 As shown, a graph neural network classification model is constructed based on fused features.

[0098] First, the fused features are used as monitoring input:

[0099] in, The input for state monitoring features includes dimensionality-reduced explicit features, implicit features, and mechanism parameter encodings. Furthermore, based on the wear amount VB on the hob's flank face, six wear levels are defined:

[0100] in, VB represents the total number of wear levels, and VB represents the wear amount on the hob's flank face. Furthermore, samples in the historical dataset are labeled with the corresponding wear level categories based on the offline measured VB values, and a labeled training dataset is constructed for model training. Furthermore, a graph is constructed based on feature correlation, using each dimension of the fused feature vector as a graph node. A feature relationship graph is built based on the correlation between features, and the Pearson correlation coefficient matrix between features is calculated.

[0101] in, Features and characteristics The correlation coefficient between them For covariance, and These are the standard deviations of the two features; Furthermore, construct the adjacency matrix of the graph based on the correlation coefficient matrix:

[0102] Where A is the adjacency matrix, and 0.6 is the relevance threshold, which retains only the connections between feature pairs with strong relevance. Furthermore, the adjacency matrix is ​​normalized:

[0103] in, For the normalized adjacency matrix, For degree matrix, Represents a node The degree; Furthermore, a graph neural network classifier is constructed to determine the wear level. The node feature update formula for the graph convolutional layer is as follows:

[0104] in, For the first The node feature matrix of the layer These are the initial node features. For the first The learnable weight matrix of the layer, For activation functions; Furthermore, after multiple layers of graph convolution, graph pooling is used to aggregate the graph representation into a global feature vector, and the probability distribution of wear level is output through fully connected layers and softmax layers:

[0105] in, For the predicted wear level probability distribution, The total number of convolutional layers in the graph. For graph pooling operations, It is a multilayer perceptron; Furthermore, define the classification loss function:

[0106] in, The number of training samples, For the sample True label one-hot encoding, To predict probabilities, the cross-entropy loss function is used to optimize the model parameters; Furthermore, during online monitoring, data is collected in a 5-second window and the confidence level is calculated:

[0107] in, To predict confidence levels, when the confidence level is below a set threshold... In such cases, a manual review mechanism may be triggered or the monitoring frequency may be increased. Furthermore, a knowledge distillation compression model is adopted:

[0108] in, For distillation losses, and The output logits of the student model and the teacher model are respectively. For temperature parameters, The KL divergence; the temperature parameter Take 3; quantize to INT8 format for deployment.

[0109] Step S3, the step of the hobbing dynamic compensation strategy based on reinforcement learning, realizes the adaptive adjustment of hobbing machining parameters, such as... Figure 4 The steps for constructing a reinforcement learning compensation system based on wear state recognition results are shown in the diagram.

[0110] First, define the adjustment range for the hobbing machining parameters:

[0111] in, for Action vector at time step, This is the amount of adjustment for the cutting speed. Adjustment amount for feed per tooth. This is the depth of cut adjustment amount. The adjustment amount is the radius compensation value for the gear forming tool; the adjustment amount is the cutting speed. The range is [-10, 10] m / min, and the feed per tooth adjustment amount The range is [-0.03, 0.03] mm / z, and the depth of cut adjustment amount is... The range is [-1, 1] mm, and the adjustment amount of the gear forming tool radius compensation value is... The range is [-0.05, 0.05] mm; Furthermore, the state space for reinforcement learning is defined, integrating the wear state of the gear forming tool, machining parameters, and feature information:

[0112] in, For state vectors, The current wear level, This refers to the amount of wear. This is the current processing parameter vector. To fuse feature vectors, This represents the increment of processing error. Furthermore, a simulation environment is constructed using a hobbing cutter mechanism model:

[0113] in, For mechanism simulation model, For the predicted cutting force, Due to machining error, This represents the increase in wear. Furthermore, a multi-objective reward function is designed:

[0114] in, For instant rewards, These are the weighting coefficients. As a processing efficiency indicator, To constrain violations and penalties; Furthermore, define the constraint penalty term:

[0115] in, The penalty coefficient is... The safe threshold for machining error is 2000N, the safe threshold for cutting force is 2000N, and the safe threshold for wear is 0.65mm. Furthermore, the DDPG algorithm is used for training, and the Actor network outputs the following strategy:

[0116] in, For policy networks, For network parameters, To explore noise, and to enhance exploration capabilities during the training phase; Furthermore, the Critic network's evaluation value:

[0117] in, For value networks, For network parameters, As a discount factor, and For target network parameters; Furthermore, policy gradient update:

[0118] in, Let be the objective function of the policy network, and optimize the policy by maximizing the cumulative reward. Further updates from the Critic network:

[0119]

[0120] Wherein, 0.99 is the target Q value, and the value network is optimized using mean squared error loss; Furthermore, soft update the target network:

[0121]

[0122] Where 0.005 is the soft update coefficient; Furthermore, deploy the policy network to output rolling cutter parameter adjustment instructions:

[0123]

[0124] in, This is the optimal parameter adjustment amount. The parameters of the converged policy network are used for training; where P(t+1) are the updated processing parameters. Furthermore, real-time acquisition of machining results feedback, including information such as cutting force, machining error, and tool wear, updates the state vector s(t+1), forming a closed-loop control chain of "prediction-compensation-feedback"; Furthermore, set security constraints:

[0125] in, This is the adjustment amount after safety constraints. This is a truncation function; Step S4, the remaining life prediction and intelligent tool change decision-making steps, realize hob life prediction and tool change optimization, such as... Figure 5 The steps for constructing a multi-model integrated lifetime prediction method are shown below.

[0126] First, define the criteria for the end of hob life:

[0127] Among them, EOL is the end of service life indicator, VBthreshold is the wear threshold, and Catastrophic Failure is the sudden failure event; Furthermore, define the remaining service life of the hob:

[0128] in, for The remaining service life at any given moment. The moment when the cutting tool reaches the end of its service life; Furthermore, a hob wear time sequence is constructed:

[0129]

[0130] in, For time-series input sequences, For the length of the history window, for The state vector at time t, To accumulate processing load; Furthermore, construct an LSTM prediction model:

[0131]

[0132] in, This represents the hidden state of the LSTM network. The remaining lifetime predicted by the LSTM model; Furthermore, construct a Transformer prediction model:

[0133]

[0134] in, For query, key-value matrix, The dimension of the key vector. For multi-head attention operation, The remaining lifetime predicted by the Transformer model; Furthermore, define lifetime prediction loss:

[0135] in, The mean absolute error is used as the optimization objective to determine the number of training samples. Furthermore, a hidden Markov model is established:

[0136]

[0137] in, for The hidden wear and tear condition at all times, Let be the state transition probability. To observe the feature vector, For the observed probability distribution; Further, Bayesian filter update:

[0138] in, The posterior probability of the state is based on all historical observations; Furthermore, multi-model integration and fusion:

[0139] in, To integrate the prediction results, 0.35, 0.45, and 0.20 are the weight coefficients of each model, respectively. Furthermore, the uncertainty is calculated:

[0140] in, The standard deviation of the remaining life forecast is used to quantify the uncertainty of the forecast. Furthermore, determine the timing for hobbing tool changes:

[0141] in, The time required to complete the next workpiece is 10, which is a safety margin. Furthermore, a multi-roller collaborative optimization model:

[0142] Where K is the total number of cutting tools, To cover the cost of downtime for tool replacement, Total number of tool changes Wasting money on cutting tools The remaining life of tool i when it is replaced; Furthermore, the constraints are:

[0143] in, For knives The time required to complete the task should be considered to ensure that the tool life meets the task requirements. Furthermore, solve using dynamic programming:

[0144] in, For the state value function, Let this be the tool change decision vector. For immediate costs; Furthermore, the optimal tool change schedule for the hob is generated.

[0145] Example 2 To facilitate understanding of the present invention, the following is a specific machining example of machining a small-module external gear in the output stage of an RV reducer, where the gear forming tool is a hob.

[0146] A gear forming tool wear online compensation and life prediction system includes a hob wear latent feature extraction module, a hob condition online monitoring and wear level discrimination module, a hob dynamic compensation strategy module, and a hob remaining life prediction intelligent tool replacement decision module. The hob wear latent feature extraction module further includes: Based on the established mechanism-data fusion digital twin dynamic model of the hobbing cutter, the mechanism parameter space is extracted and multi-source sensor signals are fused.

[0147] First, multi-source sensor signals are simultaneously acquired during the gear hobbing process:

[0148] in, This is the cutting force signal. It is a vibration signal. For acoustic emission signals, For temperature signals, It is a power signal; Furthermore, a Hampel filter is used to remove outliers:

[0149] in, This is the filtered signal. This represents the absolute deviation of the median. The threshold coefficient; the threshold coefficient The usual value is 3; Furthermore, time-domain, frequency-domain, and time-frequency-domain features are extracted, and a 3-level wavelet packet decomposition is performed using db4 wavelets to construct an explicit feature matrix. ; Furthermore, principal component analysis is used for dimensionality reduction, retaining 95% of the cumulative variance:

[0150] in, Principal component analysis transformation matrix, It is an explicit characteristic matrix. These are the explicit features after dimensionality reduction; Furthermore, a variational autoencoder is constructed to extract latent features, and the encoder maps to the latent space:

[0151]

[0152]

[0153] in, Using latent variable vectors as latent features, For encoder parameters, and The mapping function of the neural network outputs the mean and standard deviation of the latent variables respectively. For element-wise multiplication, It is standard normally distributed random noise; Furthermore, we define the variational autoencoder loss function:

[0154] in, The reconstruction loss measures the error between the reconstructed features and the original features. The KL divergence loss is used, and the latent variable distribution is constrained to approximate a standard normal distribution. The balance coefficient; Take 0.5; Furthermore, the parameter space is extracted from the digital twin mechanism model of the hobbing cutter. ,in For geometric parameter vectors, For physical property parameter vectors, For behavioral feature parameter vectors, The degeneracy law function; the geometric parameter vector The physical characteristic parameter vector includes the number of teeth, outer diameter, module, pressure angle, and helix angle of the hob. Including elastic modulus, hardness, coefficient of friction, and critical temperature; the behavioral characteristic parameter vector Including the wear coefficient and thermal deformation coefficient simulated by ANSYS, the degradation law These are parameters for the empirical wear model; Furthermore, the mechanism parameters are encoded through the multilayer perceptron:

[0155] in, Encode the mechanism parameters into a vector; For mapping functions; Furthermore, the dominant features, latent features, and mechanistic parameter codes are fused together:

[0156] in, The feature vectors are fused, incorporating explicit physical quantities, deep implicit patterns, and mechanistic knowledge. Furthermore, a physical consistency loss is introduced:

[0157] in, To measure the difference between latent variables extracted through data-driven methods and latent variables from mechanism simulation, we consider the physical consistency constraint error. These are the implicit variables corresponding to the wear state in the simulation; Furthermore, establish the total loss function:

[0158] in The physical consistency weighting coefficient is used to balance the variational autoencoder loss and the physical consistency constraint; the physical consistency weighting coefficient Take 0.3; Furthermore, the gradient descent method is used to simultaneously optimize the parameters of the variational autoencoder and the mechanistic parameter mapping network, so that the extracted latent features can accurately represent the data distribution and conform to physical laws, thereby enhancing the interpretability and generalization ability of the features. After training, the latent variable z obtained by the encoder for each sample is the latent feature of gear forming tool wear, which is used for subsequent state monitoring and life prediction.

[0159] The online monitoring and wear level determination module for the hob condition is the same as the online monitoring and wear level determination step S2 in Embodiment 1; The dynamic compensation strategy module for the hobbing cutter is the same as the step S3 of the reinforcement learning-based dynamic compensation strategy for the hobbing cutter in Embodiment 1. The hob remaining life prediction and intelligent tool change decision module is the same as the hob remaining life prediction and intelligent tool change decision step S4 in Example 1.

[0160] In this embodiment, a hobbing cutter full life cycle intelligent management system was constructed, realizing integrated management of wear latent feature extraction, online status monitoring, dynamic parameter compensation, and remaining life prediction.

Claims

1. A method for online wear compensation and life prediction of gear forming tools, characterized in that, Includes the following steps: Step S1, the step of extracting the hidden features of wear on gear forming tools, includes: Multi-source sensor signals are collected and dominant features are extracted. The dominant features are reduced to core features by principal component analysis. A variational autoencoder is constructed to encode the core features into latent features. Mechanism parameters are mapped to feature vectors and fused with core features and latent features to obtain fused features. Step S2, the online monitoring and wear level determination of gear forming tool condition, includes: An online monitoring model for gear forming tool status is constructed based on fused features. Wear level is defined according to wear amount, a graph neural network classifier is constructed, and feature aggregation of core features is achieved through graph convolution to output wear level probability. Data is collected in real time and wear status is determined. Step S3, the dynamic compensation strategy for gear forming tools, includes: A dynamic compensation mechanism is constructed based on the wear state. A reinforcement learning simulation environment is built using a mechanism model. A multi-objective reward function is designed. The optimal compensation strategy is trained using a deep deterministic policy gradient algorithm. Parameter adjustment instructions are output in real time and feedback is collected. Step S4, the remaining life prediction and intelligent tool change decision for gear forming tools, includes: A model relating wear and remaining life is established. Multi-model integration is used to fuse prediction results and assess uncertainty. Based on the prediction results, a safety margin is set to determine the timing of tool changing for a single tool. For multi-tool scenarios, a multi-objective optimization model is established, and dynamic programming is used to solve for the optimal tool changing plan.

2. The method according to claim 1, characterized in that, Step S1 specifically includes the following steps: Step S101, the step of multi-source sensor signal acquisition and explicit feature extraction, in which: Using the parameter space of the digital twin mechanism model of the gear forming tool that has been constructed, multi-source sensor signals are collected synchronously; the collected raw signals are preprocessed, and multi-dimensional features are extracted from the preprocessed signals. All features of multiple sensor channels are spliced ​​together to form an explicit feature matrix. The explicit feature matrix is ​​then reduced to a core feature matrix by principal component analysis. The multi-source sensing signals include cutting force signals, vibration acceleration signals, acoustic emission signals, temperature signals, and spindle power signals; the multi-dimensional features include time-domain statistical features, frequency-domain energy features, and time-frequency domain features based on wavelet packet decomposition; wherein the time-domain features include mean, standard deviation, root mean square value, kurtosis, and peak value; the frequency-domain features include total power, center frequency, and bandwidth obtained through fast Fourier transform; and the time-frequency domain features include energy extraction of each frequency band using wavelet packet decomposition. Step S102, the step of extracting latent features and mapping mechanistic parameters based on variational autoencoders, in which: A variational autoencoder model is constructed to extract wear latent features, and the encoder maps the core features to the latent space. The parameter space of the digital twin mechanism model is mapped to the feature space, and the mechanism parameters are encoded using a multilayer perceptron. The core features, latent features, and mechanism parameter encodings are fused to obtain a fused feature vector, as follows: in, To fuse feature vectors, It is an explicit characteristic matrix. The matrix represents the core features, where z is a latent feature. Mechanism parameter encoding; The parameter space includes geometric parameter vectors, physical property parameter vectors, behavioral characteristic parameter vectors, and degradation law functions; Step S103, the step of physical consistency loss constraint and joint optimization, in which: The simulation output of tool wear state generated by the digital twin mechanism model is used as a supervision signal, and the obtained wear state is mapped to the latent space through an encoder to obtain the corresponding latent variables. The physical consistency loss is defined as follows: in, For physical consistency constraint error, Latent variables extracted for data-driven models The latent variables are mapped to the latent space from the simulation output of the mechanistic model. Establish a total loss function for joint optimization: in, This is the physical consistency weighting coefficient. The loss function of the variational autoencoder is... For physical consistency constraint error, This is the total loss function; The gradient descent method is used to simultaneously optimize the parameters of the variational autoencoder and the mechanistic parameter mapping network. After training, the latent variables obtained by the encoder for each sample are the latent features of gear forming tool wear.

3. The method according to claim 2, characterized in that, Step S2 specifically includes the following steps: Step S201, the steps of defining wear level and constructing monitoring feature input, specifically include: The fusion feature vector extracted in step S1 is used as the input of the condition monitoring model to construct the monitoring feature input vector; the wear level is defined according to the tool back face wear amount VB, and a wear level classification standard is established; the samples in the historical dataset are labeled as the corresponding wear level categories according to the offline measured wear amount VB, and a labeled training dataset is constructed for model training. Step S202, the steps for constructing a graph structure based on feature correlation, specifically include: Each dimension of the fused feature vector is used as a graph node. A feature relationship graph is constructed based on the correlation between features. The Pearson correlation coefficient matrix between features is calculated. The adjacency matrix of the graph is constructed based on the correlation coefficient matrix, and the adjacency matrix is ​​normalized. Step S203, the steps for constructing the graph neural network classification model, specifically includes: A graph neural network classifier is constructed to determine wear level. After multiple layers of graph convolution, graph pooling is used to aggregate the graph representation into a global feature vector. The probability distribution of wear level is then output through fully connected layers and softmax layers. in, For the predicted wear level probability distribution, The total number of convolutional layers in the graph. For graph pooling operations, It is a multilayer perceptron. For the first The node feature matrix of the layer, where Softmax is the Softmax classification layer; Define a classification loss function for model training: in, The number of training samples, For the sample True label one-hot encoding, To predict probabilities; Step S204, the steps for online monitoring implementation and lightweight model deployment, specifically include: During the online monitoring phase, sensor data is collected at fixed time intervals Δt, with each sampling window having a length of Tw. The data within the window is processed by feature extraction according to step S1. The fused features are then input into the graph neural network classifier, which outputs the wear level prediction result for the current moment. Knowledge distillation technology is used to compress the graph neural network into a lightweight student model. The student model is trained using soft labels from the teacher model. The lightweight model is then quantized, converting floating-point parameters into INT8 or INT16 format.

4. The method according to claim 3, characterized in that, Step S3 specifically includes the following steps: Step S301, the steps of designing the compensation control system and defining the state-action space, specifically includes: Based on the wear state of the gear forming tool identified in step S2, a dynamic compensation control mechanism is constructed; the adjustment range of each control quantity is set to construct the action space; and the state space of reinforcement learning is defined by integrating the wear state of the gear forming tool, machining parameters, and feature information. Step S302, the steps for constructing a reinforcement learning simulation environment based on a mechanism model, specifically include: A simulation environment for reinforcement learning training is constructed using the digital twin mechanism model from step S1. The mechanism model calculates the machining result at the next moment based on the current state and action. A multi-objective reward function is designed, taking into account machining quality, gear forming tool life, and machining efficiency. Constraints and penalties are defined. Step S303, the steps of deep deterministic policy gradient algorithm and policy network training, specifically include: The intelligent agent is trained using a deep deterministic policy gradient algorithm, and an Actor-Critic network architecture is constructed. The Actor network outputs a deterministic policy, the specific expression of which is as follows: in, The action taken by the intelligent agent at time step t; For policy networks, For network parameters, For state vectors, To explore noise; The Critic network evaluates the value of state-action pairs using the following expression: in, For value networks, For network parameters, For state vectors, As a discount factor, and For target network parameters; The value of state-action pairs evaluated for the Critic network; For immediate reward at time step t; The next state output by the target Actor network The optimal action under the given conditions; For the target Critic network; Expressing expectations; The Actor network parameters are updated using policy gradients, the Critic network parameters are updated using temporal difference errors, and the target network parameters are updated using a soft update method. Step S304, the steps for implementing closed-loop control and establishing safety constraint mechanisms, specifically include: After training, the strategy network is deployed to the actual machining scenario. Based on the state monitoring results of step S2, the current state of the gear forming tool is predicted. Adjustment instructions are sent to the CNC system for execution to update the actual machining parameters. Machining results are collected in real time to update the state vector. A safety constraint mechanism is set to limit the adjustment amount output by the strategy network. When the cutting force exceeds the safety threshold, the wear reaches the limit value, or the machining error exceeds the tolerance range, the protection mechanism is triggered to switch to the preset safety parameters or stop machining.

5. The method according to claim 4, characterized in that, Step S4 specifically includes the following steps: Step S401, the steps of defining the life of gear forming tools and constructing the timing sequence, specifically include: Define a tool life termination criterion: the tool life is determined to end when the tool flank wear reaches a threshold or a sudden failure occurs; define the remaining tool life; and construct a tool wear time series as input to the life prediction model. Step S402, the steps of the time-series lifetime prediction model based on deep learning, specifically include: A lifetime prediction model based on a long short-term memory network is constructed, which captures the temporal evolution pattern of wear and tear through a recurrent neural network; a self-attention lifetime prediction model based on a Transformer is constructed, which captures long-range dependencies in the time series through a multi-head attention mechanism; and a lifetime prediction loss function is defined. Step S403, the probabilistic lifetime prediction and multi-model fusion based on Bayesian update, specifically includes: A Hidden Markov Model (HMM) is established to describe the stochastic evolution of the wear state of gear forming tools, defining state transition probabilities and observation probabilities. Bayesian filtering is used to update the posterior state probabilities online. Based on the posterior state probabilities, the probability distribution of the remaining lifetime is estimated, yielding the lifetime prediction value from the Bayesian model. A multi-model ensemble method is employed to fuse the results of three prediction models, and the prediction uncertainty is calculated. Step S404, the steps of intelligent tool change decision optimization, specifically include: Based on the remaining tool life prediction results, a safety margin is set to determine the timing of single-tool tool changes, and tool changes are recommended when conditions are met. For multi-tool collaborative machining scenarios, a multi-objective optimization model for tool change decisions is established; constraints are set; a dynamic programming algorithm is used to solve for the optimal tool change plan; an optimal tool change schedule is generated, and a recommended tool change scheme is output. The dynamic programming solution for the optimal tool change plan is specifically expressed as follows: in, For the state value function, Let this be the tool-changing decision vector. For immediate costs.

6. A system for online wear compensation and life prediction of gear forming tools, characterized in that, It includes a gear forming tool wear latent feature extraction module, a gear forming tool online status monitoring and wear level discrimination module, a gear forming tool dynamic compensation strategy module, and a gear forming tool remaining life prediction and intelligent tool replacement decision module; The gear forming tool wear latent feature extraction module specifically includes: Multi-source sensor signals and dominant features are collected. A variational autoencoder is constructed to encode the dominant features into latent features. Mechanism parameters are mapped into feature vectors and fused with dominant and latent features to obtain fused features. The online monitoring and wear level determination module for gear forming tools also includes: An online monitoring model for gear forming tool status is constructed based on fused features. Wear level is defined according to wear amount. A graph neural network classifier is constructed. Feature aggregation is achieved through graph convolution and wear level probability is output. Data is collected in real time and wear status is determined. The dynamic compensation strategy module for gear forming tools also includes: A dynamic compensation mechanism is constructed based on the wear state. A reinforcement learning simulation environment is built using a mechanism model. A multi-objective reward function is designed. The optimal compensation strategy is trained using a deep deterministic policy gradient algorithm. Parameter adjustment instructions are output in real time and feedback is collected. The gear forming tool remaining life prediction and intelligent tool change decision module also includes: A model relating wear and remaining life is established. Multi-model integration is used to fuse prediction results and assess uncertainty. Based on the prediction results, a safety margin is set to determine the timing of tool changing for a single tool. For multi-tool scenarios, a multi-objective optimization model is established, and dynamic programming is used to solve for the optimal tool changing plan.

7. The system according to claim 6, characterized in that, The gear forming tool wear latent feature extraction module further includes a multi-source sensor signal acquisition and explicit feature extraction submodule, a latent feature extraction and mechanism parameter mapping submodule, and a physical consistency loss constraint and joint optimization submodule; The multi-source sensor signal acquisition and explicit feature extraction submodule further includes: Utilizing the parameter space of the established digital twin mechanism model, multi-source sensor signals are simultaneously acquired during the gear forming process, and a set of multi-source sensor signals is constructed: in, for A collection of multi-source sensor signals at any given time. This is the cutting force signal. It is a vibration acceleration signal. For acoustic emission signals, For temperature signals, Main spindle power signal; The acquired multi-source sensor signals are preprocessed by using a Hampel filter to remove outliers. The specific expression is as follows: in, The signal before filtering. This is the filtered signal. This represents the absolute deviation of the median. This is the threshold coefficient; This represents the median of all samples within the current sliding window. Multi-dimensional feature extraction is performed on the preprocessed signal. All features obtained from the multi-dimensional feature extraction are concatenated to form a dominant feature matrix. Principal component analysis is then used to perform dimensionality reduction on the dominant feature matrix. in, Principal component analysis transformation matrix; It is an explicit characteristic matrix. The core feature matrix; The latent feature extraction and mechanism parameter mapping submodule further includes: A variational autoencoder model is constructed to extract wear latent features. The encoder maps the core features to the latent space, and the decoder reconstructs the original features. The parameter space of the digital twin mechanism model is mapped to the feature space, and the mechanism parameters are encoded by a multilayer perceptron. The core features, latent features and mechanism parameter encoded vectors are then fused. The physical consistency loss constraint and joint optimization submodule further includes: The simulation output of tool wear state is generated using a digital twin mechanism model as a supervision signal, and the obtained wear state is mapped to the latent space through an encoder to obtain the corresponding latent variables. The gradient descent method is used to simultaneously optimize the parameters of the variational autoencoder and the mechanistic parameter mapping network. After training, the latent variables obtained by the encoder for each sample are the latent features of gear forming tool wear.

8. The system according to claim 7, characterized in that, The online monitoring and wear level determination module for gear forming tool status also includes a wear level definition and monitoring feature input construction submodule, a graph structure construction submodule, a graph neural network classification model construction submodule, and an online monitoring implementation and lightweight model deployment submodule. The wear level definition and monitoring feature input construction submodule specifically includes: The fusion features extracted by the gear forming tool wear latent feature extraction module are used as input to the condition monitoring model. The wear level is defined according to the wear amount VB on the tool back face. The samples in the historical dataset are labeled as the corresponding wear level categories according to the VB value measured offline. A labeled training dataset is constructed for model training. The graph structure construction submodule specifically includes: Each dimension of the fused feature vector is used as a graph node. A feature relationship graph is constructed based on the correlation between features. The Pearson correlation coefficient matrix between features is calculated. An adjacency matrix of the graph is constructed based on the correlation coefficient matrix. The adjacency matrix is ​​then normalized. The graph neural network classification model construction submodule specifically includes: A graph neural network classifier is constructed to determine the wear level. After multiple layers of graph convolution, graph pooling is used to aggregate the graph representation into a global feature vector. The probability distribution of the wear level is output through fully connected layers and softmax layers. A classification loss function is defined for model training. The online monitoring implementation and lightweight model deployment submodule specifically includes: During the online monitoring phase, at fixed time intervals Slide to acquire sensor data, with each sampling window having a length of [length missing]. The data within the window is processed according to step S1 to extract features. After obtaining the fused features, they are input into the graph neural network classifier, which outputs the wear level prediction result at the current moment. The confidence level of the prediction result is calculated using the following expression: in, To predict confidence levels, when the confidence level is below a set threshold... In such cases, a manual review mechanism may be triggered or the monitoring frequency may be increased. This represents the probability distribution of the predicted wear level. Knowledge distillation is used to compress the graph neural network into a lightweight student model. The student model is trained using soft labels from the teacher model, as shown in the following expression: in, For distillation loss, KL divergence; Softmax is the Softmax classification layer; and The output logits of the student model and the teacher model are respectively. For temperature parameters, The lightweight model is quantized, converting floating-point parameters into INT8 or INT16 format.

9. The system according to claim 8, characterized in that, The aforementioned gear forming tool dynamic compensation strategy module includes a compensation control mechanism design and state-action space definition submodule, a reinforcement learning simulation environment construction submodule, a deep deterministic policy gradient algorithm and policy network training submodule, and a closed-loop control implementation and safety constraint mechanism submodule. The compensation control mechanism design and state-action space definition submodule specifically includes: Based on the wear status of the gear forming tool identified by the online monitoring and wear level discrimination module, a dynamic compensation control mechanism is constructed, defining the control variables as the adjustment amounts of the machining parameters; setting the adjustment range of each control amount and constructing the action space; defining the state space of the reinforcement learning algorithm. The reinforcement learning simulation environment construction submodule specifically includes: A simulation environment for reinforcement learning training is constructed using a digital twin mechanism model of a gear forming tool wear latent feature extraction module. The mechanism model calculates the machining result at the next moment based on the current state and action. A multi-objective reward function is designed, comprehensively considering machining quality, gear forming tool life, and machining efficiency. Constraint penalty terms are defined. The deep deterministic policy gradient algorithm and policy network training submodule specifically also include: A deep deterministic policy gradient algorithm is used to train the intelligent agent and construct an Actor-Critic network architecture. The Actor network outputs a deterministic policy, the Critic network evaluates the value of state-action pairs, the Actor network parameters are updated by policy gradient, the Critic network parameters are updated by temporal difference error, and the target network parameters are updated by soft update method. The closed-loop control implementation and safety constraint mechanism submodule specifically includes: After training, the strategy network is deployed to the actual machining scenario. Based on the status monitoring results of the online monitoring and wear level discrimination module for gear forming tool status, the current status of the gear forming tool is predicted. The strategy network outputs parameter adjustment instructions, which are sent to the CNC system for execution to update the actual machining parameters. A safety constraint mechanism is set to limit the adjustment amount output by the strategy network. When the cutting force exceeds the safety threshold, the wear reaches the limit value, or the machining error exceeds the tolerance range, the protection mechanism is triggered, switching to the preset safety parameters or stopping machining.

10. The system according to claim 9, characterized in that, The gear forming tool remaining life prediction and intelligent tool change decision module includes a gear forming tool life definition and time sequence construction submodule, a time sequence life prediction submodule, a probabilistic life prediction and multi-model fusion submodule, and an intelligent tool change decision optimization submodule. The gear forming tool life definition and timing sequence construction submodule specifically includes: Define the tool life termination criterion: when the tool flank wear reaches a threshold or a sudden failure occurs, the tool life is determined to end; define the remaining tool life and construct the tool wear time sequence as the input of the life prediction submodule; The time-series lifetime prediction submodule also includes: A lifetime prediction model based on a long short-term memory network is constructed, and the wear and tear time-series evolution pattern is captured by a recurrent neural network. A self-attention lifetime prediction model based on a Transformer is constructed, and the long-range dependency in the time series is captured by a multi-head attention mechanism. A lifetime prediction loss function is defined. The probabilistic lifetime prediction and multi-model fusion submodule further includes: A Hidden Markov Model (HMM) is established to describe the stochastic evolution of the wear state of gear forming tools. State transition probabilities and observation probabilities are defined. Bayesian filtering is used to update the posterior probabilities of the states online. Based on the posterior probabilities, the probability distribution of the remaining lifetime is estimated, yielding the lifetime prediction value of the Bayesian model. The results of three prediction models are fused using a multi-model ensemble method to calculate the prediction uncertainty. The intelligent tool-changing decision optimization submodule further includes: Based on the remaining tool life prediction results, a safety margin is set to determine the timing of single-tool tool changes; for multi-tool collaborative machining scenarios, a multi-objective optimization model for tool change decisions is established; a dynamic programming algorithm is used to solve for the optimal tool change plan, generate the optimal tool change schedule, and output a tool change suggestion scheme.

Images