Feeding speed intelligent optimization method and system based on instruction domain data fusion and multi-task deep reinforcement learning

Abstract

The invention belongs to the related technical field of numerical control milling, and discloses a feeding speed intelligent optimization method and system based on instruction domain data fusion and multi-task deep reinforcement learning. The method comprises the following steps: taking the feeding speed as an independent variable, and taking the maximum sum of the machining efficiency and the tool wear state as a target function; the method comprises the following steps: S1, setting the number of iterations n, and randomly generating an initial value of the feeding speed; s2, predicting the cutting force by using the current feeding speed, taking the feeding speed and the predicted cutting force as the input of an intelligent agent, and obtaining a new feeding speed after the feeding speed and the predicted cutting force pass through the intelligent agent; the intelligent agent comprises an actor network and a critic network, the actor network calculates and obtains a new feeding speed according to the input of the intelligent agent, and the critic network evaluates the output of the actor network; and S3, if n is equal to n + 1, returning to the step S2 until n reaches the maximum number of iterations. According to the invention, the problems of environment simplification, single target and neglect of parameter correlation in the existing feed speed optimization method are solved.

Description

Technical Field

[0001] This invention belongs to the field of CNC milling technology, and more specifically, relates to an intelligent feed rate optimization method and system based on instruction domain data fusion and multi-task deep reinforcement learning. Background Technology

[0002] In the field of high-end equipment manufacturing, optimizing batch milling parameters is crucial for ensuring workpiece quality and improving production efficiency. Traditional methods rely on human experience or single-objective trial-and-error approaches, which struggle to address the need for coordinated optimization of multiple objectives, such as surface roughness, machining efficiency, and energy consumption. Especially in dynamic machining environments, fixed parameters can easily lead to batch quality fluctuations, resulting in unstable machining quality and hindering efficiency improvements. Furthermore, traditional reinforcement learning methods often oversimplify the actual machining process to reduce modeling costs when constructing virtual environments, neglecting information such as cutting forces and tool vibrations. This prevents the optimization model from accurately reflecting the actual machining process, leading to low estimation accuracy and making the optimization results difficult to apply to actual production. Moreover, existing deep reinforcement learning (DRL) research largely focuses on single-task optimization, while actual machining often requires multi-objective coordinated optimization. Simultaneously, existing multi-objective optimization research often relies on multiple independent estimation models, failing to fully consider the interactions between various influencing factors when solving the optimization problem, resulting in optimization results lacking comprehensiveness and accuracy. Furthermore, when using deep reinforcement learning to solve parameter optimization problems, insufficient attention is paid to the correlation between state parameters, leading to unstable model results and difficulty in achieving reliable optimization of processing parameters.

[0003] In existing technologies, Lee et al. combined support vector regression and random forest to construct a cyber-physical system to predict machining quality; Wang et al. used residual neural networks to predict CNC machining quality and efficiency; and Li et al. used convolutional neural networks to associate multidimensional process signals with machining quality. However, these methods lack adaptability to dynamic machining environments and are difficult to adjust parameters in real time. Jiang et al. used LSTM-based DQN strategy to compensate for machining contour errors; Liu et al. used meta-DQN to control workpiece deformation; and Lu et al. used DRL to guide the selection of machining parameters. However, existing DRL methods oversimplify the actual machining environment to reduce modeling costs, ignoring key physical parameters such as cutting force and tool vibration. Moreover, they focus on single-task optimization and cannot cope with multi-objective coupled scenarios. Sun et al. established the relationship between multi-stage production variables and quality indicators based on random forests, and Chen et al. achieved spatiotemporal modeling of ultrasonic welding data through multi-task learning. However, these methods use multiple independent estimation models and do not consider the coupling relationship between various optimization objectives, resulting in a lack of comprehensiveness in the optimization results. Xu Guangda proposed the concept and analysis method of instruction domain, and Gao Zhiqiang associated G through instruction domain mapping. Liu Lu adjusts the feed rate based on the spindle current in the command domain, but existing command domain research has not been combined with multi-task DRL, making it impossible to achieve intelligent dynamic optimization of parameters.

[0004] Currently, existing technologies suffer from three major drawbacks: 1) Simplified simulation environments neglect key parameters such as cutting force and tool vibration, leading to significant deviations between the optimized model and actual machining; 2) Single optimization objective, with multi-objective optimization relying on independent models and failing to consider the coupling relationship between objectives; 3) Insufficient modeling of state parameter correlations, resulting in unstable optimization processes and low reliability of results. Therefore, there is an urgent need for a feed rate optimization method that can integrate actual machining data, consider multi-objective collaboration, and adapt to dynamic environments. Summary of the Invention

[0005] To address the aforementioned deficiencies or improvement needs of existing technologies, this invention provides an intelligent feed rate optimization method and system based on instruction domain data fusion and multi-task deep reinforcement learning, which solves the problems of simplified environment, single objective, and neglect of parameter correlation in existing feed rate optimization methods.

[0006] To achieve the above objectives, according to one aspect of the present invention, a method for intelligent optimization of feed rate based on instruction domain data fusion and multi-task deep reinforcement learning is provided, the method comprising the following steps: Using feed rate as the independent variable, the objective function is to maximize the sum of machining efficiency and tool wear. G ( x )as follows:

[0007] Wherein, quality is the tool wear state, efficiency is the machining efficiency, and w1 and w2 are the weight values ​​of tool wear state and machining efficiency, respectively. The MATD3 algorithm is used to solve for the maximum objective function and its corresponding optimal feed rate by following these steps: S1 sets the number of iterations n and randomly generates the initial value of the feed rate; S2 uses the current feed rate to predict the cutting force, and uses the feed rate and the predicted cutting force as inputs to the agent. After passing through the agent, a new feed rate is obtained. The agent includes an actor network and a critic network. The actor network calculates the new feed rate based on the input of the agent, and the critic network evaluates the output of the actor network. S3 n=n+1, return to step S2, until n reaches the maximum number of iterations.

[0008] More preferably, the constraints of the objective function are as follows:

[0009] in, x is the rate of change of the parameter; x is the feed rate. , These are the upper and lower bounds of the feed rate constraint, where f represents the feed rate. This indicates the machine tool feed rate limit, where 'a' is the feed rate variation factor, 'R' is the space between feed rate and state, and 'b' is the state formed by combining feed rate and cutting force.

[0010] More preferably, the loss function of the agent is as follows:

[0011] in, Here, t is the sigmoid function, T is the maximum number of iterations, and q is the processing status index. This is an estimate of the processing time. This refers to the variation in processing parameters. The cutting force is calculated using an empirical model, h is the tool wear coefficient, and q is the cutting force. t The maximum permissible quality index value, This is the penalty index.

[0012] More preferably, the BmmoE-TE model is used to predict the cutting force using the current feed rate.

[0013] More preferably, in step S1, the initial value of the feed rate is randomly generated using the DRL method.

[0014] More preferably, after n is greater than a preset number of times, in step S2, noise is added to the current speed to change the current feed speed.

[0015] According to another aspect of the present invention, a feed rate intelligent optimization system based on instruction domain data fusion and multi-task deep reinforcement learning is provided. The system includes an actuator for executing the above-described feed rate intelligent optimization method based on instruction domain data fusion and multi-task deep reinforcement learning.

[0016] According to another aspect of the present invention, a computer storage medium is provided, on which a computer program is stored, the computer program being used to implement the above-described intelligent feed rate optimization method based on instruction domain data fusion and multi-task deep reinforcement learning.

[0017] According to another aspect of the present invention, a computer program product is provided, comprising a computer program that, when processed and executed, implements the above-described intelligent feed speed optimization method based on instruction domain data fusion and multi-task deep reinforcement learning.

[0018] In summary, the technical solutions conceived by this invention have the following beneficial effects compared with the prior art: 1. This invention constructs a multi-objective optimization model, relies on the instruction domain to complete the fusion between data and program segments, and improves the MATD3 algorithm to balance the current tool wear state and machining efficiency, thereby realizing multi-objective optimization of milling feed rate. This solves the problems of simplified environment, single objective and neglect of parameter correlation in existing feed rate optimization methods.

[0019] 2. This invention employs an implicit balance reward mechanism, which means that instead of directly decomposing or weighting multiple objectives, it designs a reward function containing multi-dimensional feedback signals. This allows the agent to autonomously learn objective balance strategies through interaction with the environment. By dynamically adjusting the composition of the reward signals, it achieves collaborative optimization of multiple objectives.

[0020] 3. This invention employs a joint optimization of Deep Reinforcement Learning (DRL) and Multi-Agent Dual-Delay Deep Deterministic Policy Gradient Algorithm (MATD3). DRL first performs a global parameter search to achieve preliminary optimization. The improved MATD3 algorithm incorporates policy noise, noise attenuation mechanisms, and multi-head attention mechanisms. It performs fine-tuning based on the preliminary optimization results and updates the network by combining an implicit balance reward mechanism, thereby improving optimization efficiency and accuracy. Attached Figure Description

[0021] Figure 1This is an overall structural diagram constructed according to a preferred embodiment of the present invention.

[0022] Figure 2 This is an instruction domain data synchronization diagram constructed according to a preferred embodiment of the present invention.

[0023] Figure 3 This is a flowchart of the overall algorithm constructed according to a preferred embodiment of the present invention.

[0024] Figure 4 This is a structural diagram of the improved MATD3 algorithm constructed according to a preferred embodiment of the present invention.

[0025] Figure 5 This is a flowchart illustrating the feed rate optimization constructed according to a preferred embodiment of the present invention.

[0026] Figure 6 This is an improved average reward comparison chart constructed according to a preferred embodiment of the present invention.

[0027] Figure 7 A finished product drawing of a machined part constructed according to a preferred embodiment of the present invention.

[0028] Figure 8 Vibration data before and after optimization for improving processing efficiency, constructed according to a preferred embodiment of the present invention.

[0029] Figure 9 Vibration data before and after optimization for extending tool life, constructed according to a preferred embodiment of the present invention. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0031] like Figure 1As shown, this implementation case first pre-sets markers in the system and synchronously records command data along with data such as vibration acceleration, real-time feed rate, and power acquired during the machining process according to the machining data sampling cycle, thus forming composite data with command markers. Next, using an empirical model of cutting force, the relationship between tool wear and cutting force, and the relationship between vibration acceleration and cutting force, a positive correlation between vibration acceleration and tool wear is established, as well as a relationship between vibration acceleration and tool wear state. Then, optimization objectives and constraints are established, and an implicit balanced reward mechanism is used, along with DRL and an improved MATD3 for joint optimization. DRL first performs a global parameter search to achieve preliminary optimization, and the improved MATD3 (incorporating strategy noise, noise attenuation mechanisms, and multi-head attention mechanisms) performs fine-tuning based on the preliminary optimization results. Combined with the implicit balanced reward mechanism, the network is updated, and the optimized feed rate for each line of G-code is output.

[0032] The specific process is as follows: 1. Set up markers in the system in advance and follow the processing data sampling cycle. By synchronously recording the instruction data N(k) with data such as vibration acceleration, real-time feed rate, and power acquired during the machining process, machining data corresponding to the instruction data row number can be obtained. Converting the row number into specific instruction data forms composite data with instruction markers. This process is as follows: Figure 2 As shown. After obtaining the instruction field data, the data segments under specific working conditions can be processed using the instruction line numbers, and then the corresponding data segments can be analyzed and optimized in more detail.

[0033] 2. Next, using the empirical model of cutting force, the relationship between tool wear and cutting force, and the relationship between vibration acceleration and cutting force, a positive correlation between vibration acceleration and tool wear is established, as well as the relationship between vibration acceleration and tool wear state. The empirical model of cutting force is as follows:

[0034] Where F is the cutting force (unit: N); Kc is the cutting force coefficient, which is related to the degree of tool wear; the greater the tool wear, the larger the Kc value; a p Here, is the depth of cut (in mm); f is the feed rate (in mm / r); and F0 is the initial cutting force, the magnitude of which is related to the tool sharpness. In other words, for continuous cutting with the same tool, since other parameters remain essentially constant, a larger cutting force indicates a larger Kc and more severe tool wear. The relationship between vibration acceleration and cutting force is:

[0035] in, Indicates the acceleration of tool vibration; Indicates the equivalent mass of the tool vibration system; This represents the system's frequency response function, which reflects the dynamic characteristics of the machine tool and cutting tool.

[0036] This leads to an indirect relationship between vibration acceleration and tool wear:

[0037] Therefore, the current tool wear condition can be assessed by measuring the tool vibration acceleration during the machining process.

[0038] 3. Establish optimization objectives and optimization constraints. The constraints involved in this invention include parameter change rate constraints, state space constraints, and action space constraints.

[0039] The optimization objective and constraints in this invention include the following parts: (1) Optimization Objective: The optimization objective in this paper consists of machining efficiency and tool wear during machining. Machining efficiency can be specifically expressed as machining time, and tool wear can be expressed as tool vibration acceleration. The formula is as follows:

[0040] Wherein, quality represents the tool wear condition, efficiency represents the machining efficiency, and w1 and w2 represent the weight values ​​of tool wear condition and machining efficiency, respectively.

[0041] (2) Parameter change rate constraint: The parameter change rate constraint is a constraint set for the feed rate being optimized. Its purpose is to limit the feed rate from being too large or too small, thereby reducing the search range during optimization to a certain extent and ensuring the stability of the machining process and the reliability of the parameters. Its expression is:

[0042] in, This represents the rate of change of the parameter, where x represents the feed rate. , These represent the upper and lower bounds of the feed rate constraint, respectively.

[0043] (3) State-space constraints and motion-space constraints: When using the MATD3 algorithm, state space and motion space are involved. The motion space represents the combination of changes in the optimized parameter. In this paper, the optimized parameter is the feed rate, so the motion space is represented by the combination of feed rates for different G-code lines. The state space consists of the output of the motion space and the predicted tool wear. The state-space constraints and motion-space constraints are as follows:

[0044]

[0045] The range and dimensions of the optimized action are specified. Here, x represents the feed rate; b is the state (composed of the feed rate and the cutting force); and R is the space between the feed rate and the state (composed of the feed rate and the cutting force), i.e., the feed rate space or the state space.

[0046] (4) Optimize motion constraints: Optimize motion constraints refer to optimizing the rate of change during the execution of the motion. The optimized motion here is the same as the motion in the motion space, which is a combination of feed rates for different G-code lines. The constraint formula is:

[0047] Where 'a' represents the multiple of the feed rate variation.

[0048] (5) Machine tool hardware constraints: Considering the limitations of the machine tool's hardware on the feed rate, the machine tool hardware constraint formula for the feed rate can be given as follows:

[0049] Based on the above, the complete optimization objective and constraints can be derived:

[0050] in, x is the rate of change of the parameter; x is the feed rate. , These are the upper and lower bounds of the feed rate constraint, where f represents the feed rate. This indicates the machine tool feed rate limit, where 'a' is the feed rate variation factor, 'R' is the space between feed rate and state, and 'b' is the state formed by combining feed rate and cutting force.

[0051] 4. An implicit balancing reward mechanism is used, which means that instead of explicitly decomposing or weighting multiple objectives, a reward function containing multi-dimensional feedback signals is designed to allow the agent to autonomously learn an objective balancing strategy through interaction with the environment. The reward mechanism in this invention includes basic rewards and constraint penalties. Basic rewards include processing status and efficiency, where processing status is mainly related to tool vibration during processing, and efficiency is related to the overall processing time. Constraint penalties include parameter change rate penalties and physical constraint penalties, where the parameter change rate is used to limit unreasonable situations where parameters are too large or too small. The core mechanism of implicit balancing is dynamic feedback and self-learning. This method does not require explicit objective decomposition; it integrates processing status, efficiency, and constraints into a single reward signal through an integrated formula. The agent autonomously learns a balancing strategy by maximizing the total reward. For example, when parameter adjustment leads to an improvement in processing status but a decrease in efficiency, the quality weight in the reward function dynamically increases, guiding the agent to prioritize compensating for efficiency loss in subsequent iterations. If the cutting force exceeds the limit, a penalty is triggered, forcing the agent to adjust parameters to meet physical constraints, thereby automatically finding a balance point between quality, efficiency, and safety. Furthermore, the reward mechanism described in this paper can dynamically adjust priorities based on iteration. During the iteration process, the reward mechanism of this invention gradually increases the weight of the processing state as the iteration progresses. The specific reward formula is as follows:

[0052] in: Here, t is the sigmoid function, T is the maximum number of iterations, and q is the processing status index. This is an estimate of the processing time. This refers to the variation in processing parameters. The value is calculated using an empirical model of cutting force. This represents the action sequence of the most recent 5 time steps (time window from t-4 to t), where h is the tool wear coefficient, and q is the tool wear coefficient. t The maximum permissible quality index value, This is the penalty index.

[0053] 5. A joint optimization approach using DRL and an improved MATD3 algorithm is employed. DRL first performs a global parameter search to achieve initial optimization. The improved MATD3 algorithm (incorporating policy noise, noise attenuation mechanisms, and multi-head attention mechanisms) is then fine-tuned based on the initial optimization results. An implicit balanced reward mechanism is used to update the network and output the optimized feed rate per line of G code. The overall flowchart of the algorithm is shown below. Figure 3 The MATD3 algorithm structure diagram is shown below. Figure 4 .

[0054] The overall machining parameter (feed rate) optimization process is as follows: Figure 5As shown, before optimizing the machining parameters, experimental data on feed rate, spindle power, tool vibration, and spindle speed were first obtained through milling experiments. Then, machining data with an instruction domain was generated through instruction domain mapping. Subsequently, the machining data was normalized to eliminate the influence of different data dimensions and magnitudes on the optimization process.

[0055] Next, the agent is initialized, and agent parameters such as learning rate, number of hidden layers, action dimension, and state dimension are extracted from the parameter file.

[0056] Then, the process enters the agent optimization phase. In the agent optimization phase, it is first determined whether it is in the initial exploration phase (whether the number of iterations is less than the preset value of 1000). If it is in the initial exploration phase, the feed rate is randomly generated. If it is not in the initial exploration phase, the feed rate is output by the agent.

[0057] Here, the action refers to the feed rate in each line of G-code. The action is then input into the BmmoE-TE model for cutting force prediction. In this model, a gated expert network is used, and different expert networks are used for prediction for different working conditions.

[0058] After obtaining the predicted cutting force value, the predicted cutting force value is concatenated with the action to form a new state and output. In this process, the new state is used as the input to the Actor network and the Critic network, and the reward is calculated for the new state. When calculating the reward value, the optimization objective formula listed above is used. The calculated reward has two uses: one is to evaluate the quality of the generated action. The higher the reward, the better the action. The other is to store the reward and the new state in the experience replay buffer established by the agent. The data in the experience replay buffer is used to train the Critic network. The Critic network is updated by comparing the reward output by the environment with the Q value, so that the Q value is closer to the actual cumulative reward.

[0059] The Actor network takes the state as input and outputs actions (feed rate per line of G-code). The Critic network takes the state as input and the actions generated by the Actor network as input, and outputs a Q-value, which is used to evaluate the value of the state-action pair. The Q-value provides the Actor network with optimization direction, making it want to output actions with higher Q-values. After obtaining the Q-value, the Actor network's parameters are backpropagated to drive the update of the Actor network.

[0060] During the exploration phase, random actions are used to accumulate experience, and the actions output by the Actor network are not adopted. However, the experience accumulated during the exploration phase is stored in the experience replay buffer and indirectly participates in the subsequent training of the Actor and Critic networks. After the exploration phase, since the experience from the experience replay buffer has been used to train the Actor and Critic networks, the actions output by the Actor network will be considered as better actions by the current policy. Exploration noise is introduced during the Actor network's action output process to introduce randomness, ensuring that the agent can still try new actions to discover the optimal action while utilizing known effective actions (feed rate per line of G code).

[0061] During the training phase, a training loop is entered, and the MATD3 training process is executed for each sample. In the first 1000 steps of training, the agent employs a random exploration strategy, randomly selecting actions within the action space to broadly explore the environment and acquire diverse experience data. After 1000 steps, a noisy action selection method is adopted. The agent adds noise to its current action selection based on the current strategy, thus utilizing existing experience while continuing to explore new actions, interacting with the environment, and obtaining rewards. The experience data generated from each interaction (including state, action, reward, next state, etc.) is stored in the experience replay buffer. The system periodically samples experience data from the experience replay buffer to train the agent. Through continuous learning and policy updates, the agent's adaptability and optimization capabilities to the environment are improved.

[0062] To ensure the optimized parameters remain within a reasonable range and prevent excessive or insufficient parameter variations from affecting processing quality and equipment safety, a 25% variation rate limit for the feed rate is applied. Finally, the optimization results are saved, and the total running time of the entire optimization process is recorded for subsequent analysis and evaluation of the optimization effect.

[0063] In optimizing processing parameters, a joint optimization strategy is adopted using Deep Reinforcement Learning (DRL) and the Multi-Agent Dual-Delay Deep Deterministic Policy Gradient Algorithm (MATD3). DRL first performs a global search on the processing parameters to achieve preliminary optimization, providing a foundation for subsequent fine-tuning. MATD3 then performs more refined parameter adjustments based on the preliminary optimization results of DRL. During the optimization process, an implicit balanced reward mechanism is incorporated to update the network, continuously optimizing the strategy and improving the accuracy and effectiveness of parameter optimization. The optimized parameters are returned after dimensionality adjustment and boundary constraint processing to ensure that the parameters meet actual processing requirements and equipment constraints.

[0064] During the training loop, a double-delay update strategy is strictly implemented: the Critic network is updated in each iteration, while the Actor and target networks are updated every certain number of steps. This strategy effectively avoids network overfitting while ensuring learning efficiency. Furthermore, policy noise is introduced during this process, and a noise attenuation mechanism is applied.

[0065] To verify the effectiveness of the instruction domain-based multi-task deep reinforcement learning feed rate optimization method proposed in this invention, the algorithm improvement effect and optimization effect were verified respectively.

[0066] In the experiment verifying the effectiveness of the algorithm improvement, comparative optimization experiments were conducted on corresponding algorithms with and without policy noise and DRL global search, with only policy noise and no DRL global search, and with and without policy noise and DRL global search. The three algorithms were identical except for the application of policy noise and DRL global search. Ten independent optimization experiments were performed on each algorithm, with each experiment running for 2500 rounds. The reward for each algorithm during the optimization process was recorded, and the average reward of the ten independent optimization experiments was calculated after the experiments. The experimental results are as follows: Figure 6 As shown, it can be observed that the average reward per round after adding noise without applying DRL global search is higher than that of the experiment without noise and DRL global search, proving that the algorithm obtains more rewards per round after adding policy noise mechanism, thus having a better optimization effect; Observing the experiment with adding noise and performing DRL global search, compared with the algorithm with only noise without DRL global search mechanism, it can consistently obtain higher rewards in each round of decision-making, proving that the algorithm obtains higher average rewards after adding DRL algorithm, thus having a better optimization effect.

[0067] In the experiment verifying the effect of optimizing the machining parameters, the actual machined workpiece used is shown below. Figure 7 The tool parameters used in the machining are shown in Table 1 below:

[0068] To evaluate the optimization results, two sets of experiments were conducted. One set focused on improving machining efficiency, with the primary optimization objective being machining time. The other set aimed to extend tool life, with the primary optimization objective being tool vibration during machining. The vibration acceleration and machining efficiency were compared in both sets of experiments. Vibration signals and machining times were collected before and after optimization. To facilitate observation and optimization, the vibration acceleration was magnified 10,000 times for enhanced visualization. The vibration acceleration comparison graph for the first set is shown below. Figure 8 The results of the first group of experiments are shown in Table 2 below:

[0069] As can be seen, after optimization, although the vibration peak value increased slightly, the machining time was significantly reduced, achieving a significant efficiency improvement at the cost of minimal vibration. The kurtosis was significantly reduced, while the root mean square of vibration increased slightly. The tool state became more stable during machining, experiencing less impact load, and the cutting force distribution was more uniform. In other words, the optimization did not sacrifice stability. The vibration acceleration comparison graph for the second set of experiments is shown below. Figure 9 The results of the second group of experiments are shown in Table 3 below:

[0070] The experimental results show that the load on the tool was indeed reduced after optimization.

[0071] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for intelligent optimization of feed rate based on instruction domain data fusion and multi-task deep reinforcement learning, characterized in that, The method includes the following steps: Using feed rate as the independent variable, the objective function is to maximize the sum of machining efficiency and tool wear. G ( x )as follows: Wherein, quality is the tool wear state, efficiency is the machining efficiency, and w1 and w2 are the weight values ​​of tool wear state and machining efficiency, respectively. The MATD3 algorithm is used to solve for the maximum objective function and its corresponding optimal feed rate by following these steps: S1 sets the number of iterations n and randomly generates the initial value of the feed rate; S2 uses the current feed rate to predict the cutting force, and uses the feed rate and the predicted cutting force as inputs to the agent. After passing through the agent, a new feed rate is obtained. The agent includes an actor network and a critic network. The actor network calculates the new feed rate based on the input of the agent, and the critic network evaluates the output of the actor network. S3 n=n+1, return to step S2, until n reaches the maximum number of iterations.

2. The intelligent feed rate optimization method based on instruction domain data fusion and multi-task deep reinforcement learning as described in claim 1, characterized in that, The constraints of the objective function are as follows: in, x is the rate of change of the parameter; x is the feed rate. , These are the upper and lower bounds of the feed rate constraint, where f represents the feed rate. This indicates the machine tool feed rate limit, where 'a' is the feed rate variation factor, 'R' is the space between feed rate and state, and 'b' is the state formed by combining feed rate and cutting force.

3. The intelligent feed rate optimization method based on instruction domain data fusion and multi-task deep reinforcement learning as described in claim 1 or 2, characterized in that, The loss function of the agent is as follows: in, Here, t is the sigmoid function, T is the maximum number of iterations, and q is the processing status index. This is an estimate of the processing time. This refers to the variation in processing parameters. The cutting force is calculated using an empirical model, h is the tool wear coefficient, and q is the cutting force. t The maximum permissible quality index value, This is the penalty index.

4. The intelligent feed rate optimization method based on instruction domain data fusion and multi-task deep reinforcement learning as described in claim 1 or 2, characterized in that, The BmmoE-TE model is used to predict the cutting force using the current feed rate.

5. The intelligent feed rate optimization method based on instruction domain data fusion and multi-task deep reinforcement learning as described in claim 1 or 2, characterized in that, In step S1, the initial value of the feed rate is randomly generated using the DRL method.

6. The intelligent feed rate optimization method based on instruction domain data fusion and multi-task deep reinforcement learning as described in claim 1 or 2, characterized in that, If n is greater than the preset number of times, in step S2, noise is added to the current speed to change the current feed speed.

7. A feed rate intelligent optimization system based on instruction domain data fusion and multi-task deep reinforcement learning, characterized in that, The system includes an actuator for executing the intelligent feed rate optimization method based on instruction domain data fusion and multi-task deep reinforcement learning as described in any one of claims 1-6.

8. A computer storage medium storing a computer program thereon, characterized in that, The computer program is used to implement the intelligent feed rate optimization method based on instruction domain data fusion and multi-task deep reinforcement learning as described in any one of claims 1-6.

9. A computer program product, characterized in that, The method includes a computer program that, when processed and executed, implements the intelligent feed rate optimization method based on instruction domain data fusion and multi-task deep reinforcement learning as described in any one of claims 1-6.

Images