Automatic generation method for weaving production process

Abstract

The invention belongs to the field of textile process automatic design, and particularly relates to a weaving production process automatic generation method. According to the invention, multi-mode full-process weaving workshop data are collected through a high-frequency sensor. After abnormal data is eliminated, filling a time sequence vacancy through a Gaussian radial basis function; according to a real weaving workshop, a virtual environment for intelligent agent optimization is built; minimization of product errors and minimization of production time are used as global optimization targets, and a deep reinforcement learning algorithm is used to enable an intelligent agent to learn an optimal production strategy reaching the targets from a historical production process; a user customizes the organization structure, the luster color, the fabric hand feeling, the stiffness and the bulkiness of the target fabric through the interactive interface module, and the intelligent agent automatically generates process parameters of all links of the weaving process according to the characteristics of the target fabric. The development efficiency and quality of textile manufacturing are improved, and the method is suitable for rapid production and optimization of a complex weaving process.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent manufacturing of textile technology, and in particular relates to a method for automatically generating a weaving production process. Background Art

[0002] With the continuous development of the textile industry, weaving technology, as a key link in textile production, faces increasing quality and efficiency requirements. Traditional weaving process generation mainly relies on manual experience, which has many limitations. In the weaving process, there are many process parameters, including warp tension, weft insertion parameters, loom speed, opening time, etc. These parameters affect each other and their combination is complex. When manually designing the weaving process, it is difficult to fully consider the complex relationship between all parameters, which often leads to insufficient process optimization. For example, the manually determined warp tension may not guarantee the flatness of the fabric when the loom is running at high speed, or the unreasonable setting of the weft insertion parameters will increase the weft breakage rate, thereby affecting the weaving efficiency and product quality. Moreover, with the continuous emergence of new fiber materials and the increasing market demand for diversified fabric functions, such as fire resistance, antibacterial, high breathability, etc., the weaving process needs to be continuously adjusted and innovated. The traditional experience-based process generation method is difficult to quickly adapt to these new changes.

[0003] On the other hand, with the development of artificial intelligence technology, deep reinforcement learning, as a method that can optimize decision-making through trial and error learning in a dynamic environment, provides a new idea for the automatic generation of weaving processes. Reinforcement learning can establish an interactive model between the intelligent agent and the weaving environment, allowing the intelligent agent to learn the optimal process generation strategy based on environmental feedback (such as product quality indicators, production efficiency data, etc.) in the process of constantly trying different combinations of process parameters, thereby overcoming the shortcomings of traditional manual methods and is expected to achieve efficient, accurate and intelligent generation of weaving processes. Summary of the invention

[0004] In view of the shortcomings of the prior art, the present invention discloses a method for automatically generating a weaving production process, which is used to improve the production quality and efficiency of textiles. According to the actual weaving workshop, a data-driven full-process virtual weaving production environment is built. A process parameter intelligent agent is constructed in the virtual environment, and the constraints and reward functions in the multi-agent optimization process are designed with minimizing the difference in target fabric properties, maximizing production quality and minimizing production time as global optimization goals. The improved Bellman equation and the greedy strategy are used to induce the agent to select the action with the maximum expected reward, and the weaving process parameters that meet the production goals are obtained in the virtual environment where the agent obtains the maximum reward, and the optimization process is recorded as empirical data. The deep neural network is used to learn the empirical data, obtain the maximum expected reward model of the agent in any environment, and quickly generate process parameters.

[0005] The technical solution of the present invention is as follows:

[0006] A method for automatically generating a weaving production process, the steps are as follows:

[0007] The first step is to collect data from the entire workshop and remove abnormal data. The collected production data from the entire workshop includes equipment operation data and mechanical characteristics data of the produced fabrics. In order to ensure the quality of the data, it is necessary to clean the collected data and remove abnormal fluctuations under the same production state. After the data is removed, the radial basis interpolation function is used to make up for the missing data in the time series. The expression is:

[0008]

[0009] Among them, x represents the point that needs to be interpolated, k i represents the coefficient to be determined, is the radial basis function, ||xx i || is the Euclidean distance from the interpolation point to the known data point.

[0010] The second step is to build a data-driven full-process virtual weaving production environment, including warping, sizing and weaving processes. Model the full-process weaving production workshop and drive the operation of the virtual workshop through the collected workshop data. Simulate the yarn load and fabric style according to the movement coordination of the weaving equipment in the virtual workshop. For real-time simulation calculation, the deformation of the yarn is calculated using the linear elastic model, and the calculation formula is as follows:

[0011]

[0012] Where F is the stress at the end of the yarn. S is the cross-sectional area of ​​the yarn. ε and k represent the normal strain and Young's modulus of the yarn, respectively. v represents the Poisson's ratio of the yarn. The normal strain of the yarn is equal to the difference between the surface linear velocity of the virtual warp beam and the weaving beam per unit time. Other parameters can be obtained through the basic properties of the yarn and fiber.

[0013] The fabric style is calculated based on the simulation data in the virtual workshop, including the fabric's fluffiness L, stiffness D and surface friction μ. Among them, L∝(h / W), h represents the thickness of the fabric, and W represents the weight of the fabric. D∝(B / W), B represents the bending stiffness of the fabric. μ∝nl, n represents the number of hair roots, and l represents the length of hair. These fabric parameters are controlled by the whole process of weaving production technology. The weaving production process parameters are constructed through a set of vectors P(v1,v2,…v m ). The degree to which each fabric parameter is affected by the fabric production process parameters is expressed as:

[0014] Gi∝∑wjg(vj)

[0015] Among them, G irepresents fabric parameters, w j represents the influence weight of production process, g(v j ) represents the influence function of production process parameters.

[0016] The third step is to build a multi-agent production process in a virtual environment And constrain the actions of the agents. According to the production process, the multi-agents are divided into warping agents, sizing agents and weaving agents. The agents continuously move in the virtual environment to achieve the optimization of the production process. The motion set can be expressed as A{-u j ,0,u j}, u represents the unit length of the agent's movement. Since the number of process parameters is m, the number of action set elements is 3 m When the agent Select Action -u j , represents the current process parameter v j It will decrease at the next moment. When the agent chooses actions 0 and u j It means that the process parameters remain unchanged and increase respectively.

[0017] In the process of interaction, the movement of the agent is restricted by the range of process parameters and process sequence. First, the movement range of each agent is set according to the boundary conditions of the process parameters in the virtual weaving environment. Secondly, in order to meet the time continuity of the weaving production process, the time interval between the start movement of agents in different processes is set. The calculation formula is as follows:

[0018]

[0019] in: and They represent the start time of the agent’s actions in two consecutive weaving processes. i represents the transmission time of fabric production materials between adjacent processes. α∈(0,1] is used to adjust the starting movement time interval of the intelligent agent. i It is a fixed physical quantity in the virtual environment. It limits the advance movement of the agent but allows the delayed movement.

[0020] The fourth step is to set the fabric style attributes of the target fabric, determine the optimization goal of the multi-agent in the virtual environment, and design the reward function according to the goal. The fabric styles that can be set include the fluffiness, stiffness and surface friction of the fabric. These attributes are controlled by related fabric parameters, and the weaving production process parameters can adjust the fabric parameters to indirectly control the fabric style attributes.

[0021] The global goal of agent optimization contains three sub-goals:

[0022] Objective 1: Minimize the fabric style attribute error of the target fabric.

[0023] Goal 2, maximize the product quality of produced fabrics.

[0024] Goal 3: Minimize the whole weaving production time.

[0025] Goal 1 stipulates that the agent must optimize towards the correct fabric style, Goal 2 stipulates that the agent must optimize towards low-defect fabrics, and Goal 3 stipulates that the agent must optimize towards high production efficiency. The reward weights are assigned according to the differences between the three target tasks, and the reward for Goal 1 > Goal 2 > Goal 3. The global goal of the agent is expressed as:

[0026]

[0027] Among them, W1, W2 and W3 represent the target weights, W1>W2>W3, M j Indicates the set fabric style parameters, F α Indicates the allowable tension of the yarn, Represents the process speed related to the weaving production process. The Sigmoid function is used to normalize the data, and the reward function is expressed as:

[0028] R t =J t (p)-J t+1 (p)

[0029] If J(P) is gradually decreasing, the agent receives a positive reward, otherwise it receives a negative reward.

[0030] The fifth step is to use the improved Bellman equation to calculate the state value of the agent, and use the greedy strategy to induce the agent to choose the action with the largest expected reward in the current environment. The calculation formula of the improved Bellman equation is as follows:

[0031]

[0032] Among them, Q(P t ,A t ) and Q(P t+1 ,A t+1 ) represent the value of the agent in the current and future environmental actions, r represents the action reward in the current environment, and λ is used to control the discount of the agent for future rewards. The smaller the discount, the more the agent pays attention to the rewards in the current environment, and the larger the discount, the more the agent pays attention to the rewards in the future environment. The discount gradually decreases with the increase in the number of interactions, indicating that the agent pays more attention to future rewards at the beginning of the optimization stage, and after accumulating a certain amount of positive rewards, the agent pays more attention to current rewards. According to the calculated environmental action value, the agent transfers the state through the greedy strategy, and the calculation formula is as follows:

[0033] π(P t )=argmaxQ(P t ,a t )

[0034] Among them, π represents the strategy made by the agent in the current environment, a t ∈A, represents an action. The process of state transition is as follows:

[0035]

[0036] Among them, γ represents the learning rate of the state transfer process. When the environment value reaches the maximum, the state P of the agent * (v1 * ,v2 * ,…v m * ) are the weaving process parameters that meet the production objectives.

[0037] The sixth step is to record the experience data of the agent's optimization process, including the environment, action and value, and use the deep neural network to learn the experience data, obtain the maximum expected reward model of the agent in any environment, and automatically generate process parameters. In the neural network model, the input data is the production process parameters of the environment where the agent is located, and the output is the value corresponding to each action of the agent in the current environment. Using the mean square error as the loss function, the expression is:

[0038]

[0039] Where b is the number of randomly drawn experience samples, y i is the value data in the experience sample, and θ is the network parameter. The network parameters are updated through the back propagation algorithm to minimize the loss function. When the loss value drops below the threshold, it means that the neural network model can predict the maximum expected reward action in any environment, helping the intelligent agent to quickly complete the decision-making selection, thereby realizing the rapid generation of weaving production process.

[0040] Compared with the prior art, the present invention has the following beneficial effects:

[0041] The present invention uses a deep reinforcement learning model to continuously learn and optimize the process parameters that meet the target fabric from a virtual weaving production environment. Compared with fabric process design based on manual experience, the design error is reduced. In the product development process, the pre-production and testing stages of the fabric are omitted, the consumption of raw materials is reduced, and the development efficiency of the fabric is improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 A flowchart automatically generated for a weaving production process provided by an embodiment of the present invention;

[0043] Figure 2 A schematic diagram of the interaction between a process parameter intelligent agent and an environment provided by an embodiment of the present invention;

[0044] Figure 3 A schematic diagram of a process parameter intelligent agent optimization process provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0045] The following will combine the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. First, data collection and data cleaning are carried out on the real weaving workshop, and a data-driven full-process virtual weaving production environment is built. A process parameter agent is constructed in the virtual environment, and the constraints and reward functions in the multi-agent optimization process are designed to minimize the difference in target fabric properties, maximize production quality and minimize production time as global optimization goals. The improved Bellman equation and the greedy strategy are used to induce the agent to select the action with the maximum expected reward, and the weaving process parameters that meet the production goals are obtained in the virtual environment where the agent obtains the maximum reward, and the optimization process is recorded as empirical data. The deep neural network is used to learn the empirical data, obtain the maximum expected reward model of the agent in any environment, and quickly generate process parameters.

[0046] The present invention provides a method for automatically generating a weaving production process, such as Figure 1 As shown, follow these steps:

[0047] Step 1: Use sensors of multiple modes to collect data from the weaving workshop, and preprocess the collected data through Gaussian radial basis function.

[0048] In an embodiment of the present invention, workshop data of warping, sizing and weaving processes are collected and preprocessed to ensure data quality. According to the characteristics of the production process, different types of sensors are selected for data collection and storage, and the monitored physical quantities include speed, acceleration, displacement, tension, temperature, humidity, flow rate, viscosity, etc. First, in terms of warping, the collection of warping tension can use a high-frequency tension sensor to ensure that the collected tension data has sufficient details. For example, the FMSLMCZ through-axis tension sensor has a collection range of up to 25000N and a maximum collection response frequency of 1000Hz. The collection of warping speed can use the TR-PA1010 high-speed gear speed sensor, and the response frequency under full load is 15000Hz. In terms of sizing, the slurry viscosity and temperature can use the JS-N online sizing sensor, the viscosity range is between 0-70%, and the temperature range can reach 300°. The slurry concentration and flow rate sensors can be ROTOTHERM liquid concentration sensor and LG20A propeller flow rate sensor respectively, which can meet the accuracy requirements of concentration and flow rate in the slurry production process. In terms of weaving, the collection of weaving speed and upper machine tension uses the same configuration as the warping data collection. Since equipment vibration may cause the monitoring data to exceed the normal range of production process parameters, abnormal production data needs to be cleaned to eliminate data that is much larger or much smaller than the adjacent data points, as well as data that exceeds the normal process parameter range. After the data is eliminated, in order to make up for the lack of data in time series, the radial basis function is used to interpolate the data. As a possible implementation method, the Gaussian function is used as the kernel function of the radial basis, and the expression is as follows:

[0049]

[0050] Among them, σ represents a hyperparameter that can control the accuracy and mean square error of interpolation. Since the data collected in the workshop has a low dimension, σ is set to 0.1.

[0051] Step 2: Use Soildworks to build a full-process weaving production workshop model, and use the collected data in step 1 to drive the operation of the virtual weaving workshop to generate simulation data of yarn and fabric.

[0052] The full-process virtual weaving production environment built includes warping workshop, sizing workshop and weaving workshop. The modeling objects include the full-process production equipment, yarn and fabric in the production process. In order to reduce the rendering pressure of the virtual workshop after data-driven, the optimization function in Soildworks is used to optimize the model, including hidden surface deletion and sparse mesh operation.

[0053] The motion coordination of the whole weaving production workshop is simulated in the Unity physics engine. In the warping workshop, the equipment speed, warping length, number of warping roots and winding density parameters are input to simulate the load distribution of the yarn during the warping process. The sizing workshop inputs the sizing concentration, sizing speed and drying temperature to simulate the viscosity and squeezing force of the sizing. The weaving workshop inputs the weaving speed, warp and weft density, opening time and height to simulate the tension at the end of the yarn.

[0054] The yarn load and fabric style are simulated according to the motion coordination of the weaving equipment in the virtual workshop. For real-time simulation calculation, the deformation of the yarn is calculated using the linear elastic model. Taking 20-count pure cotton yarn as an example, the yarn cross-sectional area is 3×10 -8 m 2 , Young's modulus is 0.2MPa, Poisson's ratio is 0.3, and the tension at the end of the yarn is calculated as follows:

[0055] F=238ΔV·Δt

[0056] Where F(N) is the tension at the end of the yarn. ΔV(m / s) and Δt(s) represent the speed difference and unit time at both ends of the yarn, respectively. The fabric style is calculated based on the simulation data in the virtual workshop, including the fabric's fluffiness L, stiffness D, and surface friction μ. Where L∝(h / W), h represents the thickness of the fabric, and W represents the fabric weight. D∝(B / W), B represents the bending stiffness of the fabric. μ∝nl, n represents the number of hair roots, and l represents the length of hair. These fabric parameters are controlled by the weaving production process of the entire process. The weaving production process parameters are constructed through a set of vectors P(v1,v2,…v m ). In order to improve the timeliness of the solution, the hyperbolic tangent function is used to normalize the weaving process parameters, and the normalized data is weighted summed. The calculation expression is:

[0057]

[0058] Among them, G i represents fabric parameters, w j represents the influence weight of production process, v j Indicates production process parameters.

[0059] Step 3: Construct a weaving production process intelligent agent in the virtual environment, and constrain the joint actions of the intelligent agent according to the range of production process parameters and weaving production sequence in the virtual environment in step 2.

[0060] Constructing production process multi-agent in virtual environment And constrain the actions of the agents. According to the production process, the multi-agents are divided into warping agents, sizing agents and weaving agents. The movement set of the agents in the virtual environment can be expressed as A{-u j ,0,uj}, u represents the unit length of the agent's movement. Select Action -u j , represents the current process parameter v j It will decrease at the next moment. When the agent chooses actions 0 and u j It means that the process parameters remain unchanged and increase respectively.

[0061] In the virtual environment, 15 production process parameters are selected for adjustment, and these parameters are controlled by individual agents. All agents perform random state transitions through a set of motion sets. As a possible implementation, Table 1 shows the motion sets and motion ranges of the 15 agents.

[0062] Table 1 Movement range and movement set of multi-agents

[0063]

[0064] In addition to meeting the constraints of the process parameter range, these agents should also meet the requirements of the production sequence. In order to meet the time continuity of the weaving production process, the time interval between the start movements of different process agents is set, and the constraints that must be met between two consecutive processes are in

[0065] and They represent the start time of the agent’s actions in two consecutive weaving processes. i It indicates the transmission time of fabric production materials between adjacent processes.

[0066] Step 4: Set the properties of the target fabric and specify the process parameters in step 3. The multi-agent takes the minimum target fabric property difference, maximum product quality and minimum production time as the optimization goals, and sets the agent's reward function according to the goals.

[0067] like Figure 2 As shown in the figure, the global goal of the agent optimization contains three sub-goals: Goal 1 is to minimize the error of the fabric style attributes of the target fabric, Goal 2 is to maximize the product quality of the produced fabric, and Goal 3 is to minimize the production time. The fabric style of the target fabric includes fluffiness, stiffness and surface friction. These properties can be controlled by related fabric parameters, while the weaving production process parameters can adjust the fabric parameters to indirectly control the fabric style attributes. The calculation equation of Goal 1 is as follows:

[0068]

[0069] M jRepresents the target fabric parameters, including the surface friction, bulkiness and stiffness of the fabric. Sigmoid function is used for normalization. Product quality is highly correlated with the tension at the end of the yarn. The calculation equation for target 2 is as follows:

[0070]

[0071] F α is the allowable tensile force, taking 20-count cotton yarn as an example, F α Take 2.3N. The production time is related to the speed parameters in the whole weaving process. The calculation method of target 3 is as follows:

[0072]

[0073] It indicates the process speed related to the weaving production process. According to the different tasks of each goal, the goal weight needs to be defined. The goal importance is stipulated as J1(P)>J2(P)>J3(P). Therefore, the global goal can be defined as:

[0074] J(P)=0.5J1(P)+0.3J2(P)+0.2J3(P)

[0075] For the agents in the same production workshop, they need to meet the requirement of local goal consistency, which can be expressed as:

[0076]

[0077] The reward function can be expressed as:

[0078] R t =J t (p)-J t+1 (p)

[0079] If J(P) is gradually decreasing, the agent receives a positive reward, otherwise it receives a negative reward.

[0080] Step 5: According to the reward function in step 4, the dynamically discounted Bellman equation is used to calculate the action value of the agent, and the ε-greedy strategy is used to induce the agent to find the weaving process parameters that meet the goal.

[0081] The improved Bellman equation is used to calculate the state and action value of the agent. The improved Bellman equation calculation formula is as follows:

[0082] Q(P t ,A t )=r+e -λt ·max At+1 Q(P t+1 ,A t+1 )

[0083] Among them, Q(P t ,A t ) and Q(P t+1 ,A t+1 ) represent the value of the agent in the current and future environmental actions, r represents the action reward in the current environment, and λ is used to control the discount of the agent to obtain future rewards. λ is set to 0.1. The agent pays more attention to future rewards at the beginning of the optimization. After accumulating a certain amount of positive rewards, the agent pays more attention to current rewards. According to the calculated environmental action value, the agent transfers the state through the ε-greedy strategy. The calculation formula is as follows:

[0084]

[0085] Among them, π represents the strategy made by the agent in the current environment, a, a t ∈A represents the action. ε is the probability of the agent choosing the action with the maximum value function, and ε is 0.9. The process of the agent state transition is as follows:

[0086]

[0087] When the environment value reaches the maximum, the agent is in state P * (v1 * ,v2 * ,…v m * ) are the weaving process parameters that meet the production objectives.

[0088] Step 6: Use the Pytorch framework to learn the historical optimization process of step 5, obtain the maximum expected reward model of the agent in any environment, and quickly generate process parameters.

[0089] like Figure 3 As shown in the figure, the experience data of the agent's optimization process, including environment, action and value, is recorded. The deep neural network is used to learn the experience data, obtain the maximum expected reward model of the agent in any environment, and automatically generate process parameters. In the neural network model, the input data is the production process parameters of the environment where the agent is located, and the output is the value corresponding to each action of the agent in the current environment. The mean square error is used as the loss function, and the expression is:

[0090]

[0091] Where b is the number of experience samples drawn, y iis the value data in the experience sample, and θ is the network parameter. The network parameters are updated through the back-propagation algorithm to minimize the loss function. In order to increase the efficiency of data use, the experience replay area is set, and 64 experience data are repeatedly randomly selected to update the network parameters. Two neural networks with exactly the same structure are set up, and the Q network and the target network are used for parameter update and action value function calculation respectively to ensure the stability of training. The parameters in the Q network are copied to the target network every 5 time steps. When the loss value drops below the threshold, it means that the neural network model can predict the maximum expected reward action in any environment, help the intelligent agent quickly complete the decision-making selection, and thus realize the rapid generation of weaving production process.

Claims

1. A method for automatically generating a weaving production process, characterized in that: Here are the steps: The first step is to collect data from the entire workshop process and eliminate abnormal data. The collected data from the entire workshop process includes equipment operation status data and mechanical characteristics data of the produced fabrics. The collected data is cleaned to eliminate abnormal fluctuations under the same production status. The second step is to build a data-driven full-process virtual weaving production environment, including warping, sizing and weaving processes; model the full-process weaving production workshop, and drive the operation of the virtual workshop through the collected workshop data; simulate the yarn load and fabric style according to the movement coordination of the weaving equipment in the virtual workshop; The third step is to build a production process multi-agent in a virtual environment vi , and constrain the actions of the agents; according to the production process, the multi-agents are divided into warping agents, sizing agents and weaving agents; the agents continuously move in the virtual environment to achieve the optimization of the production process; the movement set is represented by A{-u j ,0,u j }, u represents the unit length of the agent's movement; the number of process parameters is m, and the number of action set elements is 3 m ; When agent a vj Select Action -u j , represents the current process parameter v j It will decrease at the next moment; when the agent chooses actions 0 and u j It means that the process parameters remain unchanged and increase respectively; First, set the movement range of each agent; second, set the time interval between the start movements of agents in different processes. The calculation formula is as follows: T i =α(t vi+1 -t vi ) Where: t vi and t vi+1 They represent the start time of the agent’s actions in two consecutive weaving processes; T i It indicates the transmission time of fabric production materials between adjacent processes; The fourth step is to set the fabric style attributes of the target fabric, determine the optimization goal of the multi-agent in the virtual environment, and design the reward function according to the goal; the fabric styles that can be set include the fluffiness, stiffness and surface friction of the fabric; The global goal of agent optimization contains three sub-goals: Objective 1, minimize the fabric style attribute error of the target fabric; Goal 2, maximize the product quality of produced fabrics; Goal 3, minimize the whole process weaving production time; The goal 1 stipulates that the agent must optimize towards the correct fabric style, the goal 2 stipulates that the agent must optimize towards low-defect fabrics, and the goal 3 stipulates that the agent must optimize towards high production efficiency; the reward weights are allocated according to the differences between the three target tasks, and the reward for goal 1> the reward for goal 2> the reward for goal 3; the global goal of the agent is expressed as: Among them, W1, W2 and W3 represent the target weights, W1>W2>W3, M j Indicates the set fabric style parameters, F α Indicates the allowable tension of the yarn, V vs Indicates the relevant process speed in the weaving production process; The Sigmoid function is used to normalize the data, and the reward function is expressed as: R t =J t (p)-J t+1 (p) If J(P) is gradually decreasing, the agent receives a positive reward, otherwise, it receives a negative reward; Step 5: Use the improved Bellman equation to calculate the state value of the agent, and use the greedy strategy to induce the agent to choose the action with the largest expected reward in the current environment; The sixth step is to record the experience data of the intelligent agent's optimization process, including the environment, actions and values, use deep neural networks to learn from the experience data, obtain the maximum expected reward model of the intelligent agent in any environment, and automatically generate process parameters; in the neural network model, the input data is the production process parameters of the environment where the intelligent agent is located, and the output is the value corresponding to each action of the intelligent agent in the current environment.

2. A method for automatically generating a weaving production process according to claim 1, characterized in that: In the second step, the deformation of the yarn is calculated using a linear elastic model, and the calculation formula is as follows: Where F is the stress at the end of the yarn; S is the cross-sectional area of ​​the yarn, ε and k represent the normal strain and Young's modulus of the yarn respectively; v represents the Poisson's ratio of the yarn; the normal strain of the yarn is equal to the difference between the surface linear velocity of the virtual warp beam and the weaving beam per unit time; The simulated data in the virtual workshop calculates the fabric style, including the fabric's fluffiness L, stiffness D and surface friction μ; where L∝(h / W), h represents the fabric thickness, and W represents the fabric weight; D∝(B / W), B represents the bending stiffness of the fabric; μ∝nl, n represents the number of hairs, l represents the length of hairs; the weaving production process parameters are constructed through a set of vectors P(v1,v2,…v m ); the degree to which each fabric parameter is affected by the fabric production process parameters is expressed as: Gi∝∑wjg(vj) Among them, G i represents fabric parameters, w j represents the influence weight of production process, g(v j ) represents the influence function of production process parameters.

3. A method for automatically generating a weaving production process according to claim 1, characterized in that: In the first step, after the data is removed, the radial basis interpolation function is used to make up for the missing data in time series. The expression is: Among them, x represents the point that needs to be interpolated, k i represents the coefficient to be determined, is the radial basis function, ||xx i || is the Euclidean distance from the interpolation point to the known data point.

4. A method for automatically generating a weaving production process as claimed in claim 1, characterized in that: In the fifth step, the improved Bellman equation calculation formula is as follows: Among them, Q(P t ,A t ) and Q(P t+1 ,A t+1 ) represent the value of the agent in the current and future environment actions, r represents the action reward in the current environment, and λ is used to control the discount of the agent to obtain future rewards.

5. A method for automatically generating a weaving production process according to claim 1, characterized in that: In the fifth step, a greedy strategy is used to induce the agent, as follows: The agent transfers state through the greedy strategy, and the calculation formula is as follows: π(P t )6argmaxQ(P t ,a t ) Among them, π represents the strategy made by the agent in the current environment, a t ∈A, represents an action; the process of state transition is as follows: Among them, γ represents the learning rate of the state transfer process; when the environmental value reaches the maximum, the state P of the agent * (v1 * ,v2 * ,…v m * ) are the weaving process parameters that meet the production objectives.

6. A method for automatically generating a weaving production process according to claim 1, characterized in that: In the sixth step, the mean square error is used as the loss function, and the expression is: Where b is the number of randomly drawn experience samples, y i is the value data in the empirical sample, and θ is the network parameter.

7. A method for automatically generating a weaving production process as claimed in claim 6, characterized in that: In the sixth step, the network parameters θ are updated through the back propagation algorithm to minimize the loss function.

8. A method for automatically generating a weaving production process as claimed in claim 1, characterized in that: In the third step, α∈(0,1].

Images