Broken end dynamic capture method and system combining reinforcement learning and physical modeling

Abstract

The invention belongs to the technical field of textile, and discloses a reinforcement learning and physical modeling combined broken end dynamic capturing method and system. The method comprises the following steps: constructing a yarn microstructure evolution model; inputting real-time environment parameters of the textile workshop into the microstructure evolution model, and predicting to obtain real-time microstructure parameters of the yarn; performing nonlinear coupling feature extraction on the real-time environment parameters, and calculating an environment coupling feature vector; inputting the real-time microstructure parameters as material attributes into a yarn tension dynamical equation, and solving to obtain a predicted yarn macroscopic stress state; splicing the macroscopic stress state of the yarn and the environment coupling feature vector to form a state observation value; the reinforcement learning agent outputs a broken end risk probability value according to the state observation value; and when the broken end risk probability value exceeds a preset threshold value, generating and outputting a broken end early warning signal. According to the invention, early and accurate early warning of yarn breakage in a complex dynamic environment can be realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of textiles, in particular to a broken end dynamic capture method and system combining reinforcement learning and physical modeling. BACKGROUND

[0002] In the textile industry, yarn breakage is a common and serious quality problem, especially in high-speed and high-load production environments. Yarn breakage not only affects production efficiency, but also leads to decreased product quality and increased costs.

[0003] Currently, yarn breakage detection techniques in the textile industry mainly fall into two categories: one is a physical modeling method based on mechanical sensors, and the other is through image recognition and video monitoring technology. The mechanical sensor method usually relies on physical parameters of the yarn, such as tension, breaking strength, etc. for monitoring, but these methods have difficulty in effectively capturing the changes in the yarn's stress state and the coupling relationship between the environmental parameters when facing complex environmental factors (such as temperature, humidity, air flow, etc.) fluctuations. While image recognition and video monitoring technology can detect the occurrence of broken ends, due to the limited ability to capture changes in the microstructure of the yarn, and often relying on post-processing, it cannot provide effective early warning before the broken end occurs. In addition, existing sensors and detection systems often cannot monitor the subtle changes in the microstructure of the yarn in real time, especially the influence of environmental fluctuations on the yarn's stress state, resulting in an inability to effectively predict the occurrence of broken ends.

[0004] In summary, the existing technology has not effectively solved the problem of broken end early warning of yarn in a dynamically changing environment, especially under the combined action of multiple environmental factors, the capture of broken end precursors still has great difficulty. Therefore, there is an urgent need for a technical solution that can accurately capture the broken end precursors of yarn under complex environmental conditions to improve the accuracy of early warning and production efficiency. SUMMARY

[0005] The purpose of the present application is to provide a broken end dynamic capture method and system combining reinforcement learning and physical modeling to solve the problems raised in the background art.

[0006] The present application provides a broken end dynamic capture method combining reinforcement learning and physical modeling, comprising the following method steps: Step S1, a yarn microstructure evolution model is constructed, which can simulate the dynamic changes in the friction coefficient of the surface fibers and the inter-fiber cohesion force of the yarn caused by fluctuations in environmental parameters; wherein the environmental parameters include temperature, humidity and air flow speed; Step S2, input the real-time environmental parameters of the textile workshop into the microstructure evolution model to predict the real-time microstructure parameters of the yarn; perform nonlinear coupling feature extraction on the real-time environmental parameters to calculate the environmental coupling feature vector including the temperature and humidity synergy index and the air flow disturbance intensity spectrum; Step S3, inputting the real-time microstructure parameters as material properties into a yarn tension dynamics equation to obtain a predicted yarn macro stress state; splicing the yarn macro stress state and the environment coupling feature vector to form a state observation value of a reinforcement learning agent; Step S4, the reinforcement learning agent outputs a breakage risk probability value according to the state observation value; wherein the reinforcement learning agent learns through training that when the environment coupling feature vector changes dramatically and the yarn macro stress state lags or responds abnormally, it is determined that the coupling relationship is unstable, and the breakage risk probability value is increased; Step S5, when the breakage risk probability value exceeds a preset threshold, a breakage warning signal is generated and output.

[0007] The application also provides a breakage dynamic capture system combining reinforcement learning and physical modeling, the system comprising: a yarn microstructure evolution model construction module for constructing a yarn microstructure evolution model capable of simulating the dynamic changes of the yarn surface fiber friction coefficient and the fiber interlocking force caused by the fluctuation of environmental parameters; wherein the environmental parameters include temperature, humidity and air flow speed; an environmental parameter processing and feature extraction module for inputting real-time environmental parameters of a textile workshop into the microstructure evolution model to predict real-time microstructure parameters of the yarn; performing nonlinear coupling feature extraction on the real-time environmental parameters to calculate an environment coupling feature vector including a temperature and humidity synergy index and an air flow disturbance intensity spectrum; a yarn state prediction and fusion module for inputting the real-time microstructure parameters as material properties into a yarn tension dynamics equation to obtain a predicted yarn macro stress state; splicing the yarn macro stress state and the environment coupling feature vector to form a state observation value of a reinforcement learning agent; a reinforcement learning agent warning module for the reinforcement learning agent to output a breakage risk probability value according to the state observation value; wherein the reinforcement learning agent learns through training that when the environment coupling feature vector changes dramatically and the yarn macro stress state lags or responds abnormally, it is determined that the coupling relationship is unstable, and the breakage risk probability value is increased; a warning signal generation and output module for generating and outputting a breakage warning signal when the breakage risk probability value exceeds a preset threshold.

[0008] This invention constructs an evolutionary model from environmental parameters to yarn microstructure, inputs real-time micro parameters as dynamic attributes into the macroscopic tension equation, and constructs state observations by combining environmental coupling characteristics. It utilizes a reinforcement learning agent to identify instability precursors of environmental disturbances and yarn response mismatch, thereby achieving early and accurate warning of yarn breakage in complex dynamic environments. Attached Figure Description

[0009] Figure 1 This is a flowchart illustrating a method for dynamic capture of decapitation combining reinforcement learning and physical modeling, as disclosed in an embodiment of the present invention. Figure 2 This is a data flow diagram of a decapitation dynamic capture method combining reinforcement learning and physical modeling disclosed in an embodiment of the present invention; Figure 3 This is a schematic diagram of a decapitation dynamic capture system that combines reinforcement learning and physical modeling, as disclosed in an embodiment of the present invention. Detailed Implementation

[0010] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0011] Please see Figure 1 , Figure 2 This invention provides a method for dynamic capture of decapitation combining reinforcement learning and physical modeling, comprising the following steps: Step S1: Construct a yarn microstructure evolution model that can simulate the dynamic changes in the friction coefficient of the yarn surface fibers and the cohesion between fibers caused by fluctuations in environmental parameters; wherein, the environmental parameters include temperature, humidity and airflow velocity; In this step, in the high-speed textile production environment, yarn breakage often does not occur suddenly, but is the result of the gradual deterioration of its internal structure under continuous or drastic fluctuations in environmental parameters such as temperature, humidity, and airflow. Traditional macroscopic sensors (such as tension sensors) can only capture the state changes at the moment of breakage or near breakage, but cannot detect this gradual microscopic deterioration process.

[0012] Therefore, this invention constructs a yarn microstructure evolution model, which is a mathematical model based on the principles of materials mechanics, tribology, and textile technology. This model is used to quantitatively simulate how dynamic fluctuations in environmental parameters (temperature, humidity, airflow velocity) affect the key intrinsic properties of yarn. Specifically: Temperature affects the modulus, elasticity and thermal expansion of fiber materials. Temperature rise can cause the fiber to soften, reducing its rigidity; while temperature fluctuations can cause changes in the friction state between fibers and fibers, fibers and yarn guiding devices.

[0013] Humidity affects the moisture absorption of fibers, thereby changing the diameter, strength and friction characteristics of the fibers. Increased humidity can cause the fibers to swell, increasing the cohesion between the fibers, but too high humidity can also make the yarn heavy and prone to sticking.

[0014] Air flow in the workshop (such as air conditioning wind, turbulence generated by equipment operation) will cause continuous, random aerodynamic disturbances to the running yarn, which directly translates into lateral forces on the yarn, affecting its running trajectory and dynamic tension.

[0015] The yarn microstructure evolution model takes the above environmental parameters as input, dynamically calculates and outputs microstructure parameters including yarn surface fiber friction coefficient and inter-fiber cohesion by built-in physical laws (for example, fiber diameter change equation considering moisture absorption swelling, Coulomb friction coefficient correction formula based on temperature and humidity conditions, stripping effect model of air flow disturbance on surface fibers, etc.). Among them, the yarn surface fiber friction coefficient reflects the sliding friction state between the yarn and key contact parts such as guide hooks, rings, and wire loops, directly affecting the running resistance and wear rate of the yarn; the inter-fiber cohesion reflects the tightness of the mutual entanglement and combination of the fibers inside the yarn, determining the internal strength of the yarn and the ability to resist disintegration under dynamic load.

[0016] Step S2, input the real-time environmental parameters of the textile workshop into the microstructure evolution model to predict the real-time microstructure parameters of the yarn; nonlinear coupling feature extraction is performed on the real-time environmental parameters to calculate the environmental coupling feature vector including the temperature and humidity synergy index and the air flow disturbance intensity spectrum; In this step, the temperature T(t), humidity H(t), air flow velocity V(t) and other data collected by the textile workshop sensor network in real time are directly input into the aforementioned yarn microstructure evolution model. Based on the current environmental parameter input, the yarn microstructure evolution model calculates and outputs the corresponding real-time microstructure parameters of the yarn in real time, i.e. the real-time yarn surface fiber friction coefficient and the inter-fiber cohesion .

[0017] The influence of environmental factors on the yarn is not independent linear superposition, but there is a complex nonlinear coupling effect. For example, the synergistic effect of high temperature and high humidity can be much higher than the simple addition of the two individual effects; air flow disturbance of a certain frequency can resonate with the natural frequency of the yarn. Therefore, in this step, the original environmental parameters are processed in depth to extract the environmental coupling feature vector, as follows: Temperature and humidity synergy index: This index quantifies the intensity and pattern of synergistic fluctuations by calculating the joint fluctuation entropy and phase space trajectory correlation of temperature and humidity time series. A high temperature and humidity synergy index indicates that temperature and humidity are undergoing drastic and synchronous changes, creating strong composite stress on the yarn structure.

[0018] Airflow disturbance intensity spectrum: By performing time-frequency analysis (such as wavelet transform) on the airflow velocity signal, the energy distribution along the frequency dimension is obtained. The focus is on analyzing the energy intensity within key frequency bands (e.g., 0.5-5Hz) close to the natural frequency of yarn lateral vibration. It can be understood that this airflow disturbance intensity spectrum characterizes the dangerous components in the airflow disturbance that may trigger yarn resonance.

[0019] Step S3: Input the real-time microstructure parameters as material properties into the yarn tension dynamics equation to solve for the predicted macroscopic stress state of the yarn; concatenate the macroscopic stress state of the yarn with the environmental coupling feature vector to form the state observation value of the reinforcement learning agent; In this step, the real-time microstructure parameters predicted in step S2 are... , As a key material property, it is input into the preset yarn tension dynamics equation. This yarn tension dynamics equation is a differential equation describing the forces acting on the yarn along the spinning path, typically considering inertial forces, damping forces, elastic restoring forces, and external excitation forces. Its expression is as follows:

[0020] in, —Inertial force term. The equivalent mass of the yarn within the monitoring interval (a constant determined by the yarn linear density and the interval length); This is the acceleration of the yarn's transverse (or axial) vibration. This reflects the inertial effect caused by changes in the yarn's motion.

[0021] —Damping force term. The speed of yarn vibration; The damping coefficient is input as a real-time microstructure parameter, i.e., the fiber friction coefficient. and interfiber cohesion .

[0022] It should be noted that the damping of yarn mainly comes from the frictional slippage between the internal fibers (and) , (Strong correlation) and friction with the yarn guiding device. By adjusting the microstructure parameters , Substitute into function , which realizes the direct mapping of the real-time variation of the internal state of the yarn to the macro-dynamic damping characteristics. For example, the cohesion force The decrease may lead to the aggravation of the inter-fiber slip, which is manifested as the change of the damping coefficient .

[0023] the elastic restoring force term. is the displacement of the yarn relative to the equilibrium position; is the stiffness coefficient (or elastic coefficient), which is also a dynamic parameter, i.e., a function of the inter-fiber cohesion force .

[0024] It should be noted that the transverse or axial stiffness of the yarn is largely dependent on the tightness of the mutual combination of the fibers (i.e., the cohesion force ). The decrease of the cohesion force leads to the decrease of the ability of the yarn to resist deformation, which is manifested as the decrease of the stiffness coefficient . It can be understood that this design directly links the micro-structure strength to the macro-mechanical stiffness.

[0025] the external excitation force term. is the calculated environmental coupling feature vector, which includes the temperature and humidity synergy index, the airflow disturbance intensity spectrum, etc.

[0026] It should be noted that the right side of the equation is a complex function related to the real-time environmental coupling features. For example, the energy of a specific frequency in the airflow disturbance intensity spectrum can be used as the excitation amplitude input at this frequency. In this way, it can be ensured that the source of the model driving force corresponds to the complex disturbance in the real workshop.

[0027] After solving the above yarn tension dynamics equation by numerical methods (such as the Runge-Kutta method), the predicted yarn macro-force state is obtained. This state is represented by a set of vectors , wherein is the dynamic tension prediction value, is the vibration amplitude prediction value, is the characteristic frequency, etc. It can be understood that this yarn macro-force state reflects the mechanical behavior that the yarn system should exhibit under the joint action of the current real-time micro-structure parameters , and the environmental coupling excitation .

[0028] Then, the macro-force state vector is combined with the environmental coupling feature vector (such as , wherein is the temperature and humidity synergy index, is the specific frequency band energy feature of the airflow disturbance intensity spectrum) according to the dimension to form a higher-dimensional state observation value .

[0029] Step S4, the reinforcement learning agent outputs a broken-end risk probability value according to the state observation value; wherein the reinforcement learning agent learns through training that when the environment coupling feature vector changes dramatically and the yarn macro stress state appears to be lagging or abnormal response, it is determined that the coupling relationship is unstable, and the broken-end risk probability value is increased; In this step, before being deployed to the actual production system, the reinforcement learning agent needs to complete sufficient learning in the training environment composed of historical production data sets, high-fidelity yarn dynamic simulation platform and expert annotated risk labels. The training goal is to learn the optimal strategy to maximize the long-term cumulative reward. Among them, the reinforcement learning agent can use deep deterministic policy gradient (DDPG) or proximal policy optimization (PPO) algorithm to construct.

[0030] It should be noted that in the training process of the reinforcement learning agent, the reward function is designed to guide the agent to capture key precursors of instability. Specifically, a small amount of positive reward is given in the stable running state; when it is monitored that the environment coupling feature vector (such as the temperature and humidity synergy index, the airflow disturbance intensity spectrum key band energy) jumps dramatically (for example, its rate of change exceeds the threshold ), and at the same time, the macro stress state (such as the tension prediction value ) obtained from the yarn tension dynamics equation compared with the stable baseline state appears significant lag (for example, the response time delay exceeds ) or shows abnormal fluctuations (for example, the fluctuation variance increases sharply), a large amount of negative reward is given. This reward mechanism can force the agent to consider the dynamic matching relationship between environmental input and system response when making decisions.

[0031] When the reinforcement learning agent makes online decisions, the state observation value output by step S3 is input into its internal network, which outputs a scalar value, i.e. the broken-end risk probability value . It can be understood that the core calculation logic of this broken-end risk probability value is derived from the learning in the training: The agent continuously analyzes the relevance of the environment feature part and the macro state part in the state observation value. When it identifies that the environment feature part changes dramatically, while the macro state part presents lag or abnormal pattern, it is determined that the environment-structure-mechanical coupling relationship of the current yarn system is in an unstable state, which is a high-risk precursor of broken-end. Therefore, the agent will dynamically increase the output risk probability value . For example, when the energy of the airflow disturbance intensity spectrum in the 1-3Hz frequency band suddenly increases by 50%, and the predicted yarn tension does not respond synchronously but appears low-frequency oscillation, It rose rapidly from 0.1 to over 0.7.

[0032] Step S5: When the probability value of decapitation risk exceeds a preset threshold, a decapitation warning signal is generated and output.

[0033] In this step, the real-time decapitation risk probability value is output to the agent. Continuous monitoring will be conducted, once The value exceeds the preset threshold This immediately triggers a head breakage warning signal. (Preset threshold) It can be flexibly configured according to different yarn types, process speeds and quality requirements. For example, it can be calibrated based on the statistical distribution of risk probability values ​​in the retrospective analysis of historical yarn breakage events (such as setting it to the 95th percentile of the risk probability distribution) and combined with the tolerance for false alarms and missed alarms in actual production.

[0034] It is understandable that the generated decapitation warning signal is a structured data packet, which may contain information such as: Warning level: According to Exceeding The severity level is divided into different grades, such as Attention, Warning, and Severe. Spindle / Yarn Path Identifier: Precisely indicates the location where the risk of yarn breakage occurs. Timestamp: The precise time the warning was triggered. Key Data Snapshot: Real-time environmental parameters at the moment the warning was triggered (…). ), predicted microstructure parameters ( ), predicted macroscopic stress state ( and the final risk probability value wait.

[0035] In addition, the generated early warning signals are output through a communication interface, enabling multi-level responses. For example, the signals can drive audible and visual alarms installed near the corresponding spindles or be pushed to the operator's smart handheld terminal, achieving precise on-site alarms and manual intervention guidance; the signals can be transmitted to the main control system of the spinning machine, automatically triggering preset deceleration programs, making minor adjustments to process parameters (such as adjusting the draft ratio), or activating backup cleaning devices to attempt intervention and adjustment before the spindle breakage occurs; the signals are simultaneously uploaded to the workshop manufacturing execution system (MES) or cloud data platform for production quality traceability, equipment health status analysis, and big data analysis for process optimization, forming a closed-loop improvement.

[0036] This invention constructs an evolutionary model from environmental parameters to yarn microstructure, inputs real-time micro parameters as dynamic attributes into the macroscopic tension equation, and constructs state observations by combining environmental coupling characteristics. It utilizes a reinforcement learning agent to identify instability precursors of environmental disturbances and yarn response mismatch, thereby achieving early and accurate warning of yarn breakage in complex dynamic environments.

[0037] As an example, the nonlinear coupling feature extraction is performed on the real-time environmental parameters, and the environmental coupling feature vector including the temperature and humidity coordination index and the air flow disturbance intensity spectrum is calculated, including: In step S21, the joint fluctuation entropy and the average trajectory divergence are respectively calculated by analyzing the joint probability distribution of the temperature time series and the humidity time series and the dynamic trajectory in the reconstructed phase space, and the temperature and humidity coordination index is generated by weighted fusion of the two. In this step, the joint probability distribution is constructed for the temperature time series and the humidity time series collected in the monitoring time window. Specifically, the temperature value and the humidity value are divided into several discrete intervals, respectively, the frequency of the value combination falling in each joint interval is counted, and the temperature-humidity joint probability distribution table is formed. Based on the above joint probability distribution, the Shannon entropy is calculated as the joint fluctuation entropy representing the chaotic degree of the temperature and humidity joint state. The lower the entropy value is, the more concentrated the temperature and humidity state combination is, and the stronger the coordination is; the higher the entropy value is, the more dispersed the state combination is, and the weaker the coordination is.

[0038] For example: the temperature sequence T=[t1, t2,..., t N ] of N consecutive sampling times is obtained, and the humidity sequence H=[h1, h2,..., h N ] of N consecutive sampling times is obtained, and normalized processing is performed: , wherein μ T , μ H is the mean value, and σ T , σ H is the standard deviation.

[0039] The normalized temperature and humidity values are discretized, respectively. Let the temperature be divided into M T intervals, and the humidity be divided into M H intervals. The probability of the temperature value falling in the i-th interval and the humidity value falling in the j-th interval is calculated, i.e. the joint probability distribution:

[0040] Further, the joint Shannon entropy is calculated as the joint fluctuation entropy JWE: .

[0041] At the same time, the phase space reconstruction technology in time series analysis is used to reconstruct the dynamic system phase space of the temperature sequence and the humidity sequence, respectively, to obtain the respective dynamic evolution trajectories. The distance between the corresponding points on the two trajectories is calculated, and the average value in the whole time window is calculated to obtain the average trajectory divergence. The smaller the divergence value is, the more similar the dynamic change mode of temperature and humidity is.

[0042] Exemplarily, the phase space is reconstructed from the temperature sequence by using delay-embedding method, and a phase space vector of the temperature sequence is obtained by setting embedding dimension m and delay time τ: Similarly, a phase space vector of the humidity sequence is reconstructed .

[0043] For each time point k, the cosine distance between the phase space vectors of the temperature sequence and the humidity sequence is calculated: .

[0044] The average trajectory divergence is calculated: , wherein is the number of reconstructed phase space points.

[0045] Finally, the joint fluctuation entropy obtained in the foregoing and the average trajectory divergence are weighted and summed to generate a comprehensive temperature and humidity coordination index , wherein the weight coefficients , are determined through historical broken data regression analysis and satisfy It can be understood that the higher the temperature and humidity coordination index, the more intense and highly synchronized the composite change of temperature and humidity is.

[0046] In step S22, time-frequency energy distribution is obtained by performing time-frequency transformation on the airflow velocity sequence, and energy features corresponding to a key frequency band are extracted from the time-frequency energy distribution according to an inherent frequency range determined by the yarn tension dynamics equation, so as to constitute the airflow disturbance intensity spectrum.

[0047] As an example, the inherent frequency range is obtained by performing eigenvalue analysis on the yarn tension dynamics equation.

[0048] In this step, the airflow velocity time sequence collected in real time is analyzed by time-frequency transformation, and continuous wavelet transformation is preferably used. By continuous wavelet transformation, the one-dimensional airflow velocity signal is converted into a two-dimensional time-frequency energy distribution diagram, which can clearly show the disturbance energy intensity at different times and different frequency components.

[0049] Exemplarily, the airflow velocity sequence V=[v1, v2,..., v N ] is continuously wavelet-transformed as follows: , wherein is a Morlet wavelet basis function, a is a scale parameter (corresponding to frequency), and b is a translation parameter (corresponding to time).

[0050] To focus on the disturbance components that significantly affect yarn vibration, it is necessary to determine the natural frequency range of the yarn system. In this embodiment, the natural frequency range is obtained by linearizing the yarn tension dynamics equation and solving its eigenvalues, the frequencies corresponding to which are the main natural frequencies of the yarn system.

[0051] Illustratively: based on the yarn tension dynamics equation, linearization is performed near the equilibrium point: , where , is the damping and stiffness coefficient of the yarn in the reference state. Solve the characteristic equation: to obtain the eigenvalues: , where the imaginary part is the angular frequency .

[0052] Thus, the natural frequency of the system is: . Take the frequency range as [f0-Δf, f0+Δf], where Δf can be set to 10%-20% of f0 according to the type of yarn.

[0053] Based on the obtained natural frequency range, from the above time-frequency energy distribution diagram, the energy features in the corresponding frequency band (for example, a plurality of key frequency bands formed by expanding a certain bandwidth above and below the center of each natural frequency) are extracted, which can be the average energy, peak energy or energy variance of each key frequency band in the monitoring period. Organize these extracted energy features by frequency band, which constitutes the airflow disturbance intensity spectrum. It can be understood that the airflow disturbance intensity spectrum directly quantifies the intensity of the dangerous frequency components that may excite yarn resonance in the airflow disturbance.

[0054] Illustratively: convert the wavelet transform result W(a, b) to time-frequency energy distribution S(f, t)=|W(a, b)| 2 , where the relationship between frequency f and scale a is f=f c / a, f c is the center frequency of the wavelet.

[0055] Within the natural frequency range [f0-Δf, f0+Δf], divide the frequency band into K sub-bands, and the center frequency of each sub-band is f k . Calculate the energy features of each sub-band. Take the kth sub-band as an example, calculate: average energy: ; peak energy: ; energy variance: .

[0056] The airflow disturbance intensity spectrum is represented in the form of a vector as follows: .

[0057] Combine the above calculation results into the environmental coupling feature vector: .

[0058] As an example, when the environmental coupling feature vector changes dramatically and the yarn macroscopic stress state appears to lag or respond abnormally, it is determined that the coupling relationship is unstable, including: Real-time calculation of the change rate of the environmental coupling feature vector in adjacent time windows, when the change rate of the temperature and humidity synergy index exceeds the first threshold value, and / or the energy change rate of any key frequency band in the air flow disturbance intensity spectrum exceeds the second threshold value, it is determined that the environmental coupling feature vector changes dramatically; In this step, the change rate of each component in the environmental coupling feature vector is calculated in real time in fixed time windows (such as 10 seconds). Specifically, the relative change rate or absolute difference rate between the feature values of the current time window and the previous time window is calculated.

[0059] For the temperature and humidity synergy index , set the first threshold value . When the change rate in unit time ( ) exceeds , it is determined that the temperature and humidity synergy has a dramatic jump. It can be understood that the first threshold value can be calibrated according to the index fluctuation range of extreme weather events or the start and stop of the air conditioning system in historical data.

[0060] For the air flow disturbance intensity spectrum , set the second threshold value . Check the change rate of the energy (such as average energy) of any key frequency band in the vector. If the energy change rate of any frequency band exceeds , it is determined that the air flow disturbance intensity changes dramatically. The second threshold value can be set according to the typical frequency spectrum energy fluctuation caused by sudden changes in fan speed or opening and closing of doors and windows.

[0061] When the energy of the temperature and humidity synergy index or any key frequency band in the air flow disturbance intensity spectrum is determined to change dramatically, it is determined that the environmental coupling feature vector changes dramatically.

[0062] Compare the yarn macroscopic stress state obtained by solving the yarn tension dynamics equation with the reference macroscopic stress state calculated under stable environmental conditions based on the yarn tension dynamics equation, when the dynamic tension response in the yarn macroscopic stress state lags more than the preset delay threshold value, and / or the vibration amplitude exceeds the preset safe amplitude range, it is determined that the yarn macroscopic stress state lags or responds abnormally; In this step, a set of baseline macro stress states, denoted as , are solved by inputting steady-state environmental parameters and corresponding baseline microstructure parameters into yarn tension dynamics equations during system initial calibration or stable production period, representing the healthy mechanical performance of yarn under no severe disturbance.

[0063] The time interval experienced by the yarn from the moment when the environmental characteristics are determined to be severely changed to the moment when the yarn macro stress state (such as dynamic tension ) starts to deviate significantly (e.g., deviates from the baseline value by more than 3%) is defined as the response delay time . If exceeds the preset delay threshold (e.g., 1.5 times the normal thermal inertia response time), it is determined that the yarn macro stress state response is lagging.

[0064] If the vibration amplitude exceeds the safety amplitude range set based on the baseline amplitude and a safety factor (e.g., 1.8 times), or the characteristic frequency deviates from the baseline frequency by a certain percentage (e.g., 15%), it is determined that the yarn macro stress state is abnormal.

[0065] When the yarn macro stress state meets any one or more of the conditions of response lag or response abnormality, the flag of macro stress state lag or abnormal response is triggered.

[0066] When the above two determination conditions are met at the same time, the reinforcement learning agent determines that the coupling relationship is unstable.

[0067] The reinforcement learning agent continuously monitors the above two determination flags. When and only when the environmental coupling feature vector is severely changed and the yarn macro stress state is lagging or responding abnormally, both flags are triggered at the same time or in adjacent monitoring time windows, the agent finally determines that the environment-structure-mechanical coupling relationship of the current yarn system is unstable, i.e., it is determined to be a precursor of yarn breakage.

[0068] As an example, the reinforcement learning agent is pre-trained in the following way: Based on historical production data, a training data set is constructed, where each training sample contains a sequence of continuous state observations in a historical time period and a corresponding breakage result label; Design a training reward function: give a basic positive reward at each step without breakage, give a negative reward when the environmental coupling feature vector is severely changed and the yarn macro stress state is lagging or responding abnormally, and give a large penalty when the breakage finally occurs; Based on the training data set, the reinforcement learning agent is trained to maximize cumulative rewards until its policy converges.

[0069] In this embodiment, historical production data of the textile workshop is collected and pre-processed, including environmental sensor data (temperature, humidity, air flow speed), yarn state data (obtained indirectly or generated by simulation), and key breakage event records. The historical production data is divided into a plurality of training samples in chronological order, each training sample containing a complete information sequence within a continuous time window (e.g. 60 seconds before the breakage occurs), as follows: State observation sequence: each point in this sequence contains the complete state vector obtained by concatenating the yarn macro stress state prediction value and the environmental coupling feature vector at the corresponding time Breakage result label: whether there is a breakage at the end of the time window; for breakage samples, the window end is marked as "breakage"; for normal samples, the entire window is marked as "normal". The constructed sample set is divided into a training set, a validation set and a test set for model training, tuning and performance evaluation.

[0070] The specific design of the reward function directly determines the quality of the policy learned by the agent. The reward function in this embodiment is At each time step calculated according to the following rules: Basic positive reward: if no breakage occurs at this time step, a small constant positive reward (e.g. +0.1) is given to encourage the agent to survive longer under normal circumstances.

[0071] Risk pattern negative reward: when it is detected that the current state satisfies the unstable pattern of the coupling relationship defined above (i.e. the environmental features change dramatically while the macro stress state lags or is abnormal), a larger negative reward (e.g. -1) is given to force the agent to actively identify and be vigilant about such high-risk states even if a breakage has not yet occurred.

[0072] Breakage termination penalty: if a breakage occurs at this time step, a very large negative reward is given as a penalty (e.g. -10) and the current training round is terminated, so that the agent must prioritize avoiding breakage as the highest priority.

[0073] A deep reinforcement learning algorithm suitable for continuous state and action space, such as Proximal Policy Optimization (PPO), Deep Deterministic Policy Gradient (DDPG) or Soft Actor-Critic (SAC) algorithm, can be used to construct and train the agent. The goal of the agent is to learn a policy such that the cumulative discounted reward obtained from the environment is maximized when interacting with the simulated environment constructed from the training dataset wherein is a discount factor (0 < γ < 1).

[0074] The agent learns from successes (positive rewards) and failures (negative rewards) through repeated trials. In particular, through extensive exposure to intermediate states marked by negative rewards of risky patterns, the agent is forced to learn that, in the early stage of environment perturbation and mismatch of mechanical response, even if the absolute value of tension, etc. has not yet reached a dangerous level, it is a strong indication of a high risk of breakage. It can be understood that the above learning process is the quantification and internalization process of the experience of human experts.

[0075] The training continues until the policy of the agent tends to be stable. The criteria for convergence can be that the performance evaluated on the validation set (such as the accuracy and recall of breakage prediction) no longer significantly improves, or the average cumulative reward curve obtained by the agent in the training environment tends to be stable.

[0076] Please refer to Figure 3 The embodiment of the present application also provides a breakage dynamic capture system 200 combining reinforcement learning and physical modeling, the system comprising: A yarn microstructure evolution model construction module 2001 is configured to: construct a yarn microstructure evolution model capable of simulating the dynamic changes of the friction coefficient of the surface fibers and the inter-fiber cohesion force of the yarn caused by fluctuations in environmental parameters; wherein the environmental parameters include temperature, humidity and air flow speed; An environmental parameter processing and feature extraction module 2002 is configured to: input the real-time environmental parameters of the textile workshop into the microstructure evolution model to predict the real-time microstructure parameters of the yarn; and perform nonlinear coupling feature extraction on the real-time environmental parameters to calculate an environmental coupling feature vector including a temperature and humidity synergy index and an air flow disturbance intensity spectrum; A yarn state prediction and fusion module 2003 is configured to: input the real-time microstructure parameters as material properties into a yarn tension dynamics equation to solve for a predicted yarn macroscopic stress state; and splice the yarn macroscopic stress state and the environmental coupling feature vector to form a state observation value of a reinforcement learning agent; A reinforcement learning agent early warning module 2004 is configured to: output a breakage risk probability value by the reinforcement learning agent according to the state observation value; wherein the reinforcement learning agent learns through training that when the environmental coupling feature vector changes dramatically while the yarn macroscopic stress state lags or responds abnormally, it is determined that the coupling relationship is unstable, and the breakage risk probability value is increased; An early warning signal generation and output module 2005 is configured to: generate and output a breakage early warning signal when the breakage risk probability value exceeds a preset threshold.

[0077] As an example, the environment parameter processing and feature extraction module 2002 is specifically configured to: Calculate the joint fluctuation entropy and the average trajectory divergence by analyzing the joint probability distribution of the temperature time series and the humidity time series and the dynamic trajectory in the reconstructed phase space, respectively, and generate the temperature and humidity coordination index by weighted fusion of the two; Obtain the time-frequency energy distribution by performing time-frequency transformation on the air flow velocity sequence, and extract the energy features corresponding to the key frequency bands from the time-frequency energy distribution according to the inherent frequency range determined by the yarn tension dynamics equation, to constitute the air flow disturbance intensity spectrum.

[0078] As an example, the inherent frequency range is obtained by eigenvalue analysis on the yarn tension dynamics equation.

[0079] As an example, the reinforcement learning intelligent agent warning module 2004 is specifically configured to: Calculate the change rate of the environment coupling feature vector in adjacent time windows in real time, and determine that the environment coupling feature vector changes dramatically when the change rate of the temperature and humidity coordination index exceeds a first threshold value, and / or the energy change rate of any key frequency band in the air flow disturbance intensity spectrum exceeds a second threshold value; Compare the yarn macroscopic stress state obtained based on the yarn tension dynamics equation with the baseline macroscopic stress state calculated based on the yarn tension dynamics equation under stable environmental conditions, and determine that the yarn macroscopic stress state has hysteresis or abnormal response when the dynamic tension response delay in the yarn macroscopic stress state exceeds a preset delay threshold value, and / or the vibration amplitude exceeds a preset safe amplitude range; When the above two determination conditions are met at the same time, the reinforcement learning intelligent agent determines that the coupling relationship is unstable.

[0080] As an example, the reinforcement learning intelligent agent is pre-trained in the following way: Construct a training data set based on historical production data, wherein each training sample contains a sequence of continuous state observations in a historical time period and a corresponding break result label; Design a training reward function: give a basic positive reward at each step without break, give a negative reward when the environment coupling feature vector changes dramatically and the yarn macroscopic stress state has hysteresis or abnormal response, and give a large penalty when the break finally occurs; Based on the training data set, the reinforcement learning intelligent agent learns to maximize the cumulative reward until its strategy converges.

[0081] The above merely describes some embodiments of the present application, and does not limit the patent scope of the present application. Any equivalent structural transformation, direct / indirect application in other related technical fields, or the like, which is made based on the technical concept of the present application and the content of the specification and drawings, is included in the patent protection scope of the present application.

Claims

1. A method for dynamically capturing decapitations by combining reinforcement learning and physical modeling, characterized in that, The methods and steps include the following: Step S1: Construct a yarn microstructure evolution model that can simulate the dynamic changes in the friction coefficient of the yarn surface fibers and the cohesion between fibers caused by fluctuations in environmental parameters; wherein, the environmental parameters include temperature, humidity and airflow velocity; Step S2: Input the real-time environmental parameters of the textile workshop into the microstructure evolution model to predict the real-time microstructure parameters of the yarn; perform nonlinear coupling feature extraction on the real-time environmental parameters and calculate the environmental coupling feature vector including the temperature and humidity synergy index and the airflow disturbance intensity spectrum. Step S3: Input the real-time microstructure parameters as material properties into the yarn tension dynamics equation to solve for the predicted macroscopic stress state of the yarn; concatenate the macroscopic stress state of the yarn with the environmental coupling feature vector to form the state observation value of the reinforcement learning agent; Step S4: The reinforcement learning agent outputs the probability value of the yarn breakage risk based on the state observation value; wherein, the reinforcement learning agent learns through training that when the environmental coupling feature vector changes drastically and the macroscopic stress state of the yarn shows a lag or abnormal response, it determines that the coupling relationship is unstable and increases the probability value of the yarn breakage risk. Step S5: When the probability value of decapitation risk exceeds a preset threshold, a decapitation warning signal is generated and output.

2. The method for dynamic capture of decapitation combining reinforcement learning and physical modeling according to claim 1, characterized in that: Nonlinear coupling feature extraction is performed on real-time environmental parameters to calculate environmental coupling feature vectors, including temperature and humidity synergy index and airflow disturbance intensity spectrum, including: Step S21: By analyzing the joint probability distribution of the temperature time series and the humidity time series and their dynamic trajectories in the reconstructed phase space, the joint fluctuation entropy and the average trajectory divergence are calculated respectively, and the two are weighted and fused to generate the temperature and humidity synergy index. Step S22: The time-frequency energy distribution is obtained by performing time-frequency transformation on the airflow velocity sequence, and the energy characteristics of the corresponding key frequency bands are extracted from the time-frequency energy distribution according to the natural frequency range determined by the yarn tension dynamics equation, so as to form the airflow disturbance intensity spectrum.

3. The method for dynamic capture of decapitation combining reinforcement learning and physical modeling according to claim 2, characterized in that: The inherent frequency range is obtained by performing eigenvalue analysis on the yarn tension dynamics equation.

4. The method for dynamic capture of decapitation combining reinforcement learning and physical modeling according to claim 1, characterized in that: When the environmental coupling characteristic vector changes drastically and the macroscopic stress state of the yarn exhibits a lag or abnormal response, the coupling relationship is determined to be unstable, including: The rate of change of the environmental coupling feature vector in adjacent time windows is calculated in real time. When the rate of change of the temperature and humidity synergy index exceeds the first threshold, and / or the rate of change of energy in any key frequency band of the airflow disturbance intensity spectrum exceeds the second threshold, it is determined that the environmental coupling feature vector has changed drastically. The macroscopic stress state of the yarn obtained by solving the yarn tension dynamics equation is compared with the benchmark macroscopic stress state calculated under stable environmental conditions based on the yarn tension dynamics equation. When the dynamic tension response delay in the macroscopic stress state of the yarn exceeds a preset delay threshold and / or the vibration amplitude exceeds a preset safe range, it is determined that the macroscopic stress state of the yarn has a hysteresis or abnormal response. When both of the above conditions are met, the decision coupling relationship of the reinforcement learning agent is unstable.

5. The method for dynamic capture of decapitation combining reinforcement learning and physical modeling according to claim 1, characterized in that: Reinforcement learning agents are pre-trained in the following ways: A training dataset is constructed based on historical production data, where each training sample contains a sequence of continuous state observations for a historical time period and the corresponding label of the decapitation result. Design the training reward function: give a basic positive reward at each step where no yarn breakage occurs, give a negative reward when the environmental coupling feature vector changes drastically and the macroscopic stress state of the yarn shows a lag or abnormal response, and give a large penalty at the step where yarn breakage finally occurs. Based on the training dataset, the reinforcement learning agent learns with the goal of maximizing cumulative reward until its policy converges.

6. A decapitation dynamic capture system combining reinforcement learning and physical modeling, characterized in that: The system includes: The yarn microstructure evolution model construction module is used to: construct a yarn microstructure evolution model that can simulate the dynamic changes in the friction coefficient of the yarn surface fibers and the cohesion between fibers caused by fluctuations in environmental parameters; wherein, the environmental parameters include temperature, humidity and airflow velocity; The environmental parameter processing and feature extraction module is used to: input the real-time environmental parameters of the textile workshop into the microstructure evolution model to predict the real-time microstructure parameters of the yarn; perform nonlinear coupling feature extraction on the real-time environmental parameters, and calculate the environmental coupling feature vector including the temperature and humidity synergy index and the airflow disturbance intensity spectrum. The yarn state prediction and fusion module is used to: input the real-time microstructure parameters as material properties into the yarn tension dynamics equation, solve for the predicted macroscopic stress state of the yarn; and concatenate the macroscopic stress state of the yarn with the environmental coupling feature vector to form the state observation value of the reinforcement learning agent. The reinforcement learning agent early warning module is used to: output a breakage risk probability value based on the state observation value; wherein, the reinforcement learning agent learns through training that: when the environmental coupling feature vector changes drastically and the macroscopic stress state of the yarn shows a lag or abnormal response, it determines that the coupling relationship is unstable and increases the breakage risk probability value. The warning signal generation and output module is used to generate and output a decapitation warning signal when the probability value of the decapitation risk exceeds a preset threshold.

7. A decapitation dynamic capture system combining reinforcement learning and physical modeling according to claim 6, characterized in that: The environmental parameter processing and feature extraction module is specifically used for: By analyzing the joint probability distribution of temperature time series and humidity time series and their dynamic trajectories in the reconstructed phase space, the joint fluctuation entropy and the average trajectory divergence are calculated respectively, and the two are weighted and fused to generate the temperature and humidity synergy index. The time-frequency energy distribution is obtained by performing time-frequency transformation on the airflow velocity sequence, and the energy characteristics of the corresponding key frequency bands are extracted from the time-frequency energy distribution based on the natural frequency range determined by the yarn tension dynamics equation to form the airflow disturbance intensity spectrum.

8. A decapitation dynamic capture system combining reinforcement learning and physical modeling according to claim 7, characterized in that: The inherent frequency range is obtained by performing eigenvalue analysis on the yarn tension dynamics equation.

9. A decapitation dynamic capture system combining reinforcement learning and physical modeling according to claim 6, characterized in that: The reinforcement learning agent early warning module is specifically used for: The rate of change of the environmental coupling feature vector in adjacent time windows is calculated in real time. When the rate of change of the temperature and humidity synergy index exceeds the first threshold, and / or the rate of change of energy in any key frequency band of the airflow disturbance intensity spectrum exceeds the second threshold, it is determined that the environmental coupling feature vector has changed drastically. The macroscopic stress state of the yarn obtained by solving the yarn tension dynamics equation is compared with the benchmark macroscopic stress state calculated under stable environmental conditions based on the yarn tension dynamics equation. When the dynamic tension response delay in the macroscopic stress state of the yarn exceeds a preset delay threshold and / or the vibration amplitude exceeds a preset safe range, it is determined that the macroscopic stress state of the yarn has a hysteresis or abnormal response. When both of the above conditions are met, the decision coupling relationship of the reinforcement learning agent is unstable.

10. A decapitation dynamic capture system combining reinforcement learning and physical modeling according to claim 6, characterized in that: Reinforcement learning agents are pre-trained in the following ways: A training dataset is constructed based on historical production data, where each training sample contains a sequence of continuous state observations for a historical time period and the corresponding label of the decapitation result. Design the training reward function: give a basic positive reward at each step where no yarn breakage occurs, give a negative reward when the environmental coupling feature vector changes drastically and the macroscopic stress state of the yarn shows a lag or abnormal response, and give a large penalty at the step where yarn breakage finally occurs. Based on the training dataset, the reinforcement learning agent learns with the goal of maximizing cumulative reward until its policy converges.

Images