Mold flow field non-uniformity early warning method based on multi-source time sequence characteristics and online reinforcement learning

Abstract

The present application relates to the technical field of steel metallurgy continuous casting, and is a crystallizer flow field unevenness early warning method based on multi-source time sequence characteristics and online reinforcement learning, comprising: obtaining operation data of the crystallizer, collecting thermocouple temperature values of adjacent two rows of the crystallizer wide-face copper plate, and calculating equivalent temperature; calculating transverse temperature asymmetry, liquid surface fluctuation energy ratio and two-side heat flow difference according to the operation data; constructing a state space according to the transverse temperature asymmetry, liquid surface fluctuation energy ratio and two-side heat flow difference, and dividing early warning levels; calculating a reward function based on the current state of the crystallizer; constructing a flow field state prediction model using a dual-network architecture combining a deep Q network and a policy network; and using the flow field state prediction model to early warn the flow field unevenness state of the crystallizer. The method can continuously learn the optimal flow deviation identification strategy in real-time interaction with the crystallizer environment and output the flow deviation direction and severity level.

Description

Technical Field

[0001] This invention relates to the field of continuous casting technology in iron and steel metallurgy, and in particular to a method for early warning of non-uniformity in the flow field of a crystallizer based on multi-source temporal characteristics and online reinforcement learning. Background Technology

[0002] The flow field distribution of molten steel within the crystallizer directly affects the uniform melting of the protective slag, the stability of the meniscus, the uniformity of billet shell growth, and the tendency for slag entrapment. Non-uniform flow field (also known as flow deviation) is a common anomaly in continuous casting production, mainly caused by nodule formation in the side holes of the submerged entry nozzle, asymmetrical nozzle erosion, nozzle alignment deviation, or abnormal electromagnetic stirring. Flow deviation can lead to violent fluctuations in the liquid surface on one side, uneven melting of the protective slag, and abnormal local heat flux density, and in severe cases, can cause slag entrapment, longitudinal cracks, or even steel leakage accidents.

[0003] Existing methods for detecting flow field non-uniformity can be mainly divided into three categories: The first category is based on statistical analysis of liquid surface fluctuations in the crystallizer. It determines flow deviation by detecting the amplitude, frequency, or level difference between the two narrow faces. This method is sensitive to sensor installation location and cannot distinguish between flow deviation and liquid surface fluctuations caused by other disturbances (such as changes in casting speed or nozzle replacement), resulting in a high false alarm rate. The second category is based on pattern recognition of thermocouple temperatures on the crystallizer copper plate. It infers whether the flow field is symmetrical by comparing the temperature difference between the thermocouples on both sides of the narrow or wide face. This method requires manual setting of the temperature difference threshold, and the threshold needs frequent adjustment depending on the steel grade and casting speed, resulting in poor adaptability. The third category is based on empirical formulas from numerical simulations or water models, such as using the F-number to assess liquid surface activity. However, the F-number calculation relies on multiple empirical coefficients, cannot reflect the direction and degree of flow deviation in real time, and cannot adapt to dynamic operating conditions. Chinese patent application CN119714590A discloses a flow field detection device, but it does not involve the directional identification of flow field non-uniformity and cannot adapt to changes in operating conditions online.

[0004] In summary, existing technologies lack a method that can detect flow field inhomogeneity in real time and adaptively and output the direction of flow deviation, especially in the absence of dual-sided liquid level sensors, where the temperature distribution of a wide-face thermocouple alone cannot be used for judgment. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a method for early warning of flow field inhomogeneity in crystallizers based on multi-source temporal features and online reinforcement learning. Based on invention patent application CN121535146A, this invention provides early warning of flow field inhomogeneity in crystallizers. This invention primarily solves the problems of fixed thresholds, inability to adapt to changes in operating conditions, and inability to output the direction of flow deviation in existing technologies. By constructing two new features—wide-area temperature lateral asymmetry and instantaneous liquid surface fluctuation energy ratio—and combining them with an online reinforcement learning framework, the agent continuously learns the optimal flow deviation identification strategy through real-time interaction with the crystallizer environment, and outputs the direction and severity level of the flow deviation.

[0006] The technical means employed in this invention are as follows: A method for early warning of flow field inhomogeneity in crystallizers based on multi-source temporal features and online reinforcement learning includes: acquiring crystallizer operating data and collecting thermocouple temperature values ​​of two adjacent rows of copper plates on the wide side of the crystallizer to calculate the equivalent temperature; calculating the lateral temperature asymmetry, liquid surface fluctuation energy ratio, and heat flow difference between the two sides based on the operating data; constructing a state space based on the lateral temperature asymmetry, liquid surface fluctuation energy ratio, and heat flow difference between the two sides; calculating a reward function based on the current state of the crystallizer, the reward function including: asymmetry rationality reward, consistency reward, and stability reward; constructing a flow field state prediction model using a dual-network architecture combining a deep Q-network and a policy network; and using the flow field state prediction model to provide early warning of the flow field inhomogeneity state of the crystallizer.

[0007] Furthermore, the operating data includes thermocouple temperature values, cooling water flow rate, inlet and outlet water temperature difference, crystallizer liquid level signal, immersion depth of the submersible nozzle, pulling speed, and electromagnetic stirring current; the heat flux density on the left and right sides is calculated based on the cooling water flow rate and the inlet and outlet water temperature difference, and the liquid level fluctuation amplitude and frequency are calculated based on the crystallizer liquid level signal; The crystallizer has 11 thermocouples per row on its wide copper plate, connecting adjacent... , The thermocouple temperature values ​​in the two rows are denoted as follows: and Calculate the equivalent temperature :

[0008] in, Indicates the first The temperature values ​​of each thermocouple in the row, Indicates the first The temperature values ​​of each thermocouple in the row.

[0009] Furthermore, the lateral temperature asymmetry... Represented as:

[0010] Among them, when When the value is 1-5, it represents the 5 thermocouples on the left side. When 7-11 is used, it indicates that there are 5 thermocouples on the right side. When, it indicates the intermediate thermocouple; The energy ratio of the liquid surface fluctuation During calculation, the liquid level signal Perform a fast Fourier transform within the time window to calculate the energy ratio between the low-frequency band and the mid-frequency band; The heat flow difference on both sides Through the heat flux density on the left and right sides , calculate:

[0011] The crystallizer condition is initially assessed based on the lateral temperature asymmetry, the energy ratio of liquid surface fluctuations, and the heat flow difference between the two sides.

[0012] Furthermore, the state space , This refers to the immersion depth of the submersible nozzle. The electromagnetic stirring current generates a two-dimensional action from the intelligent agent in the state space. ,in For the probability of skew flow, The value indicates the direction of flow deviation. A negative value indicates flow deviation to the left, a positive value indicates flow deviation to the right, and the absolute value indicates the severity of the flow deviation.

[0013] Furthermore, the asymmetric reasonableness reward is: if <0.05 and agent output <0.3, the asymmetric reasonableness reward value is 0.2; if >0.12 and If the asymmetry probability is greater than 0.6, the asymmetry rationality reward value is 0.3; if the flow deviation probability contradicts the transverse temperature asymmetry, the asymmetry rationality penalty value is 0.3; the consistency reward is based on the heat flow difference between the two sides. Energy ratio of liquid surface fluctuation Based on a comprehensive judgment of the actual deflection direction, if the agent outputs... If the direction sign is consistent with the actual direction, the consistency reward is 0.2; otherwise, the consistency penalty is 0.2. The stability reward is: if the direction sign of the deflection output remains unchanged for two consecutive outputs and If the change is less than 0.1, the stability reward value is 0.1; if the direction changes frequently, the stability penalty value is 0.1. The reward function is obtained by adding the reward values ​​from the asymmetric rationality reward, consistency reward, and stability reward, and subtracting the penalty value. .

[0014] Furthermore, the characteristic is that, after the flow field state prediction model completes one state transition in each cycle, it incorporates experience ( The data is stored in the experience replay pool; every two cycles, random samples are taken from the experience pool, and the network parameters are updated using the Adam optimizer; the policy network outputs the mean and variance of the actions, and the value network evaluates the value of the state and actions; the flow field state prediction model is stably trained using soft updates.

[0015] Furthermore, the early warning of the non-uniformity of the flow field in the crystallizer specifically includes: when two consecutive early warning cycles meet the following conditions... and When the flow deviation is detected, a flow deviation warning is issued, and the direction and severity of the flow deviation are simultaneously output.

[0016] Compared with the prior art, the present invention has the following advantages: This invention provides a method for early warning of flow field inhomogeneity in crystallizers based on multi-source temporal features and online reinforcement learning. Two key features are constructed from real-time continuous casting production data: the lateral temperature asymmetry, which characterizes the degree of asymmetry in the heat flow distribution across the wide face, and the fluctuation energy ratio, which characterizes the difference in energy fluctuation at the liquid surface. A dual-action output agent is designed to output the probability and direction of flow deviation simultaneously. Finally, the model is continuously updated with production data through an online reinforcement learning mechanism, adapting to different steel grades, casting speeds, and nozzle conditions.

[0017] This invention utilizes two dimensionless features, lateral temperature asymmetry and liquid surface fluctuation energy ratio, to characterize flow field non-uniformity from two dimensions: heat distribution and liquid surface dynamics. This approach is more robust than traditional single-point temperature difference or amplitude judgment.

[0018] This invention utilizes an intelligent agent to output two-dimensional actions, while simultaneously providing the probability and direction of flow deviation (including severity), which can directly guide operators in locating problems and overcome the shortcomings of existing methods that can only determine whether flow deviation has occurred.

[0019] The model in this invention updates every 4 seconds and can automatically adapt to working conditions such as steel grade change, casting speed adjustment, and changes in sprue immersion depth without manual parameter readjustment, and has a certain tolerance for sensor drift. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1This is a flowchart of the crystallizer flow field inhomogeneity early warning method based on multi-source temporal features and online reinforcement learning in this invention. Detailed Implementation

[0022] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0023] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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. The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit the present invention or its application or use. 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.

[0024] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0025] Unless otherwise specifically stated, the relative arrangement, numerical expressions, and values ​​of the components and steps described in these embodiments do not limit the scope of the invention. It should also be understood that, for ease of description, the dimensions of the various parts shown in the drawings are not drawn to actual scale. Techniques, methods, and devices known to those skilled in the art may not be discussed in detail, but where appropriate, such techniques, methods, and devices should be considered part of the specification. In all examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not as limitations. Therefore, other examples of exemplary embodiments may have different values. It should be noted that similar reference numerals and letters in the following figures denote similar items; therefore, once an item is defined in one figure, it need not be further discussed in subsequent figures.

[0026] like Figure 1As shown, this invention provides a method for early warning of flow field inhomogeneity in crystallizers based on multi-source temporal features and online reinforcement learning, including: acquiring the operating data of the crystallizer and collecting the thermocouple temperature values ​​of two adjacent rows of the wide copper plate of the crystallizer, and calculating the equivalent temperature; in a preferred embodiment of this invention, the operating data includes thermocouple temperature values, cooling water flow rate, inlet and outlet water temperature difference, crystallizer liquid level signal, immersion depth of the submersible nozzle, pulling speed, and electromagnetic stirring current; calculating the heat flux density on the left and right sides based on the cooling water flow rate and the inlet and outlet water temperature difference, and calculating the liquid level fluctuation amplitude and frequency through the crystallizer liquid level signal.

[0027] The crystallizer's wide copper plate has 11 thermocouples per row, connecting adjacent... , The thermocouple temperature values ​​in the two rows are denoted as follows: and Calculate the equivalent temperature :

[0028] in, Indicates the first The temperature values ​​of each thermocouple in the row, Indicates the first The temperature values ​​of each thermocouple in the row.

[0029] The transverse temperature asymmetry, liquid surface fluctuation energy ratio, and heat flow difference between the two sides are calculated based on operational data. In a preferred embodiment of this invention, the transverse temperature asymmetry is... Represented as:

[0030] Among them, when When the value is 1-5, it represents the 5 thermocouples on the left side. When 7-11 is used, it indicates that there are 5 thermocouples on the right side. When, it indicates an intermediate thermocouple; this characteristic is normal under operating conditions. <0.05; can reach 0.1~0.25 when bias occurs. Using two rows of averaging can reduce misjudgments caused by local temperature anomalies in a single row. The crystallizer sensor layout structure is the same as the crystallizer sensor layout structure schematic diagram disclosed in Figure 2 of the specification of invention patent application CN121535146A.

[0031] Liquid surface fluctuation energy ratio During calculation, the liquid level signal Perform a fast Fourier transform within a 30-second time window to calculate the energy ratio between the low-frequency band (0~0.5Hz, reacting to large fluctuations caused by bias current) and the mid-frequency band (0.5~2Hz, reacting to normal surface waves); The warning indicates abnormal low-frequency fluctuations caused by flow deviation.

[0032] Heat flow difference between the two sides Through the heat flux density on both sides , calculate:

[0033] Under normal operating conditions <0.1, and can reach 0.2~0.4 when the flow is biased. The condition of the crystallizer is initially evaluated based on the lateral temperature asymmetry, the energy ratio of liquid surface fluctuation, and the heat flow difference between the two sides.

[0034] A state space is constructed based on the lateral temperature asymmetry, the energy ratio of liquid surface fluctuations, and the difference in heat flow across the two sides; in a specific implementation, as a preferred embodiment of the present invention, the state space... , This refers to the immersion depth of the submersible nozzle. The electromagnetic stirring current generates a two-dimensional action from the intelligent agent in the state space. ,in For the probability of skew flow, The value indicates the direction of flow deviation. A negative value indicates flow deviation to the left, a positive value indicates flow deviation to the right, and the absolute value indicates the severity of the flow deviation.

[0035] The reward function is calculated based on the current state of the crystallizer. The reward function includes: asymmetric rationality reward, consistency reward, and stability reward. Specifically, in a preferred embodiment of this invention, the asymmetric rationality reward is: if... <0.05 and agent output <0.3, the asymmetric reasonableness reward value is 0.2; if >0.12 and If the asymmetric rationality reward value is >0.6, the asymmetric rationality reward value is 0.3; if the eccentricity probability contradicts the transverse temperature asymmetry, the asymmetric rationality penalty value is 0.3. The consistency bonus is based on the heat flow difference between the two sides. Energy ratio of liquid surface fluctuation Based on a comprehensive assessment of the actual flow direction, when the flow deflects to the left, the heat flux on the left side is lower and the liquid level fluctuation on the right side is larger. If the intelligent agent outputs... If the sign matches the actual direction, the consistency reward is 0.2; otherwise, the consistency penalty is 0.2. The stability bonus is: if the sign of the deflection direction remains unchanged in two consecutive outputs and If the change is less than 0.1, the stability bonus value is 0.1; if the direction changes frequently, the stability penalty value is 0.1. The reward function is obtained by adding the reward values ​​from the asymmetric rationality reward, consistency reward, and stability reward, and then subtracting the penalty value. .

[0036] A flow field state prediction model is constructed using a dual-network architecture combining a deep Q-network and a policy network. In a preferred embodiment of this invention, after completing one state transition within each cycle (2 seconds), the flow field state prediction model incorporates empirical data. The data is stored in an experience replay pool with a capacity of 3000 entries. Random samples are taken from the experience pool every two cycles (4 seconds), and the network parameters are updated using the Adam optimizer. The policy network outputs the mean and variance of actions, and the value network evaluates the value of states and actions. A soft update is used to stably train the flow field state prediction model. The soft update coefficient is 0.005. To balance exploration and utilization, Gaussian noise is added with a 10% probability when actions are adopted. .

[0037] A flow field state prediction model is used to provide early warning of flow field inhomogeneity in the crystallizer. Specifically, in a preferred embodiment of this invention, early warning of flow field inhomogeneity in the crystallizer includes: when two consecutive early warning cycles meet... and When the flow deviation warning is issued, the direction of the flow deviation (left-side or right-side deviation) and the severity (mild 0.4~0.6, moderate 0.6~0.8, severe) will be output simultaneously. 0.8).

[0038] When the batch of billets is completed and subsequent processes (surface inspection, pickling) confirm defects caused by flow deviation (slag entrapment on one side, longitudinal cracks), the quality inspector marks the time and direction of the flow deviation. The system reviews the sequence of actions within that time period; if there are any missed reports (actual flow deviation but...),... If the direction is incorrect, a corrected sample is generated and given a high weight, then stored in the experience pool for subsequent online learning.

[0039] Example This embodiment uses a twin-strand slab continuous casting machine in a steel plant as the application object, equipped with 66 K-type thermocouples in 6 rows and 11 columns, an electromagnetic stirring device for the crystallizer, and an industrial-grade data acquisition system deployed on site.

[0040] S1: Data Acquisition and Feature Calculation The system reads the thermocouple temperatures from the second and third rows (11 columns) every 2 seconds, averaging the values ​​of 10 sampling points for each thermocouple within 2 seconds. The average temperature across every two columns and two rows is then calculated. This is used for subsequent feature calculations. Using a two-row average can reduce misjudgments caused by local temperature anomalies in a single row, while the temperature asymmetry caused by flow deviation is reflected in both rows, enhancing feature stability.

[0041] Simultaneously read the cooling water flow rate and temperature difference on both sides, casting speed, sprue immersion depth, and electromagnetic stirring current. Construct three core features: Calculation of transverse temperature asymmetry:

[0042] On-site statistics show that under normal working conditions <0.05, but can reach 0.1~0.25 under biased flow.

[0043] Liquid surface fluctuation energy ratio calculation: Take the liquid level signal (150 points) from the most recent 30 seconds, perform FFT, and calculate the energy in the 0~0.5Hz frequency band. With energy in the 0.5~2Hz frequency band , The normal value is about 0.8~1.2, and it is >1.5 when there is a flow deviation.

[0044] Heat flow difference between the two sides: The formula for calculating heat flux density is: ,in , , The width of the crystallizer is (m). The effective heat exchange height (m).

[0045] S2: Agent Network Structure The policy network has a 6-dimensional input layer, 32 hidden neurons (ReLU), and a 2-neuron output layer, corresponding to... mean and The mean and standard deviation are learnable parameters. After action sampling Mapped to [0,1] via Sigmoid. The values ​​are mapped to [-1, 1] using the Tanh mapping. The input layer of the value network is 8-dimensional (6-dimensional state and 2-dimensional action), the hidden layer has 32 neurons, and the output layer has 1 neuron (Q-value). The learning rate is 0.001, the discount factor is 0.95, and the soft update coefficient is 0.005.

[0046] S3: Online Learning Main Loop Execute every 2 seconds: Data acquisition, calculation Features, constructing state The policy network performs forward computation and outputs the average action value. The original motion was obtained by sampling and adding 10% Gaussian noise, along with the logarithm and standard deviation. . , Send to the early warning judgment module. If the condition is met twice consecutively... ≥0.6 and If this occurs, a flow deviation warning will be triggered, displaying the direction and severity. Instant rewards are calculated based on the current state and action. Observe the next state. , experience Store the data in the experience replay pool (capacity 3000). Perform an update every 2 periods, randomly sample 64 experiences from the experience pool, calculate the loss of the value network and policy network, update the parameters using the Adam optimizer, and perform a soft update on the target network.

[0047] S4: Lag Correction During a certain casting process, the quality inspector discovered severe slag entrainment on the right side of the billet, determining it to be a right-side flow deviation. The system reviewed the alarm records and found that the agent output at that time... =0.7、 =+0.6 (moderate right-side skew), consistent with reality, you can choose to generate a positive reward correction sample or ignore it directly. If the system output =-0.5 (left-side deviation), then a penalty correction sample is generated. It is then given high priority and stored in the experience pool so that it can be learned first in subsequent sampling.

[0048] Comparing this method with manual judgment, the accuracy of this method in identifying flow deviation reaches 89%, and the accuracy in determining direction reaches 86%. Traditional flow deviation detection methods based on liquid level fluctuation thresholds have an accuracy of approximately 62% and cannot determine direction. This invention automatically adapts to changes in electromagnetic stirring current and adjustments in nozzle immersion depth within 3-5 batches, requiring no manual intervention.

[0049] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for early warning of crystallizer flow field inhomogeneity based on multi-source temporal features and online reinforcement learning, characterized in that, include: Acquire the crystallizer's operating data and collect the thermocouple temperature values ​​of two adjacent rows on the wide copper plate of the crystallizer to calculate the equivalent temperature; Calculate the lateral temperature asymmetry, liquid surface fluctuation energy ratio, and heat flow difference between the two sides based on the operating data. The energy ratio of the liquid surface fluctuation Energy in the 0~0.5Hz frequency band With energy in the 0.5~2Hz frequency band The ratio: ； The state space is constructed based on the lateral temperature asymmetry, the energy ratio of liquid surface fluctuations, and the heat flow difference between the two sides. The reward function is calculated based on the current state of the crystallizer, and the reward function includes: asymmetric rationality reward, consistency reward and stability reward; A flow field state prediction model is constructed using a dual-network architecture combining a deep Q-network and a policy network; after completing one state transition in each cycle, the flow field state prediction model incorporates empirical ( The data is stored in the experience replay pool; every two cycles, random samples are taken from the experience pool, and the network parameters are updated using the Adam optimizer; the policy network outputs the mean and variance of the actions, and the value network evaluates the value of the state and actions; the flow field state prediction model is stably trained using soft updates. The aforementioned flow field state prediction model is used to provide early warning of the non-uniformity of the flow field in the crystallizer.

2. The method for early warning of crystallizer flow field inhomogeneity based on multi-source temporal features and online reinforcement learning according to claim 1, characterized in that, The operating data includes thermocouple temperature values, cooling water flow rate, inlet and outlet water temperature difference, crystallizer liquid level signal, immersion depth of the submersible nozzle, pulling speed, and electromagnetic stirring current; the heat flux density on the left and right sides is calculated based on the cooling water flow rate and inlet and outlet water temperature difference, and the liquid level fluctuation amplitude and frequency are calculated based on the crystallizer liquid level signal; The crystallizer has 11 thermocouples per row on its wide copper plate, connecting adjacent... , The thermocouple temperature values ​​in the two rows are denoted as follows: and Calculate the equivalent temperature : in, Indicates the first The temperature values ​​of each thermocouple in the row, Indicates the first The temperature values ​​of each thermocouple in the row.

3. The method for early warning of crystallizer flow field inhomogeneity based on multi-source temporal features and online reinforcement learning according to claim 2, characterized in that, The lateral temperature asymmetry Represented as: Among them, when When the value is 1-5, it represents the 5 thermocouples on the left side. When 7-11 is used, it indicates that there are 5 thermocouples on the right side. When, it indicates the intermediate thermocouple; The energy ratio of the liquid surface fluctuation During calculation, the liquid level signal Perform a fast Fourier transform within the time window to calculate the energy ratio between the low-frequency band and the mid-frequency band; The heat flow difference on both sides Through the heat flux density on the left and right sides , calculate: The crystallizer condition is initially assessed based on the lateral temperature asymmetry, the energy ratio of liquid surface fluctuations, and the heat flow difference between the two sides.

4. The method for early warning of crystallizer flow field inhomogeneity based on multi-source temporal features and online reinforcement learning according to claim 3, characterized in that, The state space , This refers to the immersion depth of the submersible nozzle. The electromagnetic stirring current generates a two-dimensional action from the intelligent agent in the state space. ,in For the probability of skew flow, The value indicates the direction of flow deviation. A negative value indicates flow deviation to the left, a positive value indicates flow deviation to the right, and the absolute value indicates the severity of the flow deviation.

5. The method for early warning of crystallizer flow field inhomogeneity based on multi-source temporal features and online reinforcement learning according to claim 4, characterized in that, The asymmetric reasonable reward is: if <0.05 and agent output <0.3, the asymmetric reasonableness reward value is 0.2; if >0.12 and If the asymmetric rationality reward value is >0.6, the asymmetric rationality reward value is 0.3; if the eccentricity probability contradicts the transverse temperature asymmetry, the asymmetric rationality penalty value is 0.

3. The consistency reward is based on the heat flow difference between the two sides. Energy ratio of liquid surface fluctuation Based on a comprehensive judgment of the actual deflection direction, if the agent outputs... If the sign matches the actual direction, the consistency reward value is 0.2; Otherwise, the consistency penalty is 0.2; The stability reward is: if the sign of the deflection direction remains unchanged in two consecutive outputs and If the change is less than 0.1, the stability bonus value is 0.1; If the direction changes frequently, the stability penalty value is 0.1; The reward function is obtained by adding the reward values ​​from the asymmetric rationality reward, consistency reward, and stability reward, and then subtracting the penalty value. .

6. The method for early warning of crystallizer flow field inhomogeneity based on multi-source temporal features and online reinforcement learning according to claim 1, characterized in that, The aforementioned early warning system for the non-uniform flow field in the crystallizer specifically includes: when two consecutive early warning cycles meet the following conditions... and When the flow deviation is detected, a flow deviation warning is issued, and the direction and severity of the flow deviation are simultaneously output.

Images