An austenitic stainless steel castings in-situ micro-alloying modification method based on multi-field coupling and intelligent feedback

Abstract

The application provides an austenitic stainless steel casting in-situ micro-alloying modification method based on multi-field coupling and intelligent feedback, comprising the following steps: S1: mold presetting and machine learning model deployment; S2: master alloy smelting and pouring; S3: injecting a micro-alloying agent into the mold and carrying out multi-field coupling treatment on the molten steel; S4: predicting the microstructure evolution of the casting through a machine learning model, and adjusting the micro-alloying agent injection parameters and the multi-field coupling parameters in real time; S5: carrying out solid solution treatment and aging treatment on the solidified casting to obtain a final product. The application synchronously carries out the micro-alloying process and the solidification process, the multi-field coupling effect of the pulsed electromagnetic field and the ultrasonic field, and introduces a machine learning driven intelligent feedback system, so that not only the processing steps are provided, but also an intelligent casting system which can be self-adaptive and self-optimized is constructed, and a new technical path is provided for the performance customization production of high-end castings.

Description

Technical Field

[0001] This invention relates to the field of metal material processing and casting technology, and in particular to an in-situ microalloying modification method for austenitic stainless steel castings based on multi-field coupling and intelligent feedback. Background Technology

[0002] Austenitic stainless steel is widely used due to its excellent corrosion resistance and formability. Its strength, wear resistance, and high-temperature performance can be significantly improved through microalloying (such as adding elements like Nb, V, and Ti). Traditional microalloying methods involve adding alloying elements to the molten steel during the electric arc furnace or AOD furnace melting stage. This method has the following inherent drawbacks: Severe burn-off: Highly reactive microalloying elements are easily oxidized during high-temperature smelting, resulting in low yield and unstable composition control.

[0003] Tissue coarsening: Premature addition can easily form coarse primary precipitates, which become stress concentration sources and deteriorate material properties.

[0004] Segregation tendency: During the subsequent solidification process, alloying elements are prone to macroscopic and microscopic segregation, resulting in uneven microstructure.

[0005] Process reliance on experience: The production process heavily depends on the experience of operators, resulting in poor consistency between different batches.

[0006] To address these issues, methods such as grain refinement and homogenization heat treatment have been employed, but their effectiveness is limited and they cannot fundamentally solve the problems of element segregation and burn-off. Therefore, there is an urgent need for an innovative method that can precisely control the addition and distribution of microalloying elements and achieve optimal microstructure and properties. Summary of the Invention

[0007] The purpose of this invention is to address the shortcomings of existing technologies by proposing an in-situ microalloying modification method for austenitic stainless steel castings based on multi-field coupling and intelligent feedback.

[0008] To achieve the above objectives, the present invention adopts the following technical solution: an in-situ microalloying modification method for austenitic stainless steel castings based on multi-field coupling and intelligent feedback, comprising the following steps: S1: Pre-setting of mold and deployment of machine learning model: Pre-embed microalloying agent injection channels and multi-field coupling generator probes in the hot spot or final solidification area of ​​the mold cavity; deploy a deep learning model trained based on a large amount of historical experimental data and phase diagram calculations into the casting system within the mold; S2: Master alloy melting and casting: Melting the austenitic stainless steel master alloy and casting the molten steel into the mold, where the molten steel is cooled to form a casting; S3: Inject microalloying agent into the mold and perform multi-field coupling treatment on the molten steel; Includes the following steps: S31: Real-time monitoring of temperature field data during the solidification process of the casting using sensors pre-installed in the mold; S32: When the monitoring temperature field data indicates that a specific part of the casting has entered the semi-solid region, the nano-scale micro-alloying agent is injected into the interdendritic liquid phase through the injection channel; S33: The multi-field coupling generator probe applies pulsed electromagnetic field and ultrasonic field to the solidified casting, and applies mechanical pressure field to the end of the solidified casting; S4: Predict the microstructure evolution of castings through machine learning models and adjust microalloying agent injection parameters and multi-field coupling parameters in real time; S5: Perform solution treatment and aging treatment on the solidified casting to obtain the final product.

[0009] Preferably, the mold is equipped with a casting system, which includes a high-temperature high-speed camera, a thermocouple, and the machine learning model. The mold is manufactured using traditional investment casting or 3D printing sand mold technology. At key parts of the mold, such as hot spots and the final solidification area, microalloying agent injection channels and multi-field coupling generator probes are pre-embedded.

[0010] Preferably, the sensor detects temperature field data of the molten steel in the mold, including temperature of various parts of the casting, cooling rate, and dendrite growth rate, and the sensor sends the detected temperature field data to the machine learning model.

[0011] Preferably, step S32 includes the following steps: S321: The casting system receives and identifies the temperature field data; S322: Preliminary determination of the casting system entering the semi-solid region; When the measured temperature of any of the thermocouples is lower than the known liquidus temperature T_L of the alloy for the first time, the system determines that the region where the thermocouple is located has begun to enter the semi-solid region. S323: The casting system calculates the solid fraction fs in real time and determines the optimal injection time of the microalloying agent; The casting system has a built-in thermodynamic database of the austenitic stainless steel, which includes its liquidus temperature T_L, solidus temperature T_S, and the relationship between solid fraction and temperature. Core algorithm: The system continuously records the cooling curve of the target point and calculates its instantaneous cooling rate dT / dt in real time; By substituting the current measured temperature T into the pre-stored f_s-T relationship, the theoretical solid fraction f_s at that location can be calculated in real time. The optimal injection timing is defined as: when the calculated solid fraction f_s reaches the target range, where the target range is 0.4. <f_s<0.7 ; S324: Cross-validation and decision-making for multiple sensors; The casting system only confirms and sends a start command to the microalloying agent injection channel when multiple adjacent thermocouples show that their areas have entered the preset target range f_s. If the temperatures of the thermocouples are inconsistent, the system uses a weighted average algorithm to make a decision, with thermocouples closer to the injection point having a higher weight. S325: The casting system determines the optimal injection time of the microalloying agent, and the casting system controls the injection channel to spray the nano-scale microalloying agent into the interdendritic liquid phase.

[0012] Preferably, in step S1, the deployment of the machine learning model includes the following steps: C1: Input and process time-series temperature data, visual data and process parameters into the machine learning model, and output a visual feature vector X; C2: The machine learning model acquires microstructure labels and performance labels, and outputs label Y; C3: The machine learning model constructs a data pair {x, y} from the visual feature vector X and the label Y. The machine learning model uses a hybrid deep learning model to map a function f from several data pairs {x, y}.

[0013] Preferably, step S4 includes the following: B1: The casting system collects thermocouple data X_temp-t, image frames X_vision-t from a high-temperature high-speed camera, and the currently executing process parameter settings X_params-t in real time. The hybrid deep learning model is based on the mapping function f and outputs a prediction vector Y1. X_temp-t: Data of all thermocouples within the current time window and the time period preceding it; X_vision-t: Image frames from the high-speed camera within the window of the current time and the preceding time period; X_params-t: The currently executing process parameter settings; B2: The casting system compares the predicted vector Y_pred-t with the preset target value Y_target, and the casting system dynamically adjusts the controllable parameters according to the comparison result; The preset target value Y_target is set by the user according to product requirements, for example, grain size d<20μm and yield strength YS>350MPa; B3: The casting system adjusts its strategy based on controllable parameters. Compared with the prior art, the beneficial effects of the present invention are as follows: The present invention innovatively synchronizes the microalloying process with the solidification process (in-situ microalloying), avoids element burn-off, greatly improves the yield and utilization efficiency of microalloying elements, and reduces costs.

[0014] This invention achieves extreme grain refinement, nano-sized and uniform distribution of precipitates through multi-field coupling of pulsed electromagnetic field and ultrasonic field, and effectively eliminates defects such as pores and inclusions, resulting in a significant improvement in overall mechanical properties.

[0015] This invention introduces a machine learning-driven intelligent feedback system, which enables real-time perception, prediction, and dynamic optimization of complex processes, completely eliminating reliance on human experience and ensuring extremely high consistency and reproducibility of product performance. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the steps of an in-situ microalloying modification method for austenitic stainless steel castings based on multi-field coupling and intelligent feedback, as described in this invention. Detailed Implementation

[0017] To provide a further understanding of the purpose, structure, features, and functions of the present invention, detailed descriptions are provided below with reference to specific embodiments.

[0018] A method for in-situ microalloying modification of austenitic stainless steel castings based on multi-field coupling and intelligent feedback includes the following steps: S1: Pre-setting of mold and deployment of machine learning model: Pre-embed microalloying agent injection channels and multi-field coupling generator probes in the hot spot or final solidification area of ​​the mold cavity; deploy a deep learning model trained based on a large amount of historical experimental data and phase diagram calculations into the casting system within the mold; S2: Master alloy melting and casting: Melting the austenitic stainless steel master alloy and casting the molten steel into the mold, where the molten steel is cooled to form a casting; S3: Inject microalloying agent into the mold and perform multi-field coupling treatment on the molten steel; Includes the following steps: S31: Real-time monitoring of temperature field data during the solidification process of the casting using sensors pre-installed in the mold; S32: When the monitoring temperature field data indicates that a specific part of the casting has entered the semi-solid region, the nano-scale micro-alloying agent is injected into the interdendritic liquid phase through the injection channel; S33: The multi-field coupling generator probe applies pulsed electromagnetic field and ultrasonic field to the solidified casting, and applies mechanical pressure field to the end of the solidified casting; The pulsed electromagnetic field generates Lorentz force, which violently stirs the semi-solid melt, breaks up the nascent dendrites, and makes them the nucleation cores of equiaxed crystals, while making the injected nanoparticles extremely uniformly distributed.

[0019] The ultrasonic field generates cavitation bubbles, and the shock waves generated by their collapse further refine the grains and disperse the particles. Its acoustic flow effect promotes the homogenization of the temperature field and composition field.

[0020] Applying a mechanical pressure field to the end of the casting during solidification: Before the casting is completely solidified, apply controllable mechanical pressure or use centrifugal force to specific parts of the mold to compensate for solidification shrinkage, completely eliminate micro-shrinkage porosity, and improve density.

[0021] S4: Predict the microstructure evolution of castings through machine learning models and adjust microalloying agent injection parameters and multi-field coupling parameters in real time; S5: Perform solution treatment and aging treatment on the solidified casting to obtain the final product.

[0022] Thus, this invention achieves precise and efficient utilization of microalloying elements, obtaining high-performance castings with ultrafine grains, no defects, and uniform distribution of precipitates, and realizing intelligent and digital production processes.

[0023] Preferably, the mold is equipped with a casting system, which includes a high-temperature high-speed camera, a thermocouple, and the machine learning model. The mold is manufactured using traditional investment casting or 3D printing sand mold technology. At key parts of the mold, such as hot spots and the final solidification area, microalloying agent injection channels and multi-field coupling generator probes are pre-embedded.

[0024] Preferably, the sensor detects temperature field data of the molten steel in the mold, including temperature of various parts of the casting, cooling rate, and dendrite growth rate, and the sensor sends the detected temperature field data to the machine learning model.

[0025] Preferably, step S32 includes the following steps: S321: The casting system receives and identifies the temperature field data; Multiple high-response armored K-type or S-type thermocouples are pre-embedded in key areas of the mold cavity (such as hot spots, the center of thick sections, and the final solidification zone) to form a temperature monitoring network. The thermocouples are distributed in a specific spatial array to monitor the cooling curves of the casting at different depths and locations.

[0026] S322: Preliminary determination of the casting system entering the semi-solid region; When the measured temperature of any of the thermocouples is lower than the known liquidus temperature T_L of the alloy for the first time, the system determines that the region where the thermocouple is located has begun to enter the semi-solid region. S323: The casting system calculates the solid fraction fs in real time and determines the optimal injection time of the micro-alloying agent; The primary judgment only provides a starting point, but cannot determine the solid-liquid ratio. The present invention uses a numerical back-calculation method based on the cooling curve to estimate the solid fraction fs in real time. The casting system has a built-in thermodynamic database for this austenitic stainless steel, including its liquidus temperature TL, solidus temperature TS, and the relationship between the solid fraction and temperature fs-T. The fs-T curve can be pre-calculated through the Scheil model or CALPHAD software; Core algorithm: The system continuously records the cooling curve of the target point and calculates its instantaneous cooling rate dT / dt in real time; Substitute the currently measured temperature T into the pre-stored fs-T relation formula, and the theoretical solid fraction fs at this position can be calculated in real time; The optimal injection timing is defined as: when the calculated solid fraction fs reaches the target range, the target range is 0.4 < fs < 0.7; in this interval, the dendrite skeleton has been fully formed, which can effectively "capture" the injected nanoparticles and prevent them from being washed away; at the same time, there is still enough liquid phase between the dendrites as a channel to ensure that the nanoparticles can be transported and evenly dispersed throughout the semi-solid region.

[0027] S324: Cross-validation and decision-making of multiple sensors; To prevent single-point temperature measurement failure or error, the system adopts a multi-sensor cross-validation mechanism. When multiple adjacent thermocouples all show that the area where they are located has entered the preset fs target range, the casting system finally confirms and issues a start command to the micro-alloying agent injection channel; if the temperatures of each thermocouple are inconsistent, the system uses a weighted average algorithm for decision-making, where the thermocouple closer to the injection point has a higher weight; S325: The casting system determines the optimal injection time of the micro-alloying agent, and the casting system controls the injection channel to inject the nano-scale micro-alloying agent into the liquid phase between the dendrites.

[0028] The injection operation lasts for a preset time, such as 10 - 60 seconds, or automatically terminates when the system monitors that the fs in this area > 0.8.

[0029] Preferably, in step S1, the deployment of the machine learning model includes the following steps: C1: Input and process the time-series temperature data, visual data, and process parameters in the machine learning model, and output the visual feature vector X; Time-series temperature data: From a large number of historical tests, collect the time-temperature cooling curves collected by thermocouples and extract features from them, including the following features: Basic statistical characteristics: For the cooling curve of each monitoring point, the following are calculated at its key stages (such as from liquidus temperature to solidus temperature): average cooling rate, standard deviation of cooling rate (characterizing stability), and maximum subcooling.

[0030] Differential characteristics: By numerically differentiating T(t), the instantaneous cooling rate dT / dt curve is obtained. Features can be extracted from this curve, such as: maximum cooling rate, average cooling rate, and the moment when the cooling rate changes abruptly (which may correspond to a nucleation event).

[0031] Based on the characteristics of the phase transition point, the system identifies the inflection points of the liquidus temperature (T_L) and solidus temperature (T_S) on the cooling curve. Based on this, the local solidification time t_f = t_S - t_L (the time from the start of solidification to the end of solidification) is calculated, and this time is directly related to the dendrite arm spacing.

[0032] Dimension processing: Each cooling curve is resampled or divided into fixed-length segments and converted into a fixed-dimensional feature vector F_temp; Visual data: Images from a high-temperature, high-speed camera acquired synchronously. Image processing algorithms are used to perform noise reduction, contrast enhancement, and background subtraction on each frame to highlight the region of interest. Then, features are extracted, such as dendrite tip growth rate (V_tip), secondary dendrite arm spacing (SDAS), and grain distribution.

[0033] Dendrite tip growth rate V_tip: Step 1: In consecutive image frames, track the displacement of the dendrite tip pixel using optical flow or by directly tracing it.

[0034] Step 2: Divide the displacement by the inter-frame time interval to calculate the instantaneous growth rate V_tip of the dendrite tip. Finally, take the average or maximum value over a time period as the feature.

[0035] Secondary dendrite arm spacing SDAS: Step 1: Perform edge detection on the image, such as using the Canny algorithm or skeletonization, to clearly delineate the dendrite trunk.

[0036] Step 2: Perform a line scan along the direction perpendicular to the primary dendrite arm to measure the pixel distance between adjacent secondary arms.

[0037] Step 3: Convert pixel distances to actual physical dimensions (μm) through calibration, and statistically average a large number of measurements to obtain the SDAS feature value. SDAS is strongly correlated with local solidification time (t_f) and is a key indicator for measuring the fineness of the solidified structure.

[0038] Particle distribution: Step 1: Utilize the grayscale difference between nanoparticles and liquid metal, and employ threshold segmentation or machine learning segmentation models, such as U-Net, to identify particle clusters in the image.

[0039] Step 2: Calculate features such as the number of particle clusters, average size, area ratio, and distribution uniformity index, such as entropy calculation based on image grids.

[0040] Finally, all the visual features extracted from the image sequence are combined into a single visual feature vector F_vision.

[0041] Process parameters: Record the controllable parameters for each experiment, such as the microalloying agent injection rate V_inj, pulse electromagnetic field strength B and frequency f, ultrasonic power P_us and frequency f_us; and use these process parameters as the process parameter feature vector F_params.

[0042] Finally, at a certain time t, the overall input feature vector X(t) of the system is the concatenation of the three major features: X(t) = [F_temp(t); F_vision(t); F_params(t)].

[0043] C2: The machine learning model acquires microstructure labels and performance labels, and outputs label Y; Microstructure tags: Metallographic, SEM, and EBSD analyses are performed on the samples to obtain microstructure tags such as average grain size d, precipitate size and distribution density, and orientation difference angular distribution; these measurement results are summarized into a microstructure tag vector Y_micro = [d, precipitate size and distribution density, orientation difference angular distribution]; Performance tags: Destructive testing is performed on the castings after the experiment to obtain mechanical property tags such as yield strength YS, tensile strength UTS, elongation El%, and impact toughness CVN. These test results are summarized into a performance tag vector Y_perf = [YS, UTS, El%, CVN].

[0044] Finally, for a complete experiment, the output label Y is: Y = [Y_micro, Y_perf].

[0045] C3: The machine learning model constructs a data pair {x, y} from the visual feature vector X and the label Y. Thus, we have constructed a unique data pair {x, y} for each experiment, where x represents a holographic snapshot of the process: it includes how solidification occurs (temperature profile), how the tissue grows (visual image), and how we intervene in the process (process parameters); y represents the final result of the process: the formed microstructure and the resulting mechanical properties. The machine learning model uses a hybrid deep learning model to find a mapping function f from several data pairs {x, y} such that y ≈ f(x).

[0046] Preferably, step S4 includes the following: B1: The casting system collects thermocouple data X_temp-t, image frames X_vision-t from a high-temperature high-speed camera, and the currently executing process parameter settings X_params-t in real time. The hybrid deep learning model is based on the mapping function f and outputs a prediction vector Y1. X_temp-t: Data of all thermocouples within the current time window and the time period preceding it; X_vision-t: Image frames from the high-speed camera within the window of the current time and the preceding time period; X_params-t: The currently executing process parameter settings; These real-time data are input into a pre-trained hybrid deep learning model, which outputs a prediction vector Y1 based on the mapping function f. This vector Y1 represents the final grain size and yield strength obtained if the current parameters remain unchanged until solidification is complete.

[0047] B2: The casting system compares the predicted vector Y1 with the preset target value Y_target, and the casting system dynamically adjusts the controllable parameters according to the comparison result; The preset target value Y_target is set by the user according to product requirements, for example, grain size d<20μm and yield strength YS>350MPa; The casting system incorporates a decision-making algorithm, such as a model predictive control (MPC) or reinforcement learning (RL) controller, which then begins to operate. This algorithm simulates a "virtual experiment," calculating how to adjust controllable parameters to make the final prediction approximate the target.

[0048] B3: The casting system adjusts its strategy based on controllable parameters; The decision algorithm yields the following situations, and the casting system adjusts the controllable parameters according to the decision algorithm; Scenario 1: Insufficient prediction performance. If the model predicts YS_pred = 320MPa < 350MPa under the current parameters.

[0049] Decision: The system determines that further grain refinement or precipitation strengthening effect is needed.

[0050] Implementation: Slightly increase the pulsed electromagnetic field intensity ΔB to provide stronger grain fragmentation and refinement dynamics. Slightly increase the ultrasonic power ΔP_us to enhance cavitation effects and particle dispersion capabilities. Appropriately increase the microalloying agent injection rate ΔV_inj to provide more heterogeneous nucleation sites, but excessive increase should be prevented from leading to agglomeration under model constraints.

[0051] Scenario 2: Predicted grain size is too fine, but cracks may occur. If the model predicts d_pred = 5μm, but the cooling rate is too fast, it may lead to thermal cracking.

[0052] Decision: The cooling intensity needs to be slightly reduced to prevent defects while ensuring fine grains.

[0053] Execution: Slightly reduce the pulsed electromagnetic field frequency Δf: change the stirring force mode to reduce heat input. Maintain or fine-tune the ultrasonic power.

[0054] Scenario 3: Uneven particle distribution is predicted. High-speed camera image features show a tendency for particle aggregation.

[0055] Decision: We need to enhance our ability to distribute resources.

[0056] Execution: Prioritize adjusting the ultrasonic field, briefly increasing its power P_us to break up the aggregates using a strong acoustic flow. Simultaneously fine-tune the electromagnetic field parameters to alter the melt flow pattern and aid in dispersion.

[0057] Thus, the entire casting system forms a high-frequency iterative cycle. This cycle operates at an extremely high frequency, ensuring that process parameters are constantly and dynamically optimized to cope with various complex and dynamic changes during solidification, ultimately keeping product quality stably controlled within the optimal target range. This real-time control based on feedforward prediction surpasses the limitations of traditional PID control, which relies solely on current error feedback, and is the core of ensuring consistency.

[0058] The present invention has been described by the above-described embodiments; however, these embodiments are merely examples for implementing the present invention. It must be noted that the disclosed embodiments do not limit the scope of the present invention. Conversely, any modifications and refinements made without departing from the spirit and scope of the present invention are within the scope of patent protection of the present invention.

Claims

1. A method for in-situ microalloying modification of austenitic stainless steel castings based on multi-field coupling and intelligent feedback, characterized in that: Includes the following steps: S1: Casting pre-setting and machine learning model deployment: Pre-embed microalloying agent injection channels and multi-field coupling generator probes in the hot spot or final solidification area of ​​the casting cavity; A deep learning model trained based on a large amount of historical experimental data and phase diagram calculations is deployed into the casting system within the mold. S2: Master alloy melting and casting: Melting the austenitic stainless steel master alloy and casting the molten steel into the mold, where the molten steel is cooled to form a casting; S3: Inject microalloying agent into the mold and perform multi-field coupling treatment on the molten steel; Includes the following steps: S31: Real-time monitoring of temperature field data during the solidification process of the casting using sensors pre-installed in the mold; S32: When the monitoring temperature field data indicates that a specific part of the casting has entered the semi-solid region, the nano-scale micro-alloying agent is injected into the interdendritic liquid phase through the injection channel; S33: The multi-field coupling generator probe applies pulsed electromagnetic field and ultrasonic field to the solidified casting, and applies mechanical pressure field to the end of the solidified casting; S4: Predict the microstructure evolution of castings through machine learning models and adjust microalloying agent injection parameters and multi-field coupling parameters in real time; S5: Perform solution treatment and aging treatment on the solidified casting to obtain the final product.

2. The in-situ microalloying modification method for austenitic stainless steel castings based on multi-field coupling and intelligent feedback as described in claim 1, characterized in that: The mold is equipped with a casting system, which includes a high-temperature high-speed camera, a thermocouple, and the machine learning model. The mold is manufactured using traditional investment casting or 3D printing sand mold technology. At key parts of the mold, such as hot spots and the final solidification area, there are pre-embedded microalloying agent injection channels and multi-field coupling generator probes.

3. The in-situ microalloying modification method for austenitic stainless steel castings based on multi-field coupling and intelligent feedback as described in claim 1, characterized in that: The sensor detects temperature field data of the molten steel in the mold, including temperature of various parts of the casting, cooling rate, and dendrite growth rate. The sensor sends the detected temperature field data to the machine learning model.

4. The in-situ microalloying modification method for austenitic stainless steel castings based on multi-field coupling and intelligent feedback as described in claim 2, characterized in that: Step S32 includes the following steps: S321: The casting system receives and identifies the temperature field data; S322: Preliminary determination of the casting system entering the semi-solid region; When the measured temperature of any of the thermocouples is lower than the known liquidus temperature T_L of the alloy for the first time, the system determines that the region where the thermocouple is located has begun to enter the semi-solid region. S323: The casting system calculates the solid fraction fs in real time and determines the optimal injection time of the microalloying agent; The casting system has a built-in thermodynamic database of the austenitic stainless steel, which includes its liquidus temperature T_L, solidus temperature T_S, and the relationship between solid fraction and temperature. Core algorithm: The system continuously records the cooling curve of the target point and calculates its instantaneous cooling rate dT / dt in real time; By substituting the current measured temperature T into the pre-stored f_s-T relationship, the theoretical solid fraction f_s at that location can be calculated in real time. The optimal injection timing is defined as: when the calculated solid fraction f_s reaches the target range, where the target range is 0.

4. <f_s<0.7; S324: Cross-validation and decision-making for multiple sensors; The casting system only confirms and sends a start command to the microalloying agent injection channel when multiple adjacent thermocouples show that their areas have entered the preset target range f_s. If the temperatures of the thermocouples are inconsistent, the system uses a weighted average algorithm to make a decision, with thermocouples closer to the injection point having a higher weight. S325: The casting system determines the optimal injection time of the microalloying agent, and the casting system controls the injection channel to spray the nano-scale microalloying agent into the interdendritic liquid phase.

5. The in-situ microalloying modification method for austenitic stainless steel castings based on multi-field coupling and intelligent feedback as described in claim 2, characterized in that: In step S1, the deployment of the machine learning model includes the following steps: C1: Input and process time-series temperature data, visual data and process parameters into the machine learning model, and output a visual feature vector X; C2: The machine learning model acquires microstructure labels and performance labels, and outputs label Y; C3: The machine learning model constructs a data pair {x, y} from the visual feature vector X and the label Y. The machine learning model uses a hybrid deep learning model to map a function f from several data pairs {x, y}.

6. The in-situ microalloying modification method for austenitic stainless steel castings based on multi-field coupling and intelligent feedback as described in claim 5, characterized in that: Step S4 includes the following: B1: The casting system collects thermocouple data X_temp-t, image frames X_vision-t from a high-temperature high-speed camera, and the currently executing process parameter settings X_params-t in real time. The hybrid deep learning model is based on the mapping function f and outputs a prediction vector Y1. X_temp-t: Data of all thermocouples within the current time window and the time period preceding it; X_vision-t: Image frames from the high-speed camera within the window of the current time and the preceding time period; X_params-t: The currently executing process parameter settings; B2: The casting system compares the predicted vector Y_pred-t with the preset target value Y_target, and the casting system dynamically adjusts the controllable parameters according to the comparison result; The preset target value Y_target is set by the user according to product requirements, with a grain size d < 20 μm and a yield strength YS > 350 MPa; B3: The casting system adjusts its strategy based on controllable parameters.

Images