Machine learning based perfusion process intelligent decision system

Abstract

The application provides a machine learning-based perfusion process intelligent decision system, comprising: a data sensing module that collects pressure, temperature and flow data in real time during the perfusion process, generates a process state vector after preprocessing and feature extraction, a simulation and prediction module that generates physical field simulation data inside the perfused object through a digital twin model, and combines an attention mechanism-based prediction model to predict future flow front position, defect risk, distribution uniformity and reaction completion degree, and a decision module that adopts a reinforcement learning intelligent agent trained by neurons and generates an adjustment strategy in the perfusion process based on a multi-objective reward function. The application constructs a new hybrid intelligent system by deeply integrating multi-physical field simulation and machine learning models, improves the reliability of the system in predicting and the generalization ability of the system in decision-making during the perfusion process, and improves the quality of the perfusion product.

Description

Technical Field

[0001] This invention relates to the fields of computer intelligent systems and machine learning technology, specifically to an intelligent decision-making system for the infusion process based on machine learning. Background Technology

[0002] In traditional infusion processes, precise control of process results is crucial. Traditional control methods mainly rely on PID controllers based on fixed rules or the experience of operators to set process parameters. These methods have significant limitations when facing complex nonlinear processes, multi-physics coupling, and dynamically changing operating conditions: First, they lack the ability to predict the future state of the process and cannot foresee risks such as runaway flow front and defect formation, resulting in lag in control response; second, they cannot simultaneously optimize multiple conflicting process objectives.

[0003] In recent years, although some research has attempted to introduce machine learning into process control, many problems still exist. First, simple regression models or traditional neural networks are unable to effectively capture the complex spatiotemporal dependencies between multiple parameters such as pressure, temperature, and flow rate during the perfusion process, especially lacking the ability to capture key information of long-term dynamic processes, resulting in insufficient prediction accuracy. Second, most systems only use supervised learning for single-indicator prediction or use unoptimized reinforcement learning agents. These agents lack cross-task generalization ability, rely on a large amount of single-task field training data, have low learning efficiency, and are difficult to achieve dynamic trade-offs and synergistic optimization among multiple objectives such as flow control, defect prevention, and distribution uniformity. Finally, purely data-driven models ignore a deep understanding of the internal physical field of the controlled object, resulting in poor generalization and robustness when facing working conditions or new materials that have not appeared in the training data. Therefore, there is an urgent need for an intelligent method and system that can deeply integrate physical mechanisms and data-driven learning, possess high-precision forward-looking prediction capabilities, and adaptively achieve multi-objective synergistic optimization, in order to fundamentally improve the control level of the perfusion process. Summary of the Invention

[0004] To address the technical problems mentioned in the background section, this invention proposes an intelligent decision-making system for the infusion process based on machine learning.

[0005] Therefore, the technical solution adopted by the present invention is as follows: M1: A data sensing module that collects in-real-time infusion data of the infusion process and the infused object, and preprocesses and extracts features from the infusion data to generate a process state vector characterizing the real-time state of the infusion process; the infusion data includes at least pressure data, temperature data, and flow rate data; M2: Simulation and prediction module, equipped with digital twin model and prediction model; The digital twin model is constructed based on porous media fluid mechanics. The process state vector is used as the boundary condition. By solving Darcy's law and the energy conservation equation, simulation data of the interior of the injected object is generated. The simulation data includes pressure field, flow velocity field and temperature field. The prediction model employs an encoder-decoder neural network based on an attention mechanism. The prediction model predicts future process outcomes based on the simulation data and the process state vector. The process outcomes include the flow front location, defect formation risk, medium distribution uniformity, and reaction completion rate. M3: Decision module, equipped with a meta-trained reinforcement learning agent; the agent generates an adjustment strategy during the infusion process based on the process state vector and the process result, and by maximizing a preset multi-objective reward function; the agent is constructed based on an actor and critic framework.

[0006] Furthermore, the feature extraction is performed based on a preset time window; the features include statistical features, dynamic features, extreme value features, and correlation features; the features are concatenated to obtain the process state vector.

[0007] Furthermore, the digital twin model needs to be pre-configured with the digital geometric model and material parameters of the object to be infused; The digital twin model uses the process state vector as the boundary and initial conditions for simulation. The simulation process uses the finite volume method to discretize space and solve the control equations, which include temperature control equations, flow rate control equations, and pressure control equations.

[0008] Furthermore, the pressure field is solved based on the incompressible Navier-Stokes equations and Darcy's law, as shown in the following formula.

[0009] in, For pressure field, For the pressure field gradient, For the permeability tensor of the preform, For fluid viscosity, For divergence; The velocity field is solved using Darcy's velocity formula, as follows.

[0010] in, For the velocity field; The temperature field is solved based on the energy conservation equation, as shown in the following formula.

[0011] in, For the temperature field, For the temperature field gradient, For the density of the fluid, For specific heat capacity, Thermal conductivity, As a heat source for the reaction, This represents the rate of change of temperature over time.

[0012] Furthermore, the specific steps for the prediction model to output the process results are as follows; Key feature indicators were extracted from the simulation data, including mean pressure, mean pressure gradient modulus, flow front location, mean flow velocity, maximum and minimum temperature, and temperature standard deviation. The process state vector and the feature index are concatenated to obtain an input vector, and the input vector is then input into the prediction model. The encoder uses a multi-head self-attention mechanism to capture the dependencies of the input vector and outputs a context vector; The decoder then generates a predicted value of the process result based on the context vector through a preset fully connected layer; The loss function of the prediction model as follows,

[0013] in, , , and These are the flow loss function, defect loss function, uniformity loss function, and completion loss function, respectively. This is a hyperparameter.

[0014] Furthermore, the flow loss function, uniformity loss function, and completion loss function are all mean square error; the defect loss function is a binary cross-entropy loss.

[0015] Furthermore, the intelligent agent employs a deep reinforcement learning framework based on actors and critics, consisting of a policy network and a value network; The agent includes an input layer, a hidden layer, and an output layer, wherein the hidden layer uses the ReLU activation function, and the output layer uses the tanh function to constrain the action values; The multi-objective reward function Defined as a weighted sum, the formula is as follows.

[0016] in, 、 、 and These are the weighting coefficients. 、 、 and These are respectively: flow reward, defect reward, uniformity reward, and completion reward; The steps for generating the adjustment strategy are as follows: The process state vector and the process result are concatenated to obtain a comprehensive state vector, which is then input into the agent. The policy network outputs one or more normalized action values; The action value is mapped to a specific value of one or more physical parameters during the infusion process by linear scaling; all physical parameters and their corresponding specific values ​​constitute the adjustment strategy. The preset infusion actuator performs actions according to the adjustment strategy; After the infusion actuator performs the action, the agent calculates the reward value and stores this decision-making experience as an experience tuple for updating the agent; The experience tuple includes the integrated state vector, action value, and reward value.

[0017] Furthermore, the preprocessing includes digital filtering, sensor calibration, outlier detection and compensation, and data standardization.

[0018] Compared with the prior art, the advantages of the present invention are as follows: 1. This invention employs an encoder and decoder neural network based on a multi-head self-attention mechanism to effectively capture the complex spatiotemporal dependencies among multiple parameters, thus solving the bottleneck of traditional models in modeling long-term dynamic processes and achieving high-precision, multi-step prediction of flow fronts and defect risks.

[0019] 2. This invention employs meta-trained actor and critic reinforcement learning agents, which possess rapid cross-task adaptive capabilities. Furthermore, by maximizing the weighted multi-objective reward function, it inherently achieves dynamic trade-offs and collaborative optimization among multiple objectives such as flow, defects, and uniformity. Attached Figure Description

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

[0021] Figure 1 This is a schematic diagram of the system flow of the present invention; Figure 2 This is a schematic diagram of the prediction model of the present invention; Figure 3 This is a schematic diagram of the intelligent agent of the present invention. Detailed Implementation

[0022] To achieve the above objectives, the present invention provides an intelligent decision-making system for the perfusion process based on machine learning. Please refer to the following technical solution. Figure 1 ,include: M1: A data sensing module that collects in-real-time infusion data of the infusion process and the infused object, and preprocesses and extracts features from the infusion data to generate a process state vector characterizing the real-time state of the infusion process; the infusion data includes at least pressure data, temperature data, and flow rate data; Taking the resin infusion process of composite materials as an example, various electronic sensors deployed on the object being infused collect infusion data in real time. These sensors include at least: Pressure sensor: A piezoelectric or strain gauge pressure sensor is used and installed at key locations inside the resin injection point or mold cavity, such as the inlet and outlet, to measure the dynamic pressure of the resin during the injection process. The sensor outputs an analog voltage signal with a range of 0-1 MPa to meet the typical pressure range of resin injection. Temperature sensor: A K-type thermocouple or platinum resistance temperature sensor (PT100) is used, which is embedded in the mold wall or directly contacts the preform to monitor the resin temperature and mold temperature. The sensor placement points include the injection point, the flow front prediction area and the cooling area to ensure comprehensive coverage. Flow sensor: Install an electromagnetic flow meter or turbine flow meter in the resin delivery pipeline to measure the volumetric flow rate of the resin, with a range of 0-100 mL / min, to meet the flow rate requirements of resin filling. In this embodiment, all sensor signals undergo preliminary processing through a signal conditioning circuit. The weak sensor signals are amplified and filtered before being digitized by a high-precision analog-to-digital converter (ADC) at a sampling rate of 1000 samples per second. All data is timestamped to ensure its timeliness.

[0023] The principles for arranging electronic sensors are as follows: The locations of the sampling points must be able to capture the gradient changes and extreme values ​​of the three key physical fields: pressure, temperature, and velocity. The sampling points need to cover the key stages of the infusion process, such as the beginning, development, and end, so that the model can learn the dynamics of the entire process. Focus on deploying key points in areas most prone to defects to provide direct data support for defect risk prediction; Deploying multiple sensors of the same or different types in key areas improves the reliability of measurements through data fusion and provides complementary information for the model.

[0024] Preprocessing includes digital filtering, sensor calibration, outlier detection and compensation, and data standardization; Digital filtering removes high-frequency noise by applying real-time digital filters; for example, a second-order Butterworth low-pass filter is used for pressure data. Sensor calibration is based on laboratory calibration data, converting raw digital values ​​into physical quantities; Outlier detection and imputation use statistical process control methods, such as calculating the moving average and standard deviation of each data stream, to identify and remove outliers that exceed the ±3σ range. Missing data is repaired through linear interpolation or data fusion based on adjacent sensors to ensure data continuity; Data standardization converts the physical quantities of the perfusion data into dimensionless standard values ​​to eliminate the influence of dimensions. For example, it can use minimum and maximum standardization.

[0025] in, The pressure value is the standardized value. This is the original pressure value. and These are the minimum and maximum pressure values ​​allowed by the process.

[0026] Feature extraction is performed based on a preset time window, and the features include statistical features, dynamic features, extreme value features, and correlation features. Statistical characteristics include the mean, standard deviation, skewness, and kurtosis of the data within the time window; The dynamic feature is the rate of change of each physical quantity in the injection data within the time window, in order to capture trend information; Extreme value characteristics are the maximum and minimum values ​​of each physical quantity in the perfusion data within the time window; The correlation feature is the physical coupling effect between various physical quantities within a time window, for example: Calculate the Pearson correlation coefficient between pressure and flow data within the time window; this coefficient measures the linear correlation between pressure and flow. The effect of temperature on pressure was analyzed by linear regression, and the regression slope was calculated to assess the thermodynamic effect. The process state vector is obtained by concatenating all features.

[0027] M2: Simulation and prediction module, equipped with digital twin model and prediction model; The digital twin model is constructed based on porous media fluid mechanics. The process state vector is used as the boundary condition. By solving Darcy's law and the energy conservation equation, simulation data of the interior of the injected object is generated. The simulation data includes pressure field, flow velocity field and temperature field. The prediction model employs an encoder-decoder neural network based on an attention mechanism. The prediction model predicts future process outcomes based on the simulation data and the process state vector. The process outcomes include the flow front location, defect formation risk, medium distribution uniformity, and reaction completion rate. The digital twin model is a configurable physics simulation engine that dynamically generates high-fidelity physics simulation data based on the geometry, material properties, and real-time process state vectors of the infused object. The pressure field refers to the pressure distribution at various points in a digital twin model when fluid flows through the complex channels and voids inside the object being injected. The velocity field refers to the distribution of the flow velocity and direction of a fluid at various points in space, and includes both the magnitude and direction of the velocity. Temperature field refers to the temperature distribution at various points inside the injected object and fluid.

[0028] The construction of digital twin models relies on fluid dynamics and heat conduction principles, and simulates the multi-physics coupling behavior during the injection process by solving partial differential equations; Digital twin models require pre-configuration of the digital geometric model and material parameters of the object to be infused; In this embodiment, taking composite resin injection as an example, the geometric model is imported through three-dimensional CAD data, including the cavity of the injection object, the pore structure of the preform and the sensor placement positions; Material parameters include the relationship between resin viscosity and temperature, density, specific heat capacity, and permeability tensor of the preform. These parameters are obtained through experimental calibration or databases. Once the digital twin model is configured, the process state vector will be used as the simulation boundary conditions and initial conditions. The simulation process uses the finite volume method to discretize space and solve the governing equations, which include temperature control equations, flow rate control equations, and pressure control equations. The temperature control equation, based on the incompressible Navier-Stokes equation and Darcy's law, simulates the pressure distribution of the resin in a porous medium. The formula is as follows:

[0029] in, For pressure field, For the pressure field gradient, The permeability tensor of the preform is obtained through experimental calibration or a materials database. The viscosity of the resin is a material parameter derived from the resin and is calibrated through experiments on the relationship between viscosity and temperature. For divergence, it means that at any point in the resin flow, the mass of resin flowing in is equal to the mass of resin flowing out, and there is no source or sink of resin in the entire region. A source means that something flows out of that point, and a sink means that something flows into that point. The velocity control equations are derived from the pressure field using Darcy's velocity formula, as follows:

[0030] in, For the velocity field; The velocity field is used to characterize the flow path and velocity distribution of the resin in the mold, and its simulation results are output in the form of a vector field. The temperature control equation is based on the energy conservation equation and takes into account the convective heat transfer from resin flow and the heat conduction from the mold:

[0031] in, For the temperature field, For the temperature field gradient, The density of the resin, Specific heat capacity, a material parameter derived from the resin, is determined experimentally using methods such as differential scanning calorimetry. Thermal conductivity is a material parameter obtained through experimental measurement or a database. The heat source for the reaction is determined from the reaction kinetic model if the infusion process involves a chemical reaction; otherwise, it is determined from the reaction kinetic model. =0, The rate of change of temperature over time; The simulation solution time progression uses the implicit Euler method to ensure numerical stability; The simulation step size is synchronized with the update frequency of the process state vector.

[0032] The prediction model consists of an encoder and decoder structure, and incorporates an attention-based temporal neural network architecture to enhance the encoder and decoder, enabling the prediction model to adapt to the nonlinearity and temporal dependence of the perfusion process. (See also...) Figure 2 ; The encoder uses a multi-head self-attention mechanism to capture dependencies. The attention mechanism is calculated as follows:

[0033] in, , and FThese are the query, key, and value matrices, respectively, obtained from the input sequence through a linear transformation. For the dimensions of the query and key matrix, T This is the transpose of the key matrix; The encoder's output is a context vector that contains the global dependencies of the entire input sequence; The decoder, based on the encoder's context vector, generates predicted values ​​for each process result through fully connected layers, including: Flow front position: Indicates the spatial coordinates of the resin flow front at a future moment; Defect formation risk: Represents the probability that a defect will occur at a future time; Medium distribution uniformity: an index representing the uniformity of resin distribution at future time. Reaction completion percentage: If the infusion involves a chemical reaction, this indicates the percentage of the reaction to be completed at a future time.

[0034] Loss function of prediction model as follows:

[0035] in, 、 、 and These are the flow loss function, defect loss function, uniformity loss function, and completion loss function, respectively. 、 、 and This is a hyperparameter with a value range of [0,10]. It is used to balance the importance of different tasks in the total loss. The criticality of system control is adjusted according to the prediction target. For example, in resin infusion, the formation of internal bubbles or dry spots is absolutely unacceptable because it will cause the parts to be scrapped directly. Although it is not ideal if the flow front is predicted slightly slower or faster, it can be compensated for through real-time control. Based on this business logic, we can set the relative size of the hyperparameter as follows: This represents the risk of defect formation. Setting it to a very high value, such as 5.0, indicates that reducing defect prediction error is of paramount importance. This represents the degree of reaction completion. If resin curing is involved, the degree of completion is also crucial, so it is set to 2.0. This represents the uniformity of distribution, which affects product performance; it is set to 1.5. This represents the position of the flow front. Although important, a certain margin of error is allowed, so we set it to 1.0.

[0036] The definitions of each sub-loss function are as follows: The flow loss function, uniformity loss function, and completion loss function all use mean squared error to measure the difference between the predicted value and the actual value, as shown in the following formula:

[0037] in, For the true value, For predicted values, It is the number of training samples; The defect loss function is for a binary classification problem, classifying a problem as either defective or not defective. It uses binary cross-entropy loss to evaluate the difference between the predicted probability and the true label, as shown in the following formula:

[0038] in, These are genuine defect labels. This represents the predicted probability of the defect occurring.

[0039] The key feature indicators are extracted from the simulation data, and the specific steps are as follows: Extract the mean pressure and the mean value of the pressure gradient modulus from the pressure field to characterize the driving force and drag. Extract from the velocity field: location of the flow front and average velocity; Extracted from the temperature field: highest and lowest temperatures, temperature standard deviation, to characterize the curing state and uniformity; The extracted feature indicators are concatenated with the process state vector to obtain the input vector; By inputting historical and current input vectors into the prediction model, the future flow front location, defect formation risk, medium distribution uniformity, and response completion rate can be obtained.

[0040] The prediction details for the results of each process are as follows: Flow front position: The flow front position prediction is based on the flow velocity field simulation data and the flow characteristics in the process state vector. The prediction model predicts the advance of the flow front at future moments by analyzing the spatial distribution and changing trend of the historical flow velocity field. For example, by using the vector direction and mass conservation law in the flow velocity field, combined with real-time flow data, the prediction model can infer the displacement of the flow front. The prediction output is represented by three-dimensional coordinates or relative position. Defect formation risk: Defect formation risk prediction focuses on abnormal patterns in temperature and pressure fields. The prediction model extracts temperature fluctuation and pressure anomaly features from the process state vector and combines them with temperature gradients or pressure concentration areas in the simulation data to assess the probability of defects. For example, if the temperature field shows local overheating or the pressure field changes drastically, the prediction model will output a predicted value of the probability of defect formation risk. Medium distribution uniformity: The prediction of medium distribution uniformity depends on the uniformity index of the velocity field and pressure field. The prediction model calculates the field variance or entropy value in the simulation data and combines the flow rate and pressure correlation characteristics in the process state vector to predict the future distribution uniformity. The uniformity index is defined as the coefficient of variation of the resin coverage area. The prediction model outputs a continuous value through regression analysis. The higher the value, the better the uniformity. Reaction Completion: If the infusion process involves a chemical reaction, the reaction completion prediction is based on temperature field simulation data and temperature history in the process state vector. The prediction model uses chemical kinetic principles such as the Arrhenius equation, takes temperature-time integral as input feature, predicts the reaction progress, and outputs a percentage value, representing the cumulative reaction completion from the start to a future time.

[0041] M3: Decision module, equipped with a meta-trained reinforcement learning agent; the agent generates an adjustment strategy during the infusion process based on the process state vector and the process result, and by maximizing a preset multi-objective reward function; the agent is constructed based on an actor and critic framework.

[0042] The agent employs a deep reinforcement learning framework based on actors and critics, consisting of a policy network and a value network. The policy network is the actor, and the value network is the critic network. The policy network maps the current state to deterministic actions, i.e., adjusts the policy, while the value network evaluates the long-term cumulative reward of the state and actions to guide the optimization of the policy network. The current state refers to the comprehensive state vector formed by concatenating the current process state vector and the process result. The intelligent agent consists of an input layer, a hidden layer, and an output layer. The hidden layer uses the ReLU activation function, and the output layer uses the tanh function to constrain the action values ​​to the range [-1, 1], before mapping them to the physical range of the actual adjustment policy. Please refer to [link to relevant documentation]. Figure 3 ; Multi-objective reward function Defined as a weighted sum, the formula is as follows:

[0043] in, , , and These are weighting coefficients, which are experimentally optimized based on process priority and reward value dimensions. For example, in resin injection, defect prevention is the most critical factor, therefore... The value is much greater than , and , 、 、 and These are respectively: flow reward, defect reward, uniformity reward, and completion reward; The flow weighting coefficient, typically 1.0, encourages steady progress but gives way to other more critical objectives; This is the defect weighting coefficient, with a typical value of -3.0. It represents a strong negative reward, imposing a significant penalty once the risk of a defect is predicted to increase, thus forcing the agent to prioritize avoiding defects. This is the uniformity weighting coefficient, typically 1.5, which encourages the agent to make decisions that favor uniform distribution. This is a completion weighting coefficient, with a typical value of 2.0, to encourage efficient completion of infusion and reaction.

[0044] The flow reward is based on the position of the flow front and encourages the agent to advance the flow front smoothly and close to the target trajectory. The formula is as follows:

[0045] in, This represents the real-time displacement of the flow front. It is the target displacement, set based on process requirements; To reward based on the risk of defects, and to penalize high-risk situations to prevent defects, the formula is as follows:

[0046] in, This creates a risk due to defects. It is a small constant used to prevent errors in mathematical operations and ensure numerical stability, with a value range of [0.0000001, 0.00001]. For rewards based on the uniformity of medium distribution, the uniformity of medium distribution is used directly. ,Right now .

[0047] To reward participants based on reaction completion rate and encourage efficient reaction completion, the reaction completion rate is used directly. , .

[0048] The intelligent agent generates adjustment strategies based on a preset control cycle; Input the integrated state vector into the agent; The policy network, through forward computation, outputs one or more normalized action values, with a value range of [value range missing]. ; This abstract action value needs to be mapped to a specific value of one or more physical parameters during the infusion process. This mapping process is achieved through linear scaling, such as the resin injection pressure value. The calculation formula is as follows:

[0049] in, and These are the lower and upper limits of the safe pressure range allowed by the process. Action value; All physical parameters and their corresponding specific values ​​constitute the adjustment strategy; The injection actuator performs corresponding actions based on the adjustment strategy, such as adjusting the resin injection pressure value to... ; After the actuator performs an action, the intelligent system calculates the reward value; The agent stores this decision-making experience as an experience tuple in a fixed-capacity database called the experience replay buffer, which serves as a training sample for the agent. This is intended to break the temporal correlation between data and provide diverse historical data for stable learning. Periodically sample a small batch of empirical data randomly from the replay buffer to update the agent's parameters. The update adopts a soft update strategy, which allows the agent's parameters to slowly track the changes in the online network parameters, thereby significantly improving the stability of the entire learning process.

[0050] The steps for updating an agent are as follows: The value network calculates the future cumulative reward value of the current integrated state vector based on the current reward value using the Bellman equation. Q value; Computational value network prediction Q The value is compared with a calculated target Q value, as shown in the following formula:

[0051] in, For comparison; The parameters of the value network are directly adjusted using gradient descent to minimize... ; The policy network is updated as follows: The system calculates the gradient of the Q-value relative to the action to determine the direction for action improvement; Then this gradient is backpropagated to adjust the internal parameters of the policy network, eventually enabling it to learn to output actions that yield higher Q values. At the start of training, Q The value is typically initialized to a small random number or zero, for example, in the range of [-0.1, 0.1]. After training, during a stable perfusion process, for a good movement, its Q The value might be a stable positive number, for example, in the range [0.5, 2.0]; while for a bad action that would lead to a defect, Q The value may be a negative number, for example, in the range of [-1.0, -5.0].

[0052] Example 1: This example aims to verify the accuracy of the prediction model using specific data.

[0053] Table 1 shows the average predictive performance of the prediction model for each key indicator.

[0054] The prediction model demonstrated high accuracy in predicting all key indicators, particularly in its ability to warn of defect risks and its accuracy in tracking flow frontiers, providing reliable forward-looking information for subsequent intelligent decision-making.

[0055] To evaluate the performance of the reinforcement learning agent based on the predictive model in real-world control, we compared the performance of traditional PID control and the intelligent closed-loop control of this invention under the same injection task. The test object was a composite material aircraft wing rib component with a complex geometry.

[0056] The test conditions are as follows: The object to be infused is a carbon fiber preform; the resin is epoxy resin; the controlled variable is the resin injection pressure; the objective is to minimize the infusion time and improve the uniformity of distribution while ensuring no defects.

[0057] Table 2 shows the comparison results:

[0058] Specific control process analysis: When the infusion process reached the 12th minute, the prediction model gave an early warning 8 seconds in advance that there was a risk of defect formation at the corner of the component, which was as high as 85%. Upon receiving this prediction, the agent immediately generates an adjustment strategy, dynamically reducing the injection pressure from 350 kPa to 280 kPa for approximately 15 seconds. After pressure adjustment, flow field simulation data showed that the flow velocity in the corner region tended to stabilize, and actual sensor monitoring confirmed that the flow front smoothly passed through the region, successfully avoiding the occurrence of defects. Subsequently, the agent gradually restored the pressure to the optimized level, ensuring overall infusion efficiency.

[0059] This embodiment demonstrates through specific data that the present invention not only has high prediction accuracy, but also can transform prediction information into effective control actions, thereby reducing the defect rate and improving product uniformity without sacrificing efficiency, and enhancing the accuracy, foresight and robustness of complex injection processes.

[0060] This invention proposes a machine learning-based intelligent decision-making system for the infusion process. By introducing an attention-based encoder-decoder neural network as a prediction model and employing meta-trained actor and commentator reinforcement learning agents as the decision engine, this invention constructs a computational framework integrating intelligent perception, forward prediction, and autonomous decision-making. Through the synergistic effect of the aforementioned specific machine learning models, this system fundamentally solves the problems of low control accuracy and high defect rate in the infusion process caused by the lack of time-series prediction capabilities and multi-objective dynamic optimization mechanisms in traditional methods.

[0061] In summary, this invention constructs a system architecture with an advanced machine learning model as its intelligent core. By fusing the physical field characteristics provided by the digital twin with real-time process data and inputting them into a specific neural network and reinforcement learning model, it achieves forward-looking optimization control of nonlinear, multi-objective infusion processes. Its essential advantage stems from the superior ability of the specific computational model used to handle complex time series, achieve rapid generalization, and balance multiple objectives. Thus, while improving product quality consistency, it exhibits high adaptability and robustness.

[0062] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A machine learning-based intelligent decision-making system for the perfusion process, characterized in that, include: M1: A data sensing module that collects in-real-time infusion data of the infusion process and the infused object, and preprocesses and extracts features from the infusion data to generate a process state vector characterizing the real-time state of the infusion process; the infusion data includes at least pressure data, temperature data, and flow rate data; M2: Simulation and prediction module, equipped with digital twin model and prediction model; The digital twin model is constructed based on porous media fluid mechanics. The process state vector is used as the boundary condition. By solving Darcy's law and the energy conservation equation, simulation data of the interior of the injected object is generated. The simulation data includes pressure field, flow velocity field and temperature field. The prediction model employs an encoder-decoder neural network based on an attention mechanism; the prediction model predicts future process outcomes based on the simulation data and the process state vector, the process outcomes including the flow front location, defect formation risk, medium distribution uniformity, and reaction completion rate; M3: Decision module, equipped with a reinforcement learning agent that has undergone meta-training; The agent generates an adjustment strategy during the infusion process based on the process state vector and the process result, and by maximizing a preset multi-objective reward function. The agent adopts a deep reinforcement learning framework based on actors and critics, and consists of a policy network and a value network. The steps for generating the adjustment strategy are as follows: The process state vector and the process result are concatenated to obtain a comprehensive state vector, which is then input into the agent. The policy network outputs one or more normalized action values; The action value is mapped to a specific value of one or more physical parameters during the infusion process by linear scaling; all physical parameters and their corresponding specific values ​​constitute the adjustment strategy.

2. The system according to claim 1, characterized in that, The feature extraction is performed based on a preset time window; the features include statistical features, dynamic features, extreme value features, and correlation features; the features are concatenated to obtain the process state vector.

3. The system according to claim 2, characterized in that, The digital twin model needs to be pre-configured with the digital geometric model and material parameters of the object to be infused.

4. The system according to claim 3, characterized in that, The pressure field is solved based on the incompressible Navier-Stokes equations and Darcy's law, as shown in the following formula. in, For pressure field, For the pressure field gradient, For the permeability tensor of the preform, For fluid viscosity, For divergence; The velocity field is solved using Darcy's velocity formula, as follows. in, For the velocity field; The temperature field is solved based on the energy conservation equation, as shown in the following formula. in, For the temperature field, For the temperature field gradient, For the density of the fluid, For specific heat capacity, Thermal conductivity, As a heat source for the reaction, This represents the rate of change of temperature over time.

5. The system according to claim 4, characterized in that, The specific steps by which the prediction model outputs the process results are as follows: Key feature indicators were extracted from the simulation data, including mean pressure, mean pressure gradient modulus, flow front location, mean flow velocity, maximum and minimum temperature, and temperature standard deviation. The process state vector and the feature index are concatenated to obtain an input vector, and the input vector is then input into the prediction model. The encoder uses a multi-head self-attention mechanism to capture the dependencies of the input vector and outputs a context vector; The decoder then generates a predicted value of the process result based on the context vector through a preset fully connected layer; The loss function of the prediction model as follows, in, , , and These are the flow loss function, defect loss function, uniformity loss function, and completion loss function, respectively. This is a hyperparameter.

6. The system according to claim 5, characterized in that, The flow loss function, uniformity loss function, and completion loss function are all mean square error; the defect loss function is a binary cross-entropy loss.

7. The system according to claim 6, characterized in that, The agent includes an input layer, a hidden layer, and an output layer, wherein the hidden layer uses the ReLU activation function, and the output layer uses the tanh function to constrain the action values; The multi-objective reward function Defined as a weighted sum, the formula is as follows. in, 、 、 and These are the weighting coefficients. 、 、 and These are respectively: flow reward, defect reward, uniformity reward, and completion reward; The preset infusion actuator performs actions according to the adjustment strategy; After the infusion actuator performs the action, the agent calculates the reward value and stores this decision-making experience as an experience tuple for updating the agent; The experience tuple includes the integrated state vector, action value, and reward value.

8. The system according to claim 1, characterized in that, The preprocessing includes digital filtering, sensor calibration, outlier detection and compensation, and data standardization.

Images