Adaptive prediction system for quenching process temperature field and cooling rate control parameters

Abstract

This invention discloses an adaptive prediction system for temperature field and cooling rate control parameters in quenching processes. The invention relates to the field of intelligent heat treatment production lines and industrial control systems. The system consists of a perception layer, a cognition layer, a decision layer, and an execution layer. The perception layer collects workpiece temperature and bypass medium property parameters. The cognition layer estimates the comprehensive heat transfer coefficient online using a degradation compensation neural network, and uses a physical information neural network to reconstruct the workpiece's three-dimensional full-field temperature in real time. A reduced-order model then performs second-level predictions of the coupled thermo-microstructure-mechanical fields. The decision layer uses a model-independent element learning framework to achieve rapid cross-workpiece migration and optimizes the cooling path online using a sequential quadratic programming method. The execution layer adjusts the medium flow rate through a frequency converter. This invention solves three major problems: the inability to perceive the degradation of medium cooling capacity online, low computational efficiency in real-time prediction of multi-physics fields, and insufficient generalization ability of models for multiple product scenarios.

Description

Technical Field

[0001] This invention relates to the field of intelligent heat treatment production lines and industrial control systems, specifically to an adaptive prediction system for temperature field and cooling rate control parameters in quenching processes. Background Technology

[0002] Quenching is one of the most critical strengthening processes in metal heat treatment. Its basic principle is to rapidly immerse the austenitized metal workpiece in a cooling medium and cool it down at a rate exceeding the critical cooling rate, transforming the high-temperature austenitic structure into a high-hardness martensite or bainite structure, while controlling the distribution of residual stress and the amount of geometric deformation inside the workpiece.

[0003] The current mainstream solution in the industry is a composite control scheme based on PID feedback control with a fixed cooling curve and offline finite element verification. In the offline process design stage, process engineers use finite element simulation software (such as DEFORM, DANTE, Abaqus, etc.) to establish a thermo-organism-mechanical coupled simulation model offline, calculate the theoretically optimal cooling curve, and solidify the correspondence between stirrer frequency and cooling rate into the process formula file. In the online execution stage, the programmable logic controller drives the frequency converter to control the stirrer speed according to a preset time-frequency table, and at the same time obtains temperature feedback through a small number of thermocouples, which is then slightly corrected by the PID controller. After quenching, the process formula is manually adjusted by the process engineer through quality inspection, forming a post-process closed loop.

[0004] The existing solutions described above have the following three fundamental flaws:

[0005] First, there is a lack of online sensing of the degradation of the cooling capacity of quenching media. After long-term circulation in industrial sites, quenching media (quenching oil, water-soluble quenching liquid, etc.) undergo degradation due to oxidation decomposition, increased water content, abnormal viscosity changes, and impurity contamination. This causes unpredictable dynamic drift in their comprehensive heat transfer coefficient. Existing control systems use fixed comprehensive heat transfer coefficient parameters and do not configure any online sensors for the medium status, making it impossible to detect the above drift. This leads to the continuous accumulation of cooling rate control deviations, which ultimately causes batch quality problems such as excessive residual stress in the workpiece, excessive quenching deformation, and uneven hardness. Moreover, these defects can only be detected in the post-production quality inspection stage, making it impossible to achieve feedforward prevention.

[0006] Second, the computational efficiency bottleneck of real-time prediction of multi-physics fields throughout the quenching process. During the quenching process, there is a strong coupling interaction between the thermal field, the metallographic phase transformation field and the mechanical field. Accurately describing this process requires complex three-field coupled finite element simulation. Traditional finite element simulation takes several hours or even tens of hours per simulation, which is far beyond the time scale of the quenching process itself (usually several seconds to several minutes). It is impossible to embed a real-time control closed loop, which leads to the system being blind to the actual physical state during the quenching process and having no real-time response capability to any working condition disturbance.

[0007] Third, the generalization ability of models in multi-variety, small-batch scenarios is insufficient. Modern intelligent manufacturing puts forward flexible processing requirements for quenching production lines. The types of workpieces (material, geometry, cross-sectional dimensions) change frequently. Existing machine learning temperature field prediction schemes rely heavily on a large amount of calibration data for specific workpieces. Once the type of workpiece changes, the original model becomes invalid and data needs to be collected and trained again. The calibration cost is high and the response cycle is long, which seriously restricts the flexible deployment capability of the system. Summary of the Invention

[0008] The technical problem that the invention aims to solve

[0009] The technical problem to be solved by this invention is to provide an adaptive prediction system for the temperature field and cooling rate control parameters of the quenching process, which can estimate the cooling capacity of the medium in real time, achieve three-field coupled prediction in seconds, and support rapid migration across workpieces. This system addresses the three major problems of existing quenching control systems: the inability to sense the comprehensive heat transfer coefficient of the medium online, the low efficiency of real-time prediction of multi-physics fields, and the insufficient generalization ability of the model in multi-variety and small-batch scenarios.

[0010] Technical solution

[0011] This invention provides an adaptive prediction system for temperature field and cooling rate control parameters in quenching processes, comprising a perception layer, a cognition layer, a decision-making layer, and an execution layer. The layers interact with each other in real time via industrial Ethernet or fieldbus, forming a complete perception-prediction-decision-execution closed loop.

[0012] The sensing layer includes an embedded thermocouple array, a bypass fluid characteristic sensor unit, a tank liquid temperature sensor, and a stirrer speed sensor. The thermocouple array has at least four temperature measurement points at the typical cross section with the largest temperature gradient on the workpiece surface. At least three liquid temperature sensors are evenly arranged along the height direction in the liquid phase region of the tank. All thermocouple signals are standardized by a temperature transmitter and then connected to a high-speed data acquisition card with a sampling frequency of not less than 10Hz.

[0013] The bypass fluid characteristic sensor unit is set in the bypass pipeline of the quenching tank. It includes an ultrasonic online viscometer, a high-frequency conductivity meter and a bypass liquid temperature sensor. The sensor data is accessed to the edge computing node via the OPC-UA protocol and is timestamped and fused with the workpiece temperature data.

[0014] The cognitive layer is deployed on edge computing nodes and includes a multi-source heterogeneous data fusion and preprocessing module, a medium comprehensive heat transfer coefficient online estimation and degradation compensation module, a physical information neural network three-dimensional temperature field real-time reconstruction module, and a reduced-order model thermo-organic-mechanical three-field coupling prediction module.

[0015] The online estimation and degradation compensation module for the comprehensive heat transfer coefficient of the medium constructs a medium state feature vector consisting of medium liquid temperature, dynamic viscosity, conductivity, stirring flow rate, and cumulative usage time in each control cycle.

[0016] Using the medium temperature and stirring flow rate as indexes, the nominal reference heat transfer coefficient is obtained by interpolation from a pre-stored standard comprehensive heat transfer coefficient lookup table.

[0017] Input the feature vector of the medium state into the degradation compensation neural network and output the relative offset of the comprehensive heat transfer coefficient;

[0018] The estimated composite heat transfer coefficient for the current cycle is obtained by adding the nominal reference heat transfer coefficient to the offset.

[0019] The estimated value is physically validated. If it exceeds the physical allowable range for the media type, it is trimmed to the boundary value and a media status abnormality alarm is sent.

[0020] The physical information neural network three-dimensional temperature field real-time reconstruction module embeds the heat conduction partial differential equation into the neural network loss function, and uses the sparse thermocouple measurement value and the current cycle estimated comprehensive heat transfer coefficient as joint constraints to obtain discrete approximations of the workpiece's three-dimensional full field temperature in batch inference in each control cycle, with an inference delay of no more than 50ms.

[0021] The martensite volume fraction was estimated point by point based on the reconstructed temperature field and the Johnson-Mehl-Avrami phase transformation kinetic equation.

[0022] The reduced-order model's thermo-organic-mechanical three-field coupling prediction module employs the intrinsic orthogonal decomposition-Galogen projection method.

[0023] In the offline phase, singular value decomposition was performed on the snapshots of the temperature field, martensite volume fraction field, and equivalent stress field of the parameterized finite element simulation. The dominant orthogonal basis vectors with a cumulative energy ratio of more than 99.9% were extracted, the reduced-order orthogonal basis matrix of each field was constructed, and the low-dimensional coefficient equation system was obtained by Galerkin projection.

[0024] The online phase is driven by the temperature field output by the physical information neural network. The fourth-order Runge-Kutta method is used to perform cycle-by-cycle time-step integration of the low-dimensional equation system. The single-step integration can be completed within 10ms on a regular CPU. The output includes the maximum principal stress on the surface, the stress at the body center, the degree of martensitic transformation completion, and the future temperature trend.

[0025] The decision layer includes a meta-learning cross-workpiece migration adaptation module, a process target management module, and a cooling path optimization solver. The meta-learning cross-workpiece migration adaptation module adopts a model-independent meta-learning framework, and learns and optimizes meta-initial parameters on simulation and experimental data covering multiple workpiece types in the offline stage.

[0026] In the online phase, the system reads the workpiece type code issued by the production scheduling system to determine whether it is an unknown new type. If it is a new type, it starts with the initial parameters and performs gradient updates using only a small amount of measured data collected from 5 to 10 quenching cycles. The model migration and adaptation are completed within 2 minutes, and the adaptation parameters are registered in the local model library and bound to the workpiece type code. Subsequent incremental refinement is performed as batches accumulate. The cooling path optimization solver uses the stirrer frequency sequence within the future prediction step as the decision variable, the weighted sum of temperature tracking error, maximum surface stress, and control quantity change rate as the objective function, and the constraints of inverter physical limitations, ramp rate limitations, and residual stress upper limits as constraints. It uses a sequential quadratic programming method to solve the problem, only executing the first control quantity issued, and the solution time does not exceed 500ms.

[0027] The execution layer includes a programmable logic controller (PLC), a variable frequency drive (VFD), and a quenching cooling tank. The PLC receives the stirrer frequency command output from the decision layer and drives the stirrer and circulating pump through the VFD to adjust the flow rate of the quenching medium.

[0028] The present invention has the following beneficial effects:

[0029] 1. This invention achieves real-time online estimation of the comprehensive heat transfer coefficient of quenching medium through the combined configuration of bypass multi-source sensors and degradation compensation neural networks. It transforms the dynamic offset of the comprehensive heat transfer coefficient into a perceptible closed-loop feedback quantity, completely eliminating the implicit impact of medium condition drift on product quality and effectively reducing the residual stress exceedance rate.

[0030] 2. This invention, through the fusion architecture of physical information neural network and reduced-order model, compresses the calculation delay of thermo-organism-mechanical three-field coupling prediction from hours to seconds (the comprehensive calculation time of a single control cycle does not exceed 60ms), making real-time closed-loop control based on physical simulation possible.

[0031] 3. This invention, through a model-independent meta-learning framework, enables the system to complete rapid model transfer with only 5 to 10 quenching samples when facing new workpiece types, usually within 2 minutes. This achieves truly flexible quenching control and meets the production needs of rapid changeover in small batches of multiple varieties. Attached Figure Description

[0032] Figure 1 This is a schematic diagram of the four-layer architecture of the system of the present invention and the data flow between each module;

[0033] Figure 2 This is a schematic diagram of the cycle-by-cycle execution process of the online estimation and degradation compensation module for the integrated heat transfer coefficient of the medium in this invention;

[0034] Figure 3 This is a schematic diagram of the entire main control loop process of the system of the present invention;

[0035] Figure 4 This is a schematic diagram of the three-layer hardware deployment architecture of the system of the present invention. Detailed Implementation

[0036] The following is in conjunction with the appendix Figure 1-4 The specific embodiments of the present invention will be further described below. It should be noted that the description of these embodiments is for the purpose of helping to understand the present invention, but does not constitute a limitation of the present invention.

[0037] Furthermore, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0038] System Overall Architecture

[0039] like Figure 1 As shown, the adaptive prediction system of the present invention consists of four functional layers: perception layer, cognition layer, decision layer and execution layer. The layers interact with each other in real time through industrial Ethernet or fieldbus (PROFINET / EtherCAT) to form a complete perception-prediction-decision-execution closed loop.

[0040] Perception layer: Multi-source heterogeneous data acquisition

[0041] K-type high-temperature resistant thermocouples are pre-embedded or installed in contact at key locations inside and on the surface of the workpiece. These thermocouples have a measuring range of 0 to 1300 degrees Celsius and a response time not exceeding 500 ms, forming a spatially sparse temperature sensor array. Thermocouple arrangement follows these principles:

[0042] Select a typical section with the largest temperature gradient on the workpiece surface (including the section with the largest temperature gradient and the section with abrupt change) and arrange no less than four temperature measurement points;

[0043] At least three liquid temperature sensors are evenly arranged along the height direction in the liquid phase region of the tank to monitor the temperature field distribution of the medium.

[0044] All thermocouple signals are standardized to 4 to 20mA analog signals by a temperature transmitter and then connected to a high-speed data acquisition card with a sampling frequency of not less than 10Hz.

[0045] A dedicated medium property detection circuit is set on the bypass pipeline of the quenching tank, including an ultrasonic online viscometer, a high-frequency conductivity meter and a bypass liquid temperature sensor. The ultrasonic online viscometer uses the propagation characteristics of ultrasonic waves in the fluid (sound velocity and attenuation coefficient) to calculate the dynamic viscosity of the medium in real time. The range is 1 to 5000 mPa·s, the measurement accuracy is better than ±1%, and the output refresh frequency is not less than 1Hz.

[0046] The high-frequency conductivity meter indirectly reflects the concentration and impurity content of water-soluble quenching fluid by measuring the conductivity of the medium. For quenching oil scenarios, a dielectric constant sensor is used instead.

[0047] The bypass liquid temperature sensor accurately measures the temperature of the medium at the bypass detection location, and is used to perform temperature compensation correction on the measured physical properties.

[0048] The bypass sensor data mentioned above is accessed to the edge computing node via the OPC-UA protocol to achieve timestamp alignment and fusion with the workpiece temperature data.

[0049] Cognitive Layer Module 1: Online Estimation and Degradation Compensation of Integrated Heat Transfer Coefficient of Medium

[0050] This module is one of the core innovations of this invention. Its goal is to estimate the true comprehensive heat transfer coefficient of the quenching medium under current operating conditions online in real time, and to use the offset relative to the nominal value for subsequent multiphysics prediction and control parameter compensation. The execution flow is as follows: Figure 2 As shown, the control cycle is set to 1 second.

[0051] Within each control cycle, sensor data at the current moment is collected from the sensing layer to construct a medium state feature vector. This feature vector consists of five components: the medium temperature (in degrees Celsius) measured by the bypass liquid temperature sensor in the current cycle, the dynamic viscosity of the medium (in mPa·s) measured by the ultrasonic viscometer, the medium conductivity (in mS / cm) measured by the conductivity meter, the current speed converted to flow rate (in m / s) fed back by the stirrer speed encoder, and the cumulative usage time (in hours) maintained by the system's internal timer since the last medium replacement.

[0052] Using the current cycle medium temperature and stirring flow rate as an index, the nominal reference heat transfer coefficient under the current temperature and flow rate conditions is obtained by interpolation from a pre-stored standard comprehensive heat transfer coefficient lookup table. This lookup table is obtained offline according to ASTM D6482 standard through standard silver ball cooling experiments under the initial fresh state of the medium.

[0053] The medium state feature vector is input into a pre-trained lightweight degradation compensation neural network. This network adopts a three-layer fully connected structure with 64, 32 and 16 neurons in the hidden layer, respectively. The activation function is ReLU, and the output layer adopts linear activation. The output is the relative offset of the comprehensive heat transfer coefficient. The trainable parameters of the degradation compensation neural network are obtained offline through historical medium degradation experimental datasets before system deployment and support online incremental fine-tuning.

[0054] Adding the nominal reference heat transfer coefficient to the degradation compensation offset yields the estimated overall heat transfer coefficient for the current cycle:

[0055] ;

[0056] in, Estimate the overall heat transfer coefficient for the current cycle (unit: W / (m²·K)). The nominal reference heat transfer coefficient (unit: W / (m²·K)) is obtained by interpolation from a lookup table. The relative offset of the overall heat transfer coefficient output by the degradation compensation neural network (unit: W / (m²·K)).

[0057] right Perform a physical rationality check: If the estimated value exceeds the physically possible range of the medium type (the upper and lower limits determined by the physical property manual provided by the medium manufacturer), the estimated value is cut to the corresponding boundary value, and a medium status abnormality alarm is sent to the upper-level module to prompt the operator to manually check the medium.

[0058] When the system accumulates a sufficient number of new samples during online operation, it triggers incremental fine-tuning of the degradation compensation neural network, updates the network parameters, and achieves continuous learning and adaptation to the degradation characteristics of the medium.

[0059] Cognitive Layer Module 2: Real-time Reconstruction of Three-Dimensional Temperature Field by Physical Information Neural Network

[0060] Under sparse thermocouple data conditions, direct interpolation to obtain the workpiece's full-field temperature distribution results in significant errors. This module employs a physical information neural network, embedding the partial differential equation of heat conduction into the neural network loss function using a soft constraint approach. Driven by a combination of a small number of sensor measurements and physical laws, it achieves three-dimensional full-field temperature measurement of the workpiece. High-precision real-time reconstruction, among which Let be the coordinate vector of the workpiece's internal space. For time coordinates.

[0061] The loss function of a physical information neural network consists of three parts: ;

[0062] in, The weighting coefficients for each loss term are determined through offline parameter tuning. The data loss term represents the mean square error between the network-predicted temperature and the actual temperature measured by the sparse sensors. It is used to anchor the network predictions to the true measurements. The physical residual loss term represents the sum of squared residuals at which the network-predicted temperature does not satisfy the unsteady heat conduction equation at the placement point. The embedded heat conduction equation is:

[0063] ;

[0064] in, This refers to the material density (unit: kg / m³). Specific heat capacity (unit: J / (kg·K)) Thermal conductivity (unit: W / (m·K)) This is the latent heat source term for phase change (unit: W / m³). For divergence operators, For gradient operators, The boundary condition loss term ensures that the workpiece surface temperature meets the estimated comprehensive heat transfer coefficient based on the current cycle. The convective heat transfer boundary condition for the heat transfer coefficient enables the physical information neural network reconstruction process and the online estimation of the comprehensive heat transfer coefficient of the medium to form a dynamic closed loop.

[0065] The network adopts a six-layer fully connected structure, with the input layer receiving spatial coordinates. With time coordinates There are four inputs in total. The output layer outputs the predicted temperature at this spatiotemporal point. Each hidden layer has 128 neurons, and the activation function is... (Its smoothness facilitates automatic differential calculation of partial differential equation residuals.) In the offline phase, the network utilizes a large-scale training set constructed from high-precision finite element simulation results for the target workpiece family, minimizing... After completing pre-training and obtaining pre-training parameters, during online inference, the appropriate parameters for the current workpiece type are obtained from the meta-learning module to initialize the inference instance.

[0066] In each control cycle, the current time and the set of coordinates of the regular configuration points within the workpiece geometry (total number of configuration points) are used. Typically set to 1000 to 5000, these values ​​are fed into the network in batches, and a single forward inference operation yields a discrete approximation of the entire temperature field. Inference latency is no more than 50ms on industrial GPU edge nodes.

[0067] Based on the reconstructed temperature field and combined with the Johnson-Mehl-Avrami phase transformation kinetic equation, the current martensite volume fraction is calculated point by point. The equation integral is completed by recursion according to the control period using the Euler explicit method, with minimal computational cost. The full-field temperature discrete approximation and the martensite volume fraction field are simultaneously transferred to the three-field coupled prediction module of the reduced-order model.

[0068] Cognitive Layer Module 3: Reduced-Order Model for Coupling Prediction of Thermo-Organic-Mechanical Fields

[0069] This module uses the intrinsic orthogonal decomposition-Galogen projection method to achieve reduced-order modeling, projecting the original finite element equations with dimensions of tens or even hundreds of thousands to a low-dimensional subspace spanned by dozens of orthogonal bases, thus compressing the computational scale by thousands of times.

[0070] In the offline construction phase, parametric finite element simulations covering the process parameter space (different cooling rates, different comprehensive heat transfer coefficients, and different initial temperatures) are run for the target workpiece family. A set of time snapshots of the temperature field, martensite volume fraction field, and Von Mises equivalent force field is collected. Singular value decomposition is performed on each physical field, and the top-performing fields with accumulated energy exceeding 99.9% are selected. singular vectors ( (between 20 and 50), forming the reduced-order orthogonal basis matrix for each field. subscript Corresponding to the temperature field, martensite volume fraction field, and stress field respectively, the complete nonlinear coupled equations are projected into a low-dimensional space using Galerkin projection to obtain a low-dimensional coefficient equation set.

[0071] During the online prediction phase, each time a new workpiece enters the tank, the initial discrete values ​​of the full-field temperature output by the physical information neural network are projected onto a reduced-order orthogonal basis of the temperature field. Above, calculate the initial low-dimensional coefficient vector. The initial coefficient vectors of other fields are set to zero (assuming the workpiece is fully austenitized, with no martensite and no residual stress). In each control cycle, the sparse temperature update output by the physical information neural network is used as the driving input, and the low-dimensional equations are integrated in one step using the fourth-order Runge-Kutta method to update the low-dimensional coefficient vectors of each field. The computational cost of a single-step integration is only a few tens of order matrix operations, which can be completed within 10ms on a regular CPU. This allows for the reconstruction of physical quantities at key locations using low-dimensional coefficient vectors.

[0072] pass Reconstruct the current stress field and extract the maximum principal stress on the surface. and body center position stress ,pass Extraction of martensitic transformation completion (Spatial average value), and also output the predicted temperature trend for the next five cycles. All results are passed to the cooling path optimization solver of the decision layer.

[0073] Decision Layer Module 1: Meta-learning Cross-Workpiece Transfer Adaptation

[0074] This module employs a model-independent meta-learning framework. During the meta-training phase, a meta-training dataset containing multiple workpiece types is constructed. For each workpiece type, 5 to 10 quenching process data points are provided as support set samples and corresponding validation set samples. Following the two-layer optimization framework of model-independent meta-learning, the inner loop performs several gradient updates on the support set, while the outer loop calculates the meta-gradient and updates the meta-initial parameters on the validation set. This iteration continues until convergence, yielding the optimized meta-initial parameters. This allows the network to quickly converge to parameters suitable for the new workpiece with only a small number of new samples, starting from the initial point.

[0075] During the online triggering phase, the system reads the workpiece type code (a combination of workpiece material grade and geometric feature hash value) issued by the production scheduling system to determine whether the current workpiece is a known type. If it is a known type, the system directly retrieves the corresponding adaptation parameters from the local model library. If it is an unknown new type, the system triggers the online rapid migration process: performing 5 to 10 quenching cycles on the current new workpiece type to collect a small amount of measured data, forming a new task support set, with initial parameters... Starting from this point, gradient updates are performed on the adaptable parameter layers of the physical information neural network and the reduced-order model:

[0076] ;

[0077] in, To adapt to the updated parameters after the new workpiece, The inner loop learning rate (a hyperparameter, determined through meta-training). To support the new mission set The prediction loss function on, For parameters The gradient operator, after several steps of gradient update, yields... The entire migration process typically takes no more than 2 minutes (on edge GPU nodes). Registered to the local model library, bound to the workpiece type code, and continuously updated with measured data during subsequent formal production. Incremental refinement is performed, and the model accuracy continues to improve as batches accumulate.

[0078] Decision-making module two: Adaptive optimization solution for cooling paths

[0079] Cooling path optimization employs a rolling time-domain control strategy, which, in each control cycle, considers the future... One cycle ( Typically, a stirrer frequency sequence of 10 is taken as the decision variable, and the following objective function is minimized:

[0080] ;

[0081] in, To predict the step size, These are the weighting coefficients for the three factors: temperature tracking error, maximum surface stress, and rate of change of control quantity. For the future predicted by the reduced-order model Average temperature per step (in degrees Celsius). For the target cooling curve in the future The target temperature for the step is retrieved from the process target curve library according to the current workpiece type (unit: degrees Celsius). For the future predicted by the reduced-order model Maximum surface principal stress (unit: MPa) at step. This represents the change in stirrer frequency between adjacent cycles (in Hz).

[0082] The constraints include: the agitator frequency is within the upper and lower limits allowed by the inverter's physical limits; the absolute value of the agitator frequency change between adjacent cycles does not exceed the ramp rate limit; and the maximum surface principal stress does not exceed the upper limit of the allowable stress of the process. The above optimization problem is solved using a sequential quadratic programming method, with a reduced-order model as the predictive dynamics. Only the first control variable of the optimal frequency sequence is issued and executed, and the solution time on the edge computing node does not exceed 500ms.

[0083] Hardware deployment architecture

[0084] like Figure 4 As shown, the system adopts a three-layer hardware deployment architecture. The field device layer includes a thermocouple array (workpiece and tank), an ultrasonic viscometer (bypass pipeline), a conductivity meter (bypass pipeline), a speed encoder (stirring motor), a quenching and cooling tank (workpiece and medium), and a variable frequency drive device (stirring and circulating pump). It is connected to the edge computing layer through shielded cables. The industrial edge GPU nodes (including neural processing unit acceleration units) of the edge computing layer undertake all computational tasks of physical information neural networks, reduced-order models, and meta-learning inference. The high-speed data acquisition card accesses the sensor signals at a sampling frequency of 10Hz. The programmable logic controller acts as a PROFINET master station, exchanging millisecond-level real-time data with the edge GPU nodes through PROFINET and issuing the final control frequency command to the variable frequency drive device. The production scheduling system of the cloud / upper computer layer is responsible for issuing workpiece type and target cooling curve. The model warehouse stores the initial parameters and the comprehensive heat transfer coefficient benchmark table. The SCADA monitoring upper computer realizes real-time visualization and alarms. The above functions communicate with the edge computing layer through the OPC-UA standard protocol to realize a complete information-physical fusion closed loop from sensors to actuators.

[0085] System full-process operation logic

[0086] like Figure 3 As shown, after the system starts, it reads the workpiece type code and obtains the target cooling curve from the production scheduling system to determine whether the current workpiece type is known.

[0087] If the type is known, the adapted parameters in the local model library are loaded. If the type is unknown, meta-learning is triggered for rapid migration. After several quenching data collection cycles, the parameters are adapted and registered in the library. After the workpiece is placed in the quenching tank, the system enters a cycle-by-cycle control loop, sequentially executing the following: data collection at the perception layer, online estimation of the comprehensive heat transfer coefficient and verification of physical constraints, reconstruction of the temperature field and estimation of martensite volume fraction by the physical information neural network, prediction of the three-field coupling of the reduced-order model, optimization of the cooling path, issuance of instructions to the execution layer, real-time quality assessment, and data recording. If the maximum stress exceeds the allowable threshold, an alarm is issued and the cooling intensity is forcibly reduced to notify the operator. The quenching termination condition is checked cyclically (the workpiece surface temperature is lower than the martensite transformation end temperature). When the condition is met, the workpiece is removed from the quenching tank, triggering incremental learning and updating of the parameters of the degradation compensation neural network and the physical information neural network to prepare for the next workpiece.

[0088] Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make possible changes and modifications without departing from the spirit and scope of the present invention. Therefore, any modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention shall fall within the protection scope defined by the claims of the present invention.

Claims

1. An adaptive prediction system for temperature field and cooling rate control parameters in a quenching process, characterized in that, It includes a perception layer, a cognition layer, a decision-making layer, and an execution layer that interact with data in real time via industrial Ethernet / fieldbus; The sensing layer includes an embedded thermocouple array, a bypass fluid characteristic sensor unit, a tank liquid temperature sensor, and a stirrer speed sensor. The cognitive layer is deployed on edge computing nodes and includes a multi-source heterogeneous data fusion and preprocessing module, a medium comprehensive heat transfer coefficient online estimation and degradation compensation module, a physical information neural network three-dimensional temperature field real-time reconstruction module, and a reduced-order model thermo-organic-mechanical three-field coupling prediction module. The online estimation and degradation compensation module for the comprehensive heat transfer coefficient of the medium constructs a medium state feature vector based on the medium liquid temperature, dynamic viscosity, conductivity, stirring flow rate and cumulative usage time in each control cycle. It obtains the nominal reference heat transfer coefficient by looking up a table, calculates the relative offset of the heat transfer coefficient by a degradation compensation neural network, and adds the nominal reference heat transfer coefficient and the offset to obtain the estimated comprehensive heat transfer coefficient for the current cycle. The physical information neural network three-dimensional temperature field real-time reconstruction module embeds the unsteady heat conduction partial differential equation into the neural network loss function, uses the sparse thermocouple measurement value and the current cycle estimated comprehensive heat transfer coefficient as joint constraints, and obtains the discrete approximation of the three-dimensional full field temperature of the workpiece through cycle-by-cycle reasoning, and estimates the martensite volume fraction point by point in combination with the Johnson-Mehl-Avrami phase transformation kinetic equation. The reduced-order model thermo-structure-mechanical three-field coupled prediction module is based on a low-dimensional coefficient equation system constructed by intrinsic orthogonal decomposition-Galogen projection. Driven by the temperature field output by the physical information neural network three-dimensional temperature field real-time reconstruction module, it outputs the maximum principal stress, martensite transformation completion degree and future temperature trend of the surface by cycle-by-cycle time step integration. The decision layer includes a meta-learning cross-workpiece migration adaptation module and a cooling path optimization solver. The meta-learning cross-workpiece migration adaptation module is based on a model-independent meta-learning framework. It learns and optimizes meta-initial parameters offline. After identifying new workpiece types online, it performs gradient updates using only a small number of quenching samples to complete the rapid model migration and registers the adaptation parameters to the local model library. The cooling path optimization solver uses the reduced-order model thermo-microstructure-mechanical three-field coupling prediction module as the prediction dynamics. It uses a sequential quadratic programming method to solve for the optimal stirrer frequency in each control cycle and issues it for execution using a rolling time-domain strategy. The execution layer includes a programmable logic controller, a frequency converter, and a quenching cooling tank. It receives the stirrer frequency command output by the decision layer and drives the stirrer and the circulating pump through the frequency converter to adjust the flow rate of the quenching medium.

2. The system according to claim 1, characterized in that, The bypass fluid characteristic sensor unit is installed in the bypass pipeline of the quenching tank, including an ultrasonic online viscometer, a high-frequency conductivity meter and a bypass liquid temperature sensor; The ultrasonic online viscometer uses the propagation speed and attenuation coefficient of ultrasonic waves in the fluid to calculate the dynamic viscosity of the medium in real time, with a range of 1 to 5000 mPa·s. The high-frequency conductivity meter indirectly reflects the concentration and impurity content of water-soluble quenching fluid by measuring the conductivity of the medium, and replaces the dielectric constant sensor in the quenching oil scenario. The bypass liquid temperature sensor is used to perform temperature compensation correction on the measured physical properties. Bypass sensor data is accessed to the edge computing node via the OPC-UA protocol and timestamped and fused with the workpiece thermocouple signal.

3. The system according to claim 1, characterized in that, The online estimation and degradation compensation module for the overall heat transfer coefficient of the medium also includes a physical rationality verification step: The estimated comprehensive heat transfer coefficient for the current cycle is checked. If the estimated value exceeds the upper and lower limits allowed by the physical properties of the current medium type, the estimated value is cut to the corresponding boundary value, and a medium status abnormality alarm is sent to the upper layer module. When a preset number of new samples are accumulated during online operation, the incremental fine-tuning of the degradation compensation neural network is triggered to update the network's trainable parameters, thereby enabling continuous learning and adaptation to the degradation characteristics of the medium. The degradation compensation neural network adopts a three-layer fully connected structure, with the number of neurons in the three hidden layers being 64, 32, and 16 respectively, the activation function being ReLU, and the output layer using linear activation.

4. The system according to claim 1, characterized in that, The loss function used by the physical information neural network three-dimensional temperature field real-time reconstruction module is composed of a weighted sum of data loss terms, physical residual loss terms, and boundary condition loss terms. Among them, the data loss term is the mean square error between the network-predicted temperature and the temperature measured by the sparse sensor; The physical residual loss term is the sum of squares of the residuals where the network-predicted temperature does not satisfy the unsteady heat conduction equation at the placement point; The boundary condition loss term is the sum of squares of the residuals of the convective heat transfer boundary conditions, where the workpiece surface temperature is not sufficient to estimate the overall heat transfer coefficient for the current period. The network adopts a six-layer fully connected structure, taking four quantities as input: workpiece internal spatial coordinates and time coordinates. Each hidden layer has 128 neurons, and the activation function is tanh. The latency for a single forward inference operation is no more than 50ms.

5. The system according to claim 1, characterized in that, The offline construction process of the reduced-order model thermo-organism-mechanical three-field coupled prediction module includes: Parametric finite element simulations covering different cooling rates, different comprehensive heat transfer coefficients, and different initial temperatures were run for the target workpiece family, and time snapshot sets of temperature field, martensite volume fraction field, and equivalent stress field were collected. Singular value decomposition is performed on each physical field, and singular vectors with a cumulative energy ratio of more than 99.9% are selected to form the reduced-order orthogonal basis matrix of each field. The number of selected singular vectors is between 20 and 50. The complete nonlinear coupled equations are projected onto a low-dimensional space by Galerkin projection to obtain a low-dimensional coefficient equation system. In the online phase, the fourth-order Runge-Kutta method is used to perform cycle-by-cycle time-step integration of the low-dimensional equation system, with a single-step integration time of no more than 10ms.

6. The system according to claim 1, characterized in that, The online migration process of the meta-learning cross-workpiece migration adaptation module includes: The system determines whether the current workpiece is an unknown new type by reading the workpiece type code issued by the production scheduling system; If it is an unknown new type, then starting from the initial parameters, a new task support set is formed by collecting a small amount of measured data from 5 to 10 quenchings. Then, 5 to 10 gradient updates are performed on the adaptable parameter layer of the physical information neural network three-dimensional temperature field real-time reconstruction module and the reduced-order model thermo-organism-mechanical three-field coupling prediction module. The migration process takes no more than 2 minutes. After the migration is completed, the adaptation parameters are bound to the workpiece type code and registered in the local model library, and the adaptation parameters are continuously refined incrementally with actual measurement data in subsequent formal production.

7. The system according to claim 1, characterized in that, The objective function of the cooling path optimization solver is the following weighted sum of squares: Within the prediction step, for each future time point, the square of the difference between the predicted workpiece temperature and the target cooling curve temperature, the square of the maximum surface principal stress, and the square of the change in agitator frequency in adjacent cycles are multiplied by the corresponding weighting coefficients and then summed. The constraints include the upper and lower limits of the agitator frequency being within the physical allowable range of the frequency converter, the absolute value of the change in agitator frequency between adjacent cycles not exceeding the ramp rate limit, and the maximum surface principal stress not exceeding the upper limit of the allowable process stress. Each control cycle takes only the first value of the optimal frequency sequence and executes it, with a solution time of no more than 500ms.

8. The system according to claim 1, characterized in that, The thermocouple array of the sensing layer has at least four temperature measurement points at the typical cross section with the largest temperature gradient on the workpiece surface, and at least three liquid temperature sensors are evenly arranged along the height direction in the liquid phase region of the tank. All thermocouple signals are standardized into 4 to 20mA analog signals by a temperature transmitter and then connected to a high-speed data acquisition card with a sampling frequency of not less than 10Hz. The control cycle is set to 1 second.

9. The system according to claim 1, characterized in that, The system adopts a three-tier hardware deployment architecture: The field equipment layer includes thermocouple arrays, ultrasonic viscometers, conductivity meters, speed encoders, quenching and cooling tanks, and frequency conversion drive devices; The edge computing layer includes a high-speed data acquisition card, an industrial edge GPU node, and a programmable logic controller. The industrial edge GPU node undertakes all computational tasks of physical information neural network, reduced-order model, and meta-learning inference. The programmable logic controller exchanges millisecond-level real-time data with the industrial edge GPU node through the PROFINET bus and issues control commands to the frequency conversion drive device. The cloud / host computer layer includes a production scheduling system, a model warehouse, and a SCADA monitoring host computer, which communicates with the edge computing layer through the OPC-UA standard protocol.

10. The system according to claim 1, characterized in that, In the online estimation and degradation compensation module for the comprehensive heat transfer coefficient of the medium, the degradation compensation neural network is replaced by an extended Kalman filter or an unscented Kalman filter. The comprehensive heat transfer coefficient is used as an augmented state variable for joint estimation with the workpiece temperature. The forward heat conduction model is used as the state transition equation. The measured temperature of the workpiece thermocouple and the bypass fluid characteristic sensor data are used as the observations. The estimated value of the comprehensive heat transfer coefficient is updated in real time in each control cycle through Kalman gain. In the real-time reconstruction module of the three-dimensional temperature field of the physical information neural network, the physical information neural network is replaced by a graph neural network. The finite element mesh of the workpiece is mapped into a graph structure. The measured temperature of the thermocouple is used as a partial known node feature. The semi-supervised prediction and reconstruction of the temperature of the entire field is completed through multi-layer graph convolution.

Images