Multi-sensor large casting pouring monitoring method and system

Abstract

The invention relates to the technical field of casting process monitoring, particularly discloses a multi-sensor large casting pouring monitoring method and system, and aims to solve the problems of data missing caused by a sensor blind area, distortion of a pure data interpolation model under a dynamic working condition and response delay caused by centralized processing. The system comprises a physical field sensing network, an edge side physical field reconstruction unit, a multi-level decision-making unit and an actuator control unit, a global physical field is reconstructed in real time by non-uniformly deploying sensors and fusing a simplified physical mechanism model, and rapid identification and regulation of defect risks are realized based on a hierarchical decision-making mechanism. According to the scheme, the data validity can be improved, the reconstruction precision is ensured, the system response delay is shortened, and the real-time monitoring requirement in the pouring process of large castings is met.

Description

Technical Field

[0001] This invention belongs to the field of casting process monitoring technology, specifically relating to a multi-sensor method and system for monitoring the pouring of large castings. Background Technology

[0002] In the field of large casting manufacturing, precise monitoring and control of the pouring process is a core element determining the internal quality and service performance of castings. This technology involves multiple disciplines such as casting technology, thermodynamics, fluid mechanics, and automation control, and its development level directly affects the manufacturing quality and reliability of major national equipment such as energy equipment and heavy machinery. Among them, large casting pouring monitoring technology, as a key component of casting process automation, aims to collect key physical parameters such as temperature field and flow field in real time during the pouring process through a multi-sensor system, and achieve closed-loop control of process parameters based on data analysis, thereby ensuring the density and uniformity of the internal structure of the casting.

[0003] Existing technologies generally use temperature sensors deployed in a uniform grid for data acquisition. However, when dealing with large castings with complex geometries such as deep cavities and corners with small curvature radii, the sensor deployment is affected by the physical obstruction of the mold, resulting in monitoring blind spots in some key areas such as the corners of the inner cavity and the root of the core, leading to a high rate of missed detection of porosity and shrinkage defects.

[0004] To address the issue of missing data in blind spots, existing solutions often employ purely data-driven models such as inverse distance weighting or Kriging interpolation. However, these methods completely ignore the physical laws of fluid dynamics and heat transfer during the casting process. Under dynamic conditions where the flow rate of molten metal changes drastically, the interpolation results are prone to significant errors, leading to severe distortion in the temperature field reconstruction. Meanwhile, existing monitoring systems rely on a central server for multi-source heterogeneous data fusion and decision-making. The end-to-end delay from data acquisition to control command output is relatively long, often exceeding the critical window period for casting defect formation, resulting in a persistently high failure rate of closed-loop control. Although some studies have attempted to introduce computational fluid dynamics models to improve prediction accuracy, their solution process is computationally complex and cannot meet the millisecond-level response requirements of real-time monitoring.

[0005] Therefore, how to achieve high-precision, low-delay dynamic monitoring of the entire process of large casting while ensuring the consistency of physical laws has become a technical challenge that urgently needs to be overcome in this field. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to provide a method and system for monitoring the pouring of large castings using multiple sensors. The aim is to overcome the technical contradictions in the prior art, such as the lack of data in key areas due to blind spots in sensor deployment, the reliance on pure data interpolation models that ignore physical laws and result in reconstruction distortion under dynamic working conditions, and the excessive system response delay caused by centralized data processing mode that misses the critical window period for defect control.

[0007] To achieve the above objectives, this invention constructs a complete system that integrates physical mechanism-guided perception enhancement, real-time physical field reconstruction at the edge, and multi-level collaborative decision-making. Through the collaborative work of multiple sensing nodes deployed at the casting site, edge computing units, and a central monitoring platform, it achieves comprehensive, high-precision, and low-latency monitoring of the casting process of large castings.

[0008] A large casting pouring monitoring system with multiple sensors includes a physical field sensing network, an edge-side physical field reconstruction unit, a multi-level decision-making unit, and an actuator control unit.

[0009] The physical field sensing network is responsible for collecting multi-source physical quantity data during the pouring process; the edge-side physical field reconstruction unit reconstructs the global physical field of the pouring area in real time based on the sensor network data and the embedded simplified physical mechanism model; the multi-level decision-making unit identifies defects and risks and generates control decisions based on the reconstructed physical field data; and the actuator control unit is responsible for converting decision commands into specific actuator actions.

[0010] The physical field sensing network consists of multiple sensor groups deployed on the outer wall of the mold, the gating system, and preset internal detection points. These sensor groups include at least thermocouple temperature sensor groups, vibration acceleration sensor groups, and magnetic field strength sensor groups.

[0011] The thermocouple temperature sensor array is arranged in a non-uniform topology. Its deployment density reaches 5 to 8 measuring points per square meter in key areas such as the bends, corners, and root of the core in the metal liquid flow path, while the density is 1 to 2 measuring points per square meter in flat areas.

[0012] The locations of all measurement points were determined through computational fluid dynamics simulation of the casting process. Areas with temperature gradients exceeding 10 K / (s·mm) were extracted from the simulation as high-density deployment areas to ensure priority coverage of key areas with drastic changes in physical field gradients from the source.

[0013] The vibration acceleration sensor array is mainly attached to 15 to 20 key monitoring points on the outer wall of the mold to capture structural vibration signals caused by the impact of molten metal flow. Its sampling frequency is set to 5000Hz.

[0014] The magnetic field strength sensor group uses excitation coils and induction coils pre-embedded at specific locations in the mold to indirectly sense changes in flow velocity by detecting local magnetic field distortion caused by the flow of molten metal. The coil pairs are spaced 20ms apart and are installed at the intersection of the main flow path and branches of the casting system.

[0015] The edge-side physics reconstruction unit operates in an industrial-grade edge computing device deployed in the vicinity of the pouring site. This edge-side physics reconstruction unit receives real-time data streams from the physics sensing network. The real-time data streams carry multi-source physical quantity data and execute dynamic physics reconstruction algorithms guided by physical mechanisms.

[0016] The dynamic physics field reconstruction algorithm first constructs a simplified physics field calculation kernel that integrates the basic equations of fluid dynamics and the law of heat conduction. This simplified physics field calculation kernel discretizes the casting area into two overlapping grid systems. The first set is a non-uniform sparse grid that strictly corresponds to the sensor deployment position, and the second set is a high-resolution uniform background grid that covers the entire computational domain. The grid size is 2mm×2mm×2mm.

[0017] The reconstruction process of the edge-side physics reconstruction unit is as follows:

[0018] First, on the non-uniform sparse grid node, based on the measured data of the thermocouple temperature sensor group and the magnetic field strength sensor group, the Gaussian kernel function considering spatial variability is used for initial interpolation to obtain the initial temperature estimate and the initial flow rate estimate on the non-uniform sparse grid node.

[0019] The initial estimates of temperature and flow velocity are used as the input boundary conditions and initial conditions for the simplified physics calculation kernel.

[0020] The simplified physics calculation kernel introduces a preset fluid viscosity coefficient of 0.001 Pa·s and a thermal diffusivity of 5 × 10⁻⁻⁻⁶. 6 The calculation steps are set to 2ms, and the physical field evolution is performed using m² / s and no-slip boundary conditions.

[0021] The simplified physics calculation kernel outputs predicted physical quantities for each node on a high-resolution uniform background grid at each time step, including temperature and velocity distributions.

[0022] Meanwhile, the edge-side physics reconstruction unit also includes a residual feedback module, which continuously compares the difference between the predicted values ​​calculated by the physics evolution on the sparse grid nodes and the real-time measurement values ​​of the sensors.

[0023] When the absolute value of the temperature residual exceeds 3K or the flow rate residual exceeds 0.05m / s, the residual feedback module generates a correction field based on the radial basis function. This correction field is applied to the next time step of the physical field evolution calculation through a weighted superposition method. The weight coefficient is proportional to the magnitude of the residual, which is used to dynamically correct the model error and ensure that the reconstructed physical field is consistent with the measured data at the sensor location.

[0024] The entire reconstruction process utilizes a graphics processor for parallel acceleration calculations, with the calculation time for a single full-field reconstruction controlled within 8ms.

[0025] The multi-level decision-making unit adopts a hierarchical architecture, including a rapid response layer and a deep analysis layer.

[0026] The fast response layer is integrated into the edge-side physical field reconstruction unit. Based on the reconstructed high-resolution physical field data, it performs real-time defect risk screening based on a rule base.

[0027] The rule base predefines 12 physical field characteristic patterns of defect predictors. Specifically, when the temperature gradient of a local area changes by more than 15 K / (s·mm) within 2 seconds and the predicted flow velocity of the area is less than 0.1 m / s, it is determined that there is a risk of cold shut-off in the area. Once the fast response layer identifies high-risk defect predictors, it immediately generates a first-level control command containing the location coordinates and risk level.

[0028] The deep analysis layer is deployed on a high-performance server cluster of the central monitoring platform. It receives physical field data aggregated from multiple edge-side physical field reconstruction units and early warning information from the rapid response layer.

[0029] The deep analysis layer runs a pre-trained spatiotemporal convolutional neural network model. This model takes the global temperature field, flow velocity field, and casting process parameters of the time series as input. The model structure includes 5 convolutional layers, 3 long short-term memory network layers, and 2 fully connected layers. The output is an assessment of the overall quality grade of the casting and the prediction probability of 7 types of macroscopic defects. When the prediction probability of any defect exceeds 0.85, the deep analysis layer generates a secondary control strategy or process optimization suggestion.

[0030] The actuator control unit receives instructions from multi-level decision-making units.

[0031] For the first-level control command issued by the fast response layer, the actuator control unit directly drives the corresponding fast execution mechanism, including the ladle tilting servo motor or the local heater power controller, to perform millisecond- to second-level intervention. The ladle tilting servo motor has a control accuracy of 0.1° and a response delay of less than 50ms.

[0032] For the secondary control strategy issued by the deep analysis layer, the actuator control unit coordinates multiple actuators to make comprehensive adjustments to process parameters, with an execution cycle of 2s to 5s.

[0033] Furthermore, the data processing of the magnetic field strength sensor group in the physical field sensing network includes an eddy current compensation algorithm. This eddy current compensation algorithm is based on the known relationship between the excitation current frequency of 10kHz and the conductivity of the liquid metal. It calculates and subtracts the background magnetic field interference caused by the eddy current effect of the liquid metal in real time, thereby extracting the magnetic field distortion signal caused purely by the change in flow velocity. The signal sampling rate is 2000Hz.

[0034] Furthermore, the rule base of the rapid response layer supports online updates. The deep analysis layer of the central monitoring platform analyzes historical casting data and final casting quality inspection results monthly, automatically extracts new defect predictor feature patterns, and after confirmation by engineers, distributes them to each edge-side physical field reconstruction unit through an encrypted network channel to update their local rule base.

[0035] Furthermore, the system also includes a data synchronization and clock calibration module. This module ensures that the data acquisition timestamps of all sensors in the physical field sensing network, the calculation cycle of the edge-side physical field reconstruction unit, and the command execution time of the actuator control unit are all synchronized with the high-precision BeiDou time source at the microsecond level, with the maximum clock deviation controlled within ±2μs, thereby ensuring strict timing consistency between the data flow and control flow of the entire system.

[0036] As another embodiment of the present invention, for ultra-large castings with a casting area exceeding 50m², the system adopts a distributed edge computing architecture. This distributed edge computing architecture consists of a computing cluster composed of 3 edge computing devices. Each edge computing device is responsible for monitoring an independent casting zone. There is a 2m wide overlapping monitoring zone between the casting zones. Within the overlapping monitoring zone, the sensor deployment density is doubled to 10 measurement points per m².

[0037] Each edge computing device first independently completes the physical field reconstruction of its own partition, then exchanges boundary physical field data through a high-speed backplane, and performs boundary matching correction using the Jacobi iterative method based on the sensor measurement data of the overlapping monitoring zone. After 3 to 5 iterations, the boundary matching error can be reduced.

[0038] In another embodiment of the present invention, the magnetic field strength sensor group can use a distributed magnetoresistive sensor array to replace the excitation coil and induction coil pair. The magnetoresistive sensor array contains 32 measurement nodes, which are evenly arranged on the surface of the key area of ​​the casting mold at 5ms intervals. Each node integrates a triaxial magnetoresistive sensing element with a range of ±2Gs and a resolution of 0.1mGs.

[0039] In this configuration, the computational kernel of the edge-side physics reconstruction unit introduces a magnetohydrodynamic coupling model, adds a Lorentz force term to the simplified hydrodynamic equation, and solves for the velocity field, temperature field and magnetic field distribution. The accuracy of velocity field reconstruction is improved through multi-physics coupling iteration.

[0040] Furthermore, this invention also provides a multi-sensor method for monitoring the pouring of large castings, comprising the following steps:

[0041] S1: Multi-source physical quantity data during the casting process are collected through a physical field sensing network. The physical field sensing network consists of multiple sensor groups deployed on the outer wall of the mold, the casting system, and preset internal detection points. The sensor groups include thermocouple temperature sensor groups, vibration acceleration sensor groups, and magnetic field strength sensor groups. The thermocouple temperature sensor groups are arranged according to a non-uniform topology structure. Their deployment density reaches 5 to 8 measurement points per square meter in key areas such as the bend area of ​​the molten metal flow path, the corner area of ​​the inner cavity, and the root area of ​​the core head, while the density of measurement points is 1 to 2 measurement points per square meter in flat areas. The deployment positions of all measurement points are determined by the computational fluid dynamics simulation of the casting process in advance.

[0042] S2: In the edge-side physical field reconstruction unit, real-time data streams from the physical field sensing network are received, and a dynamic physical field reconstruction algorithm based on physical mechanisms is executed to reconstruct the global physical field in real time. The dynamic physical field reconstruction algorithm constructs a simplified physical field calculation kernel that integrates the basic equations of fluid dynamics and the law of heat conduction. The simplified physical field calculation kernel discretizes the casting area into a non-uniform sparse grid that strictly corresponds to the sensor deployment position and a high-resolution uniform background grid that covers the entire computational domain. The reconstruction process includes initial interpolation on the non-uniform sparse grid nodes, setting boundary conditions and initial conditions, performing fast physical field evolution calculations, and dynamically correcting model errors through the residual feedback module.

[0043] S3: Defect risk identification and control decision generation are carried out through multi-level decision units. The multi-level decision units include a fast response layer integrated into the physical field reconstruction unit on the edge side and a deep analysis layer deployed on the central monitoring platform. The fast response layer performs real-time defect risk screening based on the rule base, and the deep analysis layer uses a spacetime convolutional neural network model to perform overall quality assessment and defect prediction.

[0044] S4: Receives instructions from multi-level decision-making units through the actuator control unit and drives the corresponding actuators to adjust process parameters. It intervenes in the first-level control instructions at the millisecond to second level and makes comprehensive adjustments to process parameters for the second-level control strategies.

[0045] In summary, this application includes at least one of the following beneficial technical effects:

[0046] 1. By adopting a non-uniform topology deployment strategy guided by computational fluid dynamics simulation, limited sensor resources are accurately delivered to the key areas where the physical field gradient changes most drastically, reducing the monitoring blind zone area and significantly improving the integrity of defect detection from the source of data acquisition, further reducing the missed detection rate of porosity and shrinkage defects.

[0047] 2. An innovative reconstruction architecture that integrates physical mechanism models and real-time sensor data is constructed at the edge. By simplifying the physical field calculation kernel and the closed-loop mechanism of residual feedback correction, high-precision and physical law-consistent deduction and reconstruction of the physical field in the blind zone is achieved. Under the dynamic working condition of sudden changes in the flow rate of molten metal, the reconstruction error is reduced, and the distortion problem of pure data interpolation model is effectively overcome.

[0048] 3. By constructing a multi-level decision-making unit that includes a fast response layer and a deep analysis layer, and by pushing the fast response layer down to the edge, millisecond-level identification and real-time intervention of critical defect risks are achieved, shortening the end-to-end response time of the system, ensuring that control actions can be completed quickly within the critical window period of defect formation, and reducing the failure rate of closed-loop control.

[0049] 4. Through precise timing synchronization and distributed edge computing architecture, the entire system achieves low-latency processing across the entire chain, from data perception and physical field reconstruction to decision execution. While meeting the stringent requirements for real-time monitoring of large castings, it provides a solid technical guarantee for improving the internal quality and manufacturing reliability of castings. Attached Figure Description

[0050] Figure 1 This is a schematic diagram of the overall technical architecture of the multi-sensor large casting pouring monitoring method and system proposed in this invention;

[0051] Figure 2 This is a schematic diagram of the core principle framework of the edge-side physical field reconstruction unit in this invention;

[0052] Figure 3 This is a logical flow diagram of the multi-level decision-making unit in this invention. Detailed Implementation

[0053] To further illustrate the technical means and effects adopted by the present invention to achieve the intended purpose, the following description is provided in conjunction with the appendix. Figures 1 to 3 The preferred embodiments are described in detail below based on specific implementations of the present invention.

[0054] Example 1

[0055] This embodiment details the specific implementation structure and workflow of a multi-sensor large casting pouring monitoring system.

[0056] Please refer to the attached document. Figure 1The overall architecture of this system includes a physical field sensing network deployed at the casting site, edge computing devices set up near the casting area, a central monitoring platform located in the monitoring center, and a control terminal connected to the casting execution mechanism. Through the coordinated operation of each unit, the system realizes dynamic reconstruction of the physical field and real-time control of defect risks throughout the casting process of large castings.

[0057] The physical field sensing network consists of a thermocouple temperature sensor group, a vibration acceleration sensor group, and a magnetic field strength sensor group.

[0058] The thermocouple temperature sensor array adopts a non-uniform topology deployment strategy. Its deployment density reaches 5 to 8 measurement points per square meter in key areas such as the bend area, inner cavity corner area and core root area of ​​the metal liquid flow path, while only 1 to 2 measurement points per square meter in flat areas.

[0059] All measurement points were determined through computational fluid dynamics simulation of the pre-executed casting process. Areas with temperature gradients exceeding 10 K / (s·mm) were extracted from the simulation as high-density deployment zones.

[0060] The thermocouple sensor is fixed to the outer wall of the mold by a high-temperature resistant ceramic sleeve. Its signal sampling frequency is set to 100Hz, the measurement accuracy is ±0.5K, and the data is transmitted to the edge computing device through shielded twisted pair cable.

[0061] The vibration acceleration sensor array uses piezoelectric accelerometers, which are magnetically attached to 15 to 20 key vibration monitoring points on the outer wall of the mold. The sensor has a range of ±50g, a frequency response range of 0.1Hz to 2000Hz, and a sampling frequency of 5000Hz. The vibration signal is converted into a voltage signal by a built-in charge amplifier and then transmitted to the edge-side data acquisition card via a coaxial cable.

[0062] The magnetic field strength sensor group consists of an excitation coil and an induction coil pair embedded in a specific position in the mold. The excitation coil is energized with a sinusoidal alternating current with a frequency of 10kHz, and the induction coil detects the magnetic field distortion signal caused by the flow of molten metal. The distance between the excitation coil and the induction coil pair is 20ms. The installation position is located at the intersection of the main flow path and the branch of the casting system. The output signal of the induction coil is amplified by a preamplifier and then acquired by a 24-bit analog-to-digital converter with a sampling rate of 2000Hz.

[0063] The magnetic field strength sensor group has a built-in eddy current compensation algorithm, which calculates the background magnetic field interference value generated by the eddy current effect in real time based on the relationship between the conductivity of the liquid metal and the excitation frequency, and subtracts the interference component from the original signal to finally output a pure flow velocity-related magnetic field distortion signal.

[0064] After detecting flow velocity-related magnetic field distortion signals, the magnetic field strength sensor array, through its built-in signal processing module, establishes a mapping relationship between magnetic field distortion and flow velocity based on magnetohydrodynamic theory (e.g., the relationship between Lorentz force and flow velocity) and pre-set parameters of the molten metal's conductivity and geometry. This mapping relationship can be a pre-trained model or analytical formula used to convert the compensated magnetic field distortion signal into a local flow velocity value of the molten metal in real time. For example, a model based on one of the following formulas can be used for the conversion: Where B is the magnetic field distortion and v is the flow velocity. Let be the electrical conductivity and g be the geometric factor.

[0065] The edge-side physics reconstruction unit is deployed on an industrial-grade edge server. Its hardware configuration includes an 8-core processor, 16GB of memory, and a high-performance graphics processor. The edge-side physics reconstruction unit receives real-time data streams from the physics sensor network and executes a dynamic physics reconstruction algorithm guided by physical mechanisms.

[0066] Please refer to the attached document. Figure 2 The reconstruction process first involves building two mesh systems:

[0067] The first set is a non-uniform sparse grid, whose nodes correspond one-to-one with the actual spatial coordinates of all sensors in the physical field sensing network.

[0068] The second set is a high-resolution uniform background mesh covering the entire casting area, with a mesh size of 2mm×2mm×2mm.

[0069] The reconstruction algorithm kernel integrates simplified basic equations of fluid dynamics and the law of heat conduction. The simplified form of the basic equations of fluid dynamics is as follows:

[0070]

[0071] Where the velocity vector is the velocity vector of the molten metal. Time is a time variable. The density of the molten metal is equal to the density constant of the molten metal. Pressure is a pressure field The kinematic viscosity coefficient is a fluid viscosity parameter. The acceleration vector of gravity is the acceleration due to gravity. .

[0072] The discrete form of the law of heat conduction is:

[0073] Where temperature is the temperature field. The thermal diffusivity is a parameter for the thermal diffusivity of a material. .

[0074] The algorithm execution flow is as follows:

[0075] First, on a non-uniform sparse grid, spatial interpolation is performed using Gaussian kernel functions based on measured data from thermocouple temperature sensor arrays and magnetic field strength sensor arrays. The bandwidth of the Gaussian kernel function is adaptively adjusted according to the sensor deployment density. The spatial interpolation results serve as the initial conditions for the physical field evolution calculation. The simplified physical field calculation kernel adopts an explicit time integration scheme based on the finite volume method (or a specific numerical method such as the finite difference method), with a time step set to 2 ms. Within each time step, the simplified physical field calculation kernel calculates the flow velocity and temperature distribution for the next moment based on the current physical field state. Preset fluid viscosity coefficients of 0.001 Pa·s and thermal diffusivity of 5 × 10⁻⁻⁻⁶ are introduced during the calculation. 6 With m² / s and no-slip boundary conditions, the graphics processor performs parallel computation of the predicted physical quantities of all nodes on a high-resolution uniform background mesh, and the time taken for a single full-field computation is controlled within 8ms.

[0076] The residual feedback module continuously monitors the difference between the predicted values ​​and the actual values ​​measured by the sensor on the sparse grid nodes.

[0077] When the absolute value of the temperature residual exceeds 3K or the velocity residual exceeds 0.05m / s, the residual feedback module generates a correction field based on the radial basis function. This correction field is applied to the physical field evolution calculation of the next time step through a weighted superposition method, and the weight coefficient is proportional to the magnitude of the residual.

[0078] This closed-loop correction mechanism ensures that the reconstructed physical field is consistent with the measured data at the sensor location, while also ensuring that the blind zone extrapolation conforms to physical laws.

[0079] The multi-level decision-making unit adopts a two-layer structure of a rapid response layer and a deep analysis layer.

[0080] The fast response layer is integrated within the edge-side physical field reconstruction unit. Based on the reconstructed high-resolution physical field data, it performs real-time defect risk screening. The rule base predefines 12 physical field characteristic patterns for defect predictors, including but not limited to cold shut risk, shrinkage cavity tendency, and slag inclusion probability. Specific judgment rules are as follows:

[0081] When a local area is detected to have a temperature gradient change rate exceeding 15 K / (s·mm) within 2 seconds, and the flow velocity in that area is below 0.1 m / s, the rule engine immediately triggers a cold shut-off risk flag.

[0082] When a region is detected to have a continuous and rapid temperature drop within 1 second, with a drop rate exceeding 20K / s, and at the same time a high-frequency vibration signal (frequency greater than 1000Hz, amplitude exceeding 2g) is captured by a vibration acceleration sensor, it is determined that there is a risk of shrinkage or porosity.

[0083] When the reconstructed velocity field shows that a certain region has formed a significant vortex, and the velocity in the core region of the vortex is consistently lower than 30% of the mainstream velocity for more than 0.5s, and the temperature distribution in that region shows abnormal fluctuations (fluctuation amplitude greater than ±5K), then it is determined that there is a risk of slag inclusion.

[0084] Once any of the above rules is triggered, a Level 1 control instruction containing location coordinates and risk level is generated. The rule base supports dynamic updates. The central monitoring platform analyzes historical pouring data monthly, extracts new feature patterns, and after confirmation by engineers, distributes them to the edge side through an encrypted network channel to update the local rule base.

[0085] The deep analysis layer is deployed on a high-performance server cluster of the central monitoring platform. It receives the aggregated results of physical field data from multiple edge nodes. The deep analysis layer runs a pre-trained spatiotemporal convolutional neural network model. The input data includes the global temperature field, flow field and casting process parameters of the time series. The neural network structure contains 5 convolutional layers, 3 long short-term memory network layers and 2 fully connected layers. The output is the casting quality grade score and the predicted probability of 7 types of macroscopic defects.

[0086] When the predicted probability exceeds 0.85, the system generates a secondary control strategy, including adjusting the pouring speed, modifying the cooling curve, or coordinating the linkage of multiple actuators.

[0087] The actuator control unit is connected to the edge side and the central monitoring platform via industrial Ethernet.

[0088] For the first-level control command issued by the fast response layer, the unit directly drives the ladle tilting servo motor with a motor control accuracy of 0.1° and a response delay of less than 50ms; at the same time, the power output of the local heater can be adjusted, with a power adjustment range of 0 to 5kW and an adjustment step of 0.1kW.

[0089] For the secondary control strategy issued by the deep analysis layer, the actuator control unit coordinates multiple actuators such as the trajectory correction of the casting robot, the speed adjustment of the cooling fan, and the opening control of the pressure reducing valve, with an execution cycle of 2s to 5s.

[0090] All actuator status feedback signals are acquired in real time through photoelectric encoders and current sensors to form closed-loop control.

[0091] The data synchronization and clock calibration module uses the time reference provided by the BeiDou satellite navigation system to achieve microsecond-level synchronization of the entire system through a precision clock protocol. The data acquisition trigger signal of the physical field sensor network, the start time of the calculation cycle of the edge-side physical field reconstruction unit, and the instruction execution timestamp of the actuator control unit are all synchronized with this reference. The transmission of the physical field sensor network adopts time-sensitive networking technology to ensure that the transmission delay of the sensor data to the edge node is stable within 100μs, and the maximum clock deviation of the system is controlled within ±2μs, ensuring strict timing consistency between the data flow and the control flow.

[0092] This embodiment achieves optimized deployment of physical field sensing networks, high-precision reconstruction of edge-side physical fields, hierarchical decision-making on defect risks, and precise control of execution mechanisms through the above-described specific implementation methods.

[0093] Example 2

[0094] Building upon Embodiment 1 above, this embodiment further provides a monitoring system implementation plan for ultra-large castings with a pouring area exceeding 50m². The core of this system lies in its use of a distributed edge computing architecture to handle larger-scale computational loads. This architecture consists of a distributed computing cluster composed of three edge computing devices, each responsible for monitoring an independent pouring zone. There is a 2m wide overlapping monitoring zone between the pouring zones. Within the overlapping monitoring zone, the sensor deployment density is doubled to 10 measurement points per m², ensuring the continuity of physical field reconstruction at the zone boundaries.

[0095] In ultra-large casting applications, the physical field sensing network has been expanded to include 480 thermocouple measuring points, 90 vibration monitoring points, and 45 magnetic field sensing units.

[0096] Thermocouple measuring points are deployed using a four-level non-uniform strategy, with a density of 12 points per square meter in the core casting area, 6 points per square meter in the transition area, 3 points per square meter in the edge area, and 10 points per square meter in the overlapping monitoring zone.

[0097] All sensors communicate with the edge computing cluster via an industrial wireless network. The wireless transmission protocol uses time diversity multiple access technology, with each communication slot lasting 50ms, to ensure the real-time performance and reliability of data transmission.

[0098] The edge-side physics reconstruction unit introduces a regional collaborative algorithm in the distributed architecture. Each edge computing device first independently completes the physics reconstruction of its own partition, and then exchanges the reconstructed boundary physics data through a high-speed backplane. The regional collaborative algorithm calculates the physics matching error at the boundary of each casting partition based on the sensor measurement data of the overlapping monitoring zone, and makes the reconstruction results of adjacent casting partitions smoothly transition at the boundary through iterative correction. The correction process adopts the Jacobi iteration method, and each iteration takes no more than 5ms. After 3 to 5 iterations, the boundary matching error can be reduced.

[0099] In the application of ultra-large castings, a multi-level decision-making unit adds a regional coordination layer. This regional coordination layer runs on the master node of the edge computing cluster, integrates the defect risk map of each casting zone in real time, identifies macroscopic defect patterns distributed across zones, and when multiple casting zones are detected to have cold shut risk characteristics at the same time, the regional coordination layer generates a global control strategy to coordinate the actuators of all zones to perform synchronous flow adjustment.

[0100] The actuator control unit is equipped with redundant control channels, so that when the main control channel fails, the backup channel can quickly take over control.

[0101] The data synchronization and clock calibration module adopts a master-slave clock synchronization mechanism in the distributed architecture. The master edge computing device obtains a precise timestamp through the Beidou receiver, and the slave edge computing device keeps the clock synchronized with the master device through a precision time protocol. The maximum clock deviation of the entire system is controlled within ±2μs, ensuring the time consistency of distributed acquisition and computing.

[0102] This embodiment effectively solves the problems of computational load and data consistency in the monitoring of ultra-large castings through a distributed edge computing architecture and regional collaborative algorithm. The system can support real-time monitoring of a casting area of ​​up to 100 square meters, and the overall reconstruction latency is controlled within 15ms, meeting the stringent monitoring requirements of ultra-large castings.

[0103] Example 3

[0104] This embodiment is an alternative to Embodiment 1, and its main difference lies in the implementation method of magnetic field sensing in the physical field sensing network and the core model of the reconstruction algorithm. Specifically, this embodiment describes a physical field reconstruction scheme based on enhanced magnetic field sensing. In the physical field sensing network, the magnetic field strength sensor group uses a distributed magnetoresistive sensor array to replace the original excitation coil and induction coil pair.

[0105] The magnetoresistive sensor array contains 32 measurement nodes, which are evenly arranged on the surface of the key area of ​​the casting at 5ms intervals. Each sensor node integrates a triaxial magnetoresistive sensing element with a range of ±2Gs and a resolution of 0.1mGs. The sensor array is connected to the edge computing device via a fieldbus, and the sampling frequency is increased to 5000Hz.

[0106] The edge-side physics field reconstruction unit improves the physics field calculation kernel to suit the characteristics of magnetoresistive sensor data. A magnetohydrodynamic coupling model is introduced into the kernel, directly linking magnetic field changes to the flow velocity field calculation. Specifically, a Lorentz force term is added to the simplified fluid dynamics equations, and its calculation formula is as follows:

[0107]

[0108] The Lorentz force is the force exerted by a magnetic field on a moving charged body. The conductivity of liquid metal is a parameter of material conductivity. The electric field strength is the induced electric field. Magnetic induction intensity is the measured magnetic field The fluid velocity is the velocity of the molten metal. .

[0109] During the reconstruction process, the three-dimensional magnetic field data measured by the magnetoresistive sensor is denoised by Kalman filtering and used as the input boundary condition for the magnetohydrodynamic equation. The physical field calculation kernel simultaneously solves for the velocity field, temperature field and magnetic field distribution, and improves the reconstruction accuracy through multi-physics coupling iteration.

[0110] The residual feedback module is extended to magnetic field residual monitoring. When the difference between the predicted and measured magnetic induction intensity exceeds 0.5 mG, the magnetic field correction field generation mechanism is triggered.

[0111] The rapid response layer of the multi-level decision-making unit has added a magnetic field anomaly detection rule. When the local magnetic field distortion rate changes by more than 20% within 0.1s and is accompanied by a sudden drop in temperature, it is determined to be a risk of interruption of molten metal flow and an emergency increase command for pouring flow is immediately triggered.

[0112] The actuator control unit activates the high-speed servo valve in response to this command, with a valve opening adjustment speed of 100% / s, and can complete flow correction within 200ms.

[0113] This embodiment details the implementation process of a multi-sensor method for monitoring the pouring of large castings. This method is applied to the pouring process of large castings and includes the following steps:

[0114] S1: Collect multi-source physical quantity data during the casting process through a physical field sensing network. The physical field sensing network includes a thermocouple temperature sensor group, a vibration acceleration sensor group, and a magnetic field strength sensor group.

[0115] The thermocouple temperature sensor array is deployed according to a non-uniform topology. In key areas such as the bends in the metal liquid flow path, the corners of the inner cavity, and the root of the core, the deployment density is 5 to 8 measuring points per square meter, while in flat areas, the density is 1 to 2 measuring points per square meter.

[0116] All measuring point locations were determined through pre-performed computational fluid dynamics simulations of the casting process. Regions with temperature gradients exceeding 10 K / (s·mm) were extracted from the simulations as high-density deployment areas. The vibration acceleration sensor group had a sampling frequency of 5000 Hz, and the magnetic field strength sensor group had a sampling rate of 2000 Hz. Eddy current compensation algorithms were used to process the data.

[0117] S2: An edge-side physics reconstruction unit is run in an industrial-grade edge computing device deployed in the vicinity of the pouring site. This unit receives real-time data streams and executes a dynamic physics reconstruction algorithm. The reconstruction algorithm first constructs a simplified physics calculation kernel, integrates the basic equations of fluid dynamics and the law of heat conduction, and discretizes the pouring area into a non-uniform sparse grid and a high-resolution uniform background grid (grid size 2mm×2mm×2mm).

[0118] On non-uniform sparse grid nodes, Gaussian kernel function is used for initial interpolation based on sensor measurement data to obtain initial estimates of temperature and flow velocity. These values ​​are used as boundary and initial conditions, simplifying the physics calculation kernel by introducing a fluid viscosity coefficient of 0.001 Pa·s and a thermal diffusivity of 5 × 10⁻⁻⁻⁶. 6 Fast physical field evolution calculations were performed using m² / s and no-slip boundary conditions, with a calculation step size of 2ms.

[0119] Meanwhile, the residual feedback module compares the predicted values ​​with the measured values. When the temperature residual exceeds 3K or the flow rate residual exceeds 0.05m / s, it generates a correction field based on the radial basis function for dynamic correction.

[0120] The entire reconstruction process utilizes parallel acceleration via the graphics processor, with the time required for a single full-field reconstruction controlled within 8ms.

[0121] S3: Defect risk identification and control decision generation are performed through multi-level decision units. The rapid response layer is integrated into the physical field reconstruction unit on the edge side. Based on the reconstructed high-resolution physical field data, the rule base is used for real-time screening.

[0122] The rule base predefines the physical field characteristic patterns of 12 defect predictors. For example, when the temperature gradient in a local area changes by more than 15 K / (s·mm) within 2 seconds and the flow rate is less than 0.1 m / s, it is judged as a cold shut-off risk and a first-level control command is generated.

[0123] The deep analysis layer is deployed on the central monitoring platform. It runs a spacetime convolutional neural network model, takes the global temperature field, flow velocity field and casting process parameters of the time series as input, and outputs the predicted probability of casting quality grade and 7 types of macro defects. When the probability exceeds 0.85, a secondary control strategy is generated.

[0124] The rule base supports online updates. The central monitoring platform analyzes historical data monthly, automatically extracts new feature patterns, and issues updates after confirmation by engineers.

[0125] S4: Receives decision instructions through the actuator control unit and drives the actuator to make adjustments.

[0126] For first-level control commands, the ladle tilting servo motor (control accuracy 0.1°, response delay less than 50ms) or the local heater power controller is directly driven to intervene at the millisecond to second level.

[0127] For the secondary control strategy, multiple actuators such as the casting robot, cooling fan and pressure reducing valve are coordinated for comprehensive adjustment, with an execution cycle of 2s to 5s.

[0128] The data synchronization and clock calibration module ensures that the entire system is synchronized with the BeiDou time source at the microsecond level, with a maximum clock deviation of ±2μs.

[0129] This embodiment achieves full-domain monitoring, high-precision reconstruction, rapid decision-making, and precise control of the large casting pouring process through the above-described methods and steps, effectively improving casting quality and production efficiency.

[0130] This embodiment further enhances the ability to capture turbulence and flow separation phenomena through enhanced magnetic field sensing and multiphysics field coupling reconstruction, making it particularly suitable for monitoring the pouring of large castings with complex internal cavity structures. While maintaining a reconstruction cycle of less than 10ms, the system improves the accuracy of flow velocity field reconstruction to within 3% relative error, achieves magnetic field anomaly detection sensitivity at the milligauss level, and has an accuracy rate of over 95% in identifying convection interruption risks.

[0131] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention. Therefore, the embodiments should be regarded as exemplary and non-limiting in all respects.

[0132] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment includes only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A multi-sensor monitoring system for large casting pouring, characterized in that, include: The physical field sensing network is used to collect multi-source physical quantity data during the casting process. It consists of multiple sensor groups deployed on the outer wall of the mold, the casting system, and preset internal detection points. The sensor groups include thermocouple temperature sensor groups, vibration acceleration sensor groups, and magnetic field strength sensor groups. The thermocouple temperature sensor groups are arranged according to a non-uniform topology. The deployment positions of all measuring points are determined by pre-performed computational fluid dynamics simulation of the casting process, and high-density deployment areas are extracted. The edge-side physics reconstruction unit runs in an industrial-grade edge computing device deployed in the vicinity of the pouring site. It is used to receive real-time data streams from the physics sensing network and execute a dynamic physics reconstruction algorithm guided by physical mechanisms. The dynamic physics reconstruction algorithm constructs a simplified physics calculation kernel that integrates the basic equations of fluid dynamics and the law of heat conduction. The simplified physics calculation kernel discretizes the pouring area into a non-uniform sparse grid that strictly corresponds to the sensor deployment location and a high-resolution uniform background grid that covers the entire computational domain. The multi-level decision-making unit adopts a hierarchical architecture, including a fast response layer integrated inside the edge-side physical field reconstruction unit and a deep analysis layer deployed on the central monitoring platform. The fast response layer is integrated inside the edge-side physical field reconstruction unit. The actuator control unit receives instructions from multi-level decision-making units and drives the corresponding actuators to adjust process parameters.

2. The large casting pouring monitoring system with multiple sensors according to claim 1, characterized in that, The reconstruction process of the edge-side physics reconstruction unit includes: On a non-uniform sparse grid node, based on the measured data of the thermocouple temperature sensor group and the magnetic field strength sensor group, the Gaussian kernel function considering spatial variability is used for initial interpolation to obtain the initial temperature estimate and the initial flow rate estimate on the non-uniform sparse grid node. The initial estimates of temperature and flow velocity are used as the input boundary conditions and initial conditions for the simplified physics calculation kernel. The simplified physics calculation kernel introduces a preset fluid viscosity coefficient of 0.001 Pa·s and a thermal diffusivity of 5 × 10⁻⁻⁻⁶. 6 The calculation of the physical field evolution is performed using m² / s and no-slip boundary conditions, with a calculation step size of 2ms. The simplified physics calculation kernel outputs predicted physical quantities for each node on a high-resolution uniform background grid at each time step, including temperature and velocity distributions.

3. The large casting pouring monitoring system with multiple sensors according to claim 2, characterized in that, The edge-side physics reconstruction unit also includes a residual feedback module; The residual feedback module continuously compares the difference between the predicted physical quantities calculated from the physical field evolution on the sparse grid nodes and the real-time measurement values ​​of the sensors. When the absolute value of the temperature residual exceeds 3K or the flow rate residual exceeds 0.05m / s, the residual feedback module generates a correction field based on the radial basis function. The correction field is applied to the next time step of the physical field evolution calculation through a weighted superposition method. The weight coefficient is proportional to the magnitude of the residual and is used to dynamically correct the model error.

4. The large casting pouring monitoring system with multiple sensors according to claim 1, characterized in that, The rule base of the fast response layer predefines 12 physical field characteristic patterns of defect predictors; If the temperature gradient in a local area changes by more than 15 K / (s·mm) within 2 seconds, and the predicted flow velocity in that local area is less than 0.1 m / s, then the local area is deemed to have a risk of cold insulation. Once the rapid response layer identifies high-risk defect predictors, it immediately generates a first-level control instruction containing location coordinates and risk level.

5. The large casting pouring monitoring system with multiple sensors according to claim 1, characterized in that, The deep analysis layer runs a pre-trained spatiotemporal convolutional neural network model; The spatiotemporal convolutional neural network model takes the global temperature field, flow velocity field and casting process parameters of the time series as input. The spatiotemporal convolutional neural network model structure includes 5 convolutional layers, 3 long short-term memory network layers and 2 fully connected layers. The output is an assessment of the overall quality grade of the casting and a predicted probability of seven types of macroscopic defects. When the predicted probability of any defect exceeds 0.85, the deep analysis layer generates secondary control strategies or process optimization suggestions.

6. The large casting pouring monitoring system with multiple sensors according to claim 1, characterized in that, The actuator control unit directly drives the corresponding fast actuator, including the ladle tilting servo motor or the local heater power controller, to intervene at the millisecond to second level in response to the first-level control command issued by the fast response layer. The ladle tilting servo motor has a control accuracy of 0.1° and a response delay of less than 50ms. For the secondary control strategy issued by the deep analysis layer, the actuator control unit coordinates multiple actuators to make comprehensive adjustments to process parameters, with an execution cycle of 2s to 5s.

7. The large casting pouring monitoring system with multiple sensors according to claim 1, characterized in that, Data processing in the magnetic field strength sensor group of the physical field sensing network includes eddy current compensation algorithms; The eddy current compensation algorithm is based on the known relationship between the excitation current frequency of 10kHz and the conductivity of the liquid metal, and calculates and subtracts the background magnetic field interference caused by the eddy current effect of the liquid metal in real time. The magnetic field distortion signal caused purely by the change in flow velocity was extracted, with a signal sampling rate of 2000 Hz.

8. A multi-sensor large casting pouring monitoring system according to claim 1, characterized in that, The rule base for the fast response layer supports online updates; The central monitoring platform's deep analysis layer analyzes historical casting data and final casting quality inspection results monthly, automatically extracting new defect predictor feature patterns. After confirmation by engineers, the rules are distributed to each edge-side physical field reconstruction unit via an encrypted network channel to update their local rule base.

9. A multi-sensor large casting pouring monitoring system according to claim 1, characterized in that, It also includes a data synchronization and clock calibration module; The data synchronization and clock calibration module ensures that the data acquisition timestamps of all sensors in the physical field sensing network, the calculation cycle of the edge-side physical field reconstruction unit, and the command execution time of the actuator control unit are all synchronized with the high-precision BeiDou time source at the microsecond level. The maximum clock deviation is controlled within ±2μs to ensure strict timing consistency between the data flow and control flow of the entire system.

10. A multi-sensor method for monitoring the pouring of large castings, characterized in that, Includes the following steps: S1: Collect multi-source physical quantity data during the casting process through a physical field sensing network. It consists of multiple sensor groups deployed on the outer wall of the mold, the casting system, and preset internal detection points. The sensor groups include thermocouple temperature sensor groups, vibration acceleration sensor groups, and magnetic field strength sensor groups. The thermocouple temperature sensor groups are arranged according to a non-uniform topology. The deployment positions of all measuring points are determined by pre-performed computational fluid dynamics simulation of the casting process. S2: In the edge-side physical field reconstruction unit, real-time data streams from the physical field sensing network are received, and a dynamic physical field reconstruction algorithm based on physical mechanism is executed to reconstruct the global physical field in real time. The dynamic physical field reconstruction algorithm constructs a simplified physical field calculation kernel that integrates the basic equations of fluid dynamics and the law of heat conduction. The simplified physical field calculation kernel discretizes the casting area into a non-uniform sparse grid that strictly corresponds to the sensor deployment position and a high-resolution uniform background grid that covers the entire computational domain. S3: Defect risk identification and control decision generation are achieved through multi-level decision units, which include a fast response layer integrated into the edge-side physical field reconstruction unit and a deep analysis layer deployed on the central monitoring platform. S4: Receives instructions from multi-level decision-making units through the actuator control unit and drives the corresponding actuators to adjust process parameters.

Images