A hierarchical control method for reconfigurable casting process modules

Abstract

This invention discloses a hierarchical control method for a reconfigurable casting process module, comprising: first, acquiring multi-source sensor data streams generated during the execution of the casting task, including key process parameters such as temperature, pressure, and flow rate. Based on this data stream, determining an initial set of operating condition features to characterize dynamic operating conditions. Subsequently, using a hierarchical decoupling control model, segmented operating condition optimization processing is performed on this feature set, and an optimized set of operating condition features is obtained through system coupling decoupling and random disturbance removal. Finally, feature reconstruction processing is performed on the optimized feature set to generate optimized control commands that precisely match the current casting process, used to adjust pouring speed, cooling rate, or pressure parameters in real time, thereby achieving closed-loop optimized control of the casting process. This invention effectively improves control accuracy and process stability under complex operating conditions.

Description

Technical Field

[0001] This invention relates to the field of industrial control, and more specifically, to a reconfigurable hierarchical control method for casting process modules. Background Technology

[0002] In metal casting production, the precision of process control directly determines the quality of the castings. Traditional control methods often rely on preset fixed process parameters or simple single-point feedback adjustments, which are difficult to adapt to the complex dynamic processes involving multiple coupled steps such as pouring, solidification, and cooling. In actual production, there are strong coupling relationships between various process parameters; for example, the pouring speed and the mold temperature field influence each other. At the same time, random disturbances in the production environment, such as power grid fluctuations and mold wear, continuously interfere with process stability. Existing technologies lack a collaborative decoupling and compensation mechanism for system coupling and random disturbances, resulting in lagging process control, poor adaptability, and an inability to meet the stringent requirements of high-end castings for forming accuracy and internal quality. Therefore, there is an urgent need for an advanced method that can sense the working conditions in real time, intelligently decouple disturbances, and dynamically optimize control commands. Summary of the Invention

[0003] The purpose of this invention is to provide a reconfigurable casting process module hierarchical control method.

[0004] In a first aspect, embodiments of the present invention provide a reconfigurable casting process module hierarchical control method, comprising:

[0005] Acquire multi-source sensor data streams generated during the execution of a metal casting task; the metal casting task includes at least one or more of the following processes: pouring, solidification, and cooling; the multi-source sensor data streams include process parameter data reflecting the state of the molten metal or casting, and the process parameter data includes one or more of the following: temperature, pressure, flow rate, displacement, and strain;

[0006] Based on the multi-source sensor data stream, an initial set of working condition features to be processed is determined; the initial set of working condition features is used to characterize the dynamic working conditions during the casting process.

[0007] Based on the hierarchical decoupling control model, the initial working condition feature set is subjected to segmented working condition optimization processing to obtain the optimized working condition feature set.

[0008] The optimized working condition feature set is subjected to feature reconstruction processing to generate optimized control instructions that match the current casting process; the optimized control instructions are used to adjust one or more of the pouring speed, cooling rate, and pressure parameters to achieve closed-loop optimized control of the casting process.

[0009] In one possible implementation, the step of performing segmented operating condition optimization processing on the initial operating condition feature set based on the hierarchical decoupling control model to obtain an optimized operating condition feature set includes:

[0010] The multi-source sensor data stream generated during the casting task execution process is acquired. The multi-source sensor data stream includes: first state data generated on the execution side during the casting task execution process, disturbance background data where the execution side is located, and remote reference data. The remote reference data is generated based on the second state data generated on the control side.

[0011] Based on the multi-source sensor data stream, an initial set of operating condition features to be processed is determined. The initial set of operating condition features includes a master state feature matrix, a reference coupling feature matrix, and a deviation feature matrix. The master state feature matrix is ​​obtained by extracting time-series features from the multi-source sensor data stream. The reference coupling feature matrix is ​​obtained by extracting time-series features from the second state data. The deviation feature matrix is ​​obtained by extracting time-series features from the decoupling deviation between the superimposed coupled response quantities of the multi-source sensor data stream and the second state data.

[0012] The initial operating condition feature set is segmented and optimized based on the hierarchical decoupling control model to obtain the optimized operating condition feature set; the operating condition optimization process includes system coupling decoupling and random disturbance removal.

[0013] In one possible implementation, the hierarchical decoupling control model includes at least two operating condition optimization sub-models; any one of the operating condition optimization sub-models is used to: perform an operating condition optimization process on any one or more feature matrices among the master state feature matrix, the reference coupling feature matrix, and the deviation feature matrix; the step of performing segmented operating condition optimization processing on the initial operating condition feature set based on the hierarchical decoupling control model to obtain an optimized operating condition feature set includes:

[0014] Based on the target operating condition tuning sub-model among at least two operating condition tuning sub-models, operating condition tuning is performed on any one or more feature matrices among the master state feature matrix, reference coupling feature matrix, and deviation feature matrix to obtain candidate feature matrices; the target operating condition tuning sub-model is any one of the at least two operating condition tuning sub-models.

[0015] Feature integration is performed on at least two candidate feature matrices to obtain an optimized working condition feature set.

[0016] In one possible implementation, the hierarchical decoupling control model includes: a first operating condition optimization sub-model and a second operating condition optimization sub-model; the second operating condition optimization sub-model includes: a dimensionality reduction component, a self-attention component, and an enhancement component;

[0017] The step of performing segmented operating condition optimization processing on the initial operating condition feature set based on the hierarchical decoupling control model yields an optimized operating condition feature set, including:

[0018] Based on the master state feature matrix, the reference coupling feature matrix, and the deviation feature matrix, the master state feature matrix is ​​subjected to initial working condition optimization processing based on the first working condition optimization sub-model to obtain the first response feature matrix.

[0019] The first response feature matrix and the deviation feature matrix are subjected to dimension reduction and superposition projection by the dimension reduction component to obtain the dimension reduction projection feature matrix.

[0020] The self-attention component is used to perform self-attention fusion processing on the reduced-dimensional projection feature matrix according to the temporal direction to obtain a self-attention optimized working condition feature set.

[0021] The self-attention optimization feature set is enhanced by an enhancement component to obtain the enhanced weight parameters for the optimization of the operating conditions.

[0022] Based on the aforementioned operating condition optimization enhancement weight parameters, the deviation feature matrix is ​​subjected to advanced operating condition optimization processing to obtain the second response feature matrix.

[0023] The first response feature matrix and the second response feature matrix are integrated to obtain the optimized operating condition feature set.

[0024] In one possible implementation, the first operating condition optimization sub-model includes: a feature embedding component, a coupling / decoupling component, and a feature reconstruction component; the step of performing initial operating condition optimization processing on the main state feature matrix based on the main state feature matrix, the reference coupling feature matrix, and the deviation feature matrix, to obtain a first response feature matrix, includes:

[0025] The feature embedding component performs feature embedding operations on the master state feature matrix, the reference coupling feature matrix, and the deviation feature matrix to obtain the embedded matrix features.

[0026] Through the coupling-decoupling component, the kernel method is used to perform feature decoupling processing on the embedded matrix features in both the time and frequency domains to obtain kernel similarity features;

[0027] The kernel similarity features are reconstructed using the feature reconstruction component to obtain the control weight tensor.

[0028] Based on the aforementioned control weight tensor, the master state feature matrix is ​​subjected to initial working condition optimization processing to obtain the first response feature matrix.

[0029] In one possible implementation, the coupling and decoupling component includes an intra-process adaptation unit and an inter-process collaboration unit. The intra-process adaptation unit is used to perform intra-process state parsing of state data, and the inter-process collaboration unit is used to perform inter-process correlation mining of state data. The embedded matrix features are distributed in a two-dimensional structure: one dimension is the continuous sampling node sequence of multi-source sensors during the casting process, and the other dimension is the total number of workstation monitoring features to be collected under a single sampling node.

[0030] The process involves using the coupling-decoupling component and a kernel method to decouple the embedded matrix features in both the time and frequency domains to obtain kernel similarity features, including:

[0031] The embedded matrix features are segmented to obtain multiple sub-matrix features; the number of sub-matrix features is equal to the number of sampling nodes contained in the continuous sampling node sequence.

[0032] Based on the intra-process adaptation unit, the target sub-matrix feature in the multiple sub-matrix features is decomposed and adapted in the frequency domain direction through the symmetric kernel similarity mechanism to obtain the local matrix feature corresponding to the target sub-matrix feature.

[0033] Based on inter-process collaborative units, a symmetric kernel similarity mechanism is used to integrate the inter-process features of multiple local matrix features in the time domain to obtain kernel similarity features.

[0034] In one possible implementation, the first operating condition tuning sub-model further includes a multi-head attention component; the method further includes:

[0035] The main state feature matrix, the reference coupling feature matrix, and the deviation feature matrix are respectively subjected to feature matrix dimensionality reduction processing to obtain the first dimensionality-reduced feature matrix, the second dimensionality-reduced feature matrix, and the third dimensionality-reduced feature matrix.

[0036] The first, second, and third dimensionality-reduced feature matrices are subjected to feature integration processing to obtain the first integrated feature matrix.

[0037] The first integrated feature matrix is ​​subjected to multi-head attention fusion by the multi-head attention component to obtain an attention fusion feature matrix, and the attention fusion feature matrix is ​​subjected to feature embedding operation by the feature embedding component to obtain an embedding matrix feature.

[0038] In one possible implementation, the operating condition optimization enhancement weight parameters include a first self-attention compensation weight vector and a second self-attention compensation weight vector. The first self-attention compensation weight vector is a weight vector after operating condition optimization processing of the deviation feature matrix, and the second self-attention compensation weight vector is a weight vector after operating condition optimization processing of the first response feature matrix.

[0039] The method further includes:

[0040] The first self-attention compensation weight vector and the deviation feature matrix are weighted and modulated to obtain the second response feature matrix;

[0041] The second self-attention compensation weight vector and the first response feature matrix are weighted and modulated to obtain the third response feature matrix.

[0042] In one possible implementation, the casting task execution process is a process of process coordination between the first control unit and the second control unit; the method further includes:

[0043] Acquire the second state data generated by the second control unit during the casting task execution process;

[0044] Based on the dynamic decoupling mechanism, the superimposed system coupling is decoupled from the second state data to obtain the superimposed coupled response quantity corresponding to the second state data.

[0045] In one possible implementation, based on the multi-source sensor data stream, the initial set of operating condition features to be processed is determined, including:

[0046] The multi-source sensor data stream is subjected to time-series feature extraction based on the wavelet packet decomposition algorithm to obtain the main state feature matrix; and the second state data is subjected to time-series feature extraction based on the wavelet packet decomposition algorithm to obtain the reference coupling feature matrix.

[0047] The decoupling compensation calculation is performed on the multi-source sensor data stream and the superimposed coupled response quantity to obtain the decoupling deviation quantity. Then, the time-series features of the decoupling deviation quantity are extracted based on the wavelet packet decomposition algorithm to obtain the deviation feature matrix.

[0048] Compared to existing technologies, the beneficial effects of this invention include: employing a reconfigurable casting process module hierarchical control method disclosed in this invention, multi-source sensor data streams generated during the casting task execution are acquired, including key process parameters such as temperature, pressure, and flow rate. Based on this data stream, an initial set of operating condition features characterizing dynamic operating conditions is determined. Subsequently, a hierarchical decoupling control model is used to perform segmented operating condition optimization processing on this feature set. Through system coupling decoupling and random disturbance removal, an optimized set of operating condition features is obtained. Finally, the optimized feature set undergoes feature reconstruction processing to generate optimized control commands that precisely match the current casting process, used to adjust pouring speed, cooling rate, or pressure parameters in real time, thereby achieving closed-loop optimized control of the casting process. This invention effectively improves control accuracy and process stability under complex operating conditions. Attached Figure Description

[0049] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of the present invention and should not be considered as limiting the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0050] Figure 1 A schematic flowchart illustrating the steps of the reconfigurable casting process module hierarchical control method provided in this embodiment of the invention;

[0051] Figure 2 A schematic block diagram of the structure of a computer device provided in an embodiment of the present invention. Detailed Implementation

[0052] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0053] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0054] In order to solve the technical problems mentioned in the background art Figure 1 This is a flowchart illustrating the hierarchical control method for reconfigurable casting process modules provided in this embodiment. The following is a detailed description of this hierarchical control method for reconfigurable casting process modules.

[0055] Step S201: Obtain multi-source sensor data streams generated during the execution of the metal casting task; the metal casting task includes at least one or more of the following: pouring process, solidification process, and cooling process; the multi-source sensor data streams include process parameter data reflecting the state of the molten metal or casting, and the process parameter data includes one or more of the following: temperature, pressure, flow rate, displacement, and strain.

[0056] Step S202: Based on the multi-source sensor data stream, determine the initial working condition feature set to be processed; the initial working condition feature set is used to characterize the dynamic working conditions in the casting process.

[0057] Step S203: Based on the hierarchical decoupling control model, the initial working condition feature set is subjected to segmented working condition optimization processing to obtain the optimized working condition feature set.

[0058] Step S204: Perform feature reconstruction processing on the optimized working condition feature set to generate optimized control instructions that match the current casting process; the optimized control instructions are used to adjust one or more of the pouring speed, cooling rate, and pressure parameters to achieve closed-loop optimized control of the casting process.

[0059] In one embodiment of the present invention, exemplarily, on a multi-station hot mold casting production line for automotive steering knuckles, an edge server deployed on-site acquires multi-source sensor data streams generated during the casting task execution process in real time via the Industrial Ethernet protocol. This casting task encompasses three core processes: pouring, solidification, and cooling. The server continuously collects data from sensors installed at the corresponding stations: in the pouring process, the server collects thermocouple temperature data (reflecting the initial temperature of the molten metal) and electromagnetic flowmeter data (reflecting the flow rate of the molten metal) at the pouring cup of the pouring machine; in the solidification process, the server collects data from multiple pressure sensors within the mold cavity (reflecting pressure changes during molten metal filling and solidification) and infrared temperature sensor data (monitoring the solidification front temperature at different parts of the casting); in the cooling process, the server collects surface temperature scanner data when the casting enters the cooling channel after demolding, as well as flow rate and temperature data from the cooling water circulation system. All these process parameter data, such as temperature, pressure, and flow rate, are integrated into a unified time-series data stream for subsequent processing.

[0060] The edge server performs in-depth processing on the aforementioned multi-source sensor data streams to construct an initial set of operating condition features that dynamically characterizes the current casting process state. First, the server employs wavelet packet decomposition to extract multi-scale features from the time-series data of core process parameters such as pouring temperature, cavity pressure, and cooling water temperature, capturing their energy distribution and variation trends across different frequency bands to form a master state feature matrix. This matrix comprehensively reflects the actual physical state during the casting process. Simultaneously, the server synchronously acquires standard process parameters for the current steering knuckle product from a remote process server, such as theoretical pouring temperature curves and ideal holding pressure curves, and performs the same time-series feature extraction on them, generating a reference coupling feature matrix as a benchmark for process optimization. Furthermore, the server calculates the deviation between the measured data from each sensor and the corresponding standard reference data in real time, and extracts features from these deviations to form a deviation feature matrix, which quantifies the degree of deviation between the current operating condition and the ideal operating condition. The initial set of operating condition features, composed of these three feature matrices, comprehensively and structurally describes the dynamic operating conditions of the casting process.

[0061] The edge server invokes a pre-trained hierarchical decoupling control model to perform fine-tuning on the initial set of operating condition features. This model comprises two core sub-models. First, the first operating condition optimization sub-model (focusing on system coupling decoupling) performs a fusion analysis of the master state, reference coupling, and deviation feature matrices. Through its internal multi-head attention mechanism, it identifies and decouples systematic mutual interference between process parameters, such as separating the coupling effect of pouring speed fluctuations on the mold temperature field. After this step, a purified first response feature matrix is ​​obtained. Subsequently, the second operating condition optimization sub-model (focusing on random disturbance removal) receives the first response feature matrix and the deviation feature matrix. Its dimensionality reduction component focuses on random disturbance features caused by power grid fluctuations, changes in ambient temperature and humidity, etc., and learns the importance differences of each feature under different casting process stages (such as filling and holding pressure) through a self-attention mechanism, ultimately generating optimization weight parameters. These weights are used to modulate the features, further filtering out random noise, strengthening key process features, and outputting the second and third response feature matrices. The server integrates these optimized feature matrices to obtain an optimized set of operating condition features with a significantly improved signal-to-noise ratio that better reflects the essential requirements of the process.

[0062] The edge server performs inverse transformation and reconstruction of the optimized operating condition feature set. It uses time-series feature inverse mapping methods such as wavelet inverse transform to reconstruct the optimized high-dimensional feature data into a sequence of actual physical parameters, such as a set of optimized pouring speed setpoints, target water flow rates for each cooling zone, and pressure setting curves for the holding stage. Simultaneously, the server combines real-time collected background disturbance data (such as current grid voltage and cumulative mold wear) with a built-in compensation strategy to fine-tune the aforementioned physical parameters, generating a final optimized control instruction set adapted to the specific operating conditions. The server then distributes these instructions in real-time through the industrial network to the pouring machine controller, hydraulic system servo valve groups, and cooling water regulating valves. The equipment controller dynamically adjusts the pouring speed, cooling rate, and pressure parameters according to the instructions, thereby achieving precise, adaptive closed-loop optimization control of the casting process and ensuring the forming accuracy and internal quality of the steering knuckle casting.

[0063] In this embodiment of the invention, the step of performing segmented operating condition optimization processing on the initial operating condition feature set based on the hierarchical decoupling control model to obtain an optimized operating condition feature set can be implemented through the following example.

[0064] The multi-source sensor data stream generated during the casting task execution process is acquired. The multi-source sensor data stream includes: first state data generated on the execution side during the casting task execution process, disturbance background data where the execution side is located, and remote reference data. The remote reference data is generated based on the second state data generated on the control side.

[0065] Based on the multi-source sensor data stream, an initial set of operating condition features to be processed is determined. The initial set of operating condition features includes a master state feature matrix, a reference coupling feature matrix, and a deviation feature matrix. The master state feature matrix is ​​obtained by extracting time-series features from the multi-source sensor data stream. The reference coupling feature matrix is ​​obtained by extracting time-series features from the second state data. The deviation feature matrix is ​​obtained by extracting time-series features from the decoupling deviation between the superimposed coupled response quantities of the multi-source sensor data stream and the second state data.

[0066] The initial operating condition feature set is segmented and optimized based on the hierarchical decoupling control model to obtain the optimized operating condition feature set; the operating condition optimization process includes system coupling decoupling and random disturbance removal.

[0067] In this embodiment of the invention, for example, it is applied to a multi-station hot mold casting production line for complex components. The execution entity is an edge server deployed on the production line site. The core configuration of the production line includes: a large-tonnage hot mold casting equipment on the execution side (equipped with core sensors such as pressure, temperature, displacement, and servo status); a remote process server on the control side (stores a process knowledge base and issues reference operating condition commands); and uncertainties in the production environment such as power grid fluctuations, temperature changes, and mold wear. The edge server interacts with each device through industrial communication protocols, configures the data sampling frequency according to the real-time requirements of the casting process, and realizes high-precision closed-loop control of the casting process, meeting the stringent requirements of key components for forming accuracy and process stability.

[0068] The edge server collects multi-source sensor data streams in real time during the casting task execution process through industrial communication protocols. The data streams specifically include three types of data: first, execution-side first state data, which is the real-time state data generated by the casting equipment during operation, covering the time-series data of the core operating parameters of the equipment; second, execution-side disturbance background data, which is the environmental interference data that affects the stable operation of the equipment, including power fluctuations, temperature and humidity changes, mold state changes, etc. in the production environment; and third, remote reference data, which is generated based on the second state data issued by the control side. The second state data is the reference operating condition command output by the control side according to the process knowledge base, which the edge server converts into a reference dataset that matches the time sequence of the execution-side data.

[0069] The edge server determines the initial set of operating condition features to be processed based on the multi-source sensor data stream, and the specific process is as follows:

[0070] First, a wavelet analysis algorithm adapted to the timing characteristics of the casting process is used to extract timing features from the first state data on the execution side. Feature information of different frequency bands is extracted according to a preset decomposition strategy to form a main state feature matrix.

[0071] Secondly, the same timing feature extraction operation is performed on the second state data issued by the control side to extract the feature information corresponding to the reference operating condition command and form a reference coupling feature matrix.

[0072] Finally, a dynamic decoupling mechanism is constructed based on the coupling relationship of casting process parameters. The second state data is decoupled into a superimposable system coupling to obtain the independent superimposable coupling response quantities corresponding to each reference parameter. The deviation between the first state data on the execution side and the superimposable coupling response quantities is calculated to obtain the decoupling deviation quantity. The decoupling deviation quantity is subjected to time-series feature extraction to form a deviation feature matrix.

[0073] The initial operating condition feature set is composed of the main state feature matrix, the reference coupling feature matrix, and the deviation feature matrix.

[0074] The edge server performs segmented operating condition optimization on the initial operating condition feature set based on a hierarchical decoupling control model. The hierarchical decoupling control model includes a first operating condition optimization sub-model for system coupling decoupling and a second operating condition optimization sub-model for random disturbance removal. The specific processing procedure is as follows:

[0075] First, the main state feature matrix, reference coupling feature matrix, and deviation feature matrix are dimensionality-reduced and integrated. The association weights of the integrated features are calculated through the multi-head attention mechanism of the first working condition tuning sub-model to obtain the attention fusion feature matrix. After enhancing the expressive power of the attention fusion feature matrix through high-dimensional embedding, the embedding matrix is ​​locally decomposed according to time steps through the intra-process adaptation unit, and then the kernel similarity feature matrix is ​​obtained through the inter-process collaboration unit in the time domain direction. Based on the kernel similarity feature, the control weights are generated and the main state feature matrix is ​​weighted and modulated to obtain the first response feature matrix after removing system-level coupling interference.

[0076] Secondly, the first response feature matrix and the deviation feature matrix are dimensionality-reduced and projected using the dimensionality reduction component of the second working condition optimization sub-model to obtain a dimensionality-reduced projection feature matrix that focuses on key disturbance information; the dimensionality-reduced projection feature matrix is ​​then fused with attention in the time domain using the self-attention component to learn the feature importance at different time steps and obtain a self-attention optimized feature set; the self-attention optimized features are then enhanced using the feature enhancement component to generate working condition optimization enhancement weight parameters corresponding to the deviation feature matrix and the first response feature matrix.

[0077] Finally, the deviation feature matrix and the first response feature matrix are weighted and modulated based on the working condition optimization enhancement weight parameters to obtain the second response feature matrix with random disturbance removed and the third response feature matrix with enhanced key process features; the first, second and third response feature matrices are integrated to obtain the optimized working condition feature set.

[0078] The edge server performs feature reconstruction processing on the optimized operating condition feature set, converts the optimized features into actual physical parameters through time-series feature inverse mapping, and corrects the physical parameters in combination with the disturbance background compensation strategy to generate optimized control operating condition data adapted to the current operating condition. The edge server sends the optimized control operating condition data to the execution side equipment controller according to the process real-time requirements. The controller adjusts the equipment operating parameters according to the optimization instructions to realize dynamic optimization of the casting operating condition.

[0079] In this embodiment, the edge server acts as the execution entity. Through the entire process of multi-source data acquisition, temporal feature extraction, hierarchical decoupling optimization, and control command generation, it realizes reconfigurable hierarchical control of the casting process. This not only solves the system-level coupling interference problem but also eliminates the influence of random disturbances, providing a stable and efficient control scheme for the casting process under complex working conditions.

[0080] In this embodiment of the invention, the hierarchical decoupling control model includes at least two operating condition optimization sub-models; any one of the operating condition optimization sub-models is used to: perform an operating condition optimization process on any one or more feature matrices among the master state feature matrix, the reference coupling feature matrix, and the deviation feature matrix; the step of performing segmented operating condition optimization processing on the initial operating condition feature set based on the hierarchical decoupling control model to obtain an optimized operating condition feature set can be implemented through the following example.

[0081] Based on the target operating condition tuning sub-model among at least two operating condition tuning sub-models, operating condition tuning is performed on any one or more feature matrices among the master state feature matrix, reference coupling feature matrix, and deviation feature matrix to obtain candidate feature matrices; the target operating condition tuning sub-model is any one of the at least two operating condition tuning sub-models.

[0082] Feature integration is performed on at least two candidate feature matrices to obtain an optimized working condition feature set.

[0083] In this embodiment of the invention, for example, this embodiment still takes a multi-station hot mold casting production line for complex components as the application scenario. The execution subject is the edge server on the production line. The hierarchical decoupling control model is configured with two working condition optimization sub-models: the first working condition optimization sub-model (system coupling decoupling) and the second working condition optimization sub-model (random disturbance stripping). The initial working condition feature set includes three types of feature matrices: the main state feature matrix (the time sequence features of real-time pressure of the casting equipment, mold temperature, and slider displacement), the reference coupling feature matrix (the time sequence features of reference pressure and reference temperature issued by the remote process server), and the deviation feature matrix (the time sequence features of the deviation between the actual working condition and the reference working condition).

[0084] Hierarchical optimization and feature integration process:

[0085] The target sub-model processes and generates a candidate feature matrix:

[0086] The edge server calls the target operating condition optimization sub-model in the order of "first system coupling decoupling, then random perturbation stripping":

[0087] Calling the first working condition optimization sub-model (one of the target sub-models): The server selects the main state feature matrix and the reference coupling feature matrix from the initial working condition feature set as input. The first sub-model first divides the main state feature matrix into several sub-matrices according to time steps, and decomposes the coupling features within the sub-matrices through the intra-process adaptation unit (such as the system-level coupling of "pressure-temperature" in the main state) to obtain local independent features; then, through the inter-process collaboration unit, the local features are integrated in the time domain (such as the timing logic of the pressure features of the current station and the previous station), and finally outputs the first candidate feature matrix, that is, the main state feature matrix after removing system-level coupling interference.

[0088] The second operating condition optimization sub-model (target sub-model two) is invoked: The server selects the deviation feature matrix from the initial operating condition feature set as input. The second sub-model projects the deviation feature matrix through a dimensionality reduction component (focusing on deviation features corresponding to key disturbances such as power grid fluctuations and ambient temperature changes), and then learns the feature importance in the time domain through a self-attention component (e.g., the deviation feature weight in the casting and forming stage is higher than that in the no-load stage), and finally outputs the second candidate feature matrix, which is the deviation feature matrix after removing random disturbances.

[0089] The candidate feature matrices are integrated to generate an optimized working condition feature set:

[0090] The edge server performs feature integration on the two candidate feature matrices mentioned above:

[0091] First, the first candidate feature matrix (the main state features decoupled from the system) and the second candidate feature matrix (the deviation features free from random disturbances) are concatenated column by column to form a temporary integrated matrix. Then, redundant features are filtered by a preset feature selection rule (such as retaining the top 70% of the dimensions with the highest feature contribution) to finally obtain the optimized operating condition feature set. This set retains the core state features of the real-time operation of the casting equipment and integrates the deviation compensation features after disturbance removal, which can be directly used for subsequent control command generation.

[0092] In this embodiment, the edge server achieves hierarchical decoupling and optimization of casting condition features through the process of "target sub-model segmentation processing → candidate matrix feature integration", which effectively solves the problem of interference of system coupling and random disturbances on process control.

[0093] In this embodiment of the invention, the hierarchical decoupling control model includes: a first operating condition optimization sub-model and a second operating condition optimization sub-model; the step of performing segmented operating condition optimization processing on the initial operating condition feature set based on the hierarchical decoupling control model to obtain an optimized operating condition feature set can be implemented through the following example.

[0094] Based on the master state feature matrix, the reference coupling feature matrix, and the deviation feature matrix, the master state feature matrix is ​​subjected to initial working condition optimization processing based on the first working condition optimization sub-model to obtain the first response feature matrix.

[0095] Based on the first response feature matrix and the deviation feature matrix, the deviation feature matrix is ​​subjected to advanced working condition optimization processing based on the second working condition optimization sub-model to obtain the second response feature matrix.

[0096] The first response feature matrix and the second response feature matrix are integrated to obtain the optimized operating condition feature set.

[0097] In this embodiment of the invention, for example, this embodiment continues the scenario of a multi-station hot mold casting production line for complex components, and the execution subject is the edge server on the production line. The hierarchical decoupling control model includes two serially connected sub-models for working condition optimization: the first working condition optimization sub-model (initial system coupling decoupling) and the second working condition optimization sub-model (advanced random disturbance stripping). The initial set of working condition features is fixed as three types of matrices: the main state feature matrix (time-series features of real-time pressure, mold temperature, and slider displacement of the casting equipment), the reference coupling feature matrix (time-series features of reference pressure and reference temperature issued by the remote process server), and the deviation feature matrix (time-series features of the deviation between the actual working condition and the reference working condition).

[0098] Segmented optimization and feature integration process:

[0099] The first sub-model initial optimization generates the first response feature matrix:

[0100] The edge server synchronously inputs the master state feature matrix, reference coupling feature matrix, and deviation feature matrix of the initial operating condition feature set into the first operating condition optimization sub-model to start the initial operating condition optimization:

[0101] The first sub-model first maps the features of the three types of matrices to a high-dimensional space through the feature embedding component, realizing cross-matrix feature association; then it calls the coupling and decoupling component, using the physical logic of the casting process as constraints (such as the coupling relationship of "pressure-temperature-displacement"), to decouple the main state feature matrix within the process (decompose the system-level coupling features of "pressure-temperature" within a single time step) and associate it between processes (integrate the displacement feature trends of continuous time steps); finally, it outputs the first response feature matrix, which is the main state feature matrix after removing system-level coupling interference. This matrix retains the core state information of the casting equipment operation and eliminates the system coupling deviation between the reference condition and the actual condition.

[0102] The second sub-model is further optimized to generate the second response feature matrix:

[0103] The edge server inputs the first response feature matrix and the deviation feature matrix from the initial operating condition feature set into the second operating condition optimization sub-model to initiate advanced operating condition optimization:

[0104] The second sub-model first performs joint dimension reduction projection on the two types of matrices through a dimension reduction component, focusing on feature dimensions strongly correlated with random disturbances (such as pressure deviation features corresponding to power grid fluctuations); then, it uses a self-attention component to assign time-domain weights to the dimension-reduced features, strengthening the deviation feature weights of key stages in casting (such as mold closing and pressure holding stages); finally, it generates operating condition optimization weight parameters through an enhancement component, performs weighted modulation on the deviation feature matrix, and outputs the second response feature matrix, which is the deviation feature matrix after removing random disturbances (such as power grid fluctuations and ambient temperature fluctuations). This matrix only retains the controllable deviation information of the process itself.

[0105] Feature integration generates an optimized feature set for operating conditions:

[0106] The edge server performs feature integration operations on the first response feature matrix and the second response feature matrix:

[0107] The server first aligns the two types of matrices by time step and then concatenates them by column to form a temporary integrated feature matrix. Then, it removes redundant feature dimensions by using preset feature selection rules (such as filtering logic based on feature contribution). Finally, it obtains an optimized operating condition feature set, which includes both the core state features of the equipment decoupled from the system and the controllable deviation features free from random disturbances. This set can be directly used to generate subsequent optimized control operating condition data.

[0108] In this embodiment, the edge server achieves hierarchical purification of initial operating condition features through the collaborative logic of a series of sub-models: "initial system decoupling → advanced disturbance stripping → feature integration," providing a high signal-to-noise ratio feature foundation for subsequent precise control.

[0109] In this embodiment of the invention, the first operating condition optimization sub-model includes: a feature embedding component, a coupling decoupling component, and a feature reconstruction component; the first operating condition optimization process is performed on the main state feature matrix based on the main state feature matrix, the reference coupling feature matrix, and the deviation feature matrix to obtain the first response feature matrix, which can be implemented through the following example.

[0110] The feature embedding component performs feature embedding operations on the master state feature matrix, the reference coupling feature matrix, and the deviation feature matrix to obtain the embedded matrix features.

[0111] Through the coupling-decoupling component, the kernel method is used to perform feature decoupling processing on the embedded matrix features in both the time and frequency domains to obtain kernel similarity features;

[0112] The kernel similarity features are reconstructed using the feature reconstruction component to obtain the control weight tensor.

[0113] Based on the aforementioned control weight tensor, the master state feature matrix is ​​subjected to initial working condition optimization processing to obtain the first response feature matrix.

[0114] In this embodiment of the invention, for example, this embodiment continues the scenario of a multi-station hot mold casting production line for complex components, and the execution subject is the edge server on the production line site; the first working condition optimization sub-model is fixedly configured with three types of functional components: feature embedding component (to realize cross-matrix feature association), coupling and decoupling component (to decouple system-level coupling based on kernel method), and feature reconstruction component (to generate control weights); the initial input is three types of feature matrices: main state feature matrix (time-series features of real-time pressure, mold temperature, and slider displacement of casting equipment), reference coupling feature matrix (time-series features of reference pressure and reference temperature issued by remote process server), and deviation feature matrix (time-series features of deviation between actual working condition and reference working condition).

[0115] The specific process of initial operating condition optimization:

[0116] Feature embedding component generates embedding matrix features

[0117] After the edge server strictly aligns the master state feature matrix, reference coupling feature matrix, and deviation feature matrix in the initial operating condition feature set according to time steps, it calls the feature embedding component to perform the embedding operation:

[0118] The server first concatenates the three types of matrices column-wise into a joint feature matrix (integrating real-time equipment status, remote reference commands, and operating condition deviations). Then, through the linear transformation layer of the feature embedding component, the joint feature matrix is ​​mapped to a pre-defined high-dimensional embedding space. This space is initialized with dimensions based on prior knowledge of the casting process (such as the physical relationship between pressure, temperature, and displacement) to ensure effective correlation between features across matrices. Finally, the embedded matrix feature is output, which retains the original information of the three initial features while strengthening the logical correlation between features in different matrices, providing a foundation for subsequent coupling and decoupling.

[0119] The coupling / decoupling component generates kernel similarity features based on a kernel method:

[0120] The edge server invokes a coupling / decoupling component, using a Gaussian kernel function as its core tool, to perform bidirectional coupling / decoupling in both the time and frequency domains on the embedded matrix features.

[0121] In the time domain, the server calculates the similarity of features at adjacent time steps in the embedding matrix using a Gaussian kernel function (e.g., the kernel similarity of pressure features at time t and t+1), identifying and decoupling systematic coupling correlations in the time dimension (e.g., the cascading interference of pressure fluctuations at the previous workstation on the temperature of the current workstation). In the frequency domain, the server first performs a short-time Fourier transform on the features of the embedding matrix to obtain frequency components, and then calculates the similarity of different frequency components using a Gaussian kernel function, decoupling the coupling features of different physical quantities within the same time step (e.g., the system-level coupling of "pressure-temperature" in the master state features). Finally, the kernel similarity feature is output, which has eliminated the systematic coupling interference in the three types of initial features, retaining only independent and valid information.

[0122] The feature reconstruction component generates a control weight tensor:

[0123] The edge server inputs the kernel similarity features into the feature reconstruction component to initiate the weight generation process:

[0124] The feature reconstruction component performs dimensionality transformation on kernel similarity features using a multilayer perceptron (MLP). First, it compresses the high-dimensional kernel similarity features to a dimension matching the master-state feature matrix. Then, it learns feature importance weights using an activation function (such as ReLU), ultimately outputting a control weight tensor. This tensor has the same dimension as the master-state feature matrix (each element corresponds to the control weight of a feature dimension at a time step in the master-state feature matrix). Higher weight values ​​indicate stronger independent validity of the corresponding feature.

[0125] Weighted modulation generates the first response feature matrix:

[0126] The edge server will perform element-wise weighted modulation of the weight tensor and the master state feature matrix:

[0127] The server, based on the one-to-one correspondence between time steps and feature dimensions, multiplies each element of the control weight tensor with the corresponding element of the master state feature matrix. It retains independent, effective features with higher weights (such as pressure features and mold temperature features in the core casting stage) and removes coupling interference features with lower weights (such as redundant pressure fluctuation features caused by systemic deviations). The final output is the first response feature matrix, which is the master state feature matrix after removing system-level coupling interference. It retains only the core independent state information of the casting equipment and can be directly used for subsequent advanced optimization.

[0128] In this embodiment, the edge server achieves the initial system coupling decoupling of the master state feature matrix through a coherent process of "feature embedding → kernel method decoupling → feature reconstruction → weighted modulation", providing a high-purity feature foundation for subsequent random perturbation stripping.

[0129] In this embodiment of the invention, the coupling and decoupling component includes an intra-process adaptation unit and an inter-process collaboration unit. The intra-process adaptation unit is used to perform intra-process state parsing of state data, and the inter-process collaboration unit is used to perform inter-process correlation mining of state data. The embedded matrix features are distributed in a two-dimensional structure: one dimension is the continuous sampling node sequence of multi-source sensors during the casting process, and the other dimension is the total number of workstation monitoring features to be collected under a single sampling node.

[0130] The method of using the coupling-decoupling component to perform feature decoupling processing on the embedded matrix features in the time and frequency domains to obtain kernel similarity features can be implemented through the following example.

[0131] The embedded matrix features are segmented to obtain multiple sub-matrix features; the number of sub-matrix features is equal to the number of sampling nodes contained in the continuous sampling node sequence.

[0132] Based on the intra-process adaptation unit, the target sub-matrix feature in the multiple sub-matrix features is decomposed and adapted in the frequency domain direction through the symmetric kernel similarity mechanism to obtain the local matrix feature corresponding to the target sub-matrix feature.

[0133] Based on inter-process collaborative units, a symmetric kernel similarity mechanism is used to integrate the inter-process features of multiple local matrix features in the time domain to obtain kernel similarity features.

[0134] In this embodiment of the invention, for example, this embodiment continues the scenario of a multi-station hot mold casting production line for complex components, and the execution subject is the edge server on the production line. The coupling and decoupling component has a built-in intra-process adaptation unit (responsible for feature coupling and decomposition within a single station) and an inter-process collaboration unit (responsible for the association of time-series features of multiple stations). The embedded matrix features are two-dimensional structured data generated by the server through feature embedding operations. One dimension corresponds to the continuous sampling node sequence of the casting process (covering the full time-series sampling points from the start of casting to the completion of pressure holding), and the other dimension corresponds to the station monitoring feature set of a single sampling node (including core process features such as casting equipment pressure, mold temperature, and slider displacement).

[0135] Feature processing of coupling / decoupling components:

[0136] Embedded matrix feature segmentation to generate sub-matrix features:

[0137] The edge server performs node-level segmentation on the two-dimensional structured embedding matrix features:

[0138] The server strictly adheres to the temporal boundaries of the continuous sampling node sequence, splitting the embedding matrix into units of "single sampling node." Each sub-matrix feature corresponds to all workstation monitoring features of a sampling node; that is, each sub-matrix feature contains the embedded information of all monitoring items such as pressure, temperature, and displacement at that sampling node. The final number of sub-matrix features is exactly the same as the total number of nodes in the continuous sampling node sequence, ensuring that the features of each time-series node are independently split.

[0139] Frequency domain coupling decomposition of the adapter unit within the process:

[0140] The edge server invokes the intra-process adaptation unit of the coupling and decoupling component to perform intra-process feature decomposition in the frequency domain for each target sub-matrix feature (i.e., the set of workstation monitoring features for a single sampling node):

[0141] The intra-process adaptation unit adopts a preset symmetric kernel similarity mechanism (a kernel function based on the physical correlation optimization of the casting process). First, it performs frequency domain transformation on the target sub-matrix features to extract the frequency components of different monitoring features. Then, it calculates the similarity between each frequency component through the symmetric kernel function, identifies and decomposes the intra-process coupling correlation of different process features within the same sampling node (such as the systematic frequency domain coupling of "pressure-temperature" under the same node, and the feature correlation caused by the hardware coupling of the equipment hydraulic system and heating system). Finally, it outputs the local matrix features corresponding to each target sub-matrix feature. This feature has been stripped of intra-process coupling interference and only retains the independent and effective information of each process feature under a single node.

[0142] Temporal correlation integration of inter-process collaborative units:

[0143] The edge server invokes the inter-process collaboration unit of the coupling and decoupling component to perform inter-process feature integration in the time domain on the local matrix features of all sampling nodes:

[0144] The inter-process collaboration unit uses the same symmetric kernel similarity mechanism. According to the temporal sequence of the continuous sampling nodes, it calculates the temporal kernel similarity of the local matrix features of adjacent sampling nodes, identifies and associates the inter-process feature logic between different nodes (such as the temporal association between the previous sampling node "slider displacement in place" and the current sampling node "casting pressure peak", which conforms to the physical logic of "pressure rises after displacement in place" in the casting process). Then, it performs temporal dimension association and integration on all local matrix features, removes invalid coupling in the temporal sequence (such as random feature interference from non-adjacent nodes), and finally outputs kernel similarity features. This feature not only removes the process coupling within a single node, but also retains the process temporal association between multiple nodes, providing a pure basic feature for subsequent feature reconstruction.

[0145] In this embodiment, the edge server achieves the transformation of embedded matrix features from "two-dimensional structured original information" to "decoupled pure features" through two-level processing of coupling and decoupling components. This ensures the independence of process features while preserving the logical relationship between processes, fully adapting to the complex control requirements of the casting process.

[0146] In this embodiment of the invention, the first operating condition tuning sub-model further includes a multi-head attention component; this embodiment of the invention also provides the following implementation methods.

[0147] The main state feature matrix, the reference coupling feature matrix, and the deviation feature matrix are respectively subjected to feature matrix dimensionality reduction processing to obtain the first dimensionality-reduced feature matrix, the second dimensionality-reduced feature matrix, and the third dimensionality-reduced feature matrix.

[0148] The first, second, and third dimensionality-reduced feature matrices are subjected to feature integration processing to obtain the first integrated feature matrix.

[0149] The first integrated feature matrix is ​​subjected to multi-head attention fusion by the multi-head attention component to obtain an attention fusion feature matrix, and the attention fusion feature matrix is ​​subjected to feature embedding operation by the feature embedding component to obtain an embedding matrix feature.

[0150] In this embodiment of the invention, for example, this embodiment continues the scenario of a multi-station hot mold casting production line for complex components, and the execution subject is the edge server on the production line site; the first working condition optimization sub-model adds a multi-head attention component (responsible for cross-matrix feature association capture), which works in collaboration with the original feature embedding component; the initial input consists of three types of feature matrices: main state feature matrix (temporal features of real-time pressure, mold temperature, and slider displacement of the casting equipment), reference coupling feature matrix (temporal features of reference pressure and reference temperature issued by the remote process server), and deviation feature matrix (temporal features of deviation between the actual working condition and the reference working condition).

[0151] Multi-head attention fusion and feature embedding process:

[0152] Feature matrix dimensionality reduction:

[0153] The edge server performs dimensionality reduction operations on the three types of initial feature matrices respectively:

[0154] The server invokes a dimensionality reduction algorithm based on prior optimization of the casting process. It retains core process features strongly correlated with casting forming in the master state feature matrix, generating the first dimensionality-reduced feature matrix; it retains key reference features of remote process commands in the reference coupling feature matrix, generating the second dimensionality-reduced feature matrix; and it retains the main deviation features affecting process stability in the deviation feature matrix, generating the third dimensionality-reduced feature matrix. The dimensionality reduction process strictly follows the process physical meaning of the features, ensuring no core information is lost.

[0155] Dimensionality reduction feature integration generates the first integrated feature matrix:

[0156] Edge servers perform time-step alignment and feature integration on three types of dimensionality-reduced feature matrices:

[0157] The server uses the continuous sampling node sequence of the main state feature matrix as a benchmark to strictly align the time sequence nodes of the second and third dimensionality reduction feature matrices; then it concatenates the three types of dimensionality reduction matrices column by column to form the first integrated feature matrix. This matrix simultaneously contains three types of dimensionality reduction features: real-time device status, remote reference instructions, and operating condition deviations, providing a foundation for cross-matrix feature association capture.

[0158] Feature fusion of multi-head attention components:

[0159] The edge server invokes the multi-head attention component of the first-condition tuning sub-model to perform correlation fusion on the first integrated feature matrix:

[0160] The component employs a multi-head configuration based on the complexity of casting process feature associations. Each attention head focuses on one type of feature association dimension (e.g., the first head captures the association between master state pressure and reference pressure, and the second head captures the association between master state temperature and deviation features). Through attention weight calculation, it strengthens the feature associations that are key to process control (e.g., the deviation association between master state pressure and reference pressure during the casting holding stage) and weakens the interference of irrelevant features. Finally, it outputs an attention fusion feature matrix, which highlights key feature associations across matrices and eliminates redundant association information.

[0161] Feature embedding generates embedding matrix features:

[0162] The edge server integrates the attention-based feature matrix into the feature embedding component.

[0163] The component maps the fusion matrix to a high-dimensional space through a high-dimensional embedding space initialized based on the physical association of the casting process, thereby enhancing the expressive power of key association features. Finally, it outputs an embedding matrix feature, which not only retains the core information of the three initial features, but also strengthens the key cross-matrix association through multi-head attention, providing high-quality input for subsequent coupling and decoupling components.

[0164] In this embodiment, the edge server improves the effectiveness and relevance of the initial features through the process of "dimensionality reduction → integration → multi-head fusion → embedding", laying a precise foundation for the coupling and decoupling operation of the sub-model in the first working condition.

[0165] In this embodiment of the invention, the second operating condition optimization sub-model includes: a dimensionality reduction component, a self-attention component, and an enhancement component; the step of performing advanced operating condition optimization processing on the deviation feature matrix based on the first response feature matrix and the deviation feature matrix, and obtaining the second response feature matrix, can be implemented through the following example.

[0166] The first response feature matrix and the deviation feature matrix are subjected to dimension reduction and superposition projection by the dimension reduction component to obtain the dimension reduction projection feature matrix.

[0167] The self-attention component is used to perform self-attention fusion processing on the reduced-dimensional projection feature matrix according to the temporal direction to obtain a self-attention optimized working condition feature set.

[0168] The self-attention optimization feature set is enhanced by an enhancement component to obtain the enhanced weight parameters for the optimization of the operating conditions.

[0169] Based on the aforementioned operating condition optimization enhancement weight parameters, the deviation feature matrix is ​​subjected to advanced operating condition optimization processing to obtain the second response feature matrix.

[0170] In this embodiment of the invention, for example, this embodiment continues the scenario of a multi-station hot mold casting production line for complex components, and the execution subject is the edge server on the production line site; the second working condition optimization sub-model has built-in dimensionality reduction component (focusing on disturbance-related features), self-attention component (strengthening key features in the time domain), and enhancement component (generating disturbance compensation weights); the input is the first response feature matrix (the time sequence features of real-time equipment pressure and mold temperature that have been decoupled from the system coupling) output by the first working condition optimization sub-model and the deviation feature matrix (the time sequence features of the deviation between the actual working condition and the reference working condition) in the initial working condition feature set.

[0171] Advanced processing steps for optimizing the sub-model under the second operating condition:

[0172] Superimposable projection processing of dimensionality reduction components:

[0173] The edge server invokes the dimensionality reduction component of the second operating condition tuning sub-model to perform joint dimensionality reduction and superimposed projection on the first response feature matrix and the deviation feature matrix:

[0174] The server uses the random disturbance physical characteristics of the casting process as constraints (such as high-frequency fluctuations in pressure characteristics corresponding to power grid fluctuations, and slow drifts in mold temperature corresponding to changes in ambient temperature). Through a dimensionality reduction algorithm adapted to the disturbance characteristics, it retains feature dimensions strongly correlated with random disturbances in both types of matrices. For example, it retains the high-frequency fluctuation characteristics of pressure in the first response feature matrix and the deviation characteristics corresponding to voltage disturbances in the deviation feature matrix. Simultaneously, through an overlay projection mechanism, the disturbance characteristics of both types of matrices are mapped to the same feature space, ensuring feature fusion. The final output is a dimensionality-reduced projected feature matrix, which has filtered out irrelevant features and retains only the effective information directly related to random disturbances.

[0175] Temporal fusion processing of self-attention components:

[0176] The edge server invokes the self-attention component to perform temporal self-attention fusion on the dimensionality-reduced projected feature matrix:

[0177] The component strictly follows the continuous sampling node sequence of the casting process (covering the entire stages of preheating, forming, and holding pressure). It calculates the attention weight of features at each sampling node; for example, it assigns higher weights to perturbation features in the forming stage (a stage sensitive to casting quality) and lower weights to features in the preheating stage (a stage with less perturbation impact). Through weighted integration, it strengthens the correlation of perturbation features in key stages in the time domain and weakens redundant information in non-critical stages. The final output is from the attention-optimized feature set, which highlights the perturbation features with the greatest impact on product quality throughout the entire casting process.

[0178] Enhanced component generation condition tuning and enhanced weight parameters:

[0179] The edge server inputs the self-attention optimized feature set into the enhancement component to generate perturbation compensation weights:

[0180] The enhancement component uses a neural network trained with casting process disturbance compensation rules to perform dimensionality transformation and weight learning on the optimized feature set. For example, it learns higher compensation weights for pressure deviation features caused by power grid fluctuations and lower compensation weights for features of slow ambient temperature drift. The final output is an enhanced weight parameter for operating condition tuning, whose dimension perfectly matches the deviation feature matrix, with each element corresponding to the disturbance compensation weight of a sampling node in the deviation feature matrix.

[0181] Weighted optimization to generate the second response feature matrix:

[0182] The edge server optimizes and enhances weight parameters based on operating conditions, and performs element-wise weighted modulation on the initial deviation feature matrix:

[0183] The server maps the weight parameters to the deviation feature matrix one-to-one according to the sampling nodes, and uses weighted operations to remove the influence of random disturbances. For example, deviation features caused by power grid fluctuations are compensated and corrected with high weights, while deviation features caused by ambient temperature drift are fine-tuned with low weights. Finally, a second response feature matrix is ​​output, which has completely removed random disturbance interference and retains only process-controllable deviation features, providing pure deviation information for subsequent feature integration.

[0184] In this embodiment, the edge server achieves random disturbance removal of deviation features through the process of "dimensionality reduction and focusing → temporal domain enhancement → weight generation → weighted correction" of the second working condition optimization sub-model, which effectively improves the stability of casting process control.

[0185] In this embodiment of the invention, the operating condition optimization enhancement weight parameters include a first self-attention compensation weight vector and a second self-attention compensation weight vector. The first self-attention compensation weight vector is a weight vector after operating condition optimization processing of the deviation feature matrix, and the second self-attention compensation weight vector is a weight vector after operating condition optimization processing of the first response feature matrix. This embodiment of the invention also provides the following implementation methods.

[0186] The first self-attention compensation weight vector and the deviation feature matrix are weighted and modulated to obtain the second response feature matrix;

[0187] The second self-attention compensation weight vector and the first response feature matrix are weighted and modulated to obtain the third response feature matrix.

[0188] In this embodiment of the invention, for example, this embodiment continues the scenario of a multi-station hot mold casting production line for complex components, and the execution subject is the edge server on the production line site; the enhancement component of the second working condition optimization sub-model generates two types of self-attention compensation weight vectors based on the random disturbance characteristics of the casting process and the equipment state requirements: the first self-attention compensation weight vector (corresponding to the disturbance compensation of the deviation feature matrix) and the second self-attention compensation weight vector (corresponding to the state enhancement of the first response feature matrix); the input is the first response feature matrix (equipment pressure and temperature time sequence features that have been decoupled from the system coupling) output by the first working condition optimization sub-model and the deviation feature matrix (deviation time sequence features between the actual and reference working conditions) in the initial working condition feature set.

[0189] Weight modulation and response matrix generation process:

[0190] The enhancement component generates two types of self-attention compensation weight vectors:

[0191] The edge server invokes the enhancement component of the second operating condition tuning sub-model to generate two types of weight vectors based on the self-attention optimization operating condition feature set (which has already focused on key time-domain perturbation features):

[0192] First self-attention compensation weight vector: The component assigns compensation weights to each sampling node of the deviation feature matrix based on the degree of influence of random disturbances on casting quality. For example, high weights are assigned to deviation features in the casting stage (a stage sensitive to product precision) to strengthen disturbance compensation in this stage, and low weights are assigned to deviation features in the preheating stage (a stage with less disturbance influence). The dimension of the weight vector is consistent with the total number of sampling nodes in the deviation feature matrix to ensure node-by-node matching.

[0193] The second self-attention compensation weight vector: Based on the contribution of the equipment state to the process stability, the component assigns a strengthening weight to each sampling node of the first response feature matrix. For example, it assigns high weights to the pressure and temperature features in the pressure holding stage (where the equipment state needs to be stable) (to strengthen the stability of the state in this stage), and low weights to the features in the no-load stage (where equipment state fluctuations have no impact). The dimension of the weight vector is consistent with the total number of sampling nodes in the first response feature matrix to ensure node-by-node matching.

[0194] Both types of weight vectors are generated through neural network learning of the enhancement component, strictly following the physical logic and quality requirements of the casting process.

[0195] Weighted modulation of the first weight vector and the bias feature matrix:

[0196] The edge server performs sample-by-sample node weighted modulation of the first self-attention compensation weight vector and the bias feature matrix:

[0197] The server uses the sampling node time sequence as a benchmark, multiplying each element of the weight vector element-wise with all feature dimensions of the corresponding node in the deviation feature matrix. For example, if the weight of a node in the molding stage is 0.8, the pressure deviation and temperature deviation features in the deviation feature matrix of that node are multiplied by 0.8 to enhance the disturbance compensation effect at that stage. If the weight of a node in the preheating stage is 0.2, the corresponding deviation feature is multiplied by 0.2 to weaken the influence of irrelevant disturbances. Finally, a second response feature matrix is ​​output, which has completely removed random disturbance interference and only retains process-controllable deviation features, providing an accurate basis for subsequent quality compensation.

[0198] Weighted modulation of the second weight vector and the first response feature matrix:

[0199] The edge server performs sample-by-sample node weighted modulation of the second self-attention compensation weight vector and the first response feature matrix:

[0200] The server uses the sampling node time sequence as a basis, multiplying each element of the weight vector element-wise with all feature dimensions of the corresponding node in the first response feature matrix. For example, if the weight of a node in the pressure holding stage is 0.9, the pressure and temperature features in the first response feature matrix of that node are multiplied by 0.9 to enhance the stability of the state in this stage. If the weight of a node in the no-load stage is 0.1, the corresponding feature is multiplied by 0.1 to weaken irrelevant state fluctuations. Finally, a third response feature matrix is ​​output, which highlights the stable state characteristics of the key stages of the equipment, eliminates redundant information in non-critical stages, and provides core state basis for subsequent process control.

[0201] In this embodiment, the edge server achieves the perturbation removal of deviation features and the key enhancement of equipment status by precisely modulating two types of self-attention compensation weight vectors, providing high-purity and highly correlated basic features for the subsequent integration of the set of optimized operating conditions features.

[0202] In this embodiment of the invention, the following implementation methods are also provided.

[0203] The first response feature matrix is ​​reduced in dimensionality to obtain a first-scale compressed feature matrix, and the deviation feature matrix is ​​reduced in dimensionality to obtain a second-scale compressed feature matrix.

[0204] The first-scale compressed feature matrix and the second-scale compressed feature matrix are integrated to obtain the second integrated feature matrix.

[0205] The second integrated feature matrix is ​​subjected to dimension reduction and superimposed projection using the dimension reduction component to obtain a dimension reduction projected feature matrix.

[0206] In this embodiment of the invention, for example, this embodiment continues the application scenario of a multi-station hot mold casting production line for complex components. The execution subject is the edge server on the production line. The input is the first response feature matrix (the core state timing features of casting equipment such as pressure and temperature that have been decoupled from the system coupling) output by the first working condition optimization sub-model and the deviation feature matrix (the timing features of the deviation between the actual working condition and the reference working condition) in the initial working condition feature set. The dimensionality reduction component built into the second working condition optimization sub-model needs to complete the feature projection based on the correlation of process features. The whole process strictly follows the timing logic and disturbance physical characteristics of the casting process.

[0207] The edge server first performs dimensionality reduction calculations on the first response feature matrix and the deviation feature matrix respectively: For the first response feature matrix, the server adopts a dimensionality reduction algorithm adapted to the casting process state features. Based on the correlation between features and process stability, it retains the key feature dimensions for evaluating equipment operating status (such as the pressure time-series trend dimension reflecting casting molding accuracy and the temperature change dimension affecting mold life) and eliminates redundant feature dimensions to obtain the first-scale compressed feature matrix; For the deviation feature matrix, the server adopts a dimensionality reduction algorithm adapted to random disturbance features. Based on the correlation between deviation and disturbance physical characteristics, it retains the deviation dimensions key to disturbance compensation (such as the pressure deviation dimension reflecting power grid fluctuations and the temperature deviation dimension reflecting ambient temperature drift) and eliminates irrelevant deviation dimensions to obtain the second-scale compressed feature matrix.

[0208] Subsequently, the edge server performs integration processing on the first-scale compressed feature matrix and the second-scale compressed feature matrix: the server uses the continuous sampling node sequence of the casting process as a benchmark to strictly align the temporal nodes of the two types of matrices to ensure that the state features and deviation features of the same sampling node correspond one-to-one; then, the key feature dimensions of the two types of matrices are integrated by column splicing to form the second integrated feature matrix. This matrix contains both the compressed features of the core state of the equipment and the compressed features of the key deviations, and the two types of features have achieved temporal matching and dimension adaptation, providing a foundation for subsequent dimensionality reduction projection.

[0209] Finally, the edge server calls the dimensionality reduction component of the second working condition optimization sub-model to perform dimensionality reduction and superposition projection on the second integrated feature matrix: Based on the feature fusion requirements of the casting process, the dimensionality reduction component maps the second integrated feature matrix to a preset low-dimensional feature space, ensuring that the state features and deviation features are superpositionable and fusionable in this space; During the projection process, the component strictly retains the key correlation information of the two types of features (such as the correspondence between the fluctuation trend of the state features and the perturbation law of the deviation features), and finally outputs the dimensionality reduction projection feature matrix. This matrix has filtered out all redundant information, retaining only the effective features directly related to random perturbation compensation, and the feature dimension is completely matched with the input requirements of the subsequent self-attention component.

[0210] This embodiment effectively improves the efficiency and accuracy of feature processing by first reducing and compressing dimensions, then integrating temporally, and finally projecting and fusing. This provides a highly relevant and low-redundancy input foundation for the self-attention fusion of the sub-model in the second working condition.

[0211] In this embodiment of the invention, the step of integrating the first response feature matrix and the second response feature matrix to obtain an optimized operating condition feature set can be implemented through the following example.

[0212] By integrating the first and second response feature matrices, an optimized operating condition feature set is obtained; or,

[0213] The first response feature matrix, the second response feature matrix, and the third response feature matrix are integrated to obtain the optimized operating condition feature set.

[0214] In this embodiment of the invention, for example, this embodiment continues the application scenario of a multi-station hot mold casting production line for complex components. The execution subject is the edge server on the production line. The input is the first response feature matrix (the timing features of the core equipment state after system coupling has been removed), the second response feature matrix (the timing features of the deviation after random disturbances have been removed), and the third response feature matrix (the timing features of the key stage state have been strengthened) output by the second working condition optimization sub-model. The feature integration process strictly follows the timing logic and control requirements of the casting process and is divided into two implementation paths.

[0215] Path 1: Integration of first and second response feature matrices:

[0216] When the production line is under low-disturbance conditions (such as stable power grid and small ambient temperature fluctuations), the edge server performs integrated processing on the first response feature matrix and the second response feature matrix:

[0217] The server uses the continuous sampling node sequence of the casting process as a benchmark to strictly align the time-series nodes of the two types of matrices, ensuring that the equipment status characteristics and deviation characteristics of the same sampling node correspond one-to-one. Then, the feature dimensions of the two types of matrices are concatenated column by column to form a temporary integrated matrix. Then, through the process feature correlation filtering algorithm, the duplicate or low-contribution feature dimensions in the temporary matrix are removed (such as redundant dimensions that also reflect the equipment pressure status), and finally the optimized operating condition feature set is obtained. This set contains both the core equipment status information and accurate deviation compensation information, which can be directly used for subsequent control command generation.

[0218] Path Two: Integration of First, Second, and Third Response Feature Matrices

[0219] When the production line is under high-disturbance conditions (such as frequent power grid fluctuations or significant ambient temperature drift), the edge server adds a third response feature matrix to path one to perform integrated processing:

[0220] The server first aligns the timing nodes of the third response feature matrix with the first and second response feature matrices to ensure a complete match of the sampling nodes of the three types of matrices. Then, it concatenates the feature dimensions of the three types of matrices column by column to form a temporary integrated matrix that includes the core state of the equipment, precise deviation compensation, and enhanced state of key stages. Then, through a multi-layer feature filtering mechanism, it prioritizes retaining feature dimensions that contribute highly to process stability (such as pressure enhancement features in the holding stage and deviation compensation features in the forming stage) and eliminates redundant dimensions. Finally, it obtains an optimized set of operating condition features, which enhances the equipment state features of key processes based on path one and is more suitable for the control requirements of high-disturbance operating conditions.

[0221] Both integrated paths' feature sets meet the dimensional requirements for subsequent control command generation and retain the core feature associations of the casting process, providing a clean and efficient input foundation for achieving high-precision closed-loop control.

[0222] In this embodiment of the invention, the casting task execution process is a process of process coordination between the first control unit and the second control unit; the following implementation methods are also provided.

[0223] Acquire the second state data generated by the second control unit during the casting task execution process;

[0224] Based on the dynamic decoupling mechanism, the superimposed system coupling is decoupled from the second state data to obtain the superimposed coupled response quantity corresponding to the second state data.

[0225] In this embodiment of the invention, for example, this embodiment continues the scenario of a multi-station hot mold casting production line for complex components, and the execution subject is the edge server on the production line. The casting task execution process is a process collaboration between the first control unit (on-site casting equipment PLC, responsible for executing side process actions) and the second control unit (remote process collaboration server, responsible for issuing reference process parameters and cross-station collaboration instructions). The second control unit issues synchronization instructions and reference process parameters according to the global working conditions of the production line, and the first control unit executes actions and feeds back the status. The two achieve real-time collaboration through industrial Ethernet.

[0226] The edge server first establishes a real-time communication link with the second control unit via the Industrial Ethernet OPC UA protocol. During the casting task execution process (from the start of the preheating station to the completion of pressure holding at the current casting station), it continuously acquires the second state data generated by the second control unit. This data is dynamic reference data generated by the second control unit based on global process collaboration logic, specifically including: ① reference pressure curve timing data for the current casting station (pressure setpoints covering the entire molding and pressure holding stages); ② reference mold temperature curve timing data (temperature setpoints coupled to the pressure curve); ③ process parameters corresponding to cross-station collaborative synchronization commands (such as the pressure compensation parameters of the current station associated with the preheating station temperature compliance signal). The server performs timing alignment and format conversion on the acquired second state data to ensure that the data dimensions completely match the state data of the first control unit.

[0227] Subsequently, the edge server invokes a preset dynamic decoupling mechanism (built based on a thermo-mechanical coupling physical model of the casting process) to perform superimposed system coupling decoupling on the second-state data. The specific process is as follows: ① Extract coupling features from the second-state data; the reference pressure curve and reference temperature curve exhibit process coupling (the process requires the pressure rise rate and temperature rise rate to satisfy a fixed coupling coefficient); ② Determine the coupling coefficient matrix based on the thermo-mechanical coupling model (verified through historical process data and physical simulation); ③ Using the coupling coefficient matrix as a constraint, linearly decouple the coupling data of the reference pressure and reference temperature to obtain a pure pressure superimposed coupling response quantity (pressure setpoint timing excluding the influence of temperature coupling) and a pure temperature superimposed coupling response quantity (temperature setpoint timing excluding the influence of pressure coupling); ④ Independently decouple the cross-station collaborative parameters to obtain a collaborative command superimposed coupling response quantity (synchronization compensation parameters excluding inter-station interference). Finally, the superimposed coupling response quantity corresponding to the second-state data is a set of the above three types of independent response quantities. Each response quantity retains the timing characteristics and process significance of the original data, and there is no coupling relationship between them. It can be directly superimposed with the actual state data of the first control unit to calculate the deviation.

[0228] This process decouples the dynamic reference data of the second control unit, providing a clean reference for the subsequent generation of the deviation feature matrix and ensuring the accuracy of deviation calculation during process collaboration.

[0229] In this embodiment of the invention, the initial set of operating condition features to be processed is determined based on the multi-source sensor data stream, which can be implemented through the following example.

[0230] The multi-source sensor data stream is subjected to time-series feature extraction based on the wavelet packet decomposition algorithm to obtain the main state feature matrix; and the second state data is subjected to time-series feature extraction based on the wavelet packet decomposition algorithm to obtain the reference coupling feature matrix.

[0231] The decoupling compensation calculation is performed on the multi-source sensor data stream and the superimposed coupled response quantity to obtain the decoupling deviation quantity. Then, the time-series features of the decoupling deviation quantity are extracted based on the wavelet packet decomposition algorithm to obtain the deviation feature matrix.

[0232] In this embodiment of the invention, for example, this embodiment continues the scenario of a multi-station hot mold casting production line for complex components. The execution subject is the edge server on the production line. The input is the multi-source sensor data stream collected by the first control unit (on-site casting equipment PLC), the second state data generated by the second control unit (remote process collaboration server), and the superimposed coupled response quantity obtained by the dynamic decoupling mechanism. The generation of the initial working condition feature set strictly follows the temporal logic and feature extraction specifications of the casting process, and the entire process relies on the wavelet packet decomposition algorithm to achieve accurate extraction of multi-dimensional temporal features.

[0233] Wavelet packet decomposition extracts the main state and reference coupling feature matrix:

[0234] The edge server first performs time-series feature extraction on the multi-source sensor data stream: this data stream is real-time process data collected by the first control unit, covering continuous sampling time-series data of hydraulic system pressure, mold cavity temperature, and slider displacement during the casting stage. The sampling nodes cover the entire process from preheating to pressure holding. The server calls a preset wavelet packet decomposition algorithm, using the "db4" wavelet basis as the decomposition basis function (adapted to the non-stationary time-series features of the casting process), and sets the decomposition level to 3 levels (verified by historical process data, which can effectively separate the high-frequency impact features of pressure fluctuations from the low-frequency thermal equilibrium features of temperature changes). The sensor data of each sampling node is decomposed at multiple scales, and the time-domain energy features and frequency-domain entropy features of each decomposition frequency band are extracted (energy features reflect the fluctuation intensity of process parameters, and entropy features reflect the stability of parameters). The feature vectors of all nodes are integrated to form a master state feature matrix. The row vectors of this matrix correspond to the sampling nodes, and the column vectors correspond to the energy and entropy features at different decomposition scales, thus completely preserving the real-time operating state information of the equipment.

[0235] Subsequently, the server performs the same wavelet packet decomposition on the second-state data: the second-state data is dynamic reference data issued by the second control unit, containing reference pressure curves and reference temperature curves that time-series match the main state feature matrix. Using the same wavelet basis and decomposition level, the server extracts the time-domain energy and frequency-domain entropy features of the reference data to form a reference coupling feature matrix. The dimension of this matrix is ​​completely consistent with the main state feature matrix, ensuring dimensionality adaptation for subsequent deviation calculations.

[0236] Decoupling compensation calculation and deviation feature matrix extraction:

[0237] The edge server aligns the multi-source sensor data streams with the superimposed coupled response quantities (pure pressure and pure temperature independent response quantities obtained through dynamic decoupling) point by point according to the sampling nodes, and performs decoupling compensation calculations: for example, the actual hydraulic pressure of a sampling node is subtracted from the pure pressure superimposed coupled response quantity corresponding to that node to obtain the pressure decoupling deviation; the actual mold temperature is subtracted from the pure temperature superimposed coupled response quantity to obtain the temperature decoupling deviation; the pressure and temperature decoupling deviations of all nodes are integrated into a decoupling deviation quantity, which has been stripped of the system coupling interference of the second state data and only reflects the deviation between the actual working conditions and the independent reference benchmark.

[0238] Finally, the server performs wavelet packet decomposition on the decoupling deviation: using the same decomposition parameters as the master state feature matrix, it extracts the high-frequency fluctuation features (reflecting the instantaneous deviation caused by random disturbances) and low-frequency drift features (reflecting the cumulative deviation caused by slow disturbances) of the deviation data, and integrates them to form a deviation feature matrix. The row vectors of this matrix correspond to the sampling nodes, and the column vectors correspond to the decomposition scale features of the deviation data, thus completely preserving the operating condition deviation information caused by disturbances.

[0239] Output of the initial working condition feature set:

[0240] The edge server concatenates the master state feature matrix, reference coupling feature matrix, and deviation feature matrix according to feature dimensions to form an initial set of operating condition features to be processed. This set contains three core features: real-time equipment status, remote reference benchmark, and operating condition deviation, providing a complete input foundation for the subsequent hierarchical decoupling control model.

[0241] The entire process relies on the multi-scale feature extraction capability of wavelet packet decomposition to ensure the integrity and accuracy of the initial working condition features, and is fully adapted to the non-stationary time sequence feature processing requirements of complex casting processes.

[0242] In this embodiment of the invention, the casting task execution process is a closed loop of process execution in a real-time operating environment. The real-time operating environment includes any one of the following: multi-machine collaborative forging mode, continuous flow production mode, and single-machine independent control mode. This embodiment of the invention also provides the following implementation methods.

[0243] In a real-time operating environment, online operating condition sensor data streams are acquired, and the online operating condition sensor data streams are segmented in the time dimension according to a preset process window to obtain a set of process data segments; the set of process data segments includes multiple process data segments.

[0244] The target process data segment among the multiple process data segments is taken as a multi-source sensor data stream, and the optimized operating condition segment of the target process data segment is obtained based on the hierarchical decoupling control model.

[0245] The multiple optimized operating condition segments obtained are integrated and processed to obtain the optimized control operating condition data corresponding to the online operating condition sensing data stream.

[0246] In this embodiment of the invention, for example, this embodiment continues the scenario of a multi-station hot mold casting production line for complex components. The execution subject is the edge server on the production line site, and the real-time operating environment is a continuous flow production mode (covering four continuous stations: preheating, forming, holding pressure, and cooling, with seamless transfer of billets between stations achieved through robotic arms). The entire process strictly follows the timing logic and process window constraints of continuous flow production, and the operating conditions of key stations are optimized by relying on a hierarchical decoupling control model.

[0247] In continuous production mode, the edge server first acquires online working condition sensor data streams in real time via industrial Ethernet. This data stream consists of continuous sampling data from the core sensors at each station, including the surface temperature of the billet at the preheating station, the hydraulic system pressure / mold cavity temperature / slider displacement at the forming station, the holding pressure / mold temperature at the holding station, and the cooling water temperature at the cooling station. The sampling nodes cover the entire production process from the billet entering the preheating station to the finished product output, and the data stream format is strictly aligned with the process sequence. Subsequently, the server segments the online operating condition sensing data stream in the time sequence dimension according to the preset process windows (based on the process stages of casting, verified by the process knowledge base): the preset process windows correspond to the core process stages of each station, the preheating window is "from the billet entering the preheating station to the temperature reaching the target", the forming window is "from the slider contacting the billet to the completion of forming", the pressure holding window is "from the start of pressure holding to the end of pressure holding", and the cooling window is "from the billet entering the cooling station to the temperature dropping to the threshold". The server splits the data stream according to the window time sequence boundaries to obtain a set of process data segments, which includes four independent process data segments: preheating segment, forming segment, pressure holding segment, and cooling segment. Each data segment corresponds to the core process data of a station.

[0248] The server filters target process data segments based on process importance: the forming and holding pressure segments, which have the greatest impact on forging quality, are selected as target segments and input as multi-source sensor data streams into the hierarchical decoupling control model (i.e., the previous first and second working condition optimization sub-models); the model processes the target segment data according to the process of "system coupling decoupling → random disturbance removal", decoupling system-level coupling (such as pressure-temperature process coupling) of hydraulic pressure and mold temperature data in the forming segment, and removing random pressure deviations caused by power grid fluctuations; the holding pressure data in the holding pressure segment is used to enhance the state stability of key stages, and finally outputs the forming optimization working condition segment and the holding pressure optimization working condition segment.

[0249] Finally, the server performs integration processing on all optimized segments of the process data segment set: constrained by the timing requirements of continuous flow production (such as the slider displacement at the end of the forming segment must seamlessly match the displacement at the beginning of the holding segment), the preheating segment (basic parameter adjustment), forming optimization segment, holding optimization segment, and cooling segment (basic parameter adjustment) are spliced ​​together according to the production sequence to ensure the continuity and matching of process parameters in each segment; after integration, the optimized control condition data corresponding to the online working condition sensor data stream is obtained, and the server sends it to each station controller through industrial Ethernet to realize closed-loop optimized control of continuous flow production.

[0250] This embodiment effectively adapts to the dynamic and stability requirements of continuous flow production by segmenting key processes through process windows, optimizing core parameters through hierarchical models, and ensuring process continuity through time-series integration.

[0251] To more clearly describe the solutions provided in the embodiments of the present invention, a more specific implementation method is provided below.

[0252] This embodiment targets a multi-station hot mold casting production line for automotive steering knuckles (a critical safety component requiring a molding accuracy of ±0.1mm and internal microstructure uniformity ≥95%). The execution unit is an edge server with 40 TOPS computing power (equipped with an industrial-grade operating system, supporting 10ms-level real-time inference). The core configuration and parameters of the production line are as follows: the hot mold casting machine has a nominal force of 10000kN, a slide stroke of 500mm, a maximum speed of 100mm / s, and is installed at the molding station and connected via an industrial Ethernet OP. The system uses CUAv1.0a protocol communication; the pressure sensor is an HBMC9C type, with a range of 0-12000kN, accuracy of ±0.05%FS, and a sampling frequency of 1kHz, installed at the main hydraulic cylinder outlet; the mold temperature sensor is an Optris PI450 type, infrared, with a range of -20-1500℃, accuracy of ±1℃, and a sampling frequency of 50Hz, installed in the steering knuckle forming area on the inner wall of the mold cavity; the slider displacement sensor is a Heidenhain LC183 type, grating ruler type, with a range of 0-500mm, resolution of 0.1μm, and a sampling frequency of 200Hz, installed on the slider... The side of the block guide rail; the mains voltage sensor is a Schneider PM5350, with a range of 0-480V, an accuracy of ±0.1%, and a sampling frequency of 50Hz, installed at the entrance of the casting machine power supply cabinet; the field controller is a 1516-3PN / DP type, with a control cycle of 1ms, installed in the casting machine local control cabinet; the remote process server is equipped with two 6248R CPUs, 128GB of memory, and stores a 10-year steering knuckle process knowledge base (including material properties, forging temperature range, pressure curve, mold life curve, etc.), located in the central control room of the workshop and connected via 1Gbps fiber optic communication.

[0253] The edge server continuously collects multi-source sensor data streams at a sampling frequency of 100Hz via the OP CUA protocol. All data undergoes preprocessing through a 5th-order Butterworth low-pass filter (cutoff frequency 20Hz, removing high-frequency sensor noise): one is the execution-side first-state data, covering the main hydraulic cylinder outlet hydraulic system pressure (0-10000kN, 3000kN at t=0s, peak at 8000kN at t=5s, and pressure held down to 7500kN at t=10s), the temperature of the steering knuckle forming area on the inner wall of the mold cavity (800-1200℃, 900℃ at t=0s, 1150℃ at t=5s, and pressure reduced to 1100℃ at t=10s), and the side displacement of the slider guide rail (0-500mm, 500mm top dead center at t=0s, 300mm displacement upon contact with the blank at t=3s, and 200mm displacement upon completion of forming at t=5s), with a dimension of 1000×3 (1000 sampling nodes covering the start-up to pressure holding). The data consists of three parts: 1) Completed full-process stage; 2) Execution-side disturbance background data, including the grid voltage at the casting machine power supply cabinet inlet (360-400V, 370V fluctuation at t=2s, fluctuation amount -2.6%), and the cumulative wear of the mold based on the pressure-displacement curve fitting (0-500μm, 320μm at t=10s, mold life remaining 40%), with dimensions 1000×2; 3) Remote reference data, generated by the remote process server based on the steering knuckle process knowledge base, including the reference pressure curve (3000kN→8000kN→7500kN, fitting error <±1%), the reference temperature curve (900℃→1150℃→1100℃, satisfying the thermo-mechanical coupling relationship with the pressure curve), and the cross-station pressure compensation coefficient 1.02. The server interpolates the remote reference data to 100Hz and converts it to float32 format to ensure strict matching with the execution-side data dimensions.

[0254] Based on multi-source data streams, the edge server generates an initial set of operating condition features: First, a wavelet packet algorithm with a db4 wavelet basis and 3-level decomposition (3-level decomposition can cover high-frequency pressure impact 0-50Hz and low-frequency temperature trend 0-10Hz) is used to extract temporal features from the first state data on the execution side. Energy (reflecting fluctuation intensity) and entropy (reflecting stability) of each decomposition frequency band are extracted for pressure, temperature, and displacement, forming a 1000×18 main state feature matrix (3 types of physical quantities × 3-level decomposition × 2 types of features). Second, the same wavelet packet decomposition is performed on the reference pressure and reference temperature of the remote reference data, forming a 1000×12 reference coupling feature matrix. Finally, a dynamic decoupling mechanism is constructed based on the thermo-mechanical coupling physical model. The coupling coefficient matrix K=[[1,0.8],[0.8,1]] of pressure and temperature was obtained by fitting 100 sets of historical data. The reference pressure curve and the reference temperature curve were linearly decoupled to obtain the superimposed response of pure pressure (7080kN at t=5s) and the superimposed response of pure temperature (1086℃ at t=5s). The decoupling deviation was obtained by subtracting the first state data of the execution side from the superimposed coupling response at each node (pressure deviation 720kN, temperature deviation 44℃). The decoupling deviation was extracted using db4 wavelet basis and 4-level decomposition to extract high-frequency fluctuation (power grid interference) and low-frequency drift (mold wear) features, forming a 1000×8 deviation feature matrix. The three types of matrices together constitute the initial working condition feature set.

[0255] The edge server invokes a hierarchical decoupling control model to perform segmented optimization on the initial operating condition feature set: The first operating condition optimization sub-model includes a multi-head attention component, a feature embedding component, a coupling decoupling component, and a feature reconstruction component. First, the main state feature matrix, reference coupling feature matrix, and deviation feature matrix are reduced in dimensionality using PCA (retaining 95% variance), resulting in a 1000×12 main state dimensionality reduction matrix, a 1000×8 reference coupling dimensionality reduction matrix, and a 1000×5 deviation dimensionality reduction matrix. After aligning the timing with the sampling nodes of the main state as the reference, they are concatenated column-wise to form the first integrated feature matrix (1000×25). Subsequently, the first integrated feature matrix is ​​correlated and fused through an 8-head, 256-dimensional embedded multi-head attention component, using scaled dot product attention calculation (…). ),in Q is the query matrix, K is the key matrix, and V is the value matrix. The dimension of the Query / Key vector. The dot product similarity matrix is... The attention weights for the forming section (t=3-8s) are multiplied by 1.2 to strengthen them, and the weights for the preheating section (t=0-3s) are multiplied by 0.8 to weaken them, outputting a 1000×256 attention fusion feature matrix. A linear transformation layer maps the fusion matrix to a 256-dimensional embedding space (initialized by the process knowledge base). The embedding matrix is ​​split into 1000 sub-matrices per time step. A short-time Fourier transform (window length 256, overlap rate 50%, FFT points 512) is performed on each sub-matrix to obtain the frequency domain components. A Gaussian kernel similarity mechanism with σ=0.5 is used to calculate the similarity of the frequency domain components, decoupling the "pressure-temperature" system coupling. Output the local matrix features of 1000×256, then calculate the temporal Gaussian kernel similarity of adjacent submatrices (the similarity threshold between nodes t and t+1 is 0.6), and integrate them to obtain a kernel similarity feature matrix of 1000×256; perform dimension transformation on the kernel similarity features through an MLP structure (256→128→18) to generate a 1000×18 control weight tensor (the pressure feature weight of the forming section is 0.9, and that of the preheating section is 0.3), and perform element-wise weighted modulation on the master state feature matrix to obtain a 1000×18 first response feature matrix (the system coupling has been decoupled, and the pressure feature variance at t=5s decreases from 1200 to 300).

[0256] The second-condition optimization sub-model includes a dimensionality reduction component, a self-attention component, and an enhancement component. First, the first response feature matrix and the bias feature matrix are reduced to 20 dimensions using t-SNE (perplexity=30, iterations 1000, learning rate 200), and projected onto the same feature space (ensuring inner product > 0.8 through a superimposed projection matrix W=randn(26,20)), outputting a 1000×20 dimensionality-reduced projected feature matrix. Then, a 4-head self-attention component performs temporal fusion on the dimensionality-reduced projected feature matrix, employing sinusoidal and cosine position encoding (PE(pos,2i)=sin(pos / 10000)). (2i / d_model) ), PE(pos,2i+1)=cos(pos / 10000 (2i / d_model)(d_model=20), assign a weight of 0.8 to the forming section (t=3-8s), a weight of 0.9 to the pressure holding section (t=8-10s), and a weight of 0.2 to the preheating section, outputting a 1000×20 self-attention optimization feature set; perform dimension transformation on the optimization feature set through the MLP structure (20→128→2) to generate a 1000×8 first self-attention compensation weight vector (forming section weight 0.8, preheating section weight 0.2) and a 1000×18 second self-attention compensation weight vector (pressure holding section weight 0.9, no-load section weight 0.1), respectively perform element-wise weighted modulation on the deviation feature matrix and the first response feature matrix to obtain a 1000×8 second response feature matrix (random disturbances have been removed, and the pressure deviation caused by the power grid fluctuation at t=2s has been reduced from 150kN to 30kN) and a 1000×18 third response feature matrix (key processes have been strengthened, and the pressure feature variance of the pressure holding section has been reduced from 300 to 100).

[0257] The edge server performs feature integration on the response feature matrix: under low disturbance conditions (grid fluctuation < ±2%, ambient temperature fluctuation < ±1℃), it integrates the first and second response feature matrices, uses the ReliefF algorithm to retain the top 70% of the features with the highest contribution, and outputs a 1000×18 optimized feature set; under high disturbance conditions (grid fluctuation > ±5%, ambient temperature fluctuation > ±1℃), it integrates the first, second, and third response feature matrices, uses the ReliefF algorithm to retain the top 70% of the features with the highest contribution, and outputs a 1000×31 optimized feature set. Finally, an inverse transformation of the optimized operating condition feature set is performed using the db4 wavelet basis. Combined with the compensation strategy (for every 1% fluctuation in grid voltage, the pressure compensation coefficient is adjusted by 0.01; for every 100μm increase in mold wear, the pressure compensation coefficient is adjusted by 0.05), optimized control parameters (hydraulic pressure 7800kN, mold temperature 1130℃, slider displacement 200mm) are generated. The real-time operating condition is segmented according to the preset process windows (preheating window 0-3s, forming window 3-8s, pressure holding window 8-10s). The forming segment and the pressure holding segment are selected as the target process segments to ensure that the slider displacement error at the end of the forming segment and the displacement error at the beginning of the pressure holding segment do not exceed 0.1mm and the pressure error does not exceed ±50kN. The optimized control operating condition data of 1000×3 are integrated and sent to the field controller at a period of 10ms through the OPC UA protocol.

[0258] This invention provides a computer device 100, which includes a processor and a non-volatile memory storing computer instructions. When the computer instructions are executed by the processor, the computer device 100 executes the aforementioned reconfigurable casting process module hierarchical control method. For example... Figure 2 As shown, Figure 2This is a structural block diagram of a computer device 100 provided in an embodiment of the present invention. The computer device 100 includes a memory 111, a processor 112, and a communication unit 113. To enable data transmission or interaction, the memory 111, processor 112, and communication unit 113 are electrically connected to each other directly or indirectly. For example, these components can be electrically connected to each other through one or more communication buses or signal lines.

[0259] For illustrative purposes, the foregoing description has been made with reference to specific embodiments. However, the foregoing illustrative discussions are not intended to be exhaustive or to limit the present disclosure to the precise forms disclosed. Numerous modifications and variations are possible in accordance with the foregoing teachings. These embodiments were chosen and described in order to best illustrate the principles of the present disclosure and its practical application, thereby enabling those skilled in the art to best utilize the disclosure and to employ various embodiments with different modifications to suit a particular intended application.

Claims

1. A reconfigurable foundry process module hierarchical control method, characterized by, include: Acquire multi-source sensor data streams generated during the execution of a metal casting task; The metal casting task includes at least one or more of the following processes: pouring, solidification, and cooling; the multi-source sensor data stream includes process parameter data reflecting the state of the molten metal or casting, and the process parameter data includes one or more of the following: temperature, pressure, flow rate, displacement, and strain. Based on the multi-source sensor data stream, determine the initial set of operating condition features to be processed; The initial working condition feature set is used to characterize the dynamic working conditions during the casting process; Based on the hierarchical decoupling control model, the initial working condition feature set is subjected to segmented working condition optimization processing to obtain the optimized working condition feature set. The optimized working condition feature set is subjected to feature reconstruction processing to generate optimized control instructions that match the current casting process; the optimized control instructions are used to adjust one or more of the pouring speed, cooling rate, and pressure parameters to achieve closed-loop optimized control of the casting process; The step of performing segmented operating condition optimization processing on the initial operating condition feature set based on the hierarchical decoupling control model yields an optimized operating condition feature set, including: The multi-source sensor data stream generated during the casting task execution process is acquired. The multi-source sensor data stream includes: first state data generated on the execution side during the casting task execution process, disturbance background data where the execution side is located, and remote reference data. The remote reference data is generated based on the second state data generated on the control side. Based on the multi-source sensor data stream, an initial set of operating condition features to be processed is determined. The initial set of operating condition features includes a master state feature matrix, a reference coupling feature matrix, and a deviation feature matrix. The master state feature matrix is ​​obtained by extracting time-series features from the multi-source sensor data stream. The reference coupling feature matrix is ​​obtained by extracting time-series features from the second state data. The deviation feature matrix is ​​obtained by extracting time-series features from the decoupling deviation between the superimposed coupled response quantities of the multi-source sensor data stream and the second state data. Based on the hierarchical decoupling control model, the initial operating condition feature set is subjected to segmented operating condition optimization processing to obtain an optimized operating condition feature set; the operating condition optimization processing includes system coupling decoupling and random disturbance stripping. The hierarchical decoupling control model includes at least two operating condition optimization sub-models; any one of the operating condition optimization sub-models is used to: perform an operating condition optimization process on any one or more feature matrices among the master state feature matrix, reference coupling feature matrix, and deviation feature matrix; the segmented operating condition optimization process based on the hierarchical decoupling control model on the initial operating condition feature set to obtain an optimized operating condition feature set includes: Based on the target operating condition tuning sub-model among at least two operating condition tuning sub-models, operating condition tuning is performed on any one or more feature matrices among the master state feature matrix, reference coupling feature matrix, and deviation feature matrix to obtain candidate feature matrices; the target operating condition tuning sub-model is any one of the at least two operating condition tuning sub-models. Feature integration is performed on at least two candidate feature matrices to obtain an optimized working condition feature set; The hierarchical decoupling control model includes: a first operating condition optimization sub-model and a second operating condition optimization sub-model; the second operating condition optimization sub-model includes: a dimensionality reduction component, a self-attention component, and an enhancement component; The step of performing segmented operating condition optimization processing on the initial operating condition feature set based on the hierarchical decoupling control model yields an optimized operating condition feature set, including: Based on the master state feature matrix, the reference coupling feature matrix, and the deviation feature matrix, the master state feature matrix is ​​subjected to initial working condition optimization processing based on the first working condition optimization sub-model to obtain the first response feature matrix. The first response feature matrix and the deviation feature matrix are subjected to dimension reduction and superposition projection by the dimension reduction component to obtain the dimension reduction projection feature matrix. The self-attention component is used to perform self-attention fusion processing on the reduced-dimensional projection feature matrix according to the temporal direction to obtain a self-attention optimized working condition feature set. The self-attention optimization feature set is enhanced by an enhancement component to obtain the enhanced weight parameters for the optimization of the operating conditions. Based on the aforementioned operating condition optimization enhancement weight parameters, the deviation feature matrix is ​​subjected to advanced operating condition optimization processing to obtain the second response feature matrix. The first response feature matrix and the second response feature matrix are integrated to obtain the optimized operating condition feature set; The first operating condition optimization sub-model includes: a feature embedding component, a coupling decoupling component, and a feature reconstruction component; the first operating condition optimization is performed on the main state feature matrix based on the main state feature matrix, the reference coupling feature matrix, and the deviation feature matrix, to obtain the first response feature matrix, including: The feature embedding component performs feature embedding operations on the master state feature matrix, the reference coupling feature matrix, and the deviation feature matrix to obtain the embedded matrix features. Through the coupling-decoupling component, the kernel method is used to perform feature decoupling processing on the embedded matrix features in both the time and frequency domains to obtain kernel similarity features; The kernel similarity features are reconstructed using the feature reconstruction component to obtain the control weight tensor. Based on the aforementioned control weight tensor, the master state feature matrix is ​​subjected to initial working condition optimization processing to obtain the first response feature matrix.

2. The method of claim 1, wherein, The coupling and decoupling component includes an intra-process adaptation unit and an inter-process collaboration unit. The intra-process adaptation unit is used to perform intra-process state parsing of state data, and the inter-process collaboration unit is used to perform inter-process correlation mining of state data. The embedded matrix features are distributed in a two-dimensional structure: one dimension is the continuous sampling node sequence of multi-source sensors during the casting process, and the other dimension is the total number of workstation monitoring features to be collected under a single sampling node. The process involves using the coupling-decoupling component and a kernel method to decouple the embedded matrix features in both the time and frequency domains to obtain kernel similarity features, including: The embedded matrix features are segmented to obtain multiple sub-matrix features; the number of sub-matrix features is equal to the number of sampling nodes contained in the continuous sampling node sequence. Based on the intra-process adaptation unit, the target sub-matrix feature in the multiple sub-matrix features is decomposed and adapted in the frequency domain direction through the symmetric kernel similarity mechanism to obtain the local matrix feature corresponding to the target sub-matrix feature. Based on inter-process collaborative units, a symmetric kernel similarity mechanism is used to integrate the inter-process features of multiple local matrix features in the time domain to obtain kernel similarity features.

3. The method according to claim 1, characterized in that, The first operating condition optimization sub-model further includes a multi-head attention component; the method further includes: The main state feature matrix, the reference coupling feature matrix, and the deviation feature matrix are respectively subjected to feature matrix dimensionality reduction processing to obtain the first dimensionality-reduced feature matrix, the second dimensionality-reduced feature matrix, and the third dimensionality-reduced feature matrix. The first, second, and third dimensionality-reduced feature matrices are subjected to feature integration processing to obtain the first integrated feature matrix. The first integrated feature matrix is ​​subjected to multi-head attention fusion by the multi-head attention component to obtain an attention fusion feature matrix, and the attention fusion feature matrix is ​​subjected to feature embedding operation by the feature embedding component to obtain an embedding matrix feature.

4. The method according to claim 1, characterized in that, The operating condition optimization enhancement weight parameters include a first self-attention compensation weight vector and a second self-attention compensation weight vector. The first self-attention compensation weight vector is a weight vector after operating condition optimization processing of the deviation feature matrix, and the second self-attention compensation weight vector is a weight vector after operating condition optimization processing of the first response feature matrix. The method further includes: The first self-attention compensation weight vector and the deviation feature matrix are weighted and modulated to obtain the second response feature matrix; The second self-attention compensation weight vector and the first response feature matrix are weighted and modulated to obtain the third response feature matrix.

5. The method according to claim 4, characterized in that, The casting task execution process is a process of process coordination between the first control unit and the second control unit; the method further includes: Acquire the second state data generated by the second control unit during the casting task execution process; Based on the dynamic decoupling mechanism, the superimposed system coupling is decoupled from the second state data to obtain the superimposed coupled response quantity corresponding to the second state data.

6. The method according to claim 5, characterized in that, Based on the multi-source sensor data stream, the initial set of operating condition features to be processed is determined, including: Temporal features of the multi-source sensor data stream are extracted based on the wavelet packet decomposition algorithm to obtain the main state feature matrix; and temporal features of the second state data are extracted based on the wavelet packet decomposition algorithm to obtain the reference coupling feature matrix. The decoupling compensation calculation is performed on the multi-source sensor data stream and the superimposed coupled response quantity to obtain the decoupling deviation quantity. Then, the time-series features of the decoupling deviation quantity are extracted based on the wavelet packet decomposition algorithm to obtain the deviation feature matrix.

Images