Multi-physics field control method and system in aluminum product processing

Abstract

This application provides a multi-physics field control method and system in aluminum product processing, which involves the field of machine learning technology, including: collecting electromagnetic field parameters, electrostatic field parameters and process quality parameters in the aluminum product processing process, performing feature extraction, and constructing a physical field feature vector that characterizes the processing task state; constructing a meta-learning model based on the historical feature data of multiple processing tasks; during the processing process, the system responds to changes in working conditions, calls the meta-learning model and performs rapid fine-tuning operations, outputs prediction results such as physical field coupling trends, abnormal risk levels and adjustment parameter recommendations; and generates control recommendation information based on the prediction results. This application introduces sensor topology information, interference recognition mechanism and structured prediction output, which improves the accuracy of physical field state modeling, the model's ability to migrate between different tasks and the availability of prediction output, and is suitable for intelligent perception and dynamic optimization control of aluminum product processing.

Description

Technical Field

[0001] The present application relates to the field of machine learning technology, and more specifically, to a multi-physics field control method and system in aluminum product processing. Background Art

[0002] Aluminum processing often involves complex thermal, electrical, and magnetic fields. In particular, during smelting, casting, powder coating, and heat treatment, electromagnetic stirring, electrostatic spraying, and heat conduction often overlap, forming a highly coupled multi-physics intervention system. Existing technologies often control physical field interventions during aluminum processing using preset parameters or rule-based approaches. For example, these include stirring with a fixed-frequency electromagnetic field or using an electrostatic eliminator during electrostatic spraying to mitigate discharge risks. However, due to the large fluctuations in the processing environment, parameters such as electromagnetic interference, static charge accumulation, and powder adhesion often exhibit significant dynamics and uncertainty, making it difficult for static control strategies to balance product quality and operational stability. Furthermore, complex spatial interference relationships often exist between physical field parameters. For example, the time-varying characteristics of the electromagnetic field can induce nonlinear cross-interference on adjacent static detection devices, resulting in the inclusion of induced artifacts in the original sensor signal, disrupting the system's understanding of the actual operating conditions. Furthermore, due to differences in sensor placement, material properties, or process objectives across various aluminum processing tasks, existing modeling methods struggle to achieve rapid generalization or adaptive adjustment across different tasks.

[0003] In summary, existing technologies still face certain challenges in dealing with the interactive coupling of multiple physical fields, inductive interference identification, multi-task adaptation and real-time optimization in processing sites. There is an urgent need for a multi-physical field state modeling method with adaptive learning capabilities to more efficiently analyze, predict and assist in adjusting the complex working conditions in the aluminum product processing process. Summary of the Invention

[0004] In response to the deficiencies of the existing technology, this application provides a multi-physical field control method and system in aluminum product processing.

[0005] In a first aspect, the present application provides a multi-physics field control method in aluminum product processing, including:

[0006] Collect physical field data and process quality parameters associated with the processing task during the aluminum product processing, perform feature extraction, and obtain a physical field feature vector representing the processing task state; wherein the physical field data includes: electromagnetic field parameters and electrostatic field parameters;

[0007] Based on the historical physical field feature vectors of multiple machining tasks and their corresponding process control parameters and process quality information, a meta-learning model is constructed and trained.

[0008] In response to changes in the working conditions corresponding to the processing task during execution, the meta-learning model is called to perform a rapid fine-tuning operation based on the physical field feature vector under the current working conditions to generate a prediction result under the current working conditions; the prediction result is used to characterize the evolution trend of the physical field coupling behavior and its potential change risk during the aluminum product processing process;

[0009] Based on the prediction results, control suggestion information for the current processing task is generated.

[0010] As an optional implementation, the feature extraction to obtain the physical field feature vector representing the processing task state includes:

[0011] Performing a short-time Fourier transform on the electromagnetic field parameters to generate an energy spectrum reflecting the change of frequency over time; calculating the electromagnetic field dominant frequency, frequency drift amplitude, and high-frequency disturbance intensity based on the energy spectrum to generate an electromagnetic field frequency domain eigenvector;

[0012] The electrostatic field parameters are subjected to wavelet packet decomposition to generate a transient discharge response signal at multiple scales; characteristic points in the transient discharge response signal that meet an abnormality criterion are identified, and electrostatic features corresponding to the characteristic points that meet the abnormality criterion are eliminated to generate an electrostatic field denoising feature vector.

[0013] As an optional implementation, the feature extraction to obtain the physical field feature vector representing the processing task state further includes:

[0014] Merging the electromagnetic field frequency domain eigenvector and the electrostatic field denoising eigenvector to construct an initial physical field eigenvector set;

[0015] Based on the sensor layout topology diagram and the signal sampling time window, a correlation analysis is performed on the initial physical field feature vector set to identify interference feature components whose correlation with the induced interference in the topological connection path exceeds a preset threshold. The interference feature components are then removed from the initial physical field feature vector set to generate an effective physical field feature vector after interference removal.

[0016] As an optional implementation, the feature extraction to obtain the physical field feature vector representing the processing task state further includes:

[0017] Performing statistical analysis on the process quality parameters to extract process quality feature vectors including mean, standard deviation and change rate;

[0018] The process quality feature vector is concatenated with the physical field effective feature vector to generate a physical field feature vector.

[0019] As an optional implementation, building and training a meta-learning model based on historical physical field feature vectors of multiple processing tasks and their corresponding process control parameters and process quality information includes:

[0020] Encode the sensor layout topology diagram corresponding to each processing task and generate a topology vector that represents the difference in sensor layout for different tasks;

[0021] The topological structure vector and the physical field feature vector of the corresponding processing task are jointly input into the meta-learning model for task embedding representation;

[0022] A meta-learning algorithm with structural embedding capability is used to jointly train multiple processing tasks.

[0023] As an optional implementation manner, generating the effective eigenvector of the physical field after interference removal includes:

[0024] The sensor arrangement topology structure diagram is represented as an adjacency matrix of sensor nodes, and cross-correlation coefficients are calculated for physical field characteristic components of different rows within a signal sampling time window;

[0025] Based on the product of the cross-correlation coefficient and the path connection weight in the adjacency matrix, an interference correlation index is constructed;

[0026] Determining whether the interference correlation index exceeds a set threshold, and marking the characteristic component whose interference correlation index exceeds the set threshold as an interference characteristic component;

[0027] The interference characteristic components are removed from the initial physical field characteristic vector set to form an effective physical field characteristic vector after interference removal for training and prediction.

[0028] As an optional implementation manner, generating a prediction result under current operating conditions includes:

[0029] Based on the joint input consisting of the physical field feature vector under the current working conditions and the topological structure vector corresponding to the processing task, the gradient information relative to the meta-learning model parameters is calculated, and the parameters are fine-tuned within a preset number of iterations;

[0030] The fine-tuned meta-learning model is used for predictive reasoning of the physical field feature vector of the current processing task, and the prediction index value associated with the processing task is output. The prediction index value includes: physical field coupling strength score, potential anomaly probability, electrostatic discharge risk level and recommended physical field adjustment parameters.

[0031] As an optional implementation, the output format of the prediction index value includes:

[0032] Outputting prediction results of multiple dimensions in the form of structured data, wherein the format of the structured data includes a key-value pair format, a vector structure, or a hierarchical risk level representation;

[0033] The physical field coupling strength score is represented by a floating point interval; the electrostatic discharge risk level is represented by a multi-level classification label; the probability of potential abnormality occurrence is in the form of a percentage; and the recommended physical field adjustment parameters are represented in the form of a recommended value range or a threshold pair.

[0034] As an optional implementation, performing the fast fine-tuning operation includes:

[0035] Based on the joint input of the current physical field feature vector and the topological structure vector, the common encoder parameters in the meta-learning model are frozen, and only the task-specific decoder parameters are fine-tuned;

[0036] The fine-tuning operation adopts a finite-step gradient update strategy based on a model-independent meta-learning algorithm, and the number of update rounds is less than a preset maximum number of iterations.

[0037] In a second aspect, the present application provides a multi-physics field control system for aluminum product processing, including:

[0038] An acquisition module is used to collect physical field data and process quality parameters associated with the processing task during the aluminum product processing process, perform feature extraction, and obtain a physical field feature vector representing the processing task state; wherein the physical field data includes: electromagnetic field parameters and electrostatic field parameters;

[0039] A construction module for constructing and training a meta-learning model based on the historical physical field feature vectors of multiple machining tasks and their corresponding process control parameters and process quality information;

[0040] an adjustment module for, in response to changes in the working conditions corresponding to the processing task during execution, invoking the meta-learning model to perform rapid fine-tuning operations based on the physical field feature vector under the current working conditions, thereby generating a prediction result under the current working conditions; the prediction result is used to characterize the evolution trend of the physical field coupling behavior and its potential change risk during the aluminum product processing process;

[0041] The suggestion module is used to generate control suggestion information for the current processing task based on the prediction results.

[0042] Compared with the existing technology, this application introduces a physical field state modeling framework based on meta-learning methods, which can uniformly model multiple physical field parameters such as electromagnetic fields and electrostatic fields involved in the processing of aluminum products, and combine process quality parameters to achieve a comprehensive characterization of the processing task state. Compared with traditional rule-based or static model-based methods, the present invention uses data from multiple historical processing tasks for training in the model construction phase, and can respond to changes in current working conditions in the operation phase to quickly fine-tune model parameters, thereby achieving predictions of physical field coupling trends and potential risks under the current task state. The present application can output structured prediction results, covering dimensions such as physical field coupling strength, discharge risk level, and abnormality probability, and then assist in generating control recommendation information for the current processing task, thereby achieving dynamic perception and intelligent auxiliary adjustment of the processing process state without the need for frequent human intervention. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 A flowchart of the multi-physics field control method for aluminum product processing provided in an embodiment of the present application;

[0044] Figure 2 A flowchart of a method for obtaining a physical field characteristic vector representing a processing task state provided in an embodiment of the present application;

[0045] Figure 3 A flowchart of a method for constructing and training a meta-learning model provided in an embodiment of the present application;

[0046] Figure 4 Schematic diagram of the multi-physics field control system for aluminum product processing provided in an embodiment of the present application. DETAILED DESCRIPTION

[0047] The technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments.

[0048] This application first collects multi-source data generated during the processing of aluminum products, including electromagnetic field parameters, electrostatic field parameters, and process quality parameters associated with the processing tasks. By performing feature extraction on the above data, a physical field feature vector that can characterize the current processing task state is obtained. Subsequently, a meta-learning model with task migration capabilities is constructed and trained using the physical field feature vectors, process control parameters, and process quality information corresponding to multiple historical processing tasks. During the execution of the processing task, when the system detects that the working conditions have changed, the meta-learning model is called and quickly fine-tuned based on the physical field feature vector extracted under the current working conditions to output the prediction results under the current working conditions. The prediction results are used to reflect the evolution trend and potential change risks of the physical field coupling behavior during the processing process, and further generate control suggestion information that matches the current processing task based on the prediction results, providing intelligent support for dynamic optimization and safety prediction of the processing process.

[0049] See also Figure 1 FIG. 1 is a flow chart of a multi-physics field control method for aluminum product processing provided in an embodiment of the present application, wherein the method includes steps S101 to S104, wherein:

[0050] S101: collecting physical field data and process quality parameters associated with the processing task during the aluminum product processing, performing feature extraction, and obtaining a physical field feature vector representing the processing task state; wherein the physical field data includes: electromagnetic field parameters and electrostatic field parameters;

[0051] S102: constructing and training a meta-learning model based on historical physical field feature vectors of multiple processing tasks and their corresponding process control parameters and process quality information;

[0052] S103: In response to a change in the working condition corresponding to the processing task during execution, the meta-learning model is called to perform a rapid fine-tuning operation based on the physical field feature vector under the current working condition to generate a prediction result under the current working condition; the prediction result is used to characterize the evolution trend of the physical field coupling behavior and its potential change risk during the aluminum product processing process;

[0053] S104: Based on the prediction result, generate control suggestion information for the current processing task.

[0054] Regarding S101 above:

[0055] In specific implementations, this can be achieved through the use of multiple types of industrial sensors deployed at the processing site. The physical field data includes electromagnetic and electrostatic field parameters. Electromagnetic field parameters can include current, voltage, magnetic induction intensity, and other parameters reflecting the state of induction heating or electromagnetic stirring. Electrostatic field parameters can include surface potential, charge accumulation level, partial discharge events, and other parameters reflecting electrostatic behavior during spraying or powder processing. Process quality parameters can include melt flow, surface roughness, deposition uniformity, or process temperature curves, representing the state of the processing.

[0056] The collected data are dynamic in time series and heterogeneous in physical dimension. Therefore, in this embodiment, by setting a unified data sampling window and time step, sensor data from different sources are collected synchronously and pre-processed, such as removing missing items and normalizing dimensional differences.

[0057] After initial processing, various physical field data and process quality parameters are mapped into a unified vector space, forming a physical field feature vector representing the current machining task status. This physical field feature vector serves as an important basis for subsequent model input, containing key working conditions at the machining site at the current moment, supporting subsequent modeling and prediction processes.

[0058] For example, an aluminum alloy profile continuous extrusion production line is deployed on-site to collect multi-source data during the aluminum product processing, including electromagnetic field parameters, electrostatic field parameters, and process quality parameters associated with the current processing task.

[0059] The electromagnetic stirring zone is equipped with a current transformer, a voltage sampling module, and a magnetic induction intensity sensor, which respectively collect the input current and voltage of the stirring coil and the magnetic induction intensity signal around the coil. These sensors periodically sample data at 100ms intervals and transmit the data to the central processing unit via the edge data acquisition module.

[0060] At the aluminum electrostatic spraying station, surface potentiometers, spark discharge counters and ambient humidity sensors are installed to detect the electrostatic potential of the workpiece surface in the spraying area, the number of electrostatic discharges occurring per unit time, and air humidity information that may affect charge accumulation.

[0061] Corresponding process quality parameters are collected by a process monitoring system. These include spray coating thickness measured by an online laser thickness gauge, aluminum surface roughness measured by an optical scanning system, and coating adhesion measured by sample testing. All parameters are collected within the same time window and aligned using timestamps.

[0062] After acquisition, the data from multiple sensors is aggregated over a set time window, for example, every 5 seconds, to extract basic statistical features such as mean, maximum value, and amplitude of change. All parameters are then normalized and concatenated into a multidimensional feature vector of length N. This serves as the physical field feature vector for the current machining task, representing the current working state.

[0063] Regarding S102 above:

[0064] In the specific implementation, it mainly includes three stages: model task construction, training sample preparation and meta-model training.

[0065] First, multiple aluminum product processing tasks from different periods or production lines were modeled as independent tasks. Each processing task includes a set of historically collected physical field feature vectors, process control parameters, and process quality information. The physical field feature vectors reflect the working state, the process control parameters record the control settings at the time, such as electromagnetic frequency, spray voltage, coil current, etc., and the process quality information includes performance indicators related to the quality of the task, such as surface uniformity, defect rate, or stability score.

[0066] Next, the data from multiple tasks is divided into training and validation sets according to a certain ratio to construct a sample set to support meta-learning. In this embodiment, the meta-learning model adopts a task-based learning approach. By mining the common parameter structure between multiple processing tasks, it trains shared model parameters with cross-task generalization capabilities.

[0067] After training, the meta-learning model has the ability to quickly adapt to new processing tasks or new working conditions with only a small number of samples, laying the foundation for subsequent fine-tuning and prediction.

[0068] Exemplarily, a meta-learning model is constructed and trained based on the collected data of multiple historical processing tasks. Each processing task contains a set of historical physical field feature vectors, corresponding process control parameters and process quality information. The physical field feature vector is obtained by the above-mentioned sensor acquisition and feature extraction process, and is used to describe the physical field state in each time window during the processing. Eight aluminum alloy profile production tasks under different working conditions are selected as the meta-learning task set. Among them, there are differences between tasks in production formula, electromagnetic stirring parameter settings, spraying voltage and workshop environmental conditions, which reflects the cross-task variability of the distribution of physical field characteristics.

[0069] The sample data for each machining task is divided into a training subset and a validation subset. The training subset is used for within-task training in the meta-learning phase, while the validation subset is used for optimizing and evaluating shared parameters between tasks. Each sample data consists of a physical field feature vector and its corresponding process control parameters and process quality information.

[0070] This example employs a model-agnostic meta-learning (MAML) strategy for model training. During the initialization phase, the meta-learning model's initial parameters are optimized through alternating multi-task training, enabling it to quickly adapt to new processing tasks using a small number of samples. During training, a fixed batch of samples is used for each task, with several rounds of inner-loop training and gradients backpropagated to the meta-parameters. Ultimately, a shared model structure is obtained that generalizes to multiple tasks.

[0071] The meta-learning model after training serves as the basis for subsequent real-time fine-tuning and prediction, and has the ability to migrate across tasks and quickly adapt to different processing conditions.

[0072] Regarding S103 above:

[0073] In practice, this can be achieved by real-time monitoring of key physical field parameters and process quality indicators during the machining process. Model fine-tuning is triggered when changes in certain parameters outside of a preset fluctuation range are detected, such as sudden changes in electromagnetic power, increased frequency of local electrostatic discharge, or deviations from quality trends.

[0074] When a fine-tuning operation is triggered, the system uses the latest physical field feature vectors collected under the current working conditions as input, calls the pre-trained meta-learning model, and quickly fine-tunes it within a limited number of iterations. This fine-tuning process only requires a small amount of current task sample data to complete local adjustments to the model parameters, allowing the model to more accurately adapt to the characteristic distribution under the current working conditions.

[0075] The fine-tuned model can output prediction results under the current working conditions. The prediction results are used to characterize the evolution trend of the physical field coupling behavior during the processing, such as the direction of change in coupling strength and the trend of change in the degree of interference between fields. It can also further evaluate potential production risks, such as the electrostatic discharge risk level and the possibility of abnormal fluctuations.

[0076] Through this step, dynamic perception and trend prediction of the physical field status of the processing site can be achieved, providing data support for the subsequent generation of more targeted control suggestions.

[0077] Exemplarily, for the meta-learning model that has completed training, a real-time working condition perception and rapid fine-tuning mechanism is constructed. The system continuously collects physical field data and generates physical field feature vectors during the execution of the processing task, and at the same time sets the working condition change conditions for triggering the model fine-tuning operation. When it is detected that multiple dimensions in the physical field feature vector simultaneously have mutations that exceed the set change threshold, such as a significant increase in the number of electrostatic discharges and electromagnetic power fluctuations exceeding the set tolerance range, the system determines that the current processing conditions have changed and triggers the rapid fine-tuning operation of the meta-learning model. In the fine-tuning stage, several physical field feature vector samples collected by the current processing task within a short time window in the past are used as fine-tuning data input, the common coding layer parameters of the meta-learning model are fixed, and only the task-related parameter part is updated with a limited number of gradient steps. The maximum number of fine-tuning steps set in this embodiment is 5 steps, the optimizer adopts the Adam algorithm, and the learning rate is set to 0.001.

[0078] After fine-tuning is complete, the system uses the latest physical field feature vector as input and calls the updated meta-learning model for predictive reasoning. The resulting prediction results include a score reflecting the evolution trend of the current physical field coupling state, the possible risk level of abnormal fluctuations, and reference information for subsequent decision-making.

[0079] This step ensures that the system can quickly adjust the model when the working conditions suddenly change, maintaining the accuracy of the prediction results and adaptability to the working conditions.

[0080] Regarding S104 above:

[0081] In specific implementation, in this embodiment, the "generating control suggestion information for the current processing task based on the prediction results" mentioned in the above S104 means that the system automatically generates control suggestions with reference value based on the current working condition prediction results output after fine-tuning the meta-learning model, combined with preset process optimization goals or safe operation specifications.

[0082] The prediction results include, but are not limited to, physical field coupling strength scores, electrostatic discharge risk levels, potential anomaly occurrence probabilities, and recommended physical field adjustment parameters. Based on the threshold intervals or level classification criteria corresponding to different prediction results, the system can match corresponding recommendations from a knowledge rule library or generate optimization strategies using a recommendation model built based on empirical data.

[0083] The control suggestion information may include: electromagnetic power adjustment suggestions, such as lowering the induction frequency, electrostatic suppression suggestions such as increasing the discharge cycle, starting the electrostatic neutralization device, environmental condition adjustment suggestions, such as increasing humidity, adjusting wind speed, etc., and may also include alarm suggestions to remind operators to pay attention to the existence of high-risk conditions in a certain area.

[0084] The final generated control recommendation information can be output in the form of structured data and used as an auxiliary judgment basis for the control system, human-machine interface or operator to improve the stability, safety and product consistency of the processing process.

[0085] For example, based on the above prediction results, the system automatically generates control suggestion information corresponding to the current processing task.

[0086] The prediction results include indicators such as the physical field coupling strength score, electrostatic discharge risk level, and the probability of potential anomalies. In this embodiment, the system sets multi-dimensional recommendation generation rules. For example, when the physical field coupling strength score is below the set lower limit, indicating that the electromagnetic stirring or electrostatic control effect is insufficient, the system recommends increasing the electromagnetic coil input power or increasing the operating frequency of the static eliminator; when the electrostatic discharge risk level is "high", the system recommends activating the backup electrostatic neutralization device, turning on the environmental humidification system, or reducing the spray voltage; when the probability of anomalies exceeds the preset threshold, the system prompts the operator to check the sensor layout, screen for potential interference sources, and may recommend entering the process buffer stage for observation.

[0087] The system can display the aforementioned control suggestion information in a structured prompt format on the processing monitoring interface and push it to the industrial control system interface for the control system or operators to selectively adopt based on on-site conditions. The control suggestion information generated by this embodiment has the characteristics of strong adaptability to working conditions, fast response time, and high readability, and can assist the processing system in achieving safer and more efficient operation adjustments in dynamic environments.

[0088] See also Figure 2 FIG. 2 is a flowchart of a method for obtaining a physical field characteristic vector representing a processing task state provided by an embodiment of the present application. The method includes steps S201 to S202, wherein:

[0089] S201: performing a short-time Fourier transform on the electromagnetic field parameters to generate an energy spectrum reflecting the change of frequency over time; calculating the electromagnetic field dominant frequency, frequency drift amplitude, and high-frequency disturbance intensity based on the energy spectrum to generate an electromagnetic field frequency domain eigenvector;

[0090] S202: performing wavelet packet decomposition on the electrostatic field parameters to generate a multi-scale instantaneous discharge response signal; identifying feature points in the instantaneous discharge response signal that meet an abnormality criterion, and removing electrostatic features corresponding to the feature points that meet the abnormality criterion to generate an electrostatic field denoising feature vector.

[0091] In the actual aluminum processing process, physical field signals are often non-stationary, time-varying, and highly noisy. For example, the field intensity fluctuations generated by electromagnetic stirring have periodic components, while electrostatic discharge signals may contain transient pulses. Directly modeling the raw signals is susceptible to unstructured noise, local perturbations, and multi-frequency interference, resulting in reduced feature extraction quality, which in turn affects the model's recognition accuracy and prediction stability.

[0092] Therefore, it is necessary to design a set of signal processing methods for electromagnetic and electrostatic field parameters to extract characteristic vectors with time-frequency information representativeness, robustness and distinguishability, improve the expression ability of physical field states, and provide a higher quality data basis for subsequent model input.

[0093] In a specific implementation, performing a short-time Fourier transform on the electromagnetic field parameters to generate an energy spectrum reflecting the time-varying frequency variation involves segmenting the electromagnetic parameters according to a set time window and overlap ratio, and using a short-time Fourier transform (STFT) to map the one-dimensional time-domain signal into a two-dimensional time-frequency energy spectrum. The dominant frequency position and energy density of different time segments in the spectrum effectively characterize the dynamic characteristics of the electromagnetic field signal.

[0094] After completing the construction of the energy spectrum, the following three types of characteristic indicators are extracted: the dominant frequency of the electromagnetic field: the frequency component corresponding to the maximum frequency energy in each time window; the frequency drift amplitude: the maximum change amplitude of the dominant frequency between consecutive time windows; the high-frequency disturbance intensity: the total energy proportion of the high-frequency band (such as >500Hz) in the frequency domain.

[0095] The above three types of indicators are normalized to form the electromagnetic field frequency domain feature vector, which is used to represent the electromagnetic field behavior characteristics of the time segment.

[0096] The wavelet packet decomposition of the electrostatic field parameters refers to performing multi-scale wavelet packet decomposition (WPD) on the original electrostatic signals, such as surface potential and charge, to obtain time-series decomposition signals at multiple frequency components. This process can enhance the ability to capture local peak signals.

[0097] Subsequently, based on pre-defined anomaly criteria, such as Z-score outlier detection and threshold mutation detection, spikes that do not conform to statistical patterns are identified in the wavelet packet reconstructed signal and determined to be transient discharge anomalies. The electrostatic features corresponding to these anomalies are then removed, including replacing the abnormal segment signal with interpolated values ​​or smoothing filtering results, ultimately obtaining a denoised feature vector for the electrostatic field.

[0098] The electrostatic field denoising feature vector and the above-mentioned electromagnetic field frequency domain feature vector will be combined in the subsequent stage to form a complete physical field feature representation.

[0099] For example, a short-time Fourier transform is performed on the induced current signal collected on-site. The sampling frequency is set to 2 kHz, the time window length is set to 512 points, the window overlap is 50%, and weighted processing is performed using a Hamming window function. The resulting time-frequency energy spectrum has a resolution of 3.9 Hz. The maximum energy frequency of each time window is extracted as the dominant frequency, and its standard deviation in 10 consecutive windows is calculated as the frequency drift indicator. At the same time, the proportion of spectral energy above 500 Hz to the total energy is calculated as the high-frequency disturbance indicator.

[0100] For electrostatic discharge processing, a five-layer wavelet packet decomposition was performed on each 1000-point segment of the surface potential signal, using the Daubechies 4-basis wavelet function. A Z-score threshold of ±3 was used to identify transient spikes in the reconstructed signal. Linear interpolation was used to fill in the identified anomalous segments, forming a continuous, denoised electrostatic signal sequence. Statistical features such as mean and fluctuation amplitude were extracted from the denoised signal to form a denoised feature vector for the electrostatic field.

[0101] Finally, the electromagnetic field frequency domain feature vector and the electrostatic field denoising feature vector within this time window are spliced ​​into a physical field feature vector of length N for use in subsequent model construction and training stages.

[0102] As an optional implementation, the feature extraction to obtain the physical field feature vector representing the processing task state further includes:

[0103] Merging the electromagnetic field frequency domain eigenvector and the electrostatic field denoising eigenvector to construct an initial physical field eigenvector set;

[0104] Based on the sensor layout topology diagram and the signal sampling time window, a correlation analysis is performed on the initial physical field feature vector set to identify interference feature components whose correlation with the induced interference in the topological connection path exceeds a preset threshold. The interference feature components are then removed from the initial physical field feature vector set to generate an effective physical field feature vector after interference removal.

[0105] In actual machining operations, different types of physical field signals are often collected through distributed sensor networks. These sensors are affected by factors such as layout, electromagnetic induction, and reflection from metal structures, making them prone to problems such as inductive coupling and synchronization interference. For example, an electrostatic sensor near an electromagnetic stirring device may record periodic false signals rather than actual charge changes.

[0106] If these interfered features are not identified and eliminated, the constructed physical field feature vector will contain a large amount of low-quality or even misleading information, seriously affecting the subsequent model training and prediction results.

[0107] Therefore, this embodiment introduces the sensor topology diagram and time synchronization analysis mechanism to identify and clean the interference of the initial physical field feature vector set, construct a highly reliable physical field effective feature vector, and provide a robust data foundation for subsequent modeling.

[0108] In practice, within each sampling time window, the two types of feature vectors obtained in the previous stage are concatenated to form a unified set of initial physical field feature vectors. The concatenation order remains fixed to ensure consistency of the input structure across multiple tasks.

[0109] The sensor layout topology diagram is represented by a graph structure modeling method. Each node in the diagram corresponds to a sensor channel. The weight of the edge can be determined by factors such as spatial distance, cable path, and electromagnetic shielding status, reflecting the physical induction relationship or interference possibility between the two sensors.

[0110] For the initial set of physical field eigenvectors within each time window, the system performs correlation analysis on the eigenvalues ​​across all channels. Specifically, for each pair of sensor channels with a direct connection (i.e., non-zero edge weight) in the topological structure graph, the cross-correlation coefficient between their eigenvalues ​​is calculated, and then the interference correlation index is constructed based on the topological edge weights.

[0111] When a feature channel maintains a high correlation with multiple adjacent channels in a continuous time window, that is, exceeds the set threshold, the system determines that the channel feature may be a pseudo feature induced by the adjacent signal, that is, an "interference feature component."

[0112] The system removes the above-mentioned interference characteristic components from the initial physical field characteristic vector set, or applies weight attenuation processing to its value, and finally generates an effective physical field characteristic vector after interference removal as a more credible model input.

[0113] For example, the electromagnetic and electrostatic signals come from sensor channels at seven different locations. The system sequentially concatenates the 21-dimensional electromagnetic field frequency domain feature vector and the 18-dimensional electrostatic field denoising feature vector to construct an initial physical field feature vector of length 39.

[0114] The sensor topology is stored as a 7×7 adjacency matrix, where each element represents the likelihood of interference between channels. For example, if channels 2 and 5 share a common ground plane, the edge weight is set to 0.8; if they are physically far apart and shielded, the edge weight is set to 0.

[0115] The system calculates the Pearson cross-correlation coefficient between adjacent channels in each 5-second time window. If the correlation between a channel and at least two adjacent channels in the last three windows exceeds 0.9 and the edge weight product is higher than 0.6, the channel is marked as an interference feature channel.

[0116] The system assigns an attenuation weight to the marked interference channel eigenvalues, such as multiplying it by 0.2, or directly eliminating the dimension, and finally forms a physical field effective eigenvector of length M (M≤39), which serves as the input basis for the subsequent meta-learning model.

[0117] As an optional implementation, the feature extraction to obtain the physical field feature vector representing the processing task state further includes:

[0118] Performing statistical analysis on the process quality parameters to extract process quality feature vectors including mean, standard deviation and change rate;

[0119] The process quality feature vector is spliced ​​with the physical field effective feature vector to generate a complete physical field feature vector.

[0120] During aluminum processing, relying solely on physical field parameters like electromagnetic and electrostatic signals is insufficient to fully characterize the processing status. Process quality parameters such as coating thickness, surface roughness, and melt flow rate directly reflect product performance and process stability, and are highly relevant and valuable for engineering.

[0121] However, process quality parameters often have low sampling frequencies and large response lags. Without normalization and feature modeling, modeling accuracy can be reduced. Therefore, it is necessary to convert process quality parameters into structured statistical features and combine them with effective physical field feature vectors to form a complete feature expression vector to enhance the model's perception and prediction capabilities for production tasks.

[0122] In a specific implementation, the statistical characteristics of the multiple collected process quality parameters are calculated within a given time window to construct a process quality feature vector.

[0123] Specifically, for each process quality parameter, such as spray layer thickness, surface roughness, melt temperature or material density, the following indicators are calculated:

[0124] Mean: represents the central tendency of the parameter within the time window; Standard deviation: measures the degree of fluctuation of the parameter within the current window; Rate of change: defined as the difference between the last value and the first value of the parameter within the time window divided by the window duration, used to characterize the speed trend of parameter change.

[0125] The statistical characteristics of the process quality parameters are calculated using a sliding window method to maintain time alignment with the physical field signal, and the calculation results are normalized to ensure that all characteristics are in a unified dimension and numerical range.

[0126] Subsequently, the above-mentioned process quality characteristics are organized in vector form to form a process quality feature vector, which is then spliced ​​with the effective feature vector of the physical field to form a complete physical field feature vector for model input.

[0127] The dimensions of the complete physical field feature vectors after splicing remain consistent, and the positions of various features in the vector structure are fixed, ensuring comparability and stability in multi-task modeling and prediction.

[0128] For example, the following three process quality parameters can be collected: spray coating thickness (unit: μm), surface roughness Ra (unit: μm), and melt flow rate (unit: mm / s). All parameters are collected every 30 seconds and aligned with the 5-second physical field sampling window.

[0129] In each 5-second time window, the system performs statistical analysis on the historical process quality data of the previous 6 sampling points (a total of 180 seconds) and calculates the mean of the three parameters;

[0130] After normalization, all statistics are concatenated into a 9-dimensional process quality feature vector in a preset order.

[0131] The above 9-dimensional vector is concatenated with the effective feature vector of the physical field to form a complete physical field feature vector, which serves as the final input structure of the meta-learning model.

[0132] See also Figure 3 FIG. 1 is a flowchart of a method for constructing and training a meta-learning model according to an embodiment of the present application, including steps S301 to S303, wherein:

[0133] S301: Encode the sensor layout topology diagram corresponding to each processing task to generate a topology vector representing the difference in sensor layout for different tasks;

[0134] S302: jointly inputting the topological structure vector and the physical field feature vector of the corresponding processing task into the meta-learning model for task embedding representation;

[0135] S303: Use a meta-learning algorithm with structure embedding capabilities to jointly train multiple processing tasks.

[0136] While the same physical field sensor types may be used across different processing tasks, their spatial placement, wiring structure, and sensing paths can vary significantly. These differences directly impact the spatial characteristics and interference patterns of the collected signals. For example, in one production line, an electrostatic detector may be located near the spray gun, while in another, it may be near the powder recovery path, resulting in spatial variation in the characteristic distribution.

[0137] Directly using physical field feature vectors for model training without considering these differences in sensor structure between tasks will result in poor model generalization and transferability. Therefore, it is necessary to vectorize the sensor layout structure for each task and input it into the model along with the physical field features, allowing the model to learn the mapping relationship between how structural differences affect signal patterns.

[0138] In specific implementation, building and training the meta-learning model includes the following steps:

[0139] Step 1: Encoding of sensor layout topology diagram:

[0140] For each processing task, the system constructs a sensor layout topology diagram based on the spatial distribution, wiring, and interference relationships of its physical field sensors. This topology diagram is modeled as a graph G = (V, E), where V is the set of sensor nodes and E is the set of edges connecting them. Edge weights are determined based on the following factors: spatial distance (Euclidean distance between physical locations); line overlap (the proportion of cables running in common paths); electromagnetic shielding status (the presence of a common ground or metal frame); and historical interference records, such as frequent synchronous mutations between two nodes.

[0141] The system encodes the topological structure graph into a fixed-dimensional topological structure vector. The encoding methods may include flattening the adjacency matrix, graph neural network embedding encoding, and structural feature statistics splicing to ensure that the topological structure of each task is converted into numerical input.

[0142] Step 2: Combine inputs to construct task embedding representations:

[0143] The topological structure vector of each machining task is concatenated or input in parallel with its corresponding physical field feature vector to form a joint input vector for the current task. This joint input vector not only contains physical state information but also embeds information about the sensor layout structure, enabling the model to perceive structural differences between tasks.

[0144] Step 3: Model structure and joint training:

[0145] The meta-learning model uses training algorithms with structure-embedding capabilities, such as structure-aware multi-task meta-learning frameworks like Graph-MAML and structure-conditional encoder-decoder architectures. It jointly trains on multiple processing tasks and optimizes a shared set of initial model parameters, enabling the model to quickly adapt to new tasks while taking into account structural differences.

[0146] The training process may include intra-task training rounds, inter-task parameter update rounds, task encoding module weight sharing strategies, etc. The final trained model has both physical state perception capabilities and structural transfer and generalization capabilities.

[0147] For example, six different aluminum alloy production lines were selected as independent processing tasks. Each production line was equipped with eight physical field sensors, covering key locations such as the electromagnetic coil, power input section, spraying area, and static elimination point. The connections, electromagnetic shielding conditions, and spatial distances between the sensors were not exactly the same.

[0148] The system calculates the physical distance (in centimeters) between each pair of sensor nodes, the cable colinearity ratio, and the metal co-grounding situation to generate a 7×7 adjacency matrix, normalized to the range [0, 1]. The adjacency matrix is ​​flattened into a 49-dimensional structure vector, which serves as the topological structure vector for the processing task.

[0149] For simplicity, only seven key sensor channels are selected to construct the topology diagram, resulting in a corresponding 7×7 adjacency matrix. These key sensor channels cover locations such as the electromagnetic coil, spray head, and static eliminator, providing representative information and the potential for structural interference. The remaining channels are not included in the topology modeling but still participate in the subsequent feature stitching process. In actual deployments, all channels can be modeled to form an 8×8 adjacency matrix.

[0150] At the same time, the physical field feature vector corresponding to each task (e.g., with a dimension of 36) is collected and concatenated with the topological structure vector (e.g., with a dimension of 49) to form a joint input with a dimension of 85. The model is trained using the structure-aware MAML algorithm, with K samples from each task used in the inner loop and the outer loop optimizing shared parameters.

[0151] After training, the meta-learning model can be quickly adjusted in new tasks using a small number of samples to adapt to the differences in physical field signal structures under different wiring and layouts.

[0152] As an optional implementation manner, generating the effective eigenvector of the physical field after interference removal includes:

[0153] The sensor arrangement topology structure diagram is represented as an adjacency matrix of sensor nodes, and cross-correlation coefficients are calculated for physical field characteristic components of different rows within a signal sampling time window;

[0154] Based on the product of the cross-correlation coefficient and the path connection weight in the adjacency matrix, an interference correlation index is constructed;

[0155] Determining whether the interference correlation index exceeds a set threshold, and marking the characteristic component whose interference correlation index exceeds the set threshold as an interference characteristic component;

[0156] The interference characteristic components are removed from the initial physical field characteristic vector set to form an effective physical field characteristic vector after interference removal for training and prediction.

[0157] In a multi-sensor environment, inductive coupling or synchronous interference may occur between different sensor channels due to factors such as spatial layout, electromagnetic environment, and differences in sensor sensitivity. This interference typically manifests as short-term synchronous fluctuations or highly correlated signal components between physically unrelated sensors.

[0158] Especially in aluminum product processing scenarios, local interference from strong electromagnetic equipment or powder electrostatic spraying processes may cause some sensor channels to capture false signals that are not caused by their own working conditions.

[0159] Therefore, a method that combines spatial topological structure diagrams with signal time domain correlation analysis is needed to construct a quantifiable interference correlation index, thereby identifying and eliminating potential interference feature components and improving the robustness and accuracy of the model input.

[0160] In the specific implementation, the sensor space structure relationship of each processing task is expressed in the form of an adjacency matrix, which is recorded as , where n is the number of sensor channels, and the matrix elements The adjacency matrix represents the strength of the connection between sensor channels i and j, or the coupling weight, reflecting the combined influence of the two channels in terms of spatial arrangement, electromagnetic induction, or cable collinearity. This adjacency matrix is ​​not only used to represent the spatial structure layout but also serves as a reference for subsequent signal interference identification and feature processing.

[0161] In each time window, the system extracts the initial physical field feature vector set, combines the characteristic components of each channel in pairs, and calculates their cross-correlation coefficients. , which is used to measure the similarity of signal changes between different channels within the current window.

[0162] In order to combine structural coupling with signal synchronization, the interference correlation index is defined for:

[0163]

[0164] This product can comprehensively reflect the spatial correlation and signal coupling strength between channels.

[0165] For each sensor channel i, the system counts the distance between it and all adjacent channels. , if there are multiple channels j satisfying ,in, If the interference correlation threshold is set, it is considered that the characteristic component of channel i may be caused by structural interference and is marked as an interference characteristic component.

[0166] The system removes the eigenvalue of the channel from the initial physical field eigenvector set or lowers its weight, such as multiplying it by an attenuation factor. The final output is the effective eigenvector of the physical field after removal, which serves as the input basis for subsequent model training and prediction.

[0167] For example, the system deploys 8 physical field sensors, and the characteristics of each sensor channel are 1-dimensional real numbers, forming an 8-dimensional initial physical field feature vector. The adjacency matrix between sensors is , generated by spatial distance and historical interference records, the matrix elements range from [0, 1].

[0168] Every 5 seconds is a sampling time window, and the system calculates the cross-correlation coefficient of each pair of channels (i, j) within the window. , using the Pearson coefficient calculation method. Set the interference correlation threshold .

[0169] If for a certain channel i, there are at least two channels j that satisfy , then channel i is determined to be affected by synchronous interference. The eigenvalue corresponding to this channel is multiplied by an attenuation coefficient of 0.3 to reduce its impact on the overall model.

[0170] The length of the processed feature vector is still 8 dimensions, but its interference component has been suppressed, and it is finally input into the meta-learning model as an effective feature vector of the physical field for training and reasoning.

[0171] It should be noted that the examples are not mutually exclusive, and the sensor channels can be regarded as one-dimensional or multi-dimensional features in different scenarios according to requirements.

[0172] As an optional implementation manner, generating a prediction result under current operating conditions includes:

[0173] Based on the joint input consisting of the physical field feature vector under the current working conditions and the topological structure vector corresponding to the processing task, the gradient information relative to the meta-learning model parameters is calculated, and the parameters are fine-tuned within a preset number of iterations;

[0174] The fine-tuned meta-learning model is used for predictive reasoning of the physical field feature vector of the current processing task, and the prediction index value associated with the processing task is output. The prediction index value includes: physical field coupling strength score, potential anomaly probability, electrostatic discharge risk level and recommended physical field adjustment parameters.

[0175] During the aluminum product processing process, production sites are subject to highly dynamic operating conditions, such as changes in ambient temperature and humidity, fluctuations in power grid load, and equipment aging. These changes can cause drift in physical field behavior, making it impossible for statically trained models to maintain accurate predictions.

[0176] To improve the adaptability of the model, this implementation introduces a rapid fine-tuning mechanism driven by the current working condition characteristics, combines the physical field characteristics and topological structure information collected in real time, and performs small sample fine-tuning based on the current task status, so that the meta-learning model can quickly adapt to field changes while maintaining training experience, thereby achieving accurate prediction of the physical field coupling evolution trend and potential anomalies.

[0177] In the specific implementation, during the prediction stage, the system collects the physical field feature vector of the current window in real time, and calls the topological structure vector encoded by the corresponding processing task, and splices the two into a complete set of model input vectors.

[0178] This joint input is used to compute the gradient relationship with the parameters to be fine-tuned in the meta-learning model. During fine-tuning, the system maintains the common encoder parameters of the meta-model and only performs a limited number of gradient updates on the task-specific decoder layer parameters. Gradient updates typically use the Adam optimizer, with the number of updates limited to a set upper limit (e.g., 5) to ensure responsiveness and parameter stability.

[0179] After completing the quick fine-tuning, the system uses the current physical field feature vector as input and invokes the fine-tuned meta-learning model for inference and prediction. The model outputs the following multi-dimensional prediction indicators: Physical Field Coupling Strength Score, which reflects the changing trend of electromagnetic and electrostatic coupling; Potential Anomaly Occurrence Probability, which indicates the likelihood of an atypical state under the current operating conditions; ESD Risk Level, which classifies the discharge characteristics and risk model output; and Recommended Physical Field Adjustment Parameters, which provide guidance on appropriate settings for electromagnetic frequency, voltage, current, and other parameters.

[0180] The above-mentioned prediction index values ​​serve as the basis for generating subsequent control recommendations, and can also be presented in real time through a visual interface to assist on-site operators or the dispatching system in making judgments.

[0181] For example, the meta-learning model has been pre-trained using a set of structure-aware tasks. During the processing of a task, the system collects and generates physical field feature vectors in real time, using a data window of 5 seconds. It also reads the topological structure vector corresponding to the current task and concatenates them to form a multi-dimensional joint input vector.

[0182] If the system detects a fluctuation in operating conditions, such as an ESD frequency exceeding the historical mean by 30% for two consecutive cycles, it initiates a rapid fine-tuning process. The system uses feature samples from the three most recent data windows, freezes the encoder parameters, and performs only three gradient updates on the decoder layer. The optimizer uses Adam with a learning rate of 0.0005.

[0183] After fine-tuning is complete, the latest physical field eigenvector is used as input into the fine-tuned model, outputting the following prediction indicators: physical field coupling strength score: 0.78 (out of 1); probability of potential anomaly occurrence: 64%; electrostatic discharge risk level: medium-high risk (level 3); recommended physical field adjustment parameters: adjust the electromagnetic frequency from 80 Hz to 92 Hz and reduce the spray voltage by 5%.

[0184] The above results are displayed in real time through the processing execution system interface, and are also transmitted to the suggestion module to generate control suggestion information.

[0185] As an optional implementation, the output format of the prediction index value includes:

[0186] Outputting prediction results of multiple dimensions in the form of structured data, wherein the format of the structured data includes a key-value pair format, a vector structure, or a hierarchical risk level representation;

[0187] The physical field coupling strength score is represented by a floating point interval; the electrostatic discharge risk level is represented by a multi-level classification label; the probability of potential abnormality occurrence is in the form of a percentage; and the recommended physical field adjustment parameters are represented in the form of a recommended value range or a threshold pair.

[0188] In order to make the model prediction results easier to interpret, integrate and apply, in industrial processing scenarios, the prediction output needs to use a structured data format to facilitate human-machine interface display, control system analysis, and automatic calling of subsequent control suggestion modules.

[0189] Different prediction indicator dimensions have different data characteristics, so it is necessary to design adaptive output expressions respectively to improve the readability, usability and standardization of the prediction results.

[0190] In a specific implementation, the output format of the prediction index value is organized in a unified structured data format, and the prediction results of multiple dimensions are output in the form of key-value pairs, vector structures or multi-level classification representation.

[0191] Structured output includes the following fields and their corresponding formats:

[0192] coupling_score: represents the physical field coupling strength score, a floating-point number in the interval [0.0, 1.0], which is used to measure the degree of interaction between electromagnetic and electrostatic fields; discharge_risk_level: represents the electrostatic discharge risk level, using multi-level classification labels such as "low", "medium", and "high" or levels 0-3 for representation; anomaly_probability: represents the probability of potential anomaly occurrence, output as a percentage; recommended_parameters: represents the physical field adjustment parameters recommended by the system, organized in a dictionary structure, including recommended frequency, current, voltage, and other values, presented in the form of recommended ranges or recommended thresholds.

[0193] The output results can be packaged in JSON, XML or industrial protocol format for direct reading by the control platform, monitoring system or human-machine interface.

[0194] As an optional implementation, performing the fast fine-tuning operation includes:

[0195] Based on the joint input of the current physical field feature vector and the topological structure vector, the common encoder parameters in the meta-learning model are frozen, and only the task-specific decoder parameters are fine-tuned;

[0196] The fine-tuning operation adopts a finite-step gradient update strategy based on a model-independent meta-learning algorithm, and the number of update rounds is less than a preset maximum number of iterations.

[0197] During the meta-learning reasoning phase, fast fine-tuning should take into account the following characteristics: fast response speed, no destruction of shared parameters, and the ability to adjust to the current task distribution with a small amount of data.

[0198] Therefore, the fine-tuning strategy of this application adopts the method of freezing some parameters and only updating the task-related parts, and adopts a gradient update strategy with a limited number of steps to ensure practicality and controllability.

[0199] In a specific implementation, performing the fast fine-tuning operation includes the following steps:

[0200] Parameter freezing strategy: The system freezes the common encoder layer parameters in the meta-learning model during the fine-tuning phase to ensure that it retains the ability to be shared between tasks, and only unlocks the task-specific decoder layer parameters for optimization.

[0201] Input construction and gradient calculation: The system constructs a joint input based on the physical field feature vector and the corresponding topological structure vector under the current working conditions, and calculates the gradient between the input and the current decoder layer parameters.

[0202] Finite-step gradient update: A gradient descent strategy based on a model-independent meta-learning algorithm is used to update the decoding layer parameters for no more than a preset number of rounds. Each update round uses a small number of 3-5 groups of samples and uses the Adam optimizer for parameter adjustment.

[0203] Step constraint: To ensure response speed and prevent overfitting, the maximum number of update rounds is set to 5 in this embodiment. If the convergence condition is met early, for example, the loss function decrease rate is lower than the threshold, it will stop early.

[0204] For example, after a machining task undergoes a change in working conditions, the system uses the feature samples from the latest three time windows as a fine-tuning dataset to construct a joint input. The encoder parameters are frozen, and only the decoder, which contains a three-layer MLP structure, is updated.

[0205] The system uses the MAML algorithm to perform each round of forward propagation + gradient backpropagation operations, with a learning rate set to 0.001, a maximum number of update rounds of 5, and batch gradient descent with a batch size of 3. After fine-tuning, the model is immediately used to predict the output.

[0206] Based on the same inventive concept, the embodiments of the present application also provide a multi-physical field control system for aluminum product processing corresponding to the multi-physical field control method for aluminum product processing. Since the principle of problem solving by the system in the embodiments of the present application is similar to the multi-physical field control method for aluminum product processing mentioned above in the embodiments of the present application, the implementation of the system can refer to the implementation of the method, and the repeated parts will not be repeated.

[0207] Reference Figure 4 FIG. 1 is a schematic diagram of a multi-physics field control system for aluminum product processing provided in an embodiment of the present application. The system includes:

[0208] The acquisition module 10 is used to collect physical field data and process quality parameters associated with the processing task during the aluminum product processing, perform feature extraction, and obtain a physical field feature vector representing the processing task state; wherein the physical field data includes: electromagnetic field parameters and electrostatic field parameters;

[0209] A construction module 20 is used to construct and train a meta-learning model based on historical physical field feature vectors of multiple processing tasks and their corresponding process control parameters and process quality information;

[0210] An adjustment module 30 is configured to, in response to changes in the working conditions corresponding to the machining task during execution, invoke the meta-learning model to perform rapid fine-tuning operations based on the physical field feature vector under the current working conditions, thereby generating a prediction result under the current working conditions; the prediction result is used to characterize the evolution trend of the physical field coupling behavior and its potential change risk during the machining of aluminum products;

[0211] The suggestion module 40 is used to generate control suggestion information for the current processing task based on the prediction result.

[0212] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed in this application can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

Claims

1. A multi-physics field control method for aluminum product processing, characterized in that: include: Collect physical field data and process quality parameters associated with the processing task during the aluminum product processing, perform feature extraction, and obtain a physical field feature vector representing the processing task state; wherein the physical field data includes: electromagnetic field parameters and electrostatic field parameters; Based on the historical physical field feature vectors of multiple machining tasks and their corresponding process control parameters and process quality information, a meta-learning model is constructed and trained. In response to changes in the working conditions corresponding to the processing task during execution, the meta-learning model is called to perform a rapid fine-tuning operation based on the physical field feature vector under the current working conditions to generate a prediction result under the current working conditions; the prediction result is used to characterize the evolution trend and change risk of the physical field coupling behavior during the aluminum product processing process; Based on the prediction results, generating control suggestion information for the current processing task; The feature extraction to obtain the physical field feature vector representing the processing task state includes: Performing a short-time Fourier transform on the electromagnetic field parameters to generate an energy spectrum reflecting the change of frequency over time; calculating the electromagnetic field dominant frequency, frequency drift amplitude, and high-frequency disturbance intensity based on the energy spectrum to generate an electromagnetic field frequency domain eigenvector; Performing wavelet packet decomposition on the electrostatic field parameters to generate a multi-scale instantaneous discharge response signal; identifying feature points in the instantaneous discharge response signal that meet an abnormality criterion, and removing electrostatic features corresponding to the feature points that meet the abnormality criterion to generate an electrostatic field denoising feature vector; The feature extraction to obtain the physical field feature vector representing the processing task state further includes: Merging the electromagnetic field frequency domain eigenvector and the electrostatic field denoising eigenvector to construct an initial physical field eigenvector set; Based on the sensor layout topology diagram and the signal sampling time window, a correlation analysis is performed on the initial physical field feature vector set to identify interference feature components whose correlation with the induced interference in the topological connection path exceeds a preset threshold, and the interference feature components are removed from the initial physical field feature vector set to generate an effective physical field feature vector after interference removal; The feature extraction to obtain the physical field feature vector representing the processing task state further includes: Performing statistical analysis on the process quality parameters to extract process quality feature vectors including mean, standard deviation and change rate; The process quality feature vector is concatenated with the physical field effective feature vector to generate a physical field feature vector.

2. The multi-physical field control method in aluminum product processing according to claim 1, characterized in that: Based on the historical physical field feature vectors of multiple machining tasks and their corresponding process control parameters and process quality information, a meta-learning model is constructed and trained, including: Encode the sensor layout topology diagram corresponding to each processing task and generate a topology vector that represents the difference in sensor layout for different tasks; The topological structure vector and the physical field feature vector of the corresponding processing task are jointly input into the meta-learning model for task embedding representation; A meta-learning algorithm with structural embedding capability is used to jointly train multiple processing tasks.

3. The multi-physical field control method in aluminum product processing according to claim 2, characterized in that: The generating of the effective eigenvector of the physical field after interference removal includes: The sensor arrangement topology structure diagram is represented as an adjacency matrix of sensor nodes, and cross-correlation coefficients are calculated for physical field characteristic components of different rows within a signal sampling time window; Based on the product of the cross-correlation coefficient and the path connection weight in the adjacency matrix, an interference correlation index is constructed; Determining whether the interference correlation index exceeds a set threshold, and marking the characteristic component whose interference correlation index exceeds the set threshold as an interference characteristic component; The interference characteristic components are removed from the initial physical field characteristic vector set to form an effective physical field characteristic vector after interference removal for training and prediction.

4. The multi-physical field control method in aluminum product processing according to claim 3, characterized in that: Generating the prediction result under the current working condition includes: Based on the joint input consisting of the physical field feature vector under the current working conditions and the topological structure vector corresponding to the processing task, the gradient information relative to the meta-learning model parameters is calculated, and the parameters are fine-tuned within a preset number of iterations; The fine-tuned meta-learning model is used for predictive reasoning of the physical field feature vector of the current processing task, and the prediction index value associated with the processing task is output. The prediction index value includes: physical field coupling strength score, potential anomaly probability, electrostatic discharge risk level and recommended physical field adjustment parameters.

5. The multi-physical field control method in aluminum product processing according to claim 4, characterized in that: The output format of the predicted index value includes: Outputting prediction results of multiple dimensions in the form of structured data, wherein the format of the structured data includes a key-value pair format, a vector structure, or a hierarchical risk level representation; The physical field coupling strength score is represented by a floating point interval; the electrostatic discharge risk level is represented by a multi-level classification label; the probability of potential abnormality occurrence is in the form of a percentage; and the recommended physical field adjustment parameters are represented in the form of a recommended value range or a threshold pair.

6. The multi-physical field control method in aluminum product processing according to claim 5, characterized in that: The fast fine-tuning operation includes: Based on the joint input of the current physical field feature vector and the topological structure vector, the common encoder parameters in the meta-learning model are frozen, and only the task-specific decoder parameters are fine-tuned; The fine-tuning operation adopts a finite-step gradient update strategy based on a model-independent meta-learning algorithm, and the number of update rounds is less than a preset maximum number of iterations.

7. A multi-physics field control system for aluminum product processing, used to implement the multi-physics field control method for aluminum product processing according to any one of claims 1 to 6, characterized in that: include: An acquisition module is used to collect physical field data and process quality parameters associated with the processing task during the aluminum product processing process, perform feature extraction, and obtain a physical field feature vector representing the processing task state; wherein the physical field data includes: electromagnetic field parameters and electrostatic field parameters; A construction module for constructing and training a meta-learning model based on the historical physical field feature vectors of multiple machining tasks and their corresponding process control parameters and process quality information; an adjustment module for, in response to changes in the working conditions corresponding to the processing task during execution, invoking the meta-learning model to perform rapid fine-tuning operations based on the physical field feature vector under the current working conditions, thereby generating a prediction result under the current working conditions; the prediction result is used to characterize the evolution trend of the physical field coupling behavior and its potential change risk during the aluminum product processing process; The suggestion module is used to generate control suggestion information for the current processing task based on the prediction results.

Images