A data-driven aluminum extrusion press intelligent control method, system and aluminum extrusion press

Abstract

The embodiment of the application relates to a data-driven aluminum extrusion machine intelligent control method and system and an aluminum extrusion machine, and belongs to the technical field of electronic data processing. In the method steps, multi-source time sequence data is collected to obtain two-dimensional deep features, data distribution is detected, and a dynamic time warping oversampling is used to enhance unbalanced samples with a generative adversarial network. A domain-invariant feature model is trained to extract cross-condition robust features. State evaluation information output by the model is connected to a parameter-free predictive control framework. A multi-objective optimization function is constructed based on real-time sliding window data at each beat. An optimal control sequence is issued to an actuator after solving, so as to realize online closed-loop adjustment of extrusion speed, temperature and other parameters. The embodiment of the application combines data processing, AI and automatic control, and is self-adaptive to condition changes, can improve control precision and quality stability, and reduce energy consumption and cost.

Description

Technical Field

[0001] This application relates to a data-driven intelligent control method, system, and aluminum extrusion press, belonging to the technical field of electronic data processing. Background Technology

[0002] Aluminum extrusion is an important plastic forming process widely used in the production of complex cross-section profiles in aerospace, rail transportation, and building materials industries. The aluminum extrusion process is a complex industrial process involving the strong coupling of multiple physical fields such as mechanics, hydraulics, and thermodynamics. The core of the process is to apply pressure to a master cylinder, forcing heated aluminum ingots out of a die of a specific shape to obtain the desired cross-section profile. The quality control of the aluminum extrusion process directly determines the dimensional accuracy, mechanical properties, surface quality of the final product, as well as the energy consumption and efficiency of the production process.

[0003] Current aluminum extrusion press control strategies widely employ classic PID control schemes, which use fixed process parameters to perform single-loop setpoint control on key variables such as main cylinder pressure and heater temperature. However, these existing schemes rely on precise mathematical models and traditional control methods with fixed parameters. When faced with numerous challenges in actual industrial production, their limitations become increasingly apparent, primarily exhibiting the following drawbacks:

[0004] First, the aluminum extrusion process exhibits significant time-varying, nonlinear, and large hysteresis characteristics. Factors such as fluctuations in the composition of aluminum ingots, die wear, minute changes in extrusion cylinder temperature, and equipment aging can all cause drift in process dynamics, making it difficult for PID controllers based on fixed parameters to maintain optimal control performance, easily leading to problems such as out-of-tolerance product dimensions and surface defects.

[0005] Secondly, existing control methods lack a deep understanding of the inherent state of the process and the ability to make forward-looking decisions. Most existing technologies rely on feedback information at the current moment for control, and cannot effectively predict changes that may occur in the production process in the future. For example, during the extrusion process, the rising trend of the die temperature, the falling temperature of the aluminum rod, and the abnormal fluctuations of the main cylinder pressure may all indicate impending product quality problems or equipment failure risks. Current feedback control can only respond after the problem occurs, which is a post-event remedy and makes it difficult to achieve pre-event prevention. This limits the further improvement of aluminum product quality and the reduction of production losses.

[0006] Furthermore, with the development of industrial big data and artificial intelligence technologies, data-driven methods have provided new insights for the optimized control of complex industrial processes. However, directly applying existing data-driven models, which are typically based on simple deep learning prediction models or basic model predictive control, still faces significant challenges when applied to aluminum extrusion press control. On the one hand, aluminum extrusion production involves multiple varieties and small batches, resulting in significant differences in the distribution of process data under different batches of aluminum ingots and different dies. The control parameters and data distribution are bound to be unbalanced, which will greatly affect the generalization ability of data-driven models. A model that performs well under one operating condition may not perform well under another new operating condition, failing to meet the stringent requirements of industrial production.

[0007] In summary, the control of aluminum extrusion presses faces the following challenges: how to construct an intelligent control method that can adapt to different operating conditions, requires no precise mechanistic model, and possesses forward-looking optimization decision-making capabilities; how to comprehensively improve product quality, production efficiency, and equipment operational stability in the aluminum extrusion process; how to deeply integrate real-time data and intelligently drive data to form new solutions; how to break free from dependence on fixed models; and how to achieve sensitivity to changes in operating conditions. Existing technologies lack concrete solutions and can no longer meet the requirements, necessitating urgent improvement. Summary of the Invention

[0008] The main objective of this application is to provide a data-driven intelligent control method, system, and aluminum extruder to solve the technical problems of lacking fused data and data-driven intelligent control of aluminum extruders in the prior art.

[0009] The embodiments of this application adopt the following technical solutions:

[0010] According to one aspect of the embodiments of this application, a data-driven intelligent control method for an aluminum extrusion press is provided, comprising: acquiring multi-source time-series data of the aluminum extrusion press within a complete extrusion cycle; performing harmonic filtering and image-based feature transformation on the multi-source time-series data to generate a time-series signal structured representation image, wherein the time-series signal structured representation image is used to characterize two-dimensional depth features of the extrusion process state; detecting the data distribution of the two-dimensional depth features; identifying a first feature dataset with data imbalance due to changes in operating conditions; performing data augmentation on the first feature dataset using a generative adversarial network and an oversampling method based on dynamic time warping to generate a second feature dataset; and training a domain-invariant feature model using the second feature dataset, wherein the domain-invariant feature model, through... Anti-learning is used for training, and the trained domain-invariant feature model is used to extract corresponding invariant features from the input data. The domain-invariant feature model is used for real-time state monitoring, and the state evaluation information output by the domain-invariant feature model is combined with a parameterless predictive control framework. In each control cycle of the parameterless predictive control framework, a multi-objective optimization function is constructed based on the input and output data of the real-time sliding window. The optimization objectives of the multi-objective optimization function include unit energy consumption and aluminum extrusion efficiency, and the constraints include safety thresholds of process parameters. The optimal control sequence in the future time domain is obtained according to the multi-objective optimization function. The parameter control quantities of the optimal control sequence are sent to the aluminum extrusion press actuator, and the aluminum extrusion press actuator performs real-time control of the aluminum extrusion process.

[0011] According to at least one specific embodiment of the present application, in the process of generating the structured representation image of the time series signal: based on a sensor array deployed at key nodes of industrial equipment, multi-channel runtime timing signals are synchronously acquired at a fixed sampling period to obtain an initial one-dimensional time series data sequence; the initial one-dimensional time series data sequence is normalized and denoised by a signal preprocessing subroutine to generate a standardized time series signal; the standardized time series signal is processed using a Gram angle field image generation algorithm, and a first type of structured representation image of the time series signal is generated by mapping each data point to a polar coordinate system and calculating the trigonometric sum and angle field and trigonometric difference angle field between any two points.

[0012] According to at least one specific embodiment of the present application, in the process of generating the structured representation image of the time series signal: based on the sensor array arranged at key nodes of industrial equipment, multi-channel runtime timing signals are synchronously acquired at a fixed sampling period to obtain an initial one-dimensional time series data sequence; the initial one-dimensional time series data sequence is normalized and denoised by a signal preprocessing subroutine to generate a standardized time series signal; the standardized time series signal is processed by a Markov transition field image generation algorithm, and the signal value domain is discretized into multiple state partitions, and the transition probabilities between state partitions are statistically analyzed along the time axis to generate a second type of structured representation image of the time series signal.

[0013] According to at least one specific embodiment of the present application, the step of detecting the data distribution of the two-dimensional depth features, identifying a first feature dataset with data imbalance caused by changes in working conditions, and using a generative adversarial network and an oversampling method based on dynamic time warping to augment the first feature dataset and generate a second feature dataset further includes: based on a preset feature space distribution evaluation model, detecting the clustering of data points of the input two-dimensional depth features, identifying a first feature region where the number of samples is below a threshold due to changes in working conditions by calculating the sample density ratio of each cluster; constructing a data augmentation branch of the generative adversarial network, using a generator to receive random sampling vectors from the first feature region, and generating synthetic feature samples through a deconvolution layer, using a discriminator to compare the distribution differences between the real features and the synthetic feature samples, and iteratively generating a first type of augmented features by alternately optimizing the parameters of the generator and the discriminator; using a dynamic time warping algorithm to calculate the optimal planning path between each synthetic feature sample and its nearest neighbor samples, performing linear interpolation on the optimal planning path to generate a second type of augmented features; merging the first type of augmented features and the second type of augmented features into the original two-dimensional depth features, and generating a balanced second feature dataset through feature space re-standardization processing.

[0014] According to at least one specific embodiment of the present application, the step of training a domain-invariant feature model using the second feature dataset, wherein the domain-invariant feature model is trained through adversarial learning, and the trained domain-invariant feature model is used to extract corresponding invariant features from input data, further includes: initializing the structural parameters of the domain-invariant feature model based on the preprocessed second feature dataset; inputting the second feature dataset into a feature extractor through a forward propagation channel; extracting high-dimensional feature representations of the second feature dataset using a multi-layer convolutional neural network to obtain a primary feature vector; inputting the primary feature vector into a domain classifier and a task classifier respectively; and calculating the domain label score through a fully connected layer in the domain classifier. The class probability is calculated, and the predicted task label is obtained through the softmax layer in the task classifier. Using the domain classification loss function and the task classification loss function, the task classification loss value is calculated based on the difference between the predicted task label and the true task label. The weight parameters of the feature extractor and the task classifier are updated using the task classification loss value to minimize the task classification error of the domain-invariant feature model and maximize the domain discrimination accuracy of the domain classifier. The parameter update process of adversarial learning training is repeated to obtain the optimized domain-invariant feature model. The new input data to be processed is input into the trained domain-invariant feature model, and the domain-invariant feature vector is output through the forward propagation channel.

[0015] According to at least one specific embodiment of the present application, in the process of updating the weight parameters of the feature extractor and the task classifier using the task classification loss value, the network parameters of the domain classifier are pre-fixed through the gradient backpropagation mechanism; the parameters of the feature extractor and the task classifier are fixed, and the weight parameters of the domain classifier are updated using the domain discrimination loss value in the domain classification loss function, so as to maximize the domain discrimination accuracy of the domain classifier; when repeatedly executing the parameter update process of adversarial learning training, training is stopped when the weighted sum of the domain discrimination loss value and the task classification loss value reaches the convergence threshold, and the optimized domain invariant feature model is obtained.

[0016] According to at least one specific embodiment of the present application, the step of using a domain-invariant feature model for real-time state monitoring and combining the state evaluation information output by the domain-invariant feature model with a non-parametric predictive control framework, and constructing a multi-objective optimization function based on the input-output data of a real-time sliding window in each control cycle of the non-parametric predictive control framework, further includes: deploying the trained domain-invariant feature model to an online monitoring module; processing the real-time collected system input data through the domain-invariant feature model to extract domain-invariant feature vectors; calculating state evaluation information based on the domain-invariant feature vectors to generate real-time state evaluation information; and constructing an input-output data matrix based on fixed-length real-time sliding window data within each control cycle of the non-parametric predictive control framework, using the real-time state evaluation information as a reference quantity for the optimization objective to construct a multi-objective optimization function.

[0017] According to at least one specific embodiment of the present application, the status assessment information includes the aluminum extruder operating condition identification result and the aluminum extruder extrusion quality prediction result.

[0018] According to another aspect of the embodiments of this application, a data-driven intelligent control system for an aluminum extrusion press is provided, used to implement the aforementioned data-driven intelligent control method for an aluminum extrusion press, comprising: a multi-source time-series data processing module, which collects multi-source time-series data of the aluminum extrusion press within a complete extrusion cycle, performs harmonic filtering and image feature transformation on the multi-source time-series data, and generates a time-series signal structured representation image, wherein the time-series signal structured representation image is used to represent two-dimensional depth features of the extrusion process state; a second feature dataset generation module, which detects the data distribution of the two-dimensional depth features, identifies a first feature dataset with data imbalance due to changes in operating conditions, and performs data augmentation on the first feature dataset using a generative adversarial network and an oversampling method based on dynamic time warping, thereby generating a second feature dataset; and a domain-invariant feature model training module, which trains a domain-invariant feature model using the second feature dataset. The system comprises a feature model, which is trained using adversarial learning. The trained domain-invariant feature model is used to extract corresponding invariant features from the input data. A multi-objective optimization function construction module uses the domain-invariant feature model for real-time state monitoring and combines the state evaluation information output by the domain-invariant feature model with a parameterless predictive control framework. At each control cycle of the parameterless predictive control framework, a multi-objective optimization function is constructed based on the input and output data of a real-time sliding window. The optimization objectives of the multi-objective optimization function include unit energy consumption and aluminum extrusion efficiency, and the constraints include safety thresholds for process parameters. An aluminum extrusion press actuator control module obtains the optimal control sequence in the future time domain based on the multi-objective optimization function. The parameter control quantities of the optimal control sequence are then sent to the aluminum extrusion press actuator for real-time control of the aluminum extrusion process.

[0019] According to another aspect of the embodiments of this application, an aluminum extrusion press is provided, wherein the aluminum extrusion press is provided with the aforementioned data-driven intelligent control system for aluminum extrusion presses.

[0020] The beneficial technical effects of the embodiments of this application are:

[0021] This application's embodiments achieve intelligent control of an aluminum extrusion press based on control logic encompassing perception, computation, decision-making, and execution. It deeply integrates data processing, artificial intelligence, and automatic control technology to form an intelligent control method and system for aluminum extrusion presses capable of adapting to changing operating conditions. In processing raw data, harmonic filtering and image-based feature transformation elevate the raw, noisy physical signals into structured data reflecting the aluminum extrusion process. Then, through data augmentation and domain-invariant feature model training, adaptive learning is performed from finite and imbalanced data.

[0022] This application embodiment combines state assessment information with a non-parametric prediction framework, enabling proactive optimization based on future trend predictions during the intelligent control and decision-making process of aluminum extruders. This gives the extruders adaptive and self-learning capabilities, improving the control precision of the aluminum extrusion process and the quality stability of aluminum products, and enhancing the adaptability of the aluminum industry control system to different production conditions.

[0023] This application embodiment implements data augmentation based on an AI model and provides a highly adaptable model. After sufficient and diverse training, it forms intelligent decision output, which enables the aluminum extrusion process to maintain a high level of control performance. This achieves optimization of product production energy efficiency and cost reduction and efficiency improvement. It allows the aluminum extrusion process to intelligently adjust parameters such as extrusion speed and temperature while ensuring quality and safety, avoiding unnecessary energy loss and achieving the dual goals of energy saving and consumption reduction and production efficiency improvement. Attached Figure Description

[0024] To more clearly illustrate the specific implementation methods of the embodiments of this application or the technical solutions in the prior art, the drawings used in the description of the specific implementation methods or the prior art will be briefly introduced below. Obviously, the drawings described below are some implementation methods of the embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 This is a flowchart of steps S1 to S5 in an embodiment of this application.

[0026] Figure 2 This is a flowchart of the optimization technical solution provided in steps S11 to S12.

[0027] Figure 3 This is a flowchart of the optimization technical solutions provided in steps S21 to S24.

[0028] Figure 4 This is a flowchart of the optimization technical solutions provided in steps S31 to S34.

[0029] Figure 5 This is a flowchart of the optimization technical solution provided in steps S41 to S42.

[0030] Figure 6 This is the architecture diagram of a data-driven intelligent control system for aluminum extrusion presses. Detailed Implementation

[0031] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the embodiments of this application, and not all embodiments. Based on the specific implementation methods in the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the protection scope of the embodiments of this application.

[0032] The embodiments of this application have wide applications in aluminum extrusion production processes. Several typical specific application scenarios are given below:

[0033] In the production of hollow aluminum profiles with complex cross-sections, the products require high dimensional accuracy, surface finish, and consistency in mechanical properties. Frequent changes between different batches of aluminum ingots and molds are necessary during production. Existing PLC control strategies struggle to reliably guarantee quality, and changeover times are lengthy, resulting in high energy consumption. To address these issues, this application provides a technical solution for this application scenario, the main process of which can be summarized as follows:

[0034] Sensors are placed at key locations such as the extrusion cylinder, die, and main cylinder of the aluminum extrusion press to form a sensor network. Throughout a complete extrusion cycle (from feeding, heating, extrusion to discharge), dozens of time-series data points, including extrusion cylinder temperature, die temperature, aluminum rod temperature, main cylinder pressure, and extrusion speed, are collected at synchronous frequencies. The collected main cylinder pressure and vibration signals are automatically processed for harmonic filtering. For example, the fixed-frequency noise generated by the periodic displacement of the main pump is filtered out, highlighting abnormal impact signals hidden within the noise, caused by slight die wear.

[0035] After preprocessing and data cleaning, multi-source time-series data (such as temperature and pressure curves) are converted in real time into Gram angle field (GASF / GADF) images and Markov transition field (MTF) images. These curves are used to characterize the health status of the extrusion process. If the pressure curve is stable, an image with a specific texture will be generated. If abnormal pressure fluctuations occur during the aluminum extrusion process, the pressure curve will produce a completely different image pattern. Therefore, abnormal pressure conditions in the aluminum extrusion process can be identified through GASF and MTF.

[0036] When dealing with abnormal operating conditions, data equalization is required. For example, the number of high-quality feature image samples collected when producing new molds is far less than the number of samples collected when producing old molds, resulting in data imbalance. Therefore, Generative Adversarial Networks (GANs) are used to automatically generate a large number of feature images simulating the normal production state of new molds, expanding their sample library. Furthermore, for some rare abnormal state features in the early stage of new mold production, new and reasonable abnormal feature samples are generated by finding samples with similar shapes in the feature space and performing intelligent interpolation.

[0037] The domain-invariant feature model is trained using a balanced dataset. The training process involves an adversarial process involving a feature extractor, a quality classifier, and a condition discriminator. Ultimately, the domain-invariant feature model can adaptively learn to ignore the source of the data, i.e., whether the data comes from a new mold or an old mold (domain information), and focus on identifying core quality and safety-related features such as whether the product is defective or whether the pressure is abnormal. This demonstrates that the domain-invariant feature model has a certain generalization ability.

[0038] At the start of a new round of aluminum extrusion production, online real-time optimization control and execution are implemented. The trained model analyzes current data in real time, anticipates risks and predicts results, and immediately feeds the prediction results (state assessment information) into a non-parametric prediction framework. This non-parametric prediction framework does not rely on complex mathematical or physical equations; instead, it directly calls input-output data from similar cases in a historical database. Based on the latest data, it quickly solves optimization problems such as how to adjust extrusion speed and die cooling water flow in the future, determining optimization objectives to ensure quality, improve efficiency, and reduce energy consumption, while ensuring that extrusion pressure does not exceed limits and die temperature does not become too high, meeting safety threshold constraints.

[0039] After determining the optimization objective and satisfying the constraints, control commands are issued to the actuators of the aluminum extrusion press. Specific control commands may include: fine-tuning the main cylinder servo proportional valve, reducing the extrusion speed by 5%, increasing the flow rate of the die cooling system by 10%, and so on. The aluminum extrusion press actuators execute the control commands, perform the corresponding operations, collect new result data in real time, update the sliding window, and repeat the entire process of monitoring-prediction-optimization-execution in the next control cycle to achieve continuous optimization. The control cycle is calculated in multiples of seconds; N control nodes equal N seconds.

[0040] As can be seen from the implementation of the embodiments of this application in specific application scenarios, the embodiments of this application improve product quality, significantly reduce product defect rates, especially the first-piece pass rate after mold and material changes, which is greatly improved, production efficiency is increased, and energy consumption per unit product is reduced. When facing application scenarios with different batches of raw materials and mold wear, the control strategy can be automatically adjusted, reducing manual intervention and debugging time. By predicting and avoiding abnormal operating conditions, the service life of key components such as molds and main cylinders is extended. The application scenario here clearly demonstrates the complete closed loop of the embodiments of this application from data perception to intelligent decision-making and then to precise control, reflecting its enormous value in actual production processes.

[0041] like Figure 1 The data-driven intelligent control method for aluminum extrusion presses shown includes:

[0042] Step S1 involves acquiring multi-source time-series data from the aluminum extrusion press over a complete extrusion cycle. Harmonic filtering and image-based feature transformation are then performed on the multi-source time-series data to generate a structured representation image of the time-series signal. This structured representation image is used to characterize the two-dimensional depth features of the extrusion process state. The multi-source time-series data in this step can include extrusion cylinder temperature, die temperature, aluminum rod temperature, master cylinder pressure, extrusion speed, etc., from different data sources. Harmonic filtering enhances fault characteristics in the data, making them stand out from background noise.

[0043] Step S2: Detect the data distribution of two-dimensional depth features, identify the first feature dataset with data imbalance caused by changes in working conditions, and use generative adversarial network and oversampling method based on dynamic time warping to augment the first feature dataset and generate the second feature dataset.

[0044] Step S3: Train a domain-invariant feature model using the second feature. The domain-invariant feature model is trained through adversarial learning. The trained domain-invariant feature model is then used to extract corresponding invariant features from the input data. For example, in this step, the domain-invariant feature model is trained through adversarial learning between the feature extractor, the extrusion quality classifier, and the working condition discriminator. The invariant features include a high-quality domain that is insensitive to batch and die differences in aluminum bars.

[0045] Step S4: Use the domain-invariant feature model for real-time state monitoring, and combine the state evaluation information output by the domain-invariant feature model with a parameterless predictive control framework. In each control cycle of the parameterless predictive control framework, construct a multi-objective optimization function based on the input and output data of the real-time sliding window. The optimization objectives of the multi-objective optimization function include unit energy consumption and aluminum extrusion efficiency, and the constraints include safety thresholds of process parameters.

[0046] Step S5: Obtain the optimal control sequence in the future time domain based on the multi-objective optimization function; send the parameter control quantities of the optimal control sequence to the aluminum extrusion press actuator, and use the aluminum extrusion press actuator to perform real-time control of the aluminum extrusion process.

[0047] The technical solutions provided in steps S1 to S5 construct an adaptive intelligent control method for aluminum extrusion presses. This method integrates advanced signal processing, artificial intelligence, and control theory to form an aluminum extrusion press control process that is applicable to the essence of the aluminum extrusion process, adapts to production changes, and proactively optimizes decisions. Step S1 acquires multi-source time-series data to deeply perceive the aluminum extrusion press process. Through multi-source data fusion and feature engineering (harmonic filtering, image conversion, etc.), the original physical signals are enhanced into structured data that reveals the internal state of the aluminum extrusion process. Steps S2 and S3 use data augmentation techniques to solve the data imbalance problem in reality and use adversarial training mechanisms to force the model to learn to express the features of interference factors. Steps S4 and S5 realize intelligent decision output and combine it with a parameter-free predictive control framework that does not require a precise mathematical model. Through a multi-objective optimization function, the decision is transformed into the corresponding aluminum extrusion press control action.

[0048] In the logic control of steps S1 to S5, the high-quality depth features generated by harmonic filtering and image transformation in step S1 provide clear and noise-resistant raw input data for the domain-invariant feature model in step S3. This data can accurately identify minute pressure anomalies, such as those caused by slight wear of the die. In step S4, these anomalies are fed into the predictive controller in real time, enabling it to compensate for product quality defects by adjusting extrusion parameters (such as fine-tuning the extrusion speed) in step S5 before the defects actually occur. This achieves precise control of the aluminum extruder driven by data, significantly reducing the incidence of product dimensional deviations and surface defects.

[0049] Step S2 provides diverse training samples through data augmentation (GAN, DTW oversampling), resolving the uneven data distribution caused by changes in aluminum rods, dies, etc., and providing accurate training samples for model training in Step S3. The adversarial training mechanism in Step S3 utilizes these training samples to actively remove features related to specific operating conditions and strengthen invariant features related to quality and equipment status, improving the generalization ability to cope with changes. The synergy between Steps S2 and S3 enables the control method and system of the aluminum extrusion press to maintain high performance without recalibration when facing new operating conditions. The state evaluation information provided by Step S3 makes the multi-objective optimization function in Step S4 purposeful. When the model judges that the current operating condition is good, it can appropriately increase the speed to improve efficiency while ensuring quality. When an upward trend in energy consumption is detected, the heating strategy can be optimized. Step S5 ensures that the intelligent decisions provided in Step S4 are accurately executed by the aluminum extrusion press equipment.

[0050] In summary, it can be seen that the technical solutions provided in steps S1 to S5 realize the entire process of aluminum extrusion press, from status monitoring and early warning, constraint safety control to safe operation execution. The technical effects of each step are not isolated and simply superimposed, but are interconnected through the perception, cognition and decision-making execution capabilities provided by intelligent AI, realizing intelligent control of the aluminum extrusion press and fundamentally improving the overall performance of aluminum extrusion production.

[0051] like Figure 2 As shown, preferably, in step S1, the time-series signal structured representation image includes a Gram angle field image and / or a Markov transition field image. During the generation of the time-series signal structured representation image:

[0052] Step S11: Based on the sensor array deployed at key nodes of industrial equipment, multi-channel runtime timing signals are synchronously acquired at a fixed sampling period to obtain an initial one-dimensional timing data sequence.

[0053] Step S12: Normalize and denoise the initial one-dimensional time series data sequence through the signal preprocessing subroutine to generate a standardized time series signal.

[0054] Step S13: The normalized time series signal is processed using the Gram angle field image generation algorithm. By mapping each data point to the polar coordinate system and calculating the trigonometric sum and angle field and trigonometric difference angle field between any two points, a first-class time series signal structured representation image is generated.

[0055] The optimized technical solution provided in steps S11 to S13 adopts a method of synchronously acquiring multi-channel time-series signals based on sensor arrays and performing preprocessing. Step S11 ensures the synchronization and integrity of the data source, and step S12 improves the standardization and quality of the data. Through the synergistic cooperation of steps S11 and S12, the goal of providing high-quality standardized input data for subsequent image conversion is achieved, thereby realizing the accurate conversion from raw industrial signals to processable digital signals and solving the technical problem of poor characterization consistency caused by the difference in dimensions and noise interference of the raw signals.

[0056] Preferably, in step S1, the time-series signal structured representation image includes a Gram angle field image and / or a Markov transition field image. During the generation of the time-series signal structured representation image:

[0057] Step S14: Based on the sensor array deployed at key nodes of industrial equipment, multi-channel runtime timing signals are synchronously acquired at a fixed sampling period to obtain an initial one-dimensional timing data sequence.

[0058] Step S15: Normalize and denoise the initial one-dimensional time series data sequence through a signal preprocessing subroutine to generate a standardized time series signal.

[0059] Step S16: The standardized time series signal is processed using the Markov transition field image generation algorithm. By discretizing the signal value domain into multiple state partitions and statistically analyzing the transition probabilities between state partitions along the time axis, a second type of time series signal structured representation image is generated.

[0060] In the optimized technical solutions provided in steps S14 to S16, a one-dimensional time-series signal is mapped to a state transition probability image based on the Markov transition field. Relatively accurate initial data is obtained through the synchronous acquisition of multi-channel runtime time-series signals by the sensor array in step S14. In step S15, the data undergoes normalization and denoising preprocessing to ensure data quality and consistency. In step S16, the Markov transition field (MTF) generation algorithm is used to discretize the numerical range of the standardized signal into multiple state partitions, and the transition probabilities between states are accurately statistically analyzed along the time axis. Finally, a second-type time-series signal structured representation image is generated, representing the system's state evolution in image form. This achieves the goal of transforming the time dependence and state transition patterns of the dynamic process into visualized spatial structural features that can be directly processed by deep learning models, thereby realizing the technical effect of deeply mining the macroscopic operating conditions and staged evolution characteristics of industrial equipment operation.

[0061] Steps S14, S15, and S16 work together. Step S14 provides a comprehensive and synchronous raw data foundation. The preprocessing in step S15 eliminates dimensional differences and noise interference, creating conditions for accurate state division and probability statistics in step S16. Step S16 uses the clean data provided in the first two steps to transform the implicit state transition rules into image textures. The three steps work together to complete the conversion from raw signals to deep feature images, solving the technical problems of inaccurate identification of equipment operating conditions and insensitivity to changes in process stages caused by macroscopic dynamic behaviors such as equipment state transitions and mode switching. It overcomes the shortcomings of existing one-dimensional time-series data analysis methods in effectively capturing and characterizing equipment operating conditions in complex industrial processes.

[0062] like Figure 3 As shown, preferably, in step S2, the data distribution of two-dimensional depth features is detected, a first feature dataset with data imbalance caused by changes in working conditions is identified, and the first feature dataset is augmented using a generative adversarial network and an oversampling method based on dynamic time warping to generate a second feature dataset, further including:

[0063] Step S21: Based on the preset feature space distribution evaluation model, detect the clustering of data points of the input two-dimensional depth features, and identify the first feature region where the number of samples is lower than the threshold due to changes in working conditions by calculating the sample density ratio of each cluster.

[0064] Step S22: By constructing a data augmentation branch of a generative adversarial network, the generator receives random sampling vectors from the first feature region and generates synthetic feature samples through a deconvolution layer. The discriminator compares the distribution differences between the real features and the synthetic feature samples. The first type of augmented features is generated iteratively by alternately optimizing the parameters of the generator and the discriminator.

[0065] Step S23: Calculate the optimal planning path between each synthetic feature sample and its nearest neighbor samples using the dynamic time warping algorithm, and perform linear interpolation on the optimal planning path to generate the second type of enhanced features.

[0066] Step S24: The first type of enhanced features and the second type of enhanced features are merged into the original two-dimensional deep features. Through feature space re-standardization, a balanced second feature dataset is generated.

[0067] In the data augmentation optimization solutions provided in steps S21 to S24, a hybrid intelligent sampling strategy is adopted. Step S21 uses cluster analysis to accurately locate the sparse minority class feature regions. Step S22 uses a generative adversarial network (GAN) to simulate the original data distribution in the feature space and generate synthetic features that approximate reality to expand the minority class samples. Step S23 combines the dynamic time warping (DTW) algorithm to perform multiple interpolations between synthetic and real samples to further increase the diversity and rationality of the samples. Step S24 integrates the two types of augmented features with the original features and standardizes them to construct a balanced dataset. This achieves the goal of effectively solving the problem of imbalanced training data categories caused by changing working conditions at the feature level, thereby significantly improving the perception accuracy and generalization ability of subsequent artificial intelligence models under scarce and abnormal working conditions.

[0068] In the optimization solutions provided in steps S21 to S24, step S21 identifies regions of data imbalance, providing a clear target for subsequent enhancement. Steps S22 and S23 respectively implement generative and interpolation-based enhancement, forming complementary enhancement strategies. Specifically, the GAN in step S22 excels at global distribution learning, generating novel yet reasonable feature samples, while the DTW oversampling in step S23 focuses on local shape preservation, generating transitional samples based on existing samples. Steps S22 and S23 can be executed in parallel or sequentially, jointly ensuring the diversity and authenticity of the enhanced samples. Step S24 achieves data fusion and balancing. As an integration step, it integrates the enhancement results with the original data and eliminates the distribution shift that may be caused by data merging through re-standardization, ensuring the quality of the second feature dataset. This solves the technical problem that the training process of artificial intelligence models is biased towards most working conditions, resulting in poor prediction accuracy for a few working conditions or abnormal states, and weak model generalization ability due to uneven data collection frequencies under different working conditions (such as mold changing and material changing) in actual industrial environments.

[0069] like Figure 4 As shown, preferably, in step S3, a domain-invariant feature model is trained using the second feature. The domain-invariant feature model is trained through adversarial learning. The trained domain-invariant feature model is used to extract the corresponding invariant features from the input data, further including:

[0070] Step S31: Based on the preprocessed second feature dataset, initialize the structural parameters of the domain-invariant feature model, input the second feature dataset into the feature extractor through the forward propagation channel, and use a multi-layer convolutional neural network to extract the high-dimensional feature representation of the second feature dataset to obtain the primary feature vector.

[0071] Step S32: Input the primary feature vector into the domain classifier and the task classifier respectively. Calculate the classification probability of the domain label through the fully connected layer in the domain classifier, and calculate the prediction result of the task label through the softmax layer in the task classifier.

[0072] Step S33: Using the domain classification loss function and the task classification loss function, the task classification loss value is calculated based on the difference between the predicted task label and the actual task label.

[0073] Step S34: Update the weight parameters of the feature extractor and the task classifier using the task classification loss value, minimize the task classification error of the domain invariant feature model, and maximize the domain discrimination accuracy of the domain classifier.

[0074] Step S35: Repeat the parameter update process of adversarial learning training to obtain the optimized domain-invariant feature model. Input the new input data to be processed into the trained domain-invariant feature model and output the domain-invariant feature vector through the forward propagation channel.

[0075] In the optimization scheme provided in steps S31 to S35, an adversarial game training method is adopted. In step S31, the feature extractor extracts primary feature vectors from the original data. In step S32, the feature vectors are simultaneously fed into the task classifier (for the main task, such as judging the quality of vector data) and the domain classifier (for distinguishing data sources). Steps S33 and S34 construct an adversarial loop. On the one hand, the task classification loss is used to optimize the feature extractor and the task classifier to ensure that the extracted features are effective for the main task. On the other hand, the domain classification loss is used to optimize the domain classifier separately in the adversarial process to distinguish data sources. Meanwhile, the feature extractor learns the features that can recognize the domain classifier (i.e., domain-invariant features), and extracts features related to interference factors from the data, refining the feature representation that is only related to the essence of the core task. This significantly improves the generalization ability of the intelligent model in changing environments such as unknown working conditions and new production batches, so that the aluminum extrusion press can still perform stably in new working conditions.

[0076] Step S31 initializes the features. Step S32 provides bidirectional judgment between the task classifier and the domain classifier, guiding features towards the two closely related goals of completing the task and identifying the identity. Steps S33 and S34 implement adversarial optimization. The task classifier and the feature extractor are cooperative, with the common goal of minimizing the task classification error. The domain classifier and the feature extractor are adversarial; the domain classifier aims to maximize the domain discrimination accuracy, while the implicit goal of the feature extractor is to generate features that the domain classifier cannot determine the source of. Step S34 is cooperative in terms of task objectives but adversarial in terms of domain discrimination. Through this combined operation, the feature extractor is forced to continuously optimize its parameters, ultimately retaining only those features useful for completing the main task but unusable by the domain classifier. Step S35 is for iterative convergence, ensuring that the cooperative and adversarial processes in Step S34 progress towards the preset goal until an optimal balance is reached.

[0077] Steps S31 to S35 provide a model based on the training data, which solves the technical problem of significant performance degradation when encountering new working conditions in actual deployment due to the data distribution difference (domain difference) between the training data and the real application scenario.

[0078] For example, in step S34, the process of updating the weight parameters of the feature extractor and the task classifier using the task classification loss value further includes:

[0079] Step S341: The network parameters of the domain classifier are pre-fixed through the gradient backpropagation mechanism.

[0080] Step S342: Fix the parameters of the feature extractor and the task classifier, and update the weight parameters of the domain classifier using the domain discrimination loss value in the domain classification loss function, so as to maximize the domain discrimination accuracy of the domain classifier.

[0081] Step S343: When repeatedly executing the parameter update process of adversarial learning training, training is stopped when the weighted sum of the domain discrimination loss value and the task classification loss value reaches the convergence threshold, and the optimized domain invariant feature model is obtained.

[0082] In the optimization scheme provided in steps S341 to S343, an alternating fixed-parameter directional optimization strategy is adopted. Step S341 optimizes the fixed domain classifier parameters, ensuring that gradient backpropagation only updates the feature extractor and task classifier, thus guaranteeing that the extracted features are highly effective for completing the main task such as quality judgment. Step S342, while improving domain discriminative ability, fixes the parameters of the feature extractor and task classifier, updating only the domain classifier. This forces the domain classifier to acquire domain-related information with fixed features, thereby maximizing its discriminative ability. Step S343 monitors the weighted loss as a convergence condition, ensuring that the model achieves the best balance between the two objectives of main task accuracy and domain invariance. This achieves the goal of directional evolution that simultaneously satisfies the requirements of the main task and the domain invariance requirement, thereby improving the convergence speed and stability of the domain-invariant feature model training process and enhancing the model's generalization ability.

[0083] In the optimization scheme provided in steps S341 to S343, step S341 controls the fixed domain classifier and optimizes task-related parameters, while step S342 fixes the feature extractor and task classifier and optimizes the domain classifier. That is, steps S341 and S342 together constitute a complete adversarial loop. By continuously improving the discriminative power of the domain classifier, the instability in model training that might result from simultaneous updates of multiple parameters is avoided. Step S343 performs a convergence check, interrupting the adversarial loop at an appropriate time to ensure that training stops when task performance and domain invariance reach a preset balance point, preventing overfitting. Steps S341 to S343, through collaborative efforts, decompose the complex adversarial learning objective into executable and monitorable steps, ensuring the efficiency and reliability of the training process.

[0084] In the practical application of steps S341 to S343, to address potential training instability and model overfitting, further improvements can be made. The improved optimization technique can constrain the gradient norm of the feature extractor or discriminator after parameter updates, or perform spectral normalization on its weight matrix. This prevents adversarial generation from directly stopping when the training loss converges, allowing the adversarial generation process to monitor performance on an independent validation set. Training stops when the validation set performance no longer improves, and the model parameters with the best performance on the validation set are selected. Specifically, steps S344 and S345 are added:

[0085] Step S344: Based on the data of the current training batch, in the first parameter update stage, using the gradient backpropagation mechanism and with the domain classifier network weights fixed, the trainable parameters of the feature extractor and task classifier are updated using the calculated task classification loss value to obtain the minimized task classification error. In the second parameter update stage, with the feature extractor and task classifier network weights fixed, the trainable parameters of the domain classifier are updated using the calculated domain discrimination loss value to obtain the maximized domain discrimination accuracy.

[0086] Step S345: Alternately iterate through the first parameter update stage and the second parameter update stage. After each iteration, calculate the weighted sum of the domain discrimination loss value and the task classification loss value. When the weighted sum is found to be lower than the preset convergence threshold, generate the trained domain invariant feature model, or continue to train the next iteration based on the updated network parameters until the convergence condition is met.

[0087] The optimization techniques provided in steps S344 to S345 decompose the complex adversarial multi-objective optimization problem into two stable and ordered sub-problems that are solved iteratively, guiding the model to converge to a state where task performance and domain invariance are optimally balanced. Step S344 divides a single training round into two stages: the first parameter update stage fixes the domain classifier and optimizes the feature extractor and task classifier using task-specific loss to ensure the model's core task capabilities. The second parameter update stage fixes the feature extractor and task classifier and optimizes the domain classifier using domain discrimination loss to maximize its discriminative power as a strong competitor to the feature extractor. Step S345 iteratively executes the process of step S344, and calculates the weighted loss in real time and compares it with a preset threshold as an intelligent decision-making basis for dynamically terminating the training process. This achieves the goal of driving the model parameters to evolve in a controllable, stable, and automated manner, simultaneously optimally satisfying the requirements of main task accuracy and domain invariance. This significantly improves the stability and convergence efficiency of the model training process and ensures that the final domain-invariant feature model possesses both high task accuracy and strong cross-domain generalization ability.

[0088] like Figure 5 As shown, preferably, in step S4, the domain-invariant feature model is used for real-time state monitoring, and the state evaluation information output by the domain-invariant feature model is combined with a parameterless predictive control framework. At each control cycle of the parameterless predictive control framework, a multi-objective optimization function is constructed based on the input and output data of the real-time sliding window, further including:

[0089] Step S41: Deploy the trained domain-invariant feature model to the online monitoring module, process the real-time collected system input data through the domain-invariant feature model, extract the domain-invariant feature vector, calculate the state evaluation information based on the domain-invariant feature vector, and generate real-time state evaluation information.

[0090] Step S42: Within each control cycle of the non-parametric predictive control framework, an input-output data matrix is ​​constructed based on fixed-length real-time sliding window data, and a multi-objective optimization function is constructed using real-time state evaluation information as a reference quantity for the optimization objective.

[0091] The optimization solutions provided in steps S41 and S42 incorporate state assessment information as one of the optimization objectives, constructing a multi-objective optimization function. By considering both the state assessment information and the multi-objective optimization function, the system balances indicators such as control efficiency, energy consumption, and system lifespan. The quantified state indicators of the real-time state assessment information provided in step S41 give practical physical meaning to the optimization objectives of step S42, enabling the control strategy for the aluminum extruder to achieve not only short-term performance but also long-term operational reliability. Steps S41 and S42 rely on real-time data rather than precise mechanistic models, reducing system deployment costs and achieving intelligent data-driven operation. They eliminate the need for manual threshold setting or controller parameter adjustment, making them suitable for long-term autonomous operation in complex industrial scenarios.

[0092] In steps S41 and S42, the state assessment information refers to one or more quantitative indicators output by the domain-invariant feature model after analyzing real-time system data. These quantitative indicators are not the original sensor signals, but rather refined information that directly reflects the current health status or operational quality of the aluminum extrusion press. For example, state assessment information can be generated from the domain-invariant feature model, producing a value from 0 to 100, where 100 represents a perfect state and 0 represents a complete failure. In actual operating conditions, this value is approximately in the range of 20 to 80, and a decreasing trend in the value can predict performance degradation. Alternatively, state assessment information can be generated to characterize the degree of degradation, representing the gap between the current state and the understood state, such as the wear degree of the aluminum extrusion press bearings or the capacity degradation rate of the extrusion press's backup battery. The state assessment information serves as a bridge connecting the state monitoring module in step S41 and the predictive control module in step S42, transforming monitoring results from descriptive information into quantitative targets that the control module can understand and use.

[0093] In constructing a multi-objective optimization function using real-time state assessment information as a reference for the optimization objective, Willems' fundamental lemma can be used. The basic principle of Willems' fundamental lemma is that any future input-output behavior of a linear dynamic system can be linearly represented by a sufficiently long sequence of input-output data from the past that reflects its dynamic characteristics. Since the aluminum extruder approximates a pseudo-linear system within the continuous production cycle, the input-output data matrix constructed from real-time sliding window data can be directly regarded as a data-driven description, thus realizing the intelligent data-driven control logic of the aluminum extruder. The state assessment information provides a new optimization objective related to the aluminum production process cycle for the optimization problem constructed by the Willems lemma extension. This allows the state assessment information to encompass the aluminum extruder's operating condition identification results and extrusion quality prediction results. Furthermore, by using data-driven methods, the aluminum extruder can not only be controlled based on short-term performance (such as tracking accuracy) but also make intelligent decisions based on long-term state (such as equipment health), achieving intelligent closed-loop control of perception and control.

[0094] For the method steps disclosed in the above embodiments, the method steps are described as a series of actions for the purpose of simplicity. However, those skilled in the art should understand that the embodiments of this application are not limited to the described order of actions, because according to the embodiments of this application, some steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also understand that the embodiments described in the specification are all preferred embodiments, and the actions involved are not necessarily necessary for the embodiments of this application.

[0095] Any flowchart or other description of a process or method can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing a particular logical function or process. Furthermore, the scope of preferred embodiments of the invention includes additional implementations in which functions may be performed and implemented not in the order shown or discussed, including substantially simultaneously or in reverse order according to the functions involved, or in accordance with program structures such as loops, branches, etc., as will be readily understood by those skilled in the art when implementing the embodiments of this application.

[0096] like Figure 6 As shown, this application embodiment, based on the data-driven intelligent control method for aluminum extrusion presses, also provides a corresponding data-driven intelligent control system for aluminum extrusion presses, used to implement the data-driven intelligent control method for aluminum extrusion presses in any specific embodiment of this application, including:

[0097] The multi-source time-series data processing module collects multi-source time-series data from an aluminum extrusion press within a complete extrusion cycle. It performs harmonic filtering and image feature conversion on the multi-source time-series data to generate a time-series signal structured characterization image. The time-series signal structured characterization image is used to characterize the two-dimensional depth features of the extrusion process state.

[0098] The second feature dataset generation module detects the data distribution of the two-dimensional depth features, identifies the first feature dataset with data imbalance caused by changes in working conditions, and uses a generative adversarial network and an oversampling method based on dynamic time warping to augment the first feature dataset and generate the second feature dataset.

[0099] The domain-invariant feature model training module trains a domain-invariant feature model using the second feature. The domain-invariant feature model is trained through adversarial learning. After training, the domain-invariant feature model is used to extract the corresponding invariant features from the input data.

[0100] The multi-objective optimization function construction module uses the domain-invariant feature model for real-time state monitoring and combines the state evaluation information output by the domain-invariant feature model with a parameterless predictive control framework. In each control cycle of the parameterless predictive control framework, a multi-objective optimization function is constructed based on the input and output data of the real-time sliding window. The optimization objectives of the multi-objective optimization function include unit energy consumption and aluminum extrusion efficiency, and the constraints include safety thresholds of process parameters.

[0101] The aluminum extrusion press actuator control module obtains the optimal control sequence in the future time domain based on the multi-objective optimization function; it then sends the parameter control quantities of the optimal control sequence to the aluminum extrusion press actuator, thereby enabling real-time control of the aluminum extrusion process.

[0102] The implementation methods of the system described above are merely illustrative. For example, the various functional modules, units, or subsystems within the system may or may not be physically separate, or they may or may not be physical units; that is, they may be located in the same place or distributed across multiple different systems and their subsystems or modules. Those skilled in the art can select some or all of the functional modules, units, or subsystems to achieve the objectives of the embodiments of the present invention according to actual needs. Those skilled in the art can understand and implement the above-described situations without any creative effort.

[0103] This application also provides an aluminum extrusion press, which is equipped with a data-driven intelligent control system for aluminum extrusion press as described in any specific embodiment of this application.

[0104] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of the specification of the embodiments of this application.

[0105] In the description of the embodiments of this application, the reference to terms such as "an embodiment," "example," "specific example," etc., means that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0106] Furthermore, the technical solutions of the various implementation methods in this application can be combined with each other, but only if they are based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such combination of technical solutions does not exist and is not within the scope of protection claimed by the embodiments of this application.

[0107] All features disclosed in the embodiments of this application, or all steps in the disclosed methods or processes, may be combined in any way, except for mutually exclusive features and / or steps. Any feature disclosed in the specification of the embodiments of this application, unless specifically stated otherwise, may be replaced by other equivalent or similar alternative features. That is, unless specifically stated otherwise, each feature is merely one example of a series of equivalent or similar features. Throughout the specification, the same reference numerals indicate the same elements.

[0108] Those skilled in the art will understand that modules in the device of the embodiments can be adaptively changed and placed in one or more devices different from that embodiment. Modules, units, or components in the embodiments can be combined into a single module, unit, or component, and further, they can be divided into multiple sub-modules, sub-units, or sub-components. Except where at least some of such features and / or processes or units are mutually exclusive, any combination can be used to combine all features disclosed in this specification of embodiments (including the corresponding claims, abstract, and drawings) and all processes or units of any method or device so disclosed. Unless expressly stated otherwise, each feature disclosed in this specification of embodiments (including the corresponding claims, abstract, and drawings) may be replaced by an alternative feature that serves the same, equivalent, or similar purpose.

[0109] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the embodiments of this application, and are not intended to limit them. Although the embodiments of this application have been described in detail with reference to the foregoing specific embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein, and such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the specific embodiments of this application.

Claims

1. A data-driven intelligent control method for an aluminum extrusion press, characterized by, The method comprises the following steps: Collecting multi-source time series data of an aluminum extrusion machine in a complete extrusion cycle, performing harmonic filtering and image feature conversion on the multi-source time series data to generate a time series signal structured representation image, and using the time series signal structured representation image to represent two-dimensional deep features of the extrusion process state; Detecting the data distribution of the two-dimensional deep features, identifying a first feature data set caused by changes in working conditions, and using a generative adversarial network and a dynamic time warping-based oversampling method to perform data enhancement on the first feature data set to generate a second feature data set; Training a domain-invariant feature model using the second feature data set, training the domain-invariant feature model through adversarial learning, and using the trained domain-invariant feature model to extract corresponding invariant features from input data; Using the domain-invariant feature model for real-time condition monitoring, combining the condition assessment information output by the domain-invariant feature model with a parameter-free predictive control framework, constructing a multi-objective optimization function based on real-time sliding window input and output data at each control beat of the parameter-free predictive control framework, and using the multi-objective optimization function to optimize unit energy consumption and aluminum extrusion efficiency as optimization objectives and process parameter safety thresholds as constraint conditions; Obtaining an optimal control sequence in a future time domain according to the multi-objective optimization function, and sending the parameter control quantity of the optimal control sequence to an aluminum extrusion machine actuator to control the aluminum extrusion process in real time.

2. The data-driven intelligent control method for an aluminum extrusion press of claim 1, wherein, In the process of generating the time series signal structured representation image: Based on a sensor array arranged at key nodes of an industrial device, multi-channel running time series signals are synchronously collected at a fixed sampling period to obtain an initial one-dimensional time series data sequence; A normalization and denoising process is performed on the initial one-dimensional time series data sequence by a signal preprocessing subroutine to generate a standardized time series signal; The standardized time series signal is processed by a Gram angle field image generation algorithm, each data point is mapped to a polar coordinate system, and the triangular sum angle field and the triangular difference angle field between any two points are calculated to generate a first type of time series signal structured representation image.

3. The data-driven intelligent control method for an aluminum extrusion press of claim 1, wherein, In the process of generating the time series signal structured representation image: Based on a sensor array arranged at key nodes of an industrial device, multi-channel running time series signals are synchronously collected at a fixed sampling period to obtain an initial one-dimensional time series data sequence; A normalization and denoising process is performed on the initial one-dimensional time series data sequence by a signal preprocessing subroutine to generate a standardized time series signal; The standardized time series signal is processed by a Markov transition field image generation algorithm, the signal value domain is discretized into multiple state partitions, and the transition probability between state partitions along the time axis is calculated to generate a second type of time series signal structured representation image.

4. The data-driven intelligent control method for an aluminum extrusion press of claim 1, wherein, The detection of the data distribution of the two-dimensional deep features, the identification of a first feature data set caused by changes in working conditions, and the data enhancement of the first feature data set using a generative adversarial network and a dynamic time warping-based oversampling method to generate a second feature data set further comprises: Based on the preset feature space distribution evaluation model, the data point clustering of the input two-dimensional deep feature is detected, and by calculating the sample density ratio of each cluster, a first feature region with a sample number less than a threshold due to working condition changes is identified; By constructing a data enhancement branch of a generative adversarial network, a generator receives a random sampling vector from the first feature region, generates a synthetic feature sample through a deconvolution layer, a discriminator compares the distribution difference between the real feature and the synthetic feature sample, and by alternately optimizing the parameters of the generator and the discriminator, the first type of enhanced feature is iteratively generated; The optimal planning path between each synthetic feature sample and its nearest neighbor sample is calculated using the dynamic time warping algorithm, and the second type of enhanced feature is generated by performing linear interpolation on the optimal planning path; The first type of enhanced feature and the second type of enhanced feature are merged into the original two-dimensional deep feature, and through feature space re-normalization processing, a second feature dataset with balanced distribution is generated.

5. The data-driven intelligent control method for an aluminum extrusion press of claim 1, wherein, The second feature dataset is used to train a domain-invariant feature model, and the domain-invariant feature model is trained through adversarial learning. The trained domain-invariant feature model is used to extract corresponding invariant features from input data, further comprising: Based on the preprocessed second feature dataset, the structure parameters of the domain-invariant feature model are initialized, the second feature dataset is input into the feature extractor through the forward propagation channel, the high-dimensional feature representation of the second feature dataset is extracted using the multi-layer convolutional neural network, and the primary feature vector is obtained; The primary feature vector is input into the domain classifier and the task classifier respectively, the classification probability of the domain label is calculated through the fully connected layer in the domain classifier, and the prediction result of the task label is calculated through the softmax layer in the task classifier; Using the domain classification loss function and the task classification loss function, based on the difference between the prediction result of the task label and the real task label, the task classification loss value is calculated; The weight parameters of the feature extractor and the task classifier are updated using the task classification loss value, the task classification error of the domain-invariant feature model is minimized, and the domain discrimination accuracy of the domain classifier is maximized; The parameter updating process of the adversarial learning training is repeatedly executed, and the optimized domain-invariant feature model is obtained. The new input data to be processed is input into the trained domain-invariant feature model, and the domain-invariant feature vector is output through the forward propagation channel.

6. The data-driven intelligent control method for an aluminum extrusion press of claim 5, wherein, During the process of updating the weight parameters of the feature extractor and the task classifier using the task classification loss value, the network parameters of the domain classifier are fixed in advance through the gradient backpropagation mechanism; The parameters of the feature extractor and the task classifier are fixed, the weight parameters of the domain classifier are updated using the domain discrimination loss value in the domain classification loss function, and the domain discrimination accuracy of the domain classifier is maximized; When the weighted sum of the domain discrimination loss value and the task classification loss value reaches the convergence threshold, the training is stopped, and the optimized domain-invariant feature model is obtained.

7. The data-driven intelligent control method for an aluminum extrusion press of claim 1, wherein, The domain-invariant feature model is used for real-time state monitoring, and the state evaluation information output by the domain-invariant feature model is combined with a parameter-free predictive control framework, at each control beat of the parameter-free predictive control framework, a multi-objective optimization function is constructed based on real-time sliding window input-output data, further comprising: The trained domain-invariant feature model is deployed in an online monitoring module, and the real-time collected system input data is processed through the domain-invariant feature model to extract a domain-invariant feature vector, and the state evaluation information is calculated based on the domain-invariant feature vector to generate real-time state evaluation information; In each control beat in the parameter-free predictive control framework, an input-output data matrix is constructed based on fixed-length real-time sliding window data, and the real-time state evaluation information is used as a reference quantity of an optimization objective to construct a multi-objective optimization function.

8. The data-driven intelligent control method for an aluminum extrusion press of claim 7, wherein, The state evaluation information includes aluminum extrusion machine working condition recognition results and aluminum extrusion machine extrusion quality prediction results.

9. A data-driven aluminum extrusion press intelligent control system for implementing the data-driven aluminum extrusion press intelligent control method of any one of claims 1 to 8, characterized by, Comprising: A multi-source time series data processing module acquires multi-source time series data of an aluminum extrusion machine in a complete extrusion cycle, performs harmonic filtering and image feature conversion on the multi-source time series data to generate a time series signal structured representation image, and the time series signal structured representation image is used to represent two-dimensional deep features of the extrusion process state; A second feature data set generation module detects the data distribution of the two-dimensional deep features, identifies a first feature data set that has a data imbalance problem caused by working condition changes, and uses a generative adversarial network and a dynamic time warping-based oversampling method to perform data enhancement on the first feature data set to generate a second feature data set; A domain-invariant feature model training module trains a domain-invariant feature model using the second feature data set, the domain-invariant feature model is trained through adversarial learning, and the trained domain-invariant feature model is used to extract corresponding invariant features from input data; A multi-objective optimization function construction module uses the domain-invariant feature model for real-time state monitoring, and combines the state evaluation information output by the domain-invariant feature model with a parameter-free predictive control framework, at each control beat of the parameter-free predictive control framework, a multi-objective optimization function is constructed based on real-time sliding window input-output data, the optimization objectives of the multi-objective optimization function include unit energy consumption and aluminum extrusion efficiency, and the constraint conditions include safety thresholds of process parameters; An aluminum extrusion machine actuator control module obtains an optimal control sequence for a future time domain according to the multi-objective optimization function, and sends the parameter control quantity of the optimal control sequence to the aluminum extrusion machine actuator to control the aluminum extrusion process in real time through the aluminum extrusion machine actuator.

10. An aluminum extrusion press characterized by, The aluminum extrusion machine is provided with the data-driven aluminum extrusion machine intelligent control system of claim 9. The aluminum extrusion machine is provided with the data-driven aluminum extrusion machine intelligent control system of claim 9.

Images