A method and device for intelligent control of multi-process forging

Abstract

This invention provides an intelligent control method and device for multi-process forging, relating to the field of intelligent manufacturing technology. The method includes: constructing a process state sequence covering multiple forging processes; obtaining simulated forging quality corresponding to different combinations of process parameters through simulation; calculating the difference between the simulated and actual forging quality to form a perturbation variable; establishing a process state sequence sample library and training a long short-term memory network model to form a dynamic prediction model; obtaining the actual process parameters of the completed initial processes; and predicting and evaluating subsequent unexecuted processes and finding an optimized parameter set under the constraints of desired forging quality and minimizing perturbation variables using the dynamic prediction model; calculating production perturbation variables after production completion; and triggering an initial process parameter correction mechanism to re-execute the full process optimization when the quality of the produced forgings is not within the acceptable range. This invention can improve the accuracy of multi-process forging process control by combining with the actual production process.

Description

Technical Field

[0001] This invention relates to the field of intelligent manufacturing technology, specifically to an intelligent control method and device for multi-process forging. Background Technology

[0002] Multi-stage forging is widely used in the forming of large, critical components in the aerospace, automotive, and equipment manufacturing industries. It involves gradually changing the shape and internal structure of the billet through multiple continuous plastic deformation processes to obtain forged products that meet structural strength and dimensional accuracy requirements. In traditional production models, multi-stage forging process parameters are typically determined by empirical rules, trial molding results, or offline numerical simulations, and each forging step is executed according to preset parameters during production. With the development of computer-aided engineering technology, some studies have attempted to use CAE simulation models to simulate the forging process, predict the forming quality of forgings under different combinations of process parameters, and establish a mapping relationship between process parameters and forging quality by combining some production data, thereby optimizing the forging process to a certain extent. Furthermore, some technologies use data-driven methods to establish predictive models between process parameters and product quality to guide subsequent process parameter adjustments, aiming to improve forging quality stability and production efficiency.

[0003] However, in actual multi-process forging production, there is a clear temporal coupling relationship between the forging processes. The temperature, dimensional state, and material deformation state formed by the preceding processes continuously affect the forming behavior of subsequent processes. Simultaneously, random factors such as equipment condition fluctuations, ambient temperature changes, die wear, and measurement errors are common in the production process. These factors lead to non-negligible deviations between actual production results and simulation results. Existing technologies typically optimize process parameters based on ideal simulation conditions or static models, lacking a systematic characterization of the deviations between simulation results and actual production results. They also lack a mechanism to dynamically correct initial process parameters and re-optimize all process parameters based on production execution results. Consequently, if deviations or disturbances in the initial process parameters accumulate and amplify in the multi-process chain during actual production, relying solely on adjustments to subsequent process parameters is insufficient to guarantee stable and compliant final forging quality, thus affecting the overall controllability and product qualification rate of the multi-process forging process. Summary of the Invention

[0004] This invention provides a method and apparatus for intelligent control of multi-process forging, which can improve the accuracy of multi-process forging control by combining with the actual production process.

[0005] In a first aspect of the present invention, a method for intelligent control of a multi-stage forging process is provided, the method comprising: A process state sequence is constructed around multiple forging processes. The simulated forging quality corresponding to different combinations of process parameters is obtained through CAE simulation. The difference between the simulated forging quality and the actual forging quality is calculated in combination with the actual forging quality to form a perturbation variable. A long short-term memory network model is trained on a sample library of process state sequences to form a dynamic prediction model; During production execution, the actual process parameters of the completed initial processes are obtained and a known process state sequence is formed; Based on the known process state sequence, combined with the desired forging quality constraint and the disturbance variable minimization constraint, the dynamic prediction model is used to predict and evaluate the subsequent unexecuted processes, and the optimization algorithm is used to optimize the controllable process parameters of the subsequent processes to obtain the optimized parameter set. After continuing forging production using the optimized parameter set, the quality of the produced forgings is obtained. The production disturbance variable is calculated based on the difference between the quality of the produced forgings and the simulated forgings. If the quality of the produced forgings is not within the acceptable range, the initial process parameter correction mechanism is triggered to readjust the initial process parameters and re-execute the full process optimization.

[0006] In a second aspect of the invention, an intelligent control device for a multi-stage forging process is provided. The device is used to execute an intelligent control method for a multi-stage forging process as described in any of the above embodiments. The device includes an acquisition module, a processing module, and an output module, wherein: The acquisition module is used to construct a process state sequence around multiple forging processes, obtain the simulated forging quality corresponding to different combinations of process parameters through CAE simulation, and calculate the difference between the simulated forging quality and the actual forging quality in combination with the actual forging quality to form a disturbance variable. The processing module is used to train a long short-term memory network model on a sample library of process state sequences to form a dynamic prediction model; The processing module is used to acquire the actual process parameters of the completed initial process and form a known process state sequence during the production execution process; The processing module is used to combine the desired forging quality constraint and the disturbance variable minimization constraint with the known process state sequence, predict and evaluate the subsequent unexecuted processes through the dynamic prediction model, and optimize the controllable process parameters of the subsequent processes to obtain an optimized parameter set through the optimization algorithm. The output module is used to obtain the quality of the produced forgings after executing the optimized parameter set and continuing forging production, calculate the production disturbance variable based on the difference between the quality of the produced forgings and the simulated forgings, and trigger the initial process parameter correction mechanism to readjust the initial process parameters and re-execute the full process optimization when the quality of the produced forgings is not within the qualified range.

[0007] In a third aspect of the invention, an electronic device is provided, including a processor, a memory, a user interface, and a network interface, wherein the memory is used to store instructions, the user interface and the network interface are both used to communicate with other devices, and the processor is used to execute the instructions stored in the memory to cause the electronic device to perform the method as described in any of the preceding embodiments.

[0008] In a fourth aspect of the invention, a non-transitory computer-readable storage medium is provided, the computer-readable storage medium storing instructions that, when executed, perform the method as described in any of the preceding claims.

[0009] In summary, one or more technical solutions provided in the embodiments of the present invention have at least the following technical effects or advantages: This invention establishes a dynamic prediction model that reflects the actual operating state of a multi-process forging process by jointly modeling CAE simulation results with actual production results. It simultaneously introduces simulated forging quality and disturbance variables into the process state sequence. This dynamic prediction model not only depicts the relationship between process parameter combinations and forging quality but also the disturbance patterns of the actual production environment affecting forging quality. During production execution, a known process state sequence is constructed by acquiring the actual process parameters of the completed initial processes in real time. Based on the dynamic prediction model, subsequent unexecuted processes are predicted, evaluated, and their process parameters optimized, allowing for dynamic adjustment of process parameters based on the current actual production state. Simultaneously, after production is completed, a mechanism for correcting initial process parameters is triggered when quality fails to meet standards. This enables the system to re-optimize the initial process parameters based on actual production results and perform full-process optimization. Thus, the process control process continuously absorbs real production information and promptly corrects deviations between the model and process parameters, thereby improving the accuracy and stability of multi-process forging process control. Attached Figure Description

[0010] Figure 1 This is a flowchart illustrating an intelligent control method for multi-process forging disclosed in an embodiment of the present invention. Figure 2 This is a flowchart illustrating another intelligent control method for multi-process forging disclosed in an embodiment of the present invention; Figure 3 This is a schematic diagram of a blank disclosed in an embodiment of the present invention; Figure 4 This is a schematic diagram of a vertical offset position disclosed in an embodiment of the present invention; Figure 5 This is a schematic diagram of a module of an intelligent control device for multi-process forging disclosed in an embodiment of the present invention; Figure 6This is a schematic diagram of the structure of an electronic device disclosed in an embodiment of the present invention.

[0011] Explanation of reference numerals in the attached drawings: 501, acquisition module; 502, processing module; 503, output module; 601, processor; 602, communication bus; 603, user interface; 604, network interface; 605, memory. Detailed Implementation

[0012] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0013] In the description of the embodiments of the present invention, words such as "for example" or "for instance" are used to indicate examples, illustrations, or explanations. Any embodiment or design described as "for example" or "for instance" in the embodiments of the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Rather, the use of words such as "for example" or "for instance" is intended to present the relevant concepts in a specific manner.

[0014] In the description of the embodiments of the present invention, the term "multiple" means two or more. For example, multiple systems means two or more systems, and multiple screen terminals means two or more screen terminals. Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the indicated technical features. Thus, a feature defined with "first" or "second" may explicitly or implicitly include one or more of that feature. The terms "comprising," "including," "having," and variations thereof all mean "including but not limited to," unless otherwise specifically emphasized.

[0015] In multi-process forging, there is a clear temporal coupling relationship between each process. The temperature, dimensional state, and material deformation state formed by the preceding process will continuously affect the forming behavior of the subsequent process. At the same time, the production process is also affected by random factors such as equipment condition fluctuations, ambient temperature changes, mold wear, and measurement errors, which can easily lead to deviations between actual production results and simulation prediction results. Existing technologies usually optimize process parameters based on ideal simulation conditions or static models, lacking a systematic characterization of the deviation between simulation results and actual production. They also lack a mechanism to dynamically correct initial process parameters and re-optimize the process parameters of the entire process based on the production execution results. As a result, when deviations or disturbances in the initial process parameters accumulate gradually in the multi-process chain, it is difficult to guarantee the stable quality of the final forging by relying solely on the adjustment of subsequent process parameters, thus affecting the overall stability of the multi-process forging process and the product qualification rate.

[0016] This embodiment discloses an intelligent control method for multi-process forging, referring to... Figure 1 This includes the following steps S110-S150: This invention discloses an intelligent control method for a multi-process forging process, which is applied to a server. The server includes, but is not limited to, electronic devices such as mobile phones, tablets, wearable devices, and PCs (Personal Computers), and can also be a backend server running the intelligent control method for a multi-process forging process. The server can be implemented using a standalone server or a server cluster composed of multiple servers.

[0017] S110 constructs a process state sequence around multiple forging processes, obtains the simulated forging quality corresponding to different combinations of process parameters through CAE simulation, and calculates the difference between the simulated forging quality and the actual forging quality to form a disturbance variable by combining the actual forging quality.

[0018] When constructing a process state sequence around multiple forging processes, the process sequence of the multiple forging processes is first clarified during the process planning stage of the forging production line. A unique process identifier is established for each forging process, and a set of process parameter fields reflecting the forming behavior of that forging process is bound to the process identifier. The set of process parameter fields includes heating temperature, billet size, pressing speed, pressing amount, die closing stroke, die positioning offset, and equipment operating status parameters. Subsequently, a unified time axis is established during production, so that sensor data with different sampling frequencies can be mapped to the corresponding process window of the forging process. Within each process window, the set of process parameter fields is consistentized, missing data is filled according to a unified completion rule, abnormal sampled values ​​are eliminated or replaced according to a unified threshold rule, and multi-source sampled data of the same parameter are fused. After standardization, the set of process parameter fields corresponding to each forging process is converted into a process state vector. These process state vectors are then arranged sequentially according to the forging process order to form a process state sequence. This sequence allows the process state sequence to represent, in chronological order, the transmission relationship between the temperature, dimensional, and material deformation states of preceding forging processes and the forming behavior of subsequent forging processes. Specifically, the process state sequence describes the state changes of multiple forging processes in chronological order, while the process state vector describes the combination of process parameters for a single forging process at a specific moment.

[0019] After establishing the process state sequence, the simulated forging quality corresponding to different combinations of process parameters is obtained through a CAE simulation environment. First, a geometric model, material constitutive model, contact friction model, and thermal boundary model consistent with the actual forging production line are established in the CAE simulation system. This ensures that the die movement trajectory, heating conditions, and cooling conditions of each forging process are consistent with the set of process parameter fields, allowing each set of process parameter combinations to be mapped to a set of boundary conditions for the CAE simulation. Subsequently, a full-process chain simulation calculation is performed on each set of process parameter combinations. The simulation system outputs the intermediate forming state after each forging process and the final forming state after final forging. The final forming state is then evaluated based on a pre-established quality evaluation index system to obtain the simulated forging quality. The quality evaluation index system is used to uniformly map multi-dimensional quality indicators such as forging dimensional deviation, filling integrity, strain distribution uniformity, and residual stress level into quantifiable quality evaluation results, enabling comparison of forging quality under different process parameter combinations on the same evaluation scale. The simulated forging quality refers to the quality evaluation result obtained by CAE simulation calculation under the quality evaluation index system.

[0020] After obtaining the simulated forging quality, the corresponding actual forging quality is obtained through actual production testing. The difference between the simulated and actual forging quality is calculated to form a perturbation variable. Dimensional inspection, defect inspection, and microstructure inspection are performed on forgings produced according to the same combination of process parameters. The actual forging quality is calculated based on the same quality evaluation index system as in the simulation stage, ensuring a unified evaluation standard for both simulated and actual forging quality. Subsequently, the perturbation variable is calculated based on the difference between the simulated and actual forging quality. This perturbation variable characterizes the comprehensive impact of random factors such as equipment condition fluctuations, ambient temperature changes, die wear, and measurement errors on forging quality in the actual production environment. The expression for the perturbation variable is:

[0021]

[0022] Wherein, Δ represents the disturbance variable, which is used to characterize the degree of deviation between the simulated forging quality and the actual forging quality under the same combination of process parameters. The value of Δ can be positive or negative. When Δ is positive, it means that the simulated forging quality is higher than the actual production quality. When Δ is negative, it means that the simulated forging quality is lower than the actual production quality. This represents the quality of the simulated forging. Its value is calculated from the CAE simulation results under the quality evaluation index system. The calculation method is to weight and fuse multiple quality indicators according to preset weights. This represents the actual forging quality, and its value is calculated from the actual production and testing results under the same quality evaluation index system. and It is comparable on the same evaluation scale. The principle of this expression is to represent the impact of random uncertainties that are difficult to model directly in the actual production process on the quality of forgings as the difference between the simulated quality results and the actual quality results, so that the disturbance variable can serve as an important parameter to characterize the impact of random disturbances in subsequent data modeling and process optimization.

[0023] To further characterize the influence of the disturbance variable, the absolute value of the disturbance variable can also be calculated to describe the magnitude of the deviation. The expression for the disturbance magnitude is:

[0024] in, It represents the disturbance amplitude, used to describe the absolute magnitude of the deviation between the simulated forging quality and the actual forging quality, so that the disturbance intensity can still be compared on the same scale even when different production batches have different deviation directions. The value of is obtained by taking the absolute value of Δ, when The larger the value, the more significant the impact of random disturbance on the quality of the forging. The smaller the value, the higher the consistency between the simulation model and actual production. Statistical analysis of the perturbation variables and their amplitudes can provide a data foundation for training subsequent dynamic prediction models and optimizing process parameters, thereby improving the accuracy and stability of intelligent control of multi-process forging.

[0025] S120, train a long short-term memory network model on a sample library of process state sequences to form a dynamic prediction model.

[0026] In one possible implementation, a long short-term memory network model is trained on a process state sequence sample library to form a dynamic prediction model. Specifically, this includes: organizing process parameter combinations, simulated forging quality, and perturbation variables into a process state sequence according to the forging process sequence; constructing a process state sequence sample library from multiple sets of process state sequences; inputting the process state sequence sample library into the long short-term memory network model for training, enabling the long short-term memory network model to learn the temporal mapping relationship between the process state sequence, forging quality, and perturbation variables; and forming a dynamic prediction model after the model training is completed, so that the dynamic prediction model can output the predicted forging quality and the corresponding predicted perturbation variables when the process state sequence is input.

[0027] Specifically, when organizing process parameter combinations, simulated forging quality, and disturbance variables into a process state sequence according to the forging process order, a forging process sequence template is first solidified. This ensures that each forging process has a unique process identifier and a fixed process position within the process chain. A fixed set of fields is then configured for each forging process to carry the corresponding subset of process parameter fields, simulated forging quality fields, and disturbance variable fields, thereby ensuring consistent field structure across the entire process chain for the same process state sequence. The technical term "process state sequence" refers to organizing the state information of multiple forging processes in chronological order into a continuously readable temporal data structure. This allows the subset of process parameter combinations and quality-related fields from preceding forging processes to form a traceable cumulative association in subsequent forging processes through their process positions. In terms of specific organization, a single-process state record is constructed for each forging process, and these records are then concatenated according to the forging process order. This ensures that each process state sequence corresponds to a complete set of process parameter combinations, while maintaining the binding relationship between process parameter combinations, simulated forging quality, and disturbance variables within the same process state sequence.

[0028] When constructing a process state sequence sample library from multiple sets of process state sequences, multiple process state sequences are generated around different combinations of process parameters. Each process state sequence is assigned a sample identifier and a source identifier. The sample identifier is used to distinguish samples corresponding to different combinations of process parameters, and the source identifier is used to distinguish whether the sample comes from a CAE simulation batch or an actual production batch. The technical meaning of the process state sequence sample library is a set of samples used to train and validate Long Short-Term Memory (LSTM) network models. It uses process state sequences as basic storage units, and each storage unit maintains a fixed forging process sequence, a fixed set of process parameter fields, a fixed simulated forging quality field, and a fixed perturbation variable field. During the construction process, structural consistency checks are performed on all process state sequences to ensure that field naming, field meaning, field position, and the number of processes remain consistent within the sample library, thereby guaranteeing that the sample library can be read in batches and used for subsequent model training.

[0029] When training a Long Short-Term Memory (LSTM) network model by inputting a sample library of process state sequences, the library is first divided into a training subset and a validation subset. Each process state sequence in the training subset is then converted into a sequence input tensor that can be received by the LTM network model. The time steps of the sequence input tensor are aligned with the forging process sequence, and the feature vector of each time step is obtained by concatenating the process parameter field subset, the simulated forging quality field, and the perturbation variable field in a fixed order. The technical term for the LTM network model is a recurrent neural network structure with gated memory units. It selectively retains and updates historical information through input gates, forget gates, and output gates, enabling the model to capture the temporal dependencies of preceding forging processes on subsequent forging processes across multiple time steps. The technical term for the temporal mapping relationship is the functional relationship between the process state sequence and the forging quality and perturbation variables. This functional relationship simultaneously reflects the cumulative effect of process parameter combinations along the forging process chain and the propagation law of perturbation variables along the forging process chain. During training, forward propagation is performed on the sequence input tensor to obtain the predicted forging quality and predicted perturbation variables. The network parameters are iteratively updated based on the prediction error, so that the long short-term memory network model gradually converges to be able to infer the joint temporal mapping relationship between forging quality and perturbation variables based on the process state sequence.

[0030] When a dynamic prediction model is formed after model training, the network structure of the Long Short-Term Memory (LSTM) network and the trained network parameters are solidified into a reasoning-enabled model version. A model inference interface consistent with the process state sequence data structure is established, enabling the dynamic prediction model to output predicted forging quality and corresponding predicted perturbation variables when receiving any process state sequence. The technical terminology of a dynamic prediction model is a time-series prediction model deployable in a production environment. Its input is a process state sequence organized according to the forging process order, and its output is the predicted forging quality and predicted perturbation variables consistently bound to that process state sequence. In specific deployment, the dynamic prediction model is interface-bound with the process state sequence generation process, allowing the production side to directly call the dynamic prediction model to obtain the predicted forging quality and predicted perturbation variables after forming the process state sequence. The predicted output is then used for subsequent optimization of controllable process parameters and triggering of the initial process parameter correction mechanism.

[0031] In one possible implementation, the process state sequence sample library is input into the Long Short-Term Memory (LSTM) network model for training. Specifically, this includes: performing structural consistency processing on each process state sequence in the sample library to ensure a uniform structure across forging process sequences, process parameter fields, simulated forging quality fields, and perturbation variable fields; and performing length alignment processing on each process state sequence to form a processed process state sequence; constructing temporal feature vectors from the process state features corresponding to each processed process state sequence, and concatenating them according to the forging process sequence to form a process state sequence matrix; dividing the process state sequence sample library into a training subset and a validation subset, and inputting the training subset into the LTM network model to train the LTM network model according to the forging process sequence. The process sequentially reads the process state characteristics and uses a gating structure to memorize and update historical process states, thus forming a comprehensive representation of forging quality and perturbation variables at the final time step. A joint output structure for predicting forging quality and predicting perturbation variables is established at the output of the Long Short-Term Memory (LSTM) network model. The network parameters are iteratively updated by constraining the first and second biases, enabling the LTM network model to learn the joint temporal mapping relationship between the process state sequence and forging quality and perturbation variables. The first bias is the deviation between the predicted forging quality and the simulated forging quality, and the second bias is the deviation between the predicted perturbation variables and the perturbation variables. After the network parameters reach the preset convergence condition, the parameters of the LTM network model are solidified to form a dynamic prediction model.

[0032] Specifically, when performing structural consistency processing on each process state sequence in the process state sequence sample library, a unified forging process sequence template is first solidified so that each process state sequence is arranged in chronological order with the same set of process identifiers. A fixed set of process parameter fields, simulated forging quality fields, and disturbance variable fields are established for each forging process, thereby ensuring that the process state sequences generated by different batches and different combinations of process parameters are completely consistent in terms of field names, field meanings, and field positions. The technical term for structure consistency processing is to enable data from different sources or batches to be directly spliced ​​and read in batches at the structural level, avoiding different interpretations of the same semantics by the Long Short-Term Memory Network Model due to missing fields, inconsistent field order, or field naming drift. The technical term for length alignment processing is to ensure that the number of time steps in each process state sequence is consistent with the number of forging processes. When there are actual missing processes, abnormal interruptions, or incomplete sampling, a unified rule is used to fill in the missing process state with placeholder process states and write them into the missing marker field. At the same time, excessively long sequences are pruned according to the boundaries of the forging processes, so that the final processed process state sequence has a fixed time step length and a fixed field structure, so that the Long Short-Term Memory Network Model can be trained with a consistent input shape.

[0033] When constructing time-series feature vectors from the process state features corresponding to the state sequences of each processing step, a subset of the process parameter fields for each forging step is extracted and combined with the simulated forging quality field and perturbation variable field of that step in a fixed splicing order to form a single-step feature vector. This single-step feature vector is used to characterize the process state and quality state at that time step. Subsequently, the single-step feature vectors of each forging step are arranged sequentially according to the forging step order to form a process state sequence matrix. The row index of the process state sequence matrix corresponds one-to-one with the forging step order, and the column vectors maintain the same feature semantics and the same value caliber throughout the entire process chain. The technical meaning of time-series feature vector is that the state information of the process is expressed by a fixed-dimensional vector at each time step, thereby allowing the Long Short-Term Memory network model to read and learn cross-process dependencies along the time steps. The technical meaning of process state sequence matrix is ​​that the data of the entire process chain is organized into a two-dimensional tensor that can be directly consumed by the recurrent structure model, enabling the model to simultaneously capture the feature coupling within the process and the cumulative effect between processes.

[0034] When dividing the process state sequence sample library into training and validation subsets, sampling is first performed by batch identifier or process parameter combination identifier to avoid similar samples from the same batch appearing in both the training and validation subsets, which could lead to an overly optimistic evaluation. Then, the process state sequence matrix of the training subset is fed into the Long Short-Term Memory (LSTM) network model in batches, so that the LTM network model reads the process state features step by step according to the forging process sequence, and remembers and updates the historical process states through a gating structure, thereby forming a comprehensive representation of the forging quality and perturbation variables at the final time step. The technical terminology of Long Short-Term Memory (LSTM) network models refers to the use of input gates, forget gates, and output gates to control the information flow of cell states, enabling the model to retain memories of key historical states and suppress irrelevant noise across multiple time steps. The technical terminology of gating structures refers to the use of learnable gates to dynamically determine how much historical information to retain and how much new information to write at different time steps, thereby allowing the influence of the temperature state, geometric state, or deformation history of previous processes on subsequent processes to propagate across time steps through cell states. The technical terminology of the terminal time step refers to the time step corresponding to the last forging process in the sequence, whose hidden state and cell state aggregate the temporal information of the entire process chain, making it suitable as a global representation for predicting forging quality and predicting perturbation variables.

[0035] When establishing a joint output structure for predicting forging quality and predicting perturbation variables at the output end of a Long Short-Term Memory (LSTM) network model, a shared representation layer is constructed based on the hidden state at the terminal time step. This shared representation layer receives the hidden state at the terminal time step and outputs a shared feature vector. Then, a forging quality output branch and a perturbation variable output branch are set in parallel on top of the shared feature vector. The forging quality output branch generates the predicted forging quality, and the perturbation variable output branch generates the predicted perturbation variable. Both branches share the same temporal encoder parameters during the training phase to ensure consistent interpretation of the same process state sequence. The technical meaning of the joint output structure is that multiple interrelated target quantities are output simultaneously in the same model, and the correlation between the target quantities is internalized into the model parameters through shared representation. This allows the model to learn the propagation law of perturbation variables while learning the variation law of forging quality, avoiding the problems of inconsistent representations or mutual amplification of errors caused by training two models separately.

[0036] When iteratively updating network parameters by constraining the first and second biases, forward propagation is first performed on the state sequence matrix of each process step on the training subset to obtain the predicted forging quality and predicted perturbation variable. These are then compared with the simulated forging quality and perturbation variable corresponding to that sample to form the first and second biases, respectively. The first and second biases are then mapped to a joint loss, and backpropagation is performed on the joint loss, causing the gradient to be propagated back along the time step direction to the gating parameters and shared representation parameters of the Long Short-Term Memory (LSTM) network model to update the network parameters. The expression for the joint loss is:

[0037]

[0038] in, This represents the joint loss, used to simultaneously constrain the prediction error of forging quality and the prediction error of disturbance variables. The smaller the value, the more consistent the model's joint prediction of the process state sequence. This represents the weight of forging quality loss, used to balance the proportion of forging quality prediction error in the combined loss. The value range is set according to the priority of quality compliance and must be a non-negative number. This represents the loss weight of the perturbation variable, used to balance the proportion of the prediction error of the perturbation variable in the joint loss. The value range is set according to the perturbation suppression priority and must be a non-negative number. This indicates the predicted forging quality, generated from the forging quality output branch. The mass of the simulated forging is given by CAE simulation or a sample library record; This indicates the predicted perturbation variable, generated from the perturbation variable output branch; The perturbation variable is defined by the difference between the simulated forging quality and the actual forging quality and recorded in the sample library. The principle of the joint loss is to weightedly sum the squared deviations of the two types of objectives, so that the network parameter updates converge simultaneously in the direction of reducing the first deviation and reducing the second deviation. The contribution of the two types of deviations to the gradient is controlled by the weights, thereby enabling the Long Short-Term Memory network model to learn the joint temporal mapping relationship from the process state sequence to the forging quality and the perturbation variable.

[0039] After the network parameters reach the preset convergence condition, the parameters of the long short-term memory network model are solidified and a dynamic prediction model is formed. The changing trends of the joint loss, forging quality output error and perturbation variable output error on the validation subset are continuously monitored. The preset convergence condition is defined as the joint loss of the validation subset no longer significantly decreases or the decrease is lower than the preset threshold within a certain number of consecutive evaluation periods. At the same time, the forging quality output error and perturbation variable output error of the validation subset both meet the preset accuracy threshold. When the preset convergence condition is met, the current network parameters are written as stable parameters into the model version identifier and the weight update is frozen. This ensures that the model version maintains output consistency in subsequent production calls. The model inference interface corresponding to the stable parameter is bound to the data interface of the process state sequence matrix to form a dynamic prediction model that can be deployed on an industrial control computer. The technical term for preset convergence conditions is to determine the stopping criteria used to determine that the training process has reached a stable generalization ability, so as to avoid overfitting and prediction distortion in new batch production. The technical term for fixed parameters is to switch the model weights from a trainable state to a read-only state and generate a deployable version, so as to ensure that the model behavior is traceable and reproducible when called on the production side, and to facilitate the formation of a version iteration chain with subsequent incremental updates.

[0040] S130: During the production execution process, the actual process parameters of the completed initial process are obtained and a known process state sequence is formed.

[0041] During production execution, when acquiring the actual process parameters of the completed initial process, a sensor interface and equipment data interface for communication with the industrial control computer are first configured for each forging process in the forging production line. This allows each forging process to continuously record the corresponding process parameter status during execution and generate a complete process status record upon completion. Actual process parameters refer to the process parameter values ​​recorded by the equipment control system or sensor system during actual production. These include, but are not limited to, billet temperature, billet end length, vertical offset, pressing speed, pressing amount, die contact position, and other measurable parameters that characterize the execution status of the process. The initial process refers to the set of processes that have been completed in the complete forging process chain and precede subsequent processes. It is usually a series of preceding forging processes, the number of which is determined by the current production execution progress. When a forging process is completed, the industrial control computer reads the corresponding actual process parameters through the equipment communication protocol and writes them into the production record cache. This ensures that the actual process parameters are consistent with the forging process sequence in terms of time. Simultaneously, a process identifier and batch identifier are bound to each set of actual process parameters to ensure that production data from different batches are not confused during subsequent processing.

[0042] After obtaining the actual process parameters of the completed initial processes, the actual process parameters corresponding to each process are organized into a known process state sequence according to the forging process sequence. This sequence represents the complete process state information formed by the current production batch in the executed process stage. The technical term "known process state sequence" means a partial process state sequence built in real time during the forging production process as the processes are gradually completed. Its time step only covers the executed forging processes, and forging processes that have not yet been executed do not have corresponding state records in this sequence. To ensure that the known process state sequence is consistent with the input structure of the dynamic prediction model, the same data field structure as the process state sequence sample library is used during the construction process. Each forging process's corresponding sequence element includes a subset of process parameter fields, placeholders for simulated forging quality fields, and placeholders for disturbance variable fields, maintaining the field order and naming consistent with the sample library. For forging processes that have not yet been executed, no corresponding process state records are generated during the sequence construction stage. This ensures that the known process state sequence truly reflects the current production execution progress and provides the basic data structure for the subsequent dynamic prediction model to infer the final forging quality and disturbance variables under partial sequence input conditions.

[0043] S140, based on the known process state sequence, combined with the desired forging quality constraint and the disturbance variable minimization constraint, predicts and evaluates subsequent unexecuted processes through a dynamic prediction model, and optimizes the controllable process parameters of subsequent processes through an optimization algorithm to obtain an optimized parameter set.

[0044] In one possible implementation, based on the known process state sequence and combined with the desired forging quality constraint and the constraint of minimizing disturbance variables, a dynamic prediction model is used to predict and evaluate subsequent unexecuted processes. Specifically, this includes: establishing a prediction and evaluation constraint system based on the known process state sequence, wherein the prediction and evaluation constraint system consists of the desired forging quality constraint and the constraint of minimizing disturbance variables; determining the set of subsequent unexecuted processes based on the forging process position of the known process state sequence; constructing a set of candidate process parameter combinations corresponding to the set of unexecuted processes; concatenating the set of candidate process parameter combinations with the known process state sequence to form a set of candidate process state sequences; sequentially inputting the set of candidate process state sequences into the dynamic prediction model to obtain the corresponding predicted forging quality and predicted disturbance variables, and predicting and evaluating each candidate process state sequence according to the prediction and evaluation constraint system; comprehensively ranking the candidate process state sequences based on the deviation between the predicted forging quality and the desired forging quality and the magnitude of the predicted disturbance variables, and determining the candidate process state sequence that satisfies the desired forging quality constraint as the preferred process state sequence.

[0045] Specifically, when establishing a predictive evaluation constraint system based on a known process state sequence, the expected forging quality constraint, consistent with the quality evaluation index system, is first solidified in the industrial control computer. Simultaneously, a disturbance variable minimization constraint, consistent with the scope of the disturbance variable, is also solidified, allowing both types of constraints to simultaneously apply to the output of the dynamic prediction model under the same evaluation caliber. The technical terminology of the predictive evaluation constraint system is a set of constraints used for the unified evaluation of candidate process state sequences, which includes at least the expected forging quality constraint and the disturbance variable minimization constraint. The technical terminology of the expected forging quality constraint is the limitation on the target range or target value of the predicted forging quality, used to ensure that the final forging quality meets the acceptable range or target deviation requirements. The technical terminology of the disturbance variable minimization constraint is the constraint that suppresses the magnitude of the predicted disturbance variable, used to reduce the impact of random uncertainties on quality fluctuations and improve production stability. To facilitate unified scoring, the predictive evaluation constraint system is implemented as a computable set of constraint functions. This allows each candidate process state sequence, after being processed by the dynamic prediction model to output the predicted forging quality and predicted disturbance variable, to obtain quantitative results under the same rules regarding whether the constraints are satisfied and the degree of constraint violation.

[0046] When determining the set of subsequent unexecuted processes based on the forging process position of the known process state sequence, the process identifier and process position index of the last process state record in the known process state sequence are used to query the pre-fixed forging process sequence template to obtain the set of process identifiers for which no process state records have been generated yet, and this set of process identifiers is identified as the set of unexecuted processes. The technical term for "forging process position" refers to the coverage boundary of the known process state sequence in the forging process sequence template, used to define the currently executed completed process segments and unexecuted process segments. The technical term for "set of unexecuted processes" refers to the set of process identifiers located after the forging process position in the forging process sequence template, whose internal order is consistent with the forging process sequence template, ensuring temporal continuity when subsequently constructing the set of candidate process parameter combinations and the set of candidate process state sequences.

[0047] When constructing the candidate process parameter combination set corresponding to the set of unexecuted processes, firstly, a subset of controllable process parameter fields is determined for each forging process in the set of unexecuted processes. Then, parameter boundaries, discrete granularity, or sampling distribution are defined for each subset of controllable process parameter fields. The parameter boundaries are jointly determined by equipment capacity, process window, and safety constraints, and are solidified in the industrial control computer in the form of a parameter configuration table. The technical terminology of the candidate process parameter combination set refers to multiple sets of feasible process parameter combinations generated for the set of unexecuted processes under parameter boundary constraints. Each element is a complete process parameter combination fragment covering the set of unexecuted processes. Specifically, multiple sets of candidate process parameter combinations can be obtained using grid sampling, Latin hypercube sampling, or random sampling. Feasibility verification is performed on each candidate process parameter combination to eliminate combinations that violate parameter coupling constraints, ensuring that the candidate process parameter combination set only contains feasible combinations that can be mapped to equipment control commands. This guarantees that the subsequently assembled candidate process state sequence has engineering executability.

[0048] When concatenating the set of candidate process parameter combinations with the known process state sequence to form a set of candidate process state sequences, a corresponding unexecuted process state fragment is generated for each candidate process parameter combination in the set of candidate process parameter combinations. This unexecuted process state fragment is then appended to the end of the known process state sequence according to the forging process sequence, thus forming a candidate process state sequence covering the entire process chain. The technical terminology of a set of candidate process state sequences refers to a collection of multiple candidate process state sequences, each of which is determined by the known process state sequence and a set of candidate process parameter combinations. The technical terminology of concatenation refers to connecting two sequences along the time step dimension according to the forging process sequence template, maintaining consistency in field structure and meaning, so that the dynamic prediction model can read the entire process chain information under the same input interface. To ensure input structure consistency, a unified or predictive placeholder strategy is used for the simulated forging quality field and the disturbance variable field in the unexecuted process state fragment, and the placeholder strategy is kept consistent within the set of candidate process state sequences, so that the dynamic prediction model's interpretation of the unexecuted process state does not drift due to changes in placeholder caliber.

[0049] The candidate process state sequences are sequentially input into the dynamic prediction model to obtain the corresponding predicted forging quality and predicted disturbance variables. When predicting and evaluating each candidate process state sequence according to the prediction and evaluation constraint system, sequence tensor quantization is performed on each candidate process state sequence to convert it into a process state sequence matrix consistent with the input interface of the dynamic prediction model. Then, the dynamic prediction model is called to output the predicted forging quality and predicted disturbance variables, and the output results are input into the prediction and evaluation constraint system to calculate constraint satisfaction and constraint violation. The technical terminology of prediction and evaluation refers to the process of uniformly quantifying and evaluating candidate process state sequences under the prediction and evaluation constraint system. The evaluation results at least include the deviation degree of the expected forging quality constraint and the deviation degree of the disturbance variable minimization constraint. To form comparable evaluation values, the deviation degree is further mapped to a comprehensive score, the expression of which is:

[0050]

[0051] in, This represents the overall score, used to sort the candidate process state sequences. The smaller the value, the better the candidate process state sequence. This indicates the weight of forging quality deviation, used to reflect the degree of influence of the expected forging quality constraints on the comprehensive score. The value is a non-negative number and is set according to the priority of quality compliance. This represents the weight of the perturbation variable, used to reflect the degree of influence of the perturbation variable minimization constraint on the overall score. The value is a non-negative number and is set according to the perturbation suppression priority. This represents the predicted forging quality, output by the dynamic prediction model. This represents the desired forging quality, and its value is determined by the target quality requirement or the center value of the acceptable range and is consistent with the quality evaluation index system. This represents the predicted disturbance variable, output by the dynamic prediction model. The principle behind the comprehensive scoring is to uniformly map the deviation of the predicted forging quality from the expected forging quality and the magnitude of the predicted disturbance variable into a weighted and superimposed cost, so that each candidate process state sequence can be compared under the same evaluation scale, and to achieve a trade-off between the quality target and the disturbance suppression target through weighting.

[0052] When comprehensively ranking candidate process state sequences based on the deviation between predicted and desired forging quality and the magnitude of predicted disturbance variables, and determining the preferred process state sequence as one that meets the desired forging quality constraint, a feasibility screening process is first performed on the candidate process state sequences based on the desired forging quality constraint. Candidate process state sequences whose predicted forging quality does not meet the acceptable range or target deviation threshold are eliminated. Then, the remaining candidate process state sequences are ranked from smallest to largest based on their comprehensive score, and the preferred process state sequence is determined from the candidate process state sequence with the lowest comprehensive score. The technical terminology of comprehensive ranking refers to a ranking mechanism that uses both the predicted forging quality deviation and the magnitude of the predicted disturbance variable as ranking criteria. This ensures that the candidate process state sequences not only meet the desired forging quality constraint but also minimize the predicted disturbance variable while meeting the constraint. The technical term "optimized process state sequence" means that the candidate process state sequence with the best comprehensive score under the prediction and evaluation constraint system and that meets the expected forging quality constraint is used as the initial candidate for subsequent optimization algorithm search or directly as the basis for issuing controllable process parameters, so that subsequent unexecuted processes can obtain optimization guidance in terms of both quality compliance and disturbance suppression.

[0053] S150: After continuing forging production with optimized parameter set, obtain the quality of the produced forgings, calculate the production disturbance variable based on the difference between the quality of the produced forgings and the simulated forgings, and trigger the initial process parameter correction mechanism when the quality of the produced forgings is not within the qualified range to readjust the initial process parameters and re-execute the full process optimization.

[0054] In one possible implementation, after continuing forging production with the optimized parameter set, the quality of the produced forgings is obtained. Production disturbance variables are calculated based on the difference between the produced forging quality and the simulated forging quality. If the produced forging quality is not within the acceptable range, an initial process parameter correction mechanism is triggered to readjust the initial process parameters and re-execute the full process optimization. Specifically, this includes: distributing the optimized parameter set to the production equipment to drive subsequent unexecuted processes to continue forging production, and establishing a simulated forging quality corresponding to the optimized parameter set as a comparison benchmark; performing quality inspection on the completed forgings to obtain the produced forging quality, and ensuring consistent quality. The difference between the quality of the produced forgings and the simulated forgings is calculated under the evaluation index system to form a production disturbance variable. The production disturbance variable is compared with the predicted disturbance variable output by the dynamic prediction model, and the passability judgment is performed in combination with the qualified range of the produced forging quality. When the quality of the produced forgings is not within the qualified range, the initial process parameter correction mechanism is triggered, and the actual process parameters of the completed processes are frozen to construct the process state sequence constraints of the entire process. The initial process parameters are included in the optimization variable set to form the full process optimization variable set, and the candidate process parameter combination set is reconstructed based on the full process optimization variable set to perform full process optimization.

[0055] Specifically, when the optimized parameter set is sent to the production equipment to drive subsequent unexecuted processes to continue forging production, a data interface for communication with the production equipment control system is first established in the industrial control computer. This allows the optimized parameter set to be converted into a sequence of control instructions recognizable by the production equipment and written into the process control register of the production equipment step by step according to the forging process sequence. The technical term for the optimized parameter set is a complete set of process parameters covering the set of unexecuted processes, determined during the aforementioned prediction, evaluation, and optimization algorithm search process. Each set of parameters within this set corresponds one-to-one with a specific forging process and satisfies the equipment capability boundary. The production equipment refers to a hydraulic press, mechanical press, or automatic forging unit that performs the forging operation. Its control system drives subsequent unexecuted processes to continue forging production by reading parameters such as pressing speed, pressing amount, die displacement position, and billet temperature control instructions from the optimized parameter set. After the parameters are sent out, the simulated forging quality record corresponding to the optimized parameter set is simultaneously extracted from the CAE simulation database and bound to the current production batch's comparison benchmark for subsequent production result evaluation. The technical term for simulated forging quality means that the quality evaluation result obtained by CAE simulation calculation under the same combination of process parameters has the same expression as the actual quality inspection result, so that the simulated forging quality can serve as a benchmark for real production quality.

[0056] When performing quality inspection on completed forgings to obtain their quality, a testing device consistent with the quality evaluation index system is configured at the end of the production line. This allows the forgings to enter the quality inspection station after leaving the final forging process. The forgings are then inspected using dimensional measuring devices, visual inspection devices, or non-destructive testing devices, and the inspection results are recorded. The technical terminology for "production forging quality" refers to the quality evaluation value output by the testing device under a unified quality evaluation index system. Its calculation method is consistent with the simulated forging quality to ensure numerical comparability. A unified quality evaluation index system means using the same quality evaluation methods and index standards in both the simulation and inspection stages. For example, it may use indicators such as dimensional deviation, filling rate, or defect risk index, and integrate them according to pre-set weights to obtain a single quality evaluation value. After inspection, the difference between the production forging quality and the simulated forging quality is calculated in the industrial control computer to form a production disturbance variable, expressed as:

[0057]

[0058] in, This represents the production disturbance variable, used to characterize the quality deviation produced by the real production environment relative to the ideal simulation environment; This indicates the quality of the forgings produced, and its value is output by the quality inspection device. This represents the simulated forging quality, and its value is obtained from the records corresponding to the optimization parameter set in the CAE simulation database. The principle of this expression is to use the ideal simulation prediction result as the benchmark and the actual production result as the observation value. By quantifying the comprehensive deviation caused by factors such as equipment fluctuations, mold wear, and changes in environmental conditions in the production environment through the difference between the two, a production disturbance variable reflecting the actual degree of disturbance is obtained.

[0059] When comparing the consistency of production disturbance variables with the predicted disturbance variables output by the dynamic prediction model, and combining this with the acceptable quality range of forgings for compliance determination, the predicted disturbance variables output by the dynamic prediction model during the prediction and evaluation phase are first read, and a difference analysis is performed with the production disturbance variables calculated for the current batch to determine the accuracy of the predicted disturbance variables in representing the actual production disturbances. The technical term for consistency comparison is to assess the reliability of the dynamic prediction model's prediction results by comparing the magnitude of the differences between the production disturbance variables and the predicted disturbance variables, while also providing a basis for bias analysis in subsequent closed-loop learning. The acceptable quality range of forgings refers to a predefined quality interval under a unified quality evaluation index system, used to determine whether the quality of forgings meets product design requirements. The expression for compliance determination is:

[0060]

[0061] in, Indicates the quality of the forgings produced; This represents the minimum acceptable value under the quality evaluation index system; This represents the maximum acceptable value under the quality evaluation index system. If the quality of the forgings produced does not meet the above range conditions, the quality of the forgings produced is determined to be outside the acceptable range, and the process will enter the subsequent initial process parameter correction mechanism.

[0062] When the quality of forgings produced falls outside the acceptable range, triggering the initial process parameter correction mechanism and freezing the actual process parameters of completed processes to construct a process state sequence constraint for the entire process, the industrial control computer first records the actual process parameters of all executed forging processes in the current production batch and locks these actual process parameters to fixed values ​​so that they are not modified in subsequent optimization processes. The technical terminology of the initial process parameter correction mechanism is that when subsequent process optimization fails to bring the quality of forgings produced to the acceptable range, the mechanism readjusts the process parameters of the initial process to eliminate the accumulated deviations of previous processes. The technical terminology of freezing is that these actual process parameters are kept at fixed values ​​during subsequent optimization processes, thereby constructing a process state sequence constraint that reflects the actual production execution situation. Process state sequence constraints refer to constraints formed based on the actual production state of the current batch, ensuring that subsequent optimization processes are conducted under actual production conditions rather than re-assuming an ideal initial state.

[0063] When initial process parameters are incorporated into the optimization variable set to form a full-process optimization variable set, and a candidate process parameter combination set is reconstructed based on this set to perform full-process optimization, the initial process parameters are redefined as adjustable variables, expanding the scope of optimization variables from subsequent processes to the entire forging process chain. This results in the formation of the full-process optimization variable set. Technically, the full-process optimization variable set refers to a set of variables covering all controllable process parameters in the forging process, containing both initial and subsequent process parameters. Subsequently, a new candidate process parameter combination set is generated according to the parameter boundaries of the full-process optimization variable set, ensuring that each candidate process parameter combination covers the entire forging process chain. Under the condition of maintaining the process state sequence constraints, this set is input into a dynamic prediction model to perform new prediction evaluation and optimization calculations, thereby achieving full-process optimization. This allows the initial and subsequent process parameters to be adjusted synergistically to eliminate preceding deviations and improve the stability of the final forging quality.

[0064] When the quality of forgings produced is within the acceptable range, the production status is determined to be in a quality-compliant state. First, the quality of the forgings is read in the industrial control computer, and the quality judgment rules in the unified quality evaluation index system are called to perform a range judgment on the quality of the forgings, comparing it with the pre-set quality acceptable range. When the quality of the forgings meets the quality acceptable range, the production status corresponding to the current production batch is marked as a quality-compliant state, and the binding relationship between this quality-compliant state and the process status sequence, optimization parameter set, and production disturbance variables of the current production batch is recorded in the production data management system. The technical terminology of production status refers to a status identifier used to describe the overall quality status and production process stability of the current production batch, which includes at least two categories: quality-compliant state and quality-abnormal state. The technical terminology of quality-compliant state refers to the status identifier corresponding to the quality of forgings being within the quality acceptable range under the unified quality evaluation index system, indicating that the current production process parameters can stably obtain forging quality that meets the product design requirements. The judgment rule for the quality acceptable range is calculated using the following expression:

[0065]

[0066] in, This indicates the quality of the forgings produced, and its value is output by the quality inspection device. This represents the lower limit of the acceptable quality range, and its value is determined based on product design requirements and the quality evaluation index system. This represents the upper limit of the acceptable quality range, and its value is determined based on product design requirements and the quality evaluation index system. The principle behind this expression is that by comparing the quality of the produced forgings with the upper and lower limits in the quality evaluation index system, a unified judgment is made as to whether the forging quality meets the product design requirements. The judgment result then serves as the basis for selecting subsequent process parameter control strategies.

[0067] When maintaining the optimized parameter set as the current production parameter without triggering the initial process parameter correction mechanism, the industrial control computer, after confirming that the production status meets quality standards, writes the optimized parameter set used for the current production batch into the parameter locking area of ​​the production equipment control system and sets a parameter stability flag, allowing this optimized parameter set to continue to be used as the default production parameter in subsequent production batches. The technical term for the optimized parameter set is a set of process parameters covering all unexecuted processes, obtained through a dynamic prediction model and optimization algorithm optimization process. Each parameter corresponds to a specific equipment control command for a forging process. The implementation method for maintaining the optimized parameter set as the current production parameter is to keep the control command mapping relationship of this optimized parameter set unchanged in the production equipment control system, so that subsequent production batches directly call this optimized parameter set to drive the production equipment when executing the same forging process sequence, thereby maintaining the stability of the production process. The technical term "not triggering the initial process parameter correction mechanism" means that when the quality meets the standard, the system does not restart the full process optimization process, nor does it readjust the initial process parameters. Instead, it maintains the current set of optimized parameters and the current process state sequence structure, thereby avoiding the introduction of unnecessary process parameter disturbances under stable quality conditions. At the same time, it reduces production fluctuations caused by optimization calculations and equipment adjustments, and keeps the production system running continuously and stably.

[0068] In one possible implementation, after triggering the initial process parameter correction mechanism to readjust the initial process parameters and re-execute the full process optimization when the quality of the forging is not within the acceptable range, the method further includes: generating a batch identifier for each production execution; fixing the production process sequence, forging quality, and production disturbance variables under the batch identifier, and simultaneously establishing an optimization parameter set, dynamic prediction model output results, and quality judgment results bound to the batch identifier; performing structural consistency processing on the production process sequence, forging quality, and production disturbance variables to generate a new process state sequence; establishing sample weight labels for the new process state sequence; performing incremental updates on the dynamic prediction model based on the new process state sequence and corresponding sample weight labels, so that the dynamic prediction model absorbs the production disturbance information carried by the new process state sequence while maintaining the existing network structure; calculating the feasibility rate, convergence speed, prediction consistency, and disturbance suppression effect based on historical optimization decision results and automatically tuning the hyperparameters of the optimization algorithm; and linking the incremental update of the dynamic prediction model and the automatic tuning of the optimization algorithm hyperparameters under the same closed-loop trigger condition to form a closed-loop learning mechanism.

[0069] Specifically, when generating a batch identifier for each production execution, a batch identifier generation rule is established in the industrial control computer. This rule binds the batch identifier to the production line identifier, mold identifier, material batch identifier, and timestamp field simultaneously. Before each forging process chain is started, the batch identifier generation rule generates a unique batch identifier and writes it into the production instruction record. The technical term for a batch identifier is an identifier used to uniquely represent a complete production execution at the data level. Its function is to uniformly link the production process sequence, forging quality, production disturbance variables, optimization parameter set, and dynamic prediction model output results generated during the same production execution to the same index. This ensures that subsequent closed-loop learning mechanisms use the same batch identifier as the primary key for data retrieval, tracing, and version rollback, and avoids learning bias caused by mixing samples from different batches.

[0070] When solidifying the production process sequence, forging quality, and production disturbance variables under a batch identifier, a batch data package is established around the batch identifier, which only adds data without overwriting. The actual process parameters of each forging step during production are solidified into a production process sequence according to the forging process order. The quality inspection results after production are solidified into the forging quality. Simultaneously, the difference between the forging quality and the simulated forging quality corresponding to that batch is solidified into a production disturbance variable. The optimized parameter set bound to the batch identifier, the dynamic prediction model output, and the quality judgment result are written into the same batch data package. The technical term for "solidification" is writing data to a storage medium in an immutable record format and retaining the version time point, ensuring that any subsequent learning and optimization are based on the actual production data at that time, without being contaminated by subsequent overwriting updates. The technical term for the dynamic prediction model output is the predicted forging quality and predicted disturbance variables output by the dynamic prediction model during the prediction and evaluation phase of that batch, along with their related intermediate evaluation results. The technical term for the quality judgment result is the judgment label of the forging quality relative to the acceptable range, used to drive the closed-loop triggering conditions and sample weight labeling strategy.

[0071] When performing structural consistency processing on the production process sequence, forging quality, and production disturbance variables to generate a new process state sequence, the process state sequence field template is first solidified. This template covers the forging process sequence, process parameter fields, simulated forging quality fields, and disturbance variable fields. Field mapping is then performed on each forging process in the production process sequence, ensuring that actual process parameters are written to the corresponding process parameter field positions according to the field template. Simultaneously, the simulated forging quality of the batch is written to the simulated forging quality field, and the production disturbance variables of the batch are written to the disturbance variable field. This generates data records isomorphic to the process state sequence sample library. The technical term for structural consistency processing is that through field alignment, field completion, and unified field definitions, the new data is structurally identical to existing samples, allowing it to be directly received by the training interface of the dynamic prediction model. The technical term for a new process state sequence is a process state sequence constructed from the production data of that batch. It covers the entire process chain or currently available process chain segments at the time step and retains a one-to-one correspondence with the batch identifier for subsequent incremental updates and error tracking.

[0072] When establishing sample weight labels for new process state sequences, material batch identifiers, mold wear stage identifiers, equipment state interval identifiers, and environmental interval identifiers are extracted from the batch data package. A weight function related to the similarity to the current production environment is constructed, so that the sample weight label of the new process state sequence increases as its proximity to the current production environment increases. This allows the dynamic prediction model to emphasize perturbation patterns consistent with current production conditions during incremental updates. The technical meaning of sample weight label is the training weight attached to the new process state sequence, used to adjust the contribution ratio of the sample to gradient updates during loss aggregation. The technical meaning of similarity is a metric based on discrete label matching and continuous feature distance, used to characterize the degree of proximity between the production conditions of the new process state sequence and the current production conditions. The expression for sample weight label is:

[0073]

[0074] in, The sample weight label represents the state sequence of the i-th newly added process, and its value ranges from... The output range is limited and used for loss weighting; This represents a monotonic mapping function, used to map linear scores to a stable interval to avoid excessive weights that could lead to training instability. This represents the bias term, used to adjust the overall weight level; This represents the similarity gain coefficient, used to enhance the influence of similarity on weights; This represents the distance penalty coefficient, used to suppress the contribution of samples that are too far away in the current environment; This represents the similarity score obtained based on discrete label matching. Its value is calculated by the matching rules of material batch identifier, mold wear stage identifier, and equipment status interval identifier. This represents the environmental dissimilarity based on the distance of continuous features, and its value is calculated from the distance of temperature ranges, beat ranges, or other continuous environmental features. The principle behind the above expression is to determine the validity of samples by simultaneously introducing label similarity and environmental dissimilarity, so that model updates can quickly adapt to the current conditions while avoiding overfitting to samples far removed from the current conditions.

[0075] When incrementally updating the dynamic prediction model based on the newly added process state sequence and corresponding sample weight labels, the network structure of the dynamic prediction model remains unchanged. The parameters of the model version currently running on the production line are used as initial parameters. After appending the newly added process state sequence to the process state sequence sample library, a subset of samples similar to the current production environment is selected as the incremental update sample set. Small-step iterative training is performed on the incremental update sample set, and the loss is weighted according to the sample weight labels in each iteration, so that the model parameters are updated in a direction that better explains the current disturbance pattern. The technical term for incremental update is to continuously fine-tune the model parameters without changing the network structure, allowing the model to absorb the statistical characteristics of the latest production data without losing its existing temporal mapping ability. The technical term for maintaining the existing network structure is to not change the number of layers, hidden state dimensions, and joint output structure of the Long Short-Term Memory (LSTM) network model, but only update the network weight parameters. The weighted joint loss expression for incremental update is:

[0076]

[0077] in, This represents the weighted joint loss in the incremental update stage, used to simultaneously constrain the prediction error of forging quality and the prediction error of disturbance variables, and to reflect the difference in sample weights; N represents the number of samples in the incremental update sample set. The sample weight label represents the state sequence of the i-th newly added process; This represents the weight of forging quality loss, used to balance the contribution of the forging quality error term; This represents the loss weight of the perturbation variable, used to balance the contribution of the error term of the perturbation variable; This represents the predicted forging quality output by the dynamic prediction model for the i-th sample. This represents the simulated forging mass corresponding to the i-th sample; This represents the prediction perturbation variable of the dynamic prediction model for the i-th sample; Let represent the perturbation variable corresponding to the i-th sample. The principle behind the above expression is to use sample weights to adjust the influence of each sample on the gradient, so that samples that are closer to the current production conditions can help the model learn the current perturbation pattern faster, while maintaining the consistency between the prediction of forging quality and perturbation variables through joint loss.

[0078] When calculating feasibility rate, convergence speed, prediction consistency, and disturbance suppression effect based on historical optimization decision results, and automatically tuning the hyperparameters of the optimization algorithm, the following steps are first performed: First, the candidate parameter scheme set, final optimized parameter set, constraint satisfaction state, iteration algebra statistics, and corresponding production forging quality and production disturbance variables for each optimization attempt are fixed in the historical decision database using decision identifiers as indexes. These are then aggregated into an evaluation sample set within a sliding time window. Next, the feasibility rate, convergence speed, prediction consistency, and disturbance suppression effect are calculated separately and integrated into a hyperparameter score to drive automatic tuning. The technical term for feasibility rate is the proportion of feasible optimization parameter sets that meet the desired forging quality constraints under given constraints. The technical term for convergence speed is the statistical measure of the number of iteration algebras or evaluations required to reach the termination condition. The technical term for prediction consistency is the statistical consistency between the predicted disturbance variables and the production disturbance variables. The technical term for disturbance suppression effect is the concentration of production disturbance variables in the qualified state and the tail risk level. The calculation expressions for the four types of indicators are as follows:

[0079]

[0080]

[0081]

[0082]

[0083] in, The feasibility rate is determined by the percentage of times the constraints are satisfied within the evaluation sample set; M represents the number of decision samples within the evaluation sample set. This indicates an indicator function used to map whether a constraint is satisfied to 0 or 1; This represents the constraint satisfaction state of the j-th decision sample, and its value is jointly determined by the judgment rules of the expected forging quality constraint and the disturbance variable minimization constraint. The average algebraic statistics representing the convergence rate This represents the iteration algebra or evaluation number when the j-th decision reaches the termination condition; This represents the average level of the prediction consistency error. Let j represent the production disturbance variable corresponding to the j-th decision. Let represent the predicted perturbation variable corresponding to the j-th decision; This represents the overall cost of the disturbance suppression effect. This represents the cost fusion weight, used to balance the contributions of the average level, quantile level, and tail risk level to the overall cost; This represents the mean operator, used to characterize the overall disturbance level; This represents the quantile operator, used to characterize the higher-order distribution level of a perturbation variable. Preset quantile parameters; This represents the conditional tail mean operator, used to characterize the tail risk level after exceeding a quantile threshold. Consistent with the quantile operator. The principle behind the above indicators is to evaluate the contribution of hyperparameter configuration to the production objective from four complementary dimensions: feasibility, efficiency, prediction reliability, and perturbation risk. This drives the automatic tuning of hyperparameters such as population size, crossover probability, mutation probability, upper limit of evolutionary generations, and quality and perturbation weight configuration, enabling the optimization algorithm to gradually evolve towards higher feasibility, faster convergence, higher consistency, and lower perturbation risk.

[0084] When linking incremental updates of the dynamic prediction model with automatic hyperparameter tuning of the optimization algorithm under the same closed-loop trigger condition to form a closed-loop learning mechanism, the closed-loop trigger condition is defined in the industrial control computer. This condition simultaneously includes quality judgment results, production disturbance variable levels, prediction consistency error levels, and feasibility trend changes. The closed-loop trigger condition is updated at the end of each batch. When the closed-loop trigger condition meets the model offset characteristic, incremental updates of the dynamic prediction model are triggered first. When the closed-loop trigger condition meets the search degradation characteristic, automatic hyperparameter tuning of the optimization algorithm is triggered first. When both the model offset and search degradation characteristics are met, both are triggered in parallel, and a joint version identifier is set for traceability and rollback. The technical term for the closed-loop trigger condition is a set of rules used to determine when to perform learning and tuning, achieving adaptive iteration by mapping production feedback signals to operable trigger flags. The technical term for the linkage is the collaborative updating of the dynamic prediction model and the optimization algorithm under the same batch data baseline, ensuring that changes in the model output distribution and changes in the optimization algorithm's search strategy are matched, thus forming a closed-loop learning mechanism. The closed-loop triggering condition can be represented by a comprehensive triggering score as follows:

[0085]

[0086] in, This represents a comprehensive trigger score, used to uniformly characterize whether a closed-loop learning mechanism needs to be activated. This represents the trigger weight, used to balance the impact of quality anomalies, disturbance levels, prediction consistency, and feasibility on the triggering decision. This indicates the quality judgment result, and its value is obtained by comparing the quality of the produced forgings with the acceptable range. Represents production disturbance variables; This represents the statistics of prediction consistency error; This indicates the feasibility rate statistic; This represents a disturbance mapping function used to map production disturbance variables horizontally to comparable trigger contribution values; This represents the consistency mapping function, used to map the prediction consistency error to comparable trigger contribution values; This represents the feasibility mapping function, used to map the risk of decreased feasibility to a comparable trigger contribution value. The principle behind the above expression is to fuse multi-source feedback signals into a single score, and to achieve normalized contribution control of different signals through weights and the mapping function. This enables the closed-loop learning mechanism to automatically initiate the collaborative iteration of dynamic prediction model incremental updates and automatic hyperparameter tuning of the optimization algorithm when the production environment changes or quality risks increase.

[0087] Reference Figure 2 This invention first constructs a training data foundation by combining CAE simulation and actual production data. The process parameter combination, simulated forging quality, and disturbance variables are organized into a process state sequence sample library according to the forging process sequence. The process state sequence sample library is then used to train a long short-term memory network model to form a dynamic prediction model. This dynamic prediction model can output predicted forging quality and predicted disturbance variables when the process state sequence is input, thereby establishing a temporal mapping relationship between the forging process state, forging quality, and disturbance variables.

[0088] Subsequently, during actual production, the actual process parameters of the completed initial process are read, and the actual process parameters are organized into a known process state sequence according to the forging process sequence. The known process state sequence is input into the dynamic prediction model to obtain the predicted forging quality and the predicted disturbance variables. Under the constraints of the desired forging quality and the desired disturbance variables, the process parameters of the subsequent unexecuted processes are optimized to obtain the optimized parameter set that satisfies the desired forging quality constraint and minimizes the disturbance variables.

[0089] After obtaining the optimized parameter set, the optimized parameter set is sent to the production equipment to drive the subsequent unexecuted processes to continue the forging production. After the production is completed, the quality of the forging and the production disturbance variables are obtained. The success of the optimization process is determined by comparing the quality of the forging with the quality acceptance range. When the optimization fails or the quality of the forging is not within the acceptance range, the initial process parameter correction mechanism is triggered to readjust the initial process parameters. Under the new initial process parameter conditions, the process state sequence is reconstructed and the dynamic prediction model prediction and process parameter optimization process is executed again.

[0090] After successfully finding the optimal solution and obtaining the production forging quality and production disturbance variables, the production process sequence, production forging quality, and production disturbance variables are used as new samples to perform incremental learning to update the parameters of the dynamic prediction model. This allows the dynamic prediction model to absorb the disturbance information generated during the actual production process. At the same time, the hyperparameters of the optimization algorithm are tuned based on the historical optimization decision results, so that the optimization algorithm has a higher feasibility and convergence efficiency in subsequent optimization processes. Thus, a continuous iterative closed-loop learning mechanism is formed through the incremental update of the dynamic prediction model and the tuning of the optimization algorithm hyperparameters.

[0091] In one specific implementation, taking the production process of a 6082 aluminum alloy front steering knuckle as an example, the initial volume is fixed, and the dimensions of the roller are set. Based on the characteristics of the steering knuckle workpiece, a suitable-sized bar stock is first required. After a roll forging process, a billet is formed. A total of four processes are set, including roll forging, bending, pre-forging, and final forging. CAE simulation is performed using DEFORM software. Without changing the original bar stock volume, the roll forging process is adjusted so that the end lengths of the billet are 357mm, 362mm, 367mm, 372mm, 377mm, and 382mm, and the billet shape is as follows. Figure 3 As shown; for the bending process, the billet temperature is set to 400℃, 420℃, 440℃, 460℃, 480℃, and 500℃, with vertical offsets of 0mm, 10mm, 20mm, 30mm, 40mm, and 50mm, respectively. Figure 4 As shown, the pressing speeds are 200mm / s, 300mm / s, 400mm / s, 500mm / s, 600mm / s, and 700mm / s, and the pressing amounts are 75mm, 80mm, 85mm, 90mm, 95mm, and 100mm. This example optimizes the pressing amount and pressing speed during the bending process, as well as the pre-forging and final forging processes. Since the positional deviation is uncontrollable, the positional deviations for the pre-forging and final forging processes are not set. The pre-forging and final forging processes are defaulted to pressing until the upper and lower dies contact, so the pressing amount is not set. For the pre-forging process, the forging temperatures were set to 360℃, 400℃, 440℃, 480℃, 520℃, and 560℃, and the pressing speeds were set to 200mm / s, 300mm / s, 400mm / s, 500mm / s, 600mm / s, and 700mm / s. For the final forging process, the forging temperatures were set to 350℃, 400℃, 450℃, 500℃, 550℃, 600℃, and 700℃, and the pressing speeds were set to 200mm / s, 300mm / s, 400mm / s, 500mm / s, 600mm / s, and 700mm / s. The forging dimensions S1 under 300 sets of sequence samples were simulated. At the same time, actual production was carried out on these 300 sets of parameters, and the actual forging dimensions of the 300 sets were summarized. The perturbation variable Δ was obtained by comparing the actual forging dimensions with the simulated forging dimensions.

[0092] Ultimately, each training sample contains: process sequence features, simulated forging dimensions, and perturbation variables. An LSTM network is trained using this dataset of 300 samples. The innovation of this network lies in its output layer design: it contains two neurons, one for predicting the final forging dimension Sp and the other for predicting the perturbation variable value Δp. After training, the model can learn the complex nonlinear relationship between the process parameter sequence and the final quality and production perturbations.

[0093] In actual production, the initial process parameters are set as follows: billet end length is 370mm, forging temperature before pre-forging is 460℃, and vertical offset is 0mm. The system operates as follows: after the billet completes the roll forging and bending processes, the sensor collects the billet end length (368mm), forging temperature before pre-forging (458℃), and vertical offset (+1mm) in real time. The system inputs the measured state sequence of this completed process into the aforementioned LSTM model. Based on the prediction results of the LSTM model, the system then calls the built-in genetic algorithm for online optimization. Here, a weighted method is used to transform multi-objective optimization into a single-objective function, where the forging size has a weight of 0.7 and the disturbance variable has a weight of 0.3, thus balancing quality compliance and disturbance suppression. However, after optimization, the system finds that no matter how it optimizes, the workpiece cannot meet the qualification standard. At this point, the system automatically triggers the initial process parameter correction process. The system incorporates the billet end and forging temperature before pre-forging into the optimization variables, and, combined with the measured state of the completed process, restarts a global optimization covering the entire process chain. By rapidly calling the LSTM model for large-scale process prediction, the system evaluated a new solution: adjusting the billet end length from the current 370mm to 380mm and increasing the pre-forging heating temperature from 460℃ to 480℃. A subsequent production run was conducted. After the billet completed the roll forging and bending processes, sensors collected real-time data showing the billet end length at 381mm, the pre-forging temperature at 476℃, and a vertical offset of +2mm. After further optimization calculations, the system output optimized parameter combinations for the subsequent pre-forging and final forging processes: pre-forging temperature 470℃, pre-forging speed 400mm / s, final forging temperature 480℃, and final forging speed 500mm / s. These parameters were automatically converted into equipment control commands and executed.

[0094] After production is completed, the system learns from the production cycle and calculates the actual disturbance value Δ'. Regardless of whether the forgings are qualified or not, complete data for that production cycle is stored. Furthermore, the system uses the newly added data to fine-tune the LSTM model, focusing on strengthening the learning of disturbance variables. Simultaneously, the system analyzes the effectiveness of historical optimization decisions, automatically adjusts the hyperparameters of optimizers such as genetic algorithms (e.g., generation number, crossover rate) using the database, and dynamically adjusts the weight coefficients in the dual objective function to adapt to changes in production conditions.

[0095] This embodiment also discloses an intelligent control device for multi-process forging, referring to... Figure 5 The device includes an acquisition module 501, a processing module 502, and an output module 503. It is used to execute any of the above-described intelligent control methods for multi-process forging, wherein: The acquisition module 501 is used to construct a process state sequence around multiple forging processes, obtain the simulated forging quality corresponding to different combinations of process parameters through CAE simulation, and calculate the difference between the simulated forging quality and the actual forging quality in combination with the actual forging quality to form a disturbance variable. Processing module 502 is used to train a long short-term memory network model on a sample library of process state sequences to form a dynamic prediction model; Processing module 502 is used to acquire the actual process parameters of the completed initial process and form a known process state sequence during the production execution process; Processing module 502 is used to predict and evaluate subsequent unexecuted processes based on the known process state sequence, combined with the desired forging quality constraint and the disturbance variable minimization constraint, and to optimize the controllable process parameters of subsequent processes to obtain an optimized parameter set through optimization algorithm; The output module 503 is used to obtain the quality of the forging produced after the forging production continues with the optimized parameter set, calculate the production disturbance variable based on the difference between the quality of the forging produced and the simulated forging, and trigger the initial process parameter correction mechanism to readjust the initial process parameters and re-execute the full process optimization when the quality of the forging produced is not within the qualified range.

[0096] It should be noted that the above embodiments of the apparatus are only illustrated by the division of the above functional modules. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. In addition, the apparatus and method embodiments provided in the above embodiments belong to the same concept, and the specific implementation process can be found in the method embodiments, which will not be repeated here.

[0097] This embodiment also discloses an electronic device, as shown in the reference. Figure 6 The electronic device may include: at least one processor 601, at least one communication bus 602, user interface 603, network interface 604, and at least one memory 605.

[0098] The communication bus 602 is used to enable communication between these components.

[0099] The user interface 603 may include a display screen and a camera. Optionally, the user interface 603 may also include a standard wired interface and a wireless interface.

[0100] The network interface 604 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface).

[0101] The processor 601 may include one or more processing cores. The processor 601 connects to various parts of the server using various interfaces and lines, and performs various server functions and processes data by running or executing instructions, programs, code sets, or instruction sets stored in the memory 605, and by calling data stored in the memory 605. Optionally, the processor 601 may be implemented using at least one hardware form of Digital Signal Processing (DSP), Field-Programmable Gate Array (FPGA), or Programmable Logic Array (PLA). The processor 601 may integrate one or a combination of several of the following: Central Processing Unit (CPU), Graphics Processing Unit (GPU), and modem. The CPU primarily handles the operating system, user interface, and applications; the GPU is responsible for rendering and drawing the content required for display; and the modem handles wireless communication. It is understood that the modem may also not be integrated into the processor 601 and may be implemented as a separate chip.

[0102] The memory 605 may include random access memory (RAM) or read-only memory. Optionally, the memory may include a non-transitory computer-readable storage medium. The memory 605 may be used to store instructions, programs, code, code sets, or instruction sets. The memory 605 may include a program storage area and a data storage area. The program storage area may store instructions for implementing an operating system, instructions for at least one function (such as touch functionality, sound playback functionality, image playback functionality, etc.), instructions for implementing the various method embodiments described above, etc.; the data storage area may store data involved in the various method embodiments described above, etc. Optionally, the memory 605 may also be at least one storage device located remotely from the aforementioned processor 601. As a computer storage medium, the memory 605 may include an operating system, a network communication module, a user interface 603 module, and an application program for a multi-process forging intelligent control method.

[0103] exist Figure 6 In the electronic device shown, the user interface 603 is mainly used to provide an input interface for the user and to obtain the user input data; while the processor 601 can be used to call the application program of a multi-process forging intelligent control method stored in the memory 605. When executed by one or more processors 601, the electronic device executes one or more methods as described in the above embodiments.

[0104] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that the present invention is not limited to the described order of actions, as some steps can be performed in other orders or simultaneously according to the present invention. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to the present invention.

[0105] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0106] In the several embodiments provided by this invention, it should be understood that the disclosed apparatus can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some service interface; the indirect coupling or communication connection between apparatuses or units may be electrical or other forms.

[0107] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0108] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0109] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage device (CMD). Based on this understanding, the technical solution of this invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a memory 605 and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned memory 605 includes various media capable of storing program code, such as a USB flash drive, external hard drive, magnetic disk, or optical disk.

[0110] The present invention also discloses a non-transitory computer-readable storage medium storing instructions. When executed by one or more processors 601, these instructions cause an electronic device to perform one or more methods as described in the above embodiments.

[0111] The above are merely exemplary embodiments of this disclosure and should not be construed as limiting the scope of this disclosure. Any equivalent changes and modifications made in accordance with the teachings of this disclosure shall still fall within the scope of this disclosure. Those skilled in the art will readily conceive of other embodiments of this disclosure upon considering the specification and the disclosure of practical truths. This invention is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not described in this disclosure. The specification and embodiments are to be considered exemplary only, and the scope and spirit of this disclosure are defined by the claims.

Claims

1. A method for intelligent control of multi-process forging, characterized in that, The method includes: A process state sequence is constructed around multiple forging processes. The simulated forging quality corresponding to different combinations of process parameters is obtained through CAE simulation. The difference between the simulated forging quality and the actual forging quality is calculated in combination with the actual forging quality to form a perturbation variable. A long short-term memory network model is trained on a sample library of process state sequences to form a dynamic prediction model; During production execution, the actual process parameters of the completed initial processes are obtained and a known process state sequence is formed; Based on the known process state sequence, combined with the desired forging quality constraint and the disturbance variable minimization constraint, the dynamic prediction model is used to predict and evaluate the subsequent unexecuted processes, and the optimization algorithm is used to optimize the controllable process parameters of the subsequent processes to obtain the optimized parameter set. After continuing forging production using the optimized parameter set, the quality of the produced forgings is obtained. The production disturbance variable is calculated based on the difference between the quality of the produced forgings and the simulated forgings. If the quality of the produced forgings is not within the acceptable range, the initial process parameter correction mechanism is triggered to readjust the initial process parameters and re-execute the full process optimization.

2. The intelligent control method for multi-process forging as described in claim 1, characterized in that, After triggering the initial process parameter correction mechanism to readjust the initial process parameters and re-execute the full process optimization when the quality of the forgings produced is not within the acceptable range, the method further includes: Generate a batch identifier for each production execution; The production process sequence, the quality of the forgings, and the production disturbance variables are solidified under the batch identifier. At the same time, an optimization parameter set, dynamic prediction model output results, and quality judgment results bound to the batch identifier are established. The production process sequence, the quality of the forgings produced, and the production disturbance variables are subjected to structural consistency processing to generate a new process state sequence. Establish sample weight labels for the newly added process state sequence; Based on the newly added process state sequence and the corresponding sample weight label, the dynamic prediction model is incrementally updated so that the dynamic prediction model can absorb the production disturbance information carried by the newly added process state sequence while maintaining the existing network structure. Based on historical optimization decision results, the feasibility rate, convergence speed, prediction consistency, and disturbance suppression effect are calculated, and the hyperparameters of the optimization algorithm are automatically tuned. The incremental updates of the dynamic prediction model and the automatic tuning of hyperparameters of the optimization algorithm are linked under the same closed-loop triggering condition to form a closed-loop learning mechanism.

3. The intelligent control method for multi-process forging as described in claim 1, characterized in that, The training of a long short-term memory network model on a sample library of process state sequences to form a dynamic prediction model specifically includes: The process parameter combination, the simulated forging quality, and the disturbance variables are organized into a process state sequence according to the forging process sequence. Multiple sets of process state sequences are constructed into a process state sequence sample library; The sample library of process state sequences is input into the long short-term memory network model for training, so that the long short-term memory network model learns the temporal mapping relationship between the process state sequence, forging quality and perturbation variables. After the model training is completed, the dynamic prediction model is formed, which enables the dynamic prediction model to output the predicted forging quality and the corresponding predicted perturbation variables when the input process state sequence is given.

4. The intelligent control method for multi-process forging as described in claim 3, characterized in that, The step of inputting the process state sequence sample library into the Long Short-Term Memory network model for training specifically includes: The process state sequence in the process state sequence sample library is subjected to structure consistency processing to ensure that each process state sequence maintains a unified structure in the forging process sequence, process parameter field, simulated forging quality field and disturbance variable field, and the length alignment processing is performed on each process state sequence to form a process state sequence. The process state features corresponding to each of the processing process state sequences are constructed into a time-series feature vector, and then spliced ​​together according to the forging process sequence to form a process state sequence matrix; The process state sequence sample library is divided into a training subset and a validation subset. The training subset is input into the long short-term memory network model, so that the long short-term memory network model reads the process state features step by step according to the forging process sequence and remembers and updates the historical process state through a gating structure, thereby forming a comprehensive representation of the forging quality and disturbance variables at the terminal time step. A joint output structure of the predicted forging quality and the predicted perturbation variable is established at the output end of the long short-term memory network model; By iteratively updating the network parameters by constraining the first deviation and the second deviation, the long short-term memory network model learns the joint temporal mapping relationship between the process state sequence and the forging quality and the perturbation variable. The first deviation is the deviation between the predicted forging quality and the simulated forging quality, and the second deviation is the deviation between the predicted perturbation variable and the perturbation variable. Once the network parameters reach the preset convergence condition, the parameters of the long short-term memory network model are solidified, and the dynamic prediction model is formed.

5. The intelligent control method for multi-process forging as described in claim 1, characterized in that, The step of predicting and evaluating subsequent unexecuted processes based on the known process state sequence, combined with the desired forging quality constraint and the constraint of minimizing disturbance variables, through the dynamic prediction model, specifically includes: A prediction and evaluation constraint system is established based on the known process state sequence, wherein the prediction and evaluation constraint system consists of the expected forging quality constraint and the disturbance variable minimization constraint. The set of subsequent unexecuted processes is determined based on the forging process position of the known process state sequence; Construct a set of candidate process parameter combinations corresponding to the set of unexecuted processes; The candidate process parameter combination set is concatenated with the known process state sequence to form a candidate process state sequence set; The candidate process state sequence set is sequentially input into the dynamic prediction model to obtain the corresponding predicted forging quality and predicted disturbance variables, and the candidate process state sequence is predicted and evaluated according to the prediction and evaluation constraint system. The candidate process state sequences are comprehensively sorted based on the deviation between the predicted forging quality and the desired forging quality, as well as the magnitude of the predicted perturbation variable, and the candidate process state sequences that satisfy the desired forging quality constraint are determined as the preferred process state sequences.

6. The intelligent control method for multi-process forging as described in claim 5, characterized in that, After continuing forging production using the optimized parameter set, the quality of the produced forgings is obtained. Production disturbance variables are calculated based on the difference between the produced forging quality and the simulated forging quality. If the produced forging quality is not within the acceptable range, an initial process parameter correction mechanism is triggered to readjust the initial process parameters and re-execute the full process optimization. Specifically, this includes: The optimized parameter set is sent to the production equipment to drive the subsequent unexecuted processes to continue the forging production, and a simulated forging quality corresponding to the optimized parameter set is established as a comparison benchmark. The quality of the forgings produced is obtained by performing quality inspection on the forgings after production, and the difference between the quality of the forgings produced and the quality of the simulated forgings is calculated under a unified quality evaluation index system to form a production disturbance variable; The consistency of the production disturbance variable and the predicted disturbance variable output by the dynamic prediction model are compared, and the passability is determined in combination with the qualified range of the production forging quality. When the quality of the forgings produced is not within the acceptable range, the initial process parameter correction mechanism is triggered, and the actual process parameters of the completed processes are frozen to construct the process state sequence constraints for the entire process. The initial process parameters are incorporated into the set of optimization variables to form a set of full-process optimization variables, and a set of candidate process parameter combinations is reconstructed based on the set of full-process optimization variables to perform full-process optimization.

7. The intelligent control method for multi-process forging as described in claim 1, characterized in that, After obtaining the quality of the forgings produced following the execution of the optimized parameter set and continued forging production, the method further includes: When the quality of the forgings produced is within acceptable limits, the production status is determined to be in a quality-compliant state, and the optimized parameter set continues to be used as the current production parameters without triggering the initial process parameter correction mechanism.

8. The acquisition module is used in 8. An intelligent control device for multi-process forging, characterized in that, The device is used to execute the intelligent control method for a multi-process forging process as described in any one of claims 1-7. The device includes an acquisition module, a processing module, and an output module, wherein: The acquisition module is used to construct a process state sequence around multiple forging processes, obtain the simulated forging quality corresponding to different combinations of process parameters through CAE simulation, and calculate the difference between the simulated forging quality and the actual forging quality in combination with the actual forging quality to form a disturbance variable. The processing module is used to train a long short-term memory network model on a sample library of process state sequences to form a dynamic prediction model; The processing module is used to acquire the actual process parameters of the completed initial process and form a known process state sequence during the production execution process; The processing module is used to combine the desired forging quality constraint and the disturbance variable minimization constraint with the known process state sequence, predict and evaluate the subsequent unexecuted processes through the dynamic prediction model, and optimize the controllable process parameters of the subsequent processes to obtain an optimized parameter set through the optimization algorithm. The output module is used to obtain the quality of the produced forgings after executing the optimized parameter set and continuing forging production, calculate the production disturbance variable based on the difference between the quality of the produced forgings and the simulated forgings, and trigger the initial process parameter correction mechanism to readjust the initial process parameters and re-execute the full process optimization when the quality of the produced forgings is not within the qualified range.

9. An electronic device, characterized in that, The device includes a processor, a communication bus, a user interface, a network interface, and a memory. The memory is used to store instructions. The user interface and the network interface are both used to communicate with other devices. The communication bus is used to enable communication between the components within the electronic device. The processor is used to execute the instructions stored in the memory to cause the electronic device to perform the method as described in any one of claims 1-7.

10. A non-transitory computer-readable storage medium, characterized in that, The computer-readable storage medium stores instructions that, when executed, perform the method as described in any one of claims 1-7.

Images