Intelligent optimization method for metal forging process parameters

Abstract

The invention provides an intelligent optimization method for metal forging process parameters, and relates to the technical field of reinforcement learning, and the method comprises the steps: collecting multi-source process data in real time, carrying out time sequence synchronization and filtering processing, inputting a normalized feature vector into a lightweight convolutional neural network to recognize an environment state, and setting a weight mapping table in combination with expert experience and an optimization track. Multi-target reward function weighted configuration driven by state categories is realized, and a KL divergence constrained exponential smoothing mechanism is introduced to ensure smooth transition of reward weights; a soft actor-commentator algorithm is used for generating an actual process parameter instruction, closed-loop collection of feedback data is carried out, a classifier and weight configuration are driven to be subjected to continuous fine adjustment, and environment perception and decision optimization double-closed-loop self-evolution is achieved. And the global optimization stability and decision response efficiency of multi-target control are enhanced.

Description

Technical Field

[0001] This invention relates to the field of reinforcement learning technology, and in particular to an intelligent optimization method for metal forging process parameters. Background Technology

[0002] With the rapid development of industrial intelligence and the digitalization of forging processes, the forging industry is placing increasingly higher demands on the optimization of process parameters. Currently, mainstream forging process parameter optimization methods mostly employ multi-objective optimization techniques based on reinforcement learning or data-driven approaches. These methods use intelligent algorithms to automatically adjust key process parameters such as heating temperature, impact energy, and holding time, thereby improving forging quality, reducing energy consumption, and extending die life. This type of technical solution has become a major trend in the field of "self-optimization of forging process parameters" and is widely used in complex forging production processes in high-end equipment manufacturing, automotive parts, and aerospace industries. However, in the complex actual forging production environment, there are various interfering factors such as billet size fluctuations, material composition changes, and equipment performance degradation, which cause the environmental state to be in a state of dynamic change. Current technologies have not yet achieved precise identification of environmental state changes or real-time adjustment of process parameter priorities. Traditional reinforcement learning frameworks often use static weights or assume a stable environment, and the policy update rate is limited by the ability to perceive the environment, making it difficult to cope with sudden changes during billet fluctuation periods or equipment degradation stages. For example, when extreme billet fluctuations occur, focusing on energy consumption with fixed weights may lead to substandard forging quality; and if the die stress control weight is not increased in time during equipment performance degradation, there are safety hazards. Furthermore, current technologies have not disclosed effective designs for dynamic adjustment of process parameter priorities based on real-time environmental state identification, and generally lack mechanisms for smooth transition of reward weights and robustness against misjudgments in response to dynamic environmental state changes. In practical production applications, the acquisition and time-series analysis of multi-source sensor data in forging equipment are feasible. However, a systematic technological breakthrough has not yet been achieved in dynamically coupling data-driven environmental state recognition with reinforcement learning reward functions. Existing technologies mostly focus on traditional multi-objective optimization algorithms, static parameter settings, and single-objective reinforcement learning, without considering the coordinated adjustment of process objective weights under sudden changes in the production environment. Especially at the equipment level, the action space is often only subject to static physical boundary constraints, which do not support adaptive parameter adjustment with the environment, resulting in insufficient policy generalization ability and limited global optimization efficiency. Summary of the Invention

[0003] In order to solve the above-mentioned technical problems, the present invention provides an intelligent optimization method for metal forging process parameters.

[0004] The technical solution of this invention is implemented as follows: a method for intelligent optimization of metal forging process parameters, comprising: S1: Acquire multi-source process data collected in real time during the forging process. The multi-source process data includes billet size deviation, material composition fluctuation, die surface temperature gradient and main cylinder pressure response delay signal, and record their corresponding time series labels to construct the basic dataset required for environmental state perception. S2: Perform sliding window normalization on the multi-source process data to eliminate dimensional differences and time drift interference, generate a standardized time-series feature vector, and input the standardized time-series feature vector into a pre-trained lightweight convolutional neural network classifier. This classifier is obtained by offline training based on historically labeled working condition sample sets and is used to realize environmental state category discrimination. S3: Based on the discrimination results output by the lightweight convolutional neural network classifier, generate the current production environment status category label. The status category label includes 'high fluctuation billet', 'stable working condition' and 'early stage of equipment degradation'. The status category label is used as an index variable for the generation of dynamic weight vector in the subsequent multi-objective reward function. S4: Query the preset weight mapping table according to the state category label to obtain the corresponding normalized weight vector. The weight mapping table is jointly calibrated by domain expert experience and historical optimization trajectory, and is used to assign differentiated priority weights to the three optimization objectives of forging forming quality score, unit product energy consumption cost and mold stress accumulation index. S5: Based on the normalized weight vector, the components in the multi-objective reward function are weighted and fused to generate a dynamic comprehensive reward value. The weighted fusion process introduces an exponential moving average mechanism under KL divergence constraint to achieve a smooth transition between the old and new weight vectors, so as to avoid sudden changes in the control strategy due to misjudgment of the state. S6: Input the dynamic comprehensive reward value into the soft actor-critic algorithm framework to drive the strategy network update and generate a three-dimensional continuous motion vector containing the heating temperature setpoint, the energy increment per strike, and the pressure holding time adjustment. The output of the strategy network is limited by the tanh activation function and combined with the number of strokes of the GK-5000 device (16 spm) and the nominal force (50000 kN) as physical boundary constraints. S7: Execute the process parameter adjustment command defined by the three-dimensional continuous motion vector, control the forging equipment to run in the updated process window, and simultaneously collect the actual forming quality, measured energy consumption and mold stress feedback data to form the real environmental response result under closed-loop control; S8: The deviation signal between the actual environmental response result and the expected target is converted into a gradient signal that can be used to optimize the classifier's discrimination threshold. The gradient backpropagation is performed on the discrimination threshold of the output layer of the convolutional neural network classifier to adjust the priority ratio among the forging forming quality score, unit product energy consumption cost and mold stress accumulation index. This forms a dual closed-loop self-evolutionary mechanism of environmental recognition and reward adjustment to improve the adaptability and global optimization stability of the multi-objective reinforcement learning strategy under dynamic working conditions.

[0005] The present invention provides an intelligent optimization method for metal forging process parameters, which has the following beneficial effects: (1) This invention, by introducing an environmental state perception submodule and a lightweight convolutional neural network classifier, achieves real-time and accurate identification of complex forging conditions, significantly improving the adaptability of the reinforcement learning system to dynamic production environments. Utilizing sliding window normalization processing and a pre-trained CNN model, it efficiently extracts features and determines the state of multi-source process signals, outputting current condition category labels such as "high-fluctuation billet," "stable condition," or "early stage of equipment degradation" within hundreds of milliseconds, providing reliable contextual basis for subsequent decisions. This mechanism not only avoids the lag and subjectivity caused by manually setting thresholds but also ensures that the closed-loop control cycle remains unaffected through the design of separating offline training and online inference, effectively overcoming the strategy blind spot problem caused by the lag in environmental perception and the coarseness of state judgment in existing technologies.

[0006] (2) This invention proposes a dynamic multi-objective reward weighting mechanism driven by environmental state, combined with a smooth transition strategy guided by KL divergence, which effectively solves the technical problems of performance index conflict and strategy oscillation in the multi-objective optimization process. Using environmental state labels as indexes, the weight mapping table jointly calibrated by expert experience and historical optimization trajectory is called to realize the adaptive generation of normalized weights for each objective. For example, when "high fluctuation billet" is detected, the forming quality weight is automatically increased to prioritize product consistency, while under "stable working conditions", the proportion of energy-saving objectives is moderately increased, thereby realizing the fine adjustment of resource allocation. The new and old weight vectors are gradually fused by exponential moving average method to suppress weight mutations caused by transient signal noise or classification misjudgment, thereby enhancing the stability and robustness of the strategy update process. This design enables the soft actor-critic algorithm to maintain efficient exploration and convergence capabilities in complex non-steady-state environments, which is significantly better than the performance of traditional fixed weighting or multi-task parallel architecture; (3) This invention constructs a dual-closed-loop self-evolutionary optimization system integrating environmental perception, dynamic reward adjustment, and strategy reverse fine-tuning, which significantly improves the global optimization efficiency and long-term operational reliability of the process self-optimization system. After each strategy execution, actual forming results and energy consumption feedback data are actively collected to reverse correct the discrimination boundary of the environmental classifier and the configuration parameters in the weight mapping table, forming a complete closed-loop link of "perception-decision-execution-learning". This continuous evolution mechanism enables the system to continuously accumulate process knowledge and optimize decision logic over time, which is especially suitable for real-world scenarios such as state drift of long-term service equipment and the introduction of new materials. At the same time, the action space design fully considers the actual physical constraints of the GK-5000 equipment (such as nominal force of 50000kN and stroke count of 16spm), and ensures the safety and feasibility of the output action through tanh activation function and boundary clipping, avoiding the disconnect between theoretical algorithms and engineering implementation. The overall architecture does not require frequent manual intervention and complex hyperparameter tuning, and has good interpretability and scalability, providing solid technical support for the autonomous upgrading of intelligent forging production lines. Attached Figure Description

[0007] Figure 1 This is a flowchart of an intelligent optimization method for metal forging process parameters according to the present invention; Figure 2 This is a sub-flowchart of an intelligent optimization method for metal forging process parameters according to the present invention; Figure 3 This is another sub-flowchart of the intelligent optimization method for metal forging process parameters according to the present invention. Detailed Implementation

[0008] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0009] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0010] like Figure 1 As shown, this invention provides an intelligent optimization method for metal forging process parameters, specifically including: S1: Acquire multi-source process data collected in real time during the forging process. The multi-source process data includes billet size deviation, material composition fluctuation, die surface temperature gradient and main cylinder pressure response delay signal, and record their corresponding time series labels to construct the basic dataset required for environmental state perception. S2: Perform sliding window normalization on the multi-source process data to eliminate dimensional differences and time drift interference, generate a standardized time-series feature vector, and input the standardized time-series feature vector into a pre-trained lightweight convolutional neural network classifier. This classifier is obtained by offline training based on historically labeled working condition sample sets and is used to realize environmental state category discrimination. S3: Based on the discrimination results output by the lightweight convolutional neural network classifier, generate the current production environment status category label. The status category label includes 'high fluctuation billet', 'stable working condition' and 'early stage of equipment degradation'. The status category label is used as an index variable for the generation of dynamic weight vector in the subsequent multi-objective reward function. S4: Query the preset weight mapping table according to the state category label to obtain the corresponding normalized weight vector. The weight mapping table is jointly calibrated by domain expert experience and historical optimization trajectory, and is used to assign differentiated priority weights to the three optimization objectives of forging forming quality score, unit product energy consumption cost and mold stress accumulation index. S5: Based on the normalized weight vector, the components in the multi-objective reward function are weighted and fused to generate a dynamic comprehensive reward value. The weighted fusion process introduces an exponential moving average mechanism under KL divergence constraint to achieve a smooth transition between the old and new weight vectors, so as to avoid sudden changes in the control strategy due to misjudgment of the state. S6: Input the dynamic comprehensive reward value into the soft actor-critic algorithm framework to drive the strategy network update and generate a three-dimensional continuous motion vector containing the heating temperature setpoint, the energy increment per strike, and the pressure holding time adjustment. The output of the strategy network is limited by the tanh activation function and combined with the number of strokes of the GK-5000 device (16 spm) and the nominal force (50000 kN) as physical boundary constraints. S7: Execute the process parameter adjustment command defined by the three-dimensional continuous motion vector, control the forging equipment to run in the updated process window, and simultaneously collect the actual forming quality, measured energy consumption and mold stress feedback data to form the real environmental response result under closed-loop control; S8: The deviation signal between the actual environmental response result and the expected target is converted into a gradient signal that can be used to optimize the classifier's discrimination threshold. The gradient backpropagation is performed on the discrimination threshold of the output layer of the convolutional neural network classifier to adjust the priority ratio among the forging forming quality score, unit product energy consumption cost and mold stress accumulation index. This forms a dual closed-loop self-evolutionary mechanism of environmental recognition and reward adjustment to improve the adaptability and global optimization stability of the multi-objective reinforcement learning strategy under dynamic working conditions.

[0011] Step S1: Acquire multi-source process data collected in real time during the forging process. This multi-source process data includes billet size deviation, material composition fluctuation, die surface temperature gradient, and master cylinder pressure response delay signal, and records their corresponding time series labels to construct the basic dataset required for environmental state perception. Specifically, this includes: S1.1: Based on a multimodal sensor network deployed on the forging equipment, the raw input data are obtained by acquiring the billet size deviation, material composition fluctuation, die surface temperature gradient and main cylinder pressure response delay signal. The billet size deviation is acquired by a laser ranging array, the material composition fluctuation is obtained by online detection by an X-ray fluorescence spectrometer, the die surface temperature gradient is obtained by spatial sampling by an infrared thermal imaging array, and the main cylinder pressure response delay signal is obtained from the timestamp record of the high-speed pressure transmitter of the hydraulic system, so as to achieve comprehensive perception of key process disturbance sources. The input condition is a multi-modal sensor network deployed on the GK-5000 forging press. The monitored objects include billet size deviation, material composition fluctuation, die surface temperature gradient and master cylinder pressure response delay signal. The signal types are respectively from laser ranging array, X-ray fluorescence spectrometer, infrared thermal imaging array and hydraulic system high-speed pressure transmitter. A laser ranging array (parameters: ranging accuracy ±0.01mm, sampling frequency 500Hz) is used to realize non-contact spatial measurement of billet size deviation and generate continuous time series data of size deviation. Furthermore, by using an X-ray fluorescence spectrometer (parameters: energy resolution 130 eV, sampling period 1 s), real-time spectral detection of material composition fluctuations was achieved, and dynamic change curves of the mass fraction of each chemical element were obtained. Furthermore, an infrared thermal imaging array (parameters: wavelength response range 8~14μm, spatial resolution 640×480 pixels) is used to achieve spatial sampling of the temperature distribution on the mold surface, thereby obtaining a time-series data matrix of multi-point temperature gradients; Furthermore, a high-speed pressure transmitter for the hydraulic system (parameters: range 0~60MPa, sampling frequency 2kHz) is used to acquire the pressure response delay signal of the master cylinder, and combined with the timestamp recording module to generate a delay feature sequence of pressure change and control command; By synchronously working through a multimodal sensor network, the above four types of raw input signals are collected into a time-series basic dataset, enabling comprehensive perception of key process disturbance sources. For example, the laser ranging array on the GK-5000 equipment is configured with six measuring points on both sides and above and below the billet feed path to collect billet diameter deviation and end face flatness data; the X-ray fluorescence spectrometer is set to collect elemental content once per second to detect the three main components Fe, C, and Mn, and the output component fluctuation curve shows a slight decreasing trend of Fe content during ten consecutive impacts; the infrared thermal imaging array is arranged around the mold, and the captured temperature gradient data shows that the temperature difference in the mold edge area exceeds 25°C in the high-fluctuation billet scenario; the high-speed pressure transmitter of the hydraulic system records the main cylinder command signal and establishes a delay sequence with the actual pressure, and the response characteristics of the early stage of equipment degradation are determined by analyzing the delay value. In this scenario, the four types of sensor data are synchronously collected to form a high-dimensional time-series raw dataset, providing complete input for the unified time label processing in the subsequent step S1.2, and achieving significantly improved interference source detection accuracy in the environmental state perception model; S1.2: Perform time alignment preprocessing on the original input data, use the PLC synchronization clock module to attach a unified time series label to each sensor channel, eliminate the time misalignment problem caused by communication delay or sampling frequency difference, and generate a multi-source heterogeneous data group with strict time sequence consistency as the basic data unit for subsequent feature extraction and state discrimination; Based on the multi-channel acquisition results of the original input data, the unified time synchronization function of the PLC synchronous clock module (parameter: synchronization accuracy ±1ms) is used to realize the unified processing of the full timestamps of four types of signals: billet size deviation, material composition fluctuation, mold surface temperature gradient and main cylinder pressure response delay. Furthermore, by using a multi-channel communication delay estimation method (parameter: sampling frequency difference threshold ±0.5Hz), the time offset of each sensor channel during data packet transmission and reception is calculated, and the original timestamp is corrected for differences using a delay compensation algorithm to eliminate timing inconsistencies caused by fieldbus transmission delay. Furthermore, by using a unified sampling frequency resampling algorithm (parameter: target frequency 100Hz), interpolation resampling processing is performed on the corrected channel signals to achieve timing alignment of all channel signals at the same sampling frequency and ensure that the sampling point intervals of different physical quantities are consistent. Furthermore, a timestamp normalization encoding method is adopted (parameter: using UNIX time format as the only time base) to convert the timestamps of the resampled sampling points into a unified integer time label, and the label is attached to the metadata field and stored in the data queue to achieve strict time consistency guarantee when calling across modules; Furthermore, based on the rules for constructing multi-source heterogeneous feature matrices after data synchronization, the sampled values ​​of each channel under the same time label are concatenated into column vectors to generate a multi-source heterogeneous data set with strict temporal consistency, which can be directly referenced by the subsequent feature extraction and environmental state discrimination modules. By using PLC synchronous clock for unified timing, delay compensation, resampling and timestamp normalization encoding, the multi-channel raw signal results from the previous step are transformed into a multi-source heterogeneous data group with strict time consistency, thereby achieving the goal of consistent basic data for subsequent feature extraction and pattern recognition. For example, in the process optimization scenario of the GK-5000 forging press, the sampling frequency of the laser ranging array is 102Hz, the X-ray fluorescence spectrometer is 98Hz, the infrared thermal imaging array is 100Hz, the high-speed pressure transmitter is 99.5Hz, the timing accuracy of the PLC synchronization module is ±0.5ms, and the multi-channel delay detection module evaluates the communication delay of the material composition channel as 3ms, the main cylinder pressure channel as 2ms, the billet size channel as 1ms, and the die temperature channel as 2.5ms. After performing delay compensation, resampling is performed based on the target sampling frequency of 100Hz. Missing point values ​​are generated using cubic spline interpolation. The timestamps of each channel are converted into UNIX second-level integer codes and concatenated into a four-column vector, forming a multi-source heterogeneous data group with each row corresponding to a unified time label. In this dataset, the billet size deviation is controlled within ±0.05mm, the material composition fluctuation detection accuracy is 0.01%, the mold surface temperature gradient resolution is 0.1℃, and the main cylinder pressure response delay accuracy is 0.1ms. The final verification shows that after adopting the above time alignment preprocessing, the environmental state classifier has significantly improved the recognition accuracy on the offline training set, and the online inference latency is reduced to less than 50ms. S1.3: Based on the industrial edge computing node, perform preliminary filtering on the received multi-source heterogeneous data group, use a second-order Butterworth low-pass filter to suppress high-frequency noise signals, retain effective process dynamic information in the range of 0.1~10Hz, prevent noise interference from affecting the accuracy of subsequent environmental status identification, and output the denoised clean signal stream. S1.4: The denoised clean signal stream is segmented into slices according to a preset sliding window length (e.g., 60 seconds). Each data segment is marked with its corresponding time window start and end labels to form a structured time series data block, which facilitates subsequent normalization processing and feature vector construction. S1.5: Store structured time-series data blocks in a real-time database (RTDB) and attach contextual metadata such as work order number, mold ID, and shift information to build a basic dataset for environmental state awareness that can be used for offline training and online inference. This dataset serves as the input source for a lightweight convolutional neural network classifier, supporting the accurate discrimination of subsequent environmental state categories.

[0012] Step S2: The multi-source process data undergoes sliding window normalization to eliminate dimensional differences and time drift interference, generating a standardized time-series feature vector. This standardized time-series feature vector is then input into a pre-trained lightweight convolutional neural network classifier. This classifier is trained offline based on a historically labeled set of operating condition samples and is used to determine environmental state categories. Specifically, this includes: S2.1: Obtain the multi-source process data and its corresponding time series labels recorded in the previous step S1. The multi-source process data includes billet size deviation, material composition fluctuation, mold surface temperature gradient and main cylinder pressure response delay signal, which serve as the original input dataset for this step to ensure that subsequent processing has a complete and synchronous sensor information foundation. Based on the environmental state perception basic dataset constructed in the preceding step S1.5, obtain the multi-source process data records and their corresponding time series labels stored in the real-time database (RTDB) to ensure that the input dataset contains clean signals collected from the multi-dimensional sensing channels of billet, material, mold and hydraulic system. The database index retrieval method (parameters: work order number, mold ID, shift information, etc.) is used to quickly locate all multi-source process data records under the target production batch, and load data from four core input channels: billet size deviation, material composition fluctuation, mold surface temperature gradient, and master cylinder pressure response delay signal. Furthermore, by using a data integrity verification algorithm (parameters: missing value threshold 0.5%, outlier detection range ±3σ), signal missing and abnormal fluctuations in the input dataset are detected, and a set of verified valid data frames is obtained, providing a robust original sample basis for subsequent sliding window slicing; Furthermore, by using a unified timestamp mapping algorithm (parameter: PLC synchronization clock accuracy ±1ms), the time axis of each sensing channel is made consistent, and a strictly synchronized multimodal timing matrix is ​​generated to eliminate the interference of cross-channel sampling differences on subsequent feature construction. Furthermore, through signal amplitude normalization preprocessing (parameter: preliminary normalization factor based on sensor range), preliminary amplitude scale adjustment of each channel is achieved, and original time series data blocks within a consistent amplitude range are generated, laying the foundation for subsequent high-precision Z-score standardization. Through the above algorithm processing method, the original structured time series data block output in the previous step S1.5 is transformed into a complete and synchronous multi-source process dataset that can be directly used for sliding window normalization processing in the current step, so as to achieve comprehensive preparation and quality assurance of input signals. For example, during a GK-5000 forging production process, the real-time database records multi-source process data including billet size deviation signals ranging from 0.3 to 0.8 mm, material composition fluctuation signals ranging from 1.2 to 3.5% mass fraction, die surface temperature gradient signals ranging from 40 to 85℃, and main cylinder pressure response delay signals ranging from 0.08 to 0.15 s. The search criteria are configured as work order number "W20240315", die ID "MD-06", and shift information "night shift," loading all data for the corresponding batch. Applying a data integrity verification algorithm, the billet size deviation signal missing rate is 0.3%, and the proportion of die temperature gradient signal abnormal fluctuations exceeding ±2.5σ is 0.4%, both below the missing value threshold of 0.5% and the outlier detection range of ±3σ. Therefore, all data frames are considered valid. Using a unified timestamp mapping algorithm, the data from each channel is aligned to a time axis based on the PLC synchronous clock, with time alignment accuracy controlled within ±0.5 ms. A preliminary amplitude normalization factor based on the measurement range is used to normalize the billet size deviation to the maximum range of 1mm and map it to the 0~1 interval. The mold temperature gradient is also mapped to the 0~1 interval according to the range of 100℃. Other signals are processed in the same way to generate a multimodal time series matrix with consistent amplitude. The final output dataset, under the condition that the timing synchronization error does not exceed ±0.5ms and the amplitude range consistency error does not exceed 1%, enters the sliding window slicing process in the next step S2.2 to achieve high-quality input support for the environmental state recognition algorithm. S2.2: Perform a sliding time window slicing operation on the multi-source process data based on time series labels. Using a fixed-length time window (e.g., 10 seconds) and setting an overlap rate (e.g., 50%), extract local dynamic segments of signals from each channel to generate data subsequences with temporal continuity, so as to preserve the transient change characteristics in the process and enhance the response sensitivity of environmental state identification. S2.3: For each data subsequence within a sliding window, the dimensions of the process signals are normalized separately. The mean and standard deviation are calculated using the Z-score standardization method, and the original values ​​are converted into dimensionless standard fractions to eliminate the difference in amplitude magnitude caused by different sensor types and obtain a standardized numerical range under a unified scale. S2.4: The signals from each channel that have completed Z-score standardization are spliced ​​together in a time-aligned manner to form a high-dimensional temporal feature vector. Each feature vector corresponds to a sliding window segment, which includes the joint dynamic representation of four monitoring dimensions: billet, material, mold and hydraulic system, forming a structured input sample that can be used for pattern recognition. S2.5: The generated standardized temporal feature vectors are batch input into a pre-trained lightweight convolutional neural network classifier. The classifier receives a fixed-dimensional temporal tensor as input, performs forward inference to extract deep spatial-temporal feature representations, and outputs a preliminary implicit representation of the environment state, providing a model-driven feature foundation for accurate environment state category discrimination in the next step. The high-dimensional normalized temporal feature vectors formed by concatenation in the preceding step S2.4 are loaded in batches into the feature input cache unit to ensure that the input data fully matches the requirements of the lightweight convolutional neural network classifier in terms of dimension and time length. An input tensor reconstruction algorithm (parameters: batch size B is 1-128, preferably 64; temporal length T is 10-60 seconds, preferably 10 seconds; feature dimension F is 4) is used to reconstruct the standardized feature vector into a standardized temporal feature vector of the form 4×100, thereby achieving structural alignment with the sliding kernel operation of the convolutional layer; Furthermore, the loading module is initialized with convolution kernels (parameters: kernel size is 3×3, 4×4, 5×5, stride is 1, activation function is ReLU), convolution kernel weights are constructed and mapped to the input tensor to extract local spatial-temporal correlation patterns in four dimensions: billet size deviation, material composition fluctuation, mold surface temperature gradient and hydraulic delay signal, and the convolution feature map matrix is ​​obtained. Furthermore, a batch normalization processing algorithm is adopted (parameters: (Momentum = 0.9) Performs channel-wise normalization on the convolutional feature map matrix, mapping the output of each convolutional channel to a standard distribution with zero mean and unit variance, suppressing overfitting and improving gradient propagation efficiency; Furthermore, a temporal pooling algorithm (parameters: window length = 2, stride = 2, pooling type = max pooling) is applied to compress the feature map matrix in the time dimension, retain the maximum response value of the abrupt change region, generate a compressed feature map tensor, and reduce the computational burden of subsequent fully connected layers. By using the forward inference algorithm of the fully connected layer (parameters: number of hidden units = 128, activation function ReLU), weight matrix multiplication and nonlinear mapping are performed on the compressed feature map tensor to obtain the deep spatial-temporal feature representation vector; For example, under the dynamic operating conditions of the GK-5000 equipment, the batch input standardized temporal feature vectors are set with a batch size B=64, a temporal window T=10 seconds, and a feature dimension F=4, corresponding to billet size deviation (mm), material composition fluctuation (wt%), mold surface temperature gradient (°C), and hydraulic response delay (ms). The input tensor is reconstructed into a shape of 64×10×4, with a convolution kernel size of 3×3 and a stride of 1. The initial weights are generated using a normal distribution N(0,0.01). After two layers of convolution and batch normalization, the output feature map matrix has a shape of 64×8×128. After temporal pooling, it is compressed to 64×4×128 and then input into a fully connected layer to generate a deep feature representation vector of length 128. In this embodiment, the deep feature representation can stably capture the joint temporal features of billet fluctuation and equipment performance degradation, significantly improving the accuracy of subsequent state classification and the robustness of environmental perception.

[0013] like Figure 2 As shown, step S3 involves generating a current production environment state category label based on the discrimination result output by the lightweight convolutional neural network classifier. This state category label includes 'high-fluctuation billet', 'stable operating condition', and 'early stage of equipment degradation'. The state category label is then used as an index variable for generating the dynamic weight vector in the subsequent multi-objective reward function. Specifically, this includes: S3.1: Based on the standardized temporal feature vector output by the previous step S2, obtain the original output response value of the lightweight convolutional neural network classifier. The response value is the unnormalized confidence score of each candidate environmental state category, so as to construct the initial decision basis for state discrimination. Based on the standardized temporal feature vector output by the preceding step S2, a lightweight convolutional neural network classifier forward inference algorithm (parameters: kernel size 3×3, stride 1, activation function ReLU) is used to deeply extract the spatial-temporal features of multi-source process dynamic signals and generate initial values ​​of confidence vectors for each candidate environmental state category. Furthermore, through the fully connected mapping structure of the network output layer (parameter: the weight matrix dimension is N×3, where N is the number of feature channels output by the last convolutional layer, and 3 corresponds to the number of predefined categories), the high-dimensional feature representation is projected onto the score space of three categories: 'high-fluctuation billet', 'stable working condition', and 'early stage of equipment degradation', and the unnormalized confidence value of each category is obtained. Furthermore, the batch normalization method (parameters: momentum coefficient 0.9, ...) is used. The confidence scores of the fully connected layer output are scaled to suppress score fluctuations caused by input distribution drift between different batches, and a normalized candidate category response value vector is generated. Furthermore, a numerical truncation mechanism (parameter: truncation range [-10, 10]) is used to truncate each component of the batch normalized response value to prevent score overflow caused by extreme inputs or abnormal activation within the network, thereby ensuring the numerical stability of the subsequent probability normalization process. The algorithm described above transforms the standardized time-series feature vector from the previous step into unnormalized confidence scores for each candidate environmental state category, thus enabling the construction of the initial decision basis for state discrimination. For example, under stable operating conditions of 16 strokes per minute and a nominal force of 50,000 kN, the GK-5000 forging press was equipped with multi-source signals including billet size deviation (maximum 0.15 mm), die surface temperature gradient (maximum 45 °C), main cylinder pressure response delay (maximum 0.12 s), and material composition fluctuation (Fe content variation of 0.3%). These signals were processed by a sliding time window to generate a standardized temporal feature vector of dimension 4 × 100, which was then input into a lightweight convolutional neural network classifier. The last convolutional layer of the classifier outputs 64 feature channels, which are mapped by a fully connected layer to generate three unnormalized confidence scores: 8.35 for the 'high-fluctuation billet' category, 11.27 for the 'stable operating conditions' category, and -3.14 for the 'early stage of equipment degradation' category. After batch normalization adjustment and truncation, stable response values ​​of 8.21, 10.96, and -3.05 were obtained, respectively. These response values ​​will be converted into a probability distribution in the subsequent Softmax normalization step to determine the current operating condition category; S3.2: Perform Softmax normalization on the unnormalized confidence score to convert it into a probability distribution form of the environmental state posterior probability vector, where each dimension corresponds to the probability of occurrence of a predefined state category such as 'high fluctuation billet', 'stable operating condition', or 'early stage of equipment degradation', in order to quantify the uncertainty level of the current operating condition; S3.3: Based on the posterior probability vector of the environmental state, the maximum a posteriori probability criterion (MAP) is used to perform state category decision, select the category with the highest probability value as the optimal state estimation result, and generate the corresponding discrete state category label to form a clear semantic environmental identifier. Based on the environmental state posterior probability vector output in the preceding step S3.2, the maximum a posteriori probability criterion (MAP) method is adopted (parameters: environmental state category set Ω={'high fluctuation billet', 'stable operating condition', 'early stage of equipment degradation'}, posterior probability vector P) to realize the function of optimal category selection for the current production environment state; Furthermore, through a probability comparison algorithm (parameters: This allows for the retrieval of a single component with the highest probability value from the posterior probability vector, and the acquisition of its corresponding category index data or result, thereby ensuring that the state decision is statistically optimal. Furthermore, through the index mapping method (parameters: category index i, category name set Ω), the index position where the maximum probability value is located is converted into a semantic category label, and clear environmental state identification data is generated for subsequent query control of the weight mapping table; Furthermore, through the decision consistency verification algorithm (parameter: current category label) Previous period category label Threshold This allows for a quantitative assessment of the mutation risk of category switching, and identifies situations where the risk is below a threshold. Confirm the validity of the category output when the risk is higher than or equal to the threshold. The time stamp is marked as a potential abnormal output for reference by subsequent smoothing mechanisms; Furthermore, through label encoding processing (parameter: current category label) The encoding rule table R converts semantic environment state category labels into standardized discrete integer labels to be compatible with the input requirements of subsequent table lookup and matrix operation structures. The maximum a posteriori probability decision algorithm transforms the environmental state posterior probability vector from the previous step into discrete state category labels, enabling deterministic identification of the current production condition and providing accurate index data input for the generation of dynamic weight vectors. For example, in a certain batch of production, the posterior probability vector of the environmental state is P={0.15,0.75,0.10}, where each dimension corresponds to one of three operating conditions: 'highly volatile billet', 'stable operating condition', and 'early stage of equipment degradation'. The MAP decision formula is used:

[0014] in, For elements in the category set Ω, Input samples as standardized time-series feature vectors. Represents the parameters under the given observation data. The posterior probability; argmax represents finding the posterior probability. The biggest one The value of in the posterior probability vector. In the range, the maximum value is 0.75, corresponding to the category 'steady-state condition', therefore the generated discrete category label is... =Stable operating condition. Using the encoding rule table R, 'Stable operating condition' is mapped to the integer index value 1, which is used to directly query the configuration data in the first row of the weight mapping table. In this batch case, the decision consistency check result shows that the label in the previous period was also 'Stable operating condition,' and there is no significant risk of abrupt changes in state transition; therefore, the label output is valid. Ultimately, the execution effect of this step is to generate a clear and standardized category label for the current production environment, and input it as an index variable into the weight mapping module to support the adaptive adjustment of the priority of the subsequent multi-objective reward function. S3.4: Encode the discrete state category label into a unique integer index value, the index value corresponding one-to-one with the row number in the preset weight mapping table, so as to construct an addressable mapping channel from the environmental identification result to the multi-objective optimization priority configuration; S3.5: Output the integer-indexed state category label as the control enable signal for the dynamic weighting of the multi-objective reward function in the next stage, so as to drive the weight mapping module to select the appropriate normalized weight vector according to the current actual working conditions, and realize the context-aware adjustment of the optimization target priority.

[0015] like Figure 3 As shown, step S4 involves querying a preset weight mapping table based on the state category label to obtain the corresponding normalized weight vector. This weight mapping table is jointly calibrated by domain expert experience and historical optimization trajectories, and is used to assign differentiated priority weights to the three optimization objectives: forging forming quality score, unit product energy consumption cost, and mold stress accumulation index. Specifically, this includes: S4.1: Based on the current production environment status category label output from the previous step S3, use it as an index input variable, wherein the status category label includes 'high fluctuation billet', 'stable operating condition' and 'early stage of equipment degradation', which is used to trigger the multi-objective weight configuration logic under the corresponding operating condition and generate the environment-driven priority selection basis; S4.2: Construct and load a preset weight mapping table. The weight mapping table is an offline calibrated lookup table structure. Its row index is the environmental state category label. The column items include the initial weight coefficients of three optimization objectives: forging forming quality score, unit product energy consumption cost and mold stress accumulation index. This table is jointly calibrated and generated by domain expert experience and historical best process trajectory to reflect the relative importance of each objective under different interference conditions. Given that the current production environment status category label has been output by the previous step S4.1, the input object is the initial weight configuration requirements of three optimization objectives, which correspond to the forging forming quality score, unit product energy consumption cost and mold stress accumulation index, respectively. A structured lookup table construction method (parameters: row index = environment state category label, column field = initial weight coefficients of three optimization objectives) is adopted to realize the fixed mapping relationship of multi-objective weight configuration; Furthermore, by using a domain expert experience quantification method (parameters: expert scoring matrix, historical sample data of production process), the relative importance coefficients of each optimization objective under different environmental interference conditions are assigned, and a draft of the weight configuration matrix is ​​formed. Furthermore, by using the historical optimal process trajectory backtracking analysis method (parameters: performance index time series, state label series), the initial draft weight coefficients are corrected, the weight distribution characteristics that exhibit the best comprehensive performance under a specific state category are extracted, and the column values ​​of the corresponding row in the weight mapping table are updated. Furthermore, through an offline calibration algorithm (parameters: difference evaluation function, target performance boundary constraint), the updated weight mapping table is subjected to a full table consistency check and boundary constraint verification, and weight values ​​that exceed the feasible region are adjusted to the legal range. Through the above construction and loading process, the environmental state category label of the previous step is transformed into an initial weight coefficient that can be directly used for multi-objective weighted fusion calculation, thereby achieving the effect of environment-driven dynamic mapping technology to optimize target priority. For example, in the GK-5000 forging production scenario, the weight mapping table contains three rows, corresponding to three state categories: 'High-Fluctuation Billet', 'Stable Operating Condition', and 'Early Stage of Equipment Degradation'. For the 'High-Fluctuation Billet' row, the initial weight for the forging forming quality score is set to 0.6, the unit product energy consumption cost is set to 0.25, and the die stress accumulation index is set to 0.15. For the 'Stable Operating Condition' row, the initial weight for the forging forming quality score is 0.4, the unit product energy consumption cost is 0.4, and the die stress accumulation index is 0.2. For the 'Early Stage of Equipment Degradation' row, the initial weight for the forging forming quality score is 0.5, the unit product energy consumption cost is 0.2, and the die stress accumulation index is 0.3. During the mapping table generation process, initial coefficients are first assigned to each state based on an expert experience database, and then the weights are fine-tuned using a mean deviation correction algorithm, combined with historical best production section data from the past 24 months. For example, if historical data backtesting reveals that the actual sensitivity of energy consumption costs in the 'early stage of equipment degradation' state is higher than the initial expert setting, the energy consumption cost weight for that row in the mapping table is updated to 0.25, and the mold stress accumulation index weight is adjusted accordingly. In the final loading stage, the system performs an L1 norm check on each row of weights in the entire table, ensuring normalization validity through the constraint that the sum of weights equals 1, thus forming a stable and usable weight mapping table that provides reliable input for subsequent multi-objective reward function calculations. S4.3: Based on the state category labels provided in S4.1, perform an exact match query in the weight mapping table constructed in S4.2, extract the original weight coefficient vector of the corresponding row, and perform L1 norm normalization on the vector to obtain a normalized weight vector in which the sum of the weights of each objective is 1, which serves as the basic input for the weighted fusion of the multi-objective reward function. S4.4: Assign each component of the normalized weight vector output by S4.3 to the corresponding sub-objective item in the multi-objective reward function. The forging forming quality score item gets the first weight component, the unit product energy consumption cost item gets the second weight component, and the mold stress accumulation index item gets the third weight component, thereby completing the mapping transformation from environmental state to multi-objective priority. The normalized weight vector output by S4.3 is received as a direct input condition for the weight allocation of the multi-objective reward function. The vector contains three components, which correspond to the optimized weights of the forging forming quality score, the unit product energy consumption cost, and the mold stress accumulation index, respectively. An index mapping method (parameter: a table showing the correspondence between the normalized weight vector component index and the reward function sub-objective item index) is used to achieve accurate matching between weight components and reward function sub-objective items. During the matching process, data type verification is performed on each component to ensure that the weight coefficient is a floating-point number and is located in the [0,1] interval. Furthermore, through a parameter binding algorithm (parameter: weight components) , , The first component is bound to the forging quality score item of the multi-objective reward function, the second component is bound to the unit product energy consumption cost item, and the third component is bound to the mold stress accumulation index item. The binding relationship is stored by reference to ensure the synchronous consistency of weight assignment during function call. Furthermore, an initialization assignment operation is performed on the bound sub-target items using a linear weighted update rule. The operation model is as follows:

[0016] in, For a multi-objective comprehensive score, , , These are three normalized weight components. To standardize the quality score of forgings, To standardize unit energy consumption indicators, To standardize the mold stress accumulation index; Furthermore, a floating-point precision retention mechanism (parameter: retention precision set to 6 decimal places) is applied to truncate and round the above assignment operation results to ensure that the transfer of data in different computing modules does not introduce cumulative numerical drift. Through the above weight allocation and assignment operations, each component of the normalized weight vector is transformed into the effective weight coefficient of the corresponding sub-objective in the multi-objective reward function, realizing the functional mapping from environmental state recognition results to multi-objective priority configuration, and providing a precise parameter basis for subsequent dynamic weighted fusion. For example, under a certain dynamic working condition, the normalized weight vector obtained after processing by S4.3 is 0.6, 0.25, 0.15, where the first component 0.6 is assigned to the forging forming quality score, the second component 0.25 is assigned to the unit product energy consumption cost, and the third component 0.15 is assigned to the die stress accumulation index. In this scenario, the three standardized indices are 0.92, 0.78, and 0.55, respectively. Substituting these values ​​into the weighted fusion formula:

[0017] The calculated multi-objective comprehensive score is 0.8085, and the precision control output after retaining 6 decimal places is 0.808500. This result is passed to the S4.5 output stage and cached. It can be directly called in the KL divergence smoothing process in S5 to realize the strategy configuration effect of prioritizing the quality of forgings under high fluctuation billet conditions, and significantly improve the stability of the strategy and the consistency of process response under dynamic working conditions. S4.5: Output the dynamic weight configuration result generated in S4.4 to the next processing stage as the basis for the weighted fusion of the multi-objective reward function. At the same time, cache the weight vector to the local storage unit for subsequent KL divergence calculation and exponential moving average smoothing mechanism to support policy transition stability control.

[0018] Step S5: Based on the normalized weight vector, the components in the multi-objective reward function are weighted and fused to generate a dynamic comprehensive reward value. The weighted fusion process incorporates an exponential moving average mechanism under KL divergence constraints to achieve a smooth transition between the old and new weight vectors, thus avoiding sudden changes in the control strategy due to misjudgment of the state. Specifically, this includes: S5.1: Obtain the normalized weight vector output from the previous steps. The normalized weight vector includes the weight coefficients corresponding to the three optimization objectives: forging forming quality score, unit product energy consumption cost, and mold stress accumulation index. Perform initial weighting operation based on the normalized weight vector. Use a linear weighting model to perform weighted summation on each standardized multi-objective performance index to generate a preliminary comprehensive reward value, which serves as the basic output for dynamic reward adjustment. S5.2: Obtain the historical weight vector stored in the previous control cycle and input it as a reference benchmark into the KL divergence calculation module; based on the probability distribution difference between the current weight vector and the historical weight vector, use the Kullback-Leibler divergence algorithm to calculate the state deviation between the two, so as to quantify the level of sudden change risk during the environmental state switching process, and output the KL divergence value as the trigger basis for the smooth adjustment mechanism. Based on the normalized weight vector corresponding to the preliminary comprehensive reward value generated in the preceding step S5.1 and the historical weight vector cached in S4.5, an input data structure for state deviation measurement is constructed. The historical weight vector stored in the previous control cycle is read using a variable index. The memory management module is called to ensure its data integrity and timing correctness. It is then used as a reference benchmark and input to the KL divergence calculation module to establish a direct comparison channel with the current weight vector. Furthermore, a vector normalization verification algorithm (parameter: L1 norm = 1) is used to ensure that the current weight vector and the historical weight vector have the same probability distribution scale, providing a prerequisite for subsequent information theory-based difference measurement. Furthermore, the Kullback-Leibler divergence algorithm (parameter: natural logarithm base e) is used to realize the current weight vector. With historical weight vector The mathematical form of relative entropy calculation within the probability distribution space is:

[0019] in, Let KL divergence value be denoted as . This represents the i-th component of the current weight vector. Let represent the i-th component of the historical weight vector, and ln be the natural logarithm function; Furthermore, the weight distribution difference is calculated component by component using the above formula and accumulated in the numerical calculation module to obtain the overall KL divergence value, so as to quantify the degree of deviation of the probability distribution between the current environmental state and the previous cycle state. By archiving and outputting KL divergence values, the calculated deviation is used as the trigger for the smoothing adjustment mechanism, thereby enabling a quantitative assessment of the risk of sudden environmental state transitions and providing data support for subsequent adaptive threshold judgment and exponential moving average fusion. By using the information theory metric algorithm described above, the weight vector comparison results from the previous step are transformed into a quantifiable deviation index, thereby achieving an accurate measurement of the risk of sudden environmental changes and providing a scientific basis for dynamic reward smoothing strategies. For example, during a dynamic operating cycle of the GK-5000 forging press, the current normalized weight vector is [0.6, 0.25, 0.15], and the historical weight vector is [0.5, 0.35, 0.15]. The KL divergence calculation module is called. First, it is verified that the sum of the two vector components is 1, satisfying the probability distribution condition. The contribution of each component is calculated according to the formula: the contribution value of the first component is 0.6 × ln(0.6 / 0.5) = 0.6 × ln(1.2), the contribution value of the second component is 0.25 × ln(0.25 / 0.35) = 0.25 × ln(0.714), and the contribution value of the third component is 0.15 × ln(0.15 / 0.15) = 0. The values ​​of each component are calculated sequentially in the numerical calculation module and accumulated. The sum is a KL divergence value of 0.041, indicating a slight weight distribution shift in the current state compared to the previous cycle. In the threshold judgment unit, this value is lower than the preset mutation threshold of 0.05, so the exponential moving average mechanism will not be triggered, maintaining the efficiency of the control response. Finally, this deviation result is used as the input reference for the dynamic reward fusion in the next cycle, realizing a stable strategy output under slight operating condition fluctuations. S5.3: Input the KL divergence value to the adaptive smoothing threshold judgment unit and perform a comparison operation based on the preset critical threshold range; if the KL divergence value exceeds the set threshold, it is determined that there is a significant risk of state jump and the exponential moving average adjustment mechanism is activated; otherwise, the current weight vector is maintained and directly output to ensure control response efficiency and ensure that no redundant delay is introduced under stable operating conditions. S5.4: Under the activated exponential moving average adjustment mechanism, the current weight vector and the historical weight vector are fused based on the exponential moving average algorithm to generate a transitional weight vector, so as to obtain an intermediate weight state with time continuity, thereby suppressing reward oscillations caused by environmental classification misjudgment or instantaneous disturbances. Based on the smooth adjustment condition triggering result of the previous step S5.3, the normalized weight vector generated in the current control cycle and the historical weight vector cached in the previous control cycle are input as the fusion processing object. The exponential moving average (EMA) algorithm is used (parameter: smoothing factor). , range 0 < <1), change the current weight vector With historical weight vector Placed within a recursive filtering structure, it achieves gradual fusion to generate intermediate weight states that are continuous in time; Furthermore, through the formula

[0020] To achieve a weighted fusion of the current weight vector and the historical weight vector, where... The transition state weight vector, As a smoothing factor, it is adaptively adjusted according to the amplitude of the KL divergence value; Furthermore, by performing a normalization check on each component obtained from the formula, it is ensured that the weight vector after fusion still satisfies the constraint that the sum of the three is 1, and the non-negativity of each component is preserved. Furthermore, by outputting the transition state weight vector through recursive filtering, the rate of weight change in the time series is significantly suppressed, thereby reducing the oscillation of reward value caused by environmental classification misjudgment or instantaneous disturbances. Through the above exponential moving average fusion process, the current weight vector and historical weight vector determined in the previous step are transformed into weight data after a smooth transition, thereby achieving the temporal continuity of the reward function input signal and the stability of the control strategy. For example, in a single-condition switching of the GK-5000 forging press, the current cycle normalized weight vector is [0.65, 0.20, 0.15], the previous cycle weight vector is [0.50, 0.30, 0.20], the KL divergence calculation result is 0.45, which is greater than the system-set smoothing threshold of 0.40, and the smoothing factor α is adaptively set to 0.6. Substituting into the formula:

[0021] The transition state weight vector [0.59, 0.24, 0.17] is calculated component by component and checked to ensure that the sum of the three is 1, which satisfies the normalization requirement. In the subsequent weighted fusion of the reward function, this weight vector significantly mitigates the gradient of reward change caused by the high fluctuation of billet conditions. The update process of the policy network in the SAC framework remains stable, ensuring that the action commands of heating temperature, impact energy and holding time are smoothly transitioned within the physical boundary conditions of the equipment, thereby achieving the effect of continuous control of the production process. S5.5: Substitute the smoothed transition state weight vector into the multi-objective reward function structure, and re-perform a weighted fusion calculation on the three indicators of forging forming quality score, unit product energy consumption cost, and mold stress accumulation index to generate the final dynamic comprehensive reward value; the dynamic comprehensive reward value serves as the input signal of the soft actor-critic algorithm to drive the policy network update, ensuring that the reinforcement learning decision has good environmental adaptability and control stability.

[0022] Step S6: The dynamic comprehensive reward value is input into the soft actor-critic algorithm framework to drive the policy network update, generating a three-dimensional continuous motion vector containing the heating temperature setpoint, the energy increment per strike, and the pressure holding time adjustment. The output of the policy network is constrained by the tanh activation function and combined with the GK-5000 device's stroke count of 16 sppm and nominal force of 50000 kN as physical boundary constraints. Specifically, this includes: S6.1: Based on the dynamic comprehensive reward value output in the previous step S5, construct the temporal sample tuple required by the soft actor-critic (SAC) algorithm. The sample tuple includes the current production environment state category label, the process parameter action vector generated in the previous time step, the corresponding dynamic comprehensive reward value, and the environment state in the next time step. Use this sample tuple as input to fill the training data of the actor network and two independent critic networks in the SAC framework to support the subsequent policy gradient calculation based on the maximum entropy reinforcement learning principle. Based on the dynamic comprehensive reward value and the corresponding environmental state feature vector generated in the preceding step S5, a sample sequence construction method (parameters: current production environment state category label, previous action vector, dynamic comprehensive reward value, and next environmental state) is used to generate the time-series sample tuples required by the soft actor-critic algorithm. Furthermore, by using a state feature encoding method (parameters: multidimensional continuous feature vector, normalized scale, time series position index), the tensor representation of multi-source environmental features in a unified input space is realized, and state-encoded data matching the input dimension of the policy network is obtained. Furthermore, an action history cache retrieval method (parameters: previous action vector cache area, timestamp matching rule) is adopted to accurately extract the process parameter adjustment instructions of the previous execution cycle and generate state-action pairs associated with environmental state features; Furthermore, by using a reward scalar binding algorithm (parameters: dynamic comprehensive reward value, normalization coefficient λ), the reward signal is bound to the state-action pair, forming a complete training sample structure that includes value feedback; Furthermore, by utilizing the state transition recording method (parameters: next-time environmental state feature vector, termination flag), the state transition relationships in the sample tuples are determined, and a data format conforming to the soft actor-critic framework is generated. A set of four tuples, where s is the current production environment state category label, a is the process parameter action vector generated in the previous time step, and r is the dynamic comprehensive reward value. This is the feature vector of the environment state at the next time step; By using a time-series sample batch loading method, the above set of four tuples is transformed into batch training data, which is then filled into the actor network and two independent critic networks in the SAC framework, respectively, to realize the input preparation required for policy gradient calculation based on the maximum entropy reinforcement learning principle. For example, in the dynamic process optimization scenario of the GK-5000 forging press, the current production environment state category label is extracted by a convolutional neural network to form a normalized tensor with a dimension of 128. The action vector at the previous moment is (0.15, -0.08, 0.05), and the dynamic comprehensive reward value is 0.72. The environmental state feature vector at the next moment also has a dimension of 128 and includes encoded features of billet fluctuation and equipment temperature response delay. In the state encoding step, the environmental features are normalized to the interval [−1, 1] and a time position index vector is added; the action history cache is retrieved to explicitly match the timestamp of the previous cycle, and the impact energy increment and holding time adjustment ratio are extracted. When binding the reward scalar, a normalization coefficient of 1.0 is used to keep the reward value scale unchanged, and the binding result is generated. The state transition record determines the termination flag bit as 0, forming a complete state transition record. The sample tuple set contains 64 items, which are batch-input into the actor network and the critic network. After one training iteration, the stability of the policy output is significantly improved, and the action vectors generated by the policy network in the next iteration achieve the expected process optimization performance target. S6.2: Perform Q-value estimation on the critic network, using a dual-depth Q-network structure to calculate the two Q-value outputs of the current state-action pair, and select the lower value as the final Q-evaluation result to suppress excessively high estimation bias; based on the Bellman residual between the Q-value and the dynamic comprehensive reward value, calculate the mean squared error loss function of the critic network, and optimize its network weights through the backpropagation algorithm to ensure the stability and convergence of the value function estimation; For the input current production environment state category label and the process parameter action vector generated in the previous time step, a dual-depth Q-network structure (parameters: three network layers, 256 hidden units, ReLU activation function) is used to perform forward propagation, generating two independent Q-value outputs for the current state-action pair, which serve as preliminary estimation data for the value function. The two Q-value outputs are compared using a minimum value selection operator, and the lower value is selected to generate the final Q-evaluation result. This suppresses the estimation bias caused by overfitting of a single network and forms a robust basis for strategy evaluation. Furthermore, based on the Q-value estimation of the current dynamic comprehensive reward value and the environmental state at the next time step in the critic network, a Bellman residual calculation input is constructed. The difference between the target Q-value and the current Q-evaluation result is quantified as a mean squared error loss function, as shown in the following formula:

[0023] in, For mean square error loss, This refers to the batch sample size. The Q-evaluation result for the current state-action pair. The Bellman expected return is calculated by applying a discount factor to the dynamic composite reward value and the next-state target Q value. The gradient of the mean squared error loss function is calculated using the gradient backpropagation algorithm. The resulting gradient signal is then input into the optimizer (parameters: type Adam, learning rate 0.0003) to update the weight parameters of the double Q network, thereby improving the stability and convergence of the value function estimation. By combining the above-mentioned dual-network minimum selection with Bellman residual constraints, the dynamic comprehensive reward value of the previous step is effectively transformed into a stable value evaluation index required for the control strategy iteration, thereby realizing the robustness and long-term convergence effect of the reinforcement learning strategy in the dynamic forging environment. For example, in the operating environment of the GK-5000 forging press, the batch sample size is set to 64. The dynamic comprehensive reward value is obtained through weighted fusion in the preceding step S5, with a discount factor configured to 0.99. The input of the dual-depth Q-network is a vector composed of three process parameters (heating temperature, impact energy, and holding time) and environmental features. The first and second layers of the network each contain 256 hidden units, with ReLU as the activation function. The third layer directly outputs the Q-value. The two network outputs are Q1=125.7 and Q2=128.4, respectively. After the minimum value selection operator, the final Q=125.7 is obtained. The target Q-value for the next state is obtained by discount calculation and fusion with the dynamic reward, resulting in Y=126.3. The Bellman residual is -0.6, corresponding to the mean square error loss. After iteratively updating the critic network weights using the Adam optimizer, the fluctuation of the value function output is significantly reduced, and the stability of the policy network's response to changes in the environmental state is significantly enhanced in subsequent training. S6.3: Based on the updated critic network, calculate the policy gradient of the actor network: First, generate candidate action samples using the reparameterization technique, and combine the action probability distribution output by the current policy network to calculate the maximum expected reward objective function containing the entropy regularization term; perform automatic differentiation on the objective function to obtain the gradient direction of the policy network parameters, and update the actor network parameters through the Adam optimizer to achieve a policy evolution that balances exploration efficiency and long-term benefits. S6.4: The original motion vector output by the updated actor network is nonlinearly mapped using the tanh activation function to generate a normalized three-dimensional continuous motion vector. Its three dimensions correspond to the relative changes of the heating temperature setpoint, the energy increment per hit, and the pressure holding time adjustment, respectively. This normalization process ensures that the output motion is within the range of [-1, 1], providing a standardized control command basis for subsequent physical equipment adaptation. S6.5: Based on the physical characteristic parameters of the GK-5000 forging press—stroke count 16 spm and nominal force 50000 kN—a motion space inverse mapping model is constructed: the normalized three-dimensional continuous motion vector is multiplied by the adjustable range amplitude of its corresponding process parameter, and a reference operating point is superimposed to convert it into an actual executable absolute process parameter instruction; among which, the stroke count 16 spm is used to constrain the upper limit of the maximum number of strikes per unit time, and the nominal force 50000 kN is used as the absolute upper limit of pressure output to participate in the safety verification of the pressure holding stage, ensuring that the generated motion vector meets the equipment dynamic boundary conditions and process safety requirements; Based on the normalized three-dimensional continuous motion vector, the motion space inverse mapping method (parameters: adjustable range of process parameters of GK-5000 equipment, reference operation point) is used to realize the transformation of standardized control ratios into actual process instructions with physical dimensions. Furthermore, for the heating temperature dimension motion component, a linear scaling method (scaling factor: the temperature adjustment amplitude allowed by the equipment) is adopted, and the current reference temperature value is superimposed to generate the actual heating temperature setpoint, which is used to drive the heating system; Furthermore, for the motion component of the energy increment dimension of each strike, the same linear scaling method (scaling coefficient: the energy adjustment amplitude allowed by the device) is adopted, and combined with the stroke limit of GK-5000, the result is added to the baseline strike energy to generate the energy command for the strike stroke configuration per strike, while ensuring that the upper limit constraint of 16 spm is not exceeded per unit time. Furthermore, for the action component of the pressure holding time adjustment dimension, a linear scaling method (proportional coefficient: allowable pressure holding time adjustment amplitude) is adopted, and a reference pressure holding time value is added to generate the actual pressure holding stage duration command. At the same time, a nominal force of 50000kN is used to perform a safety check on the pressure of the pressure holding stage to prevent exceeding the upper limit of equipment dynamic safety. The absolute value of each action dimension is calculated using the following formula:

[0024] in, These are the converted absolute process parameter values. This is the baseline operation point value for the corresponding dimension. The allowable adjustment range of the equipment. These are the component values ​​of the normalized action vector; Furthermore, through a safety verification algorithm (parameters: number of strokes 16 sppm, nominal force 50000 kN), the joint compliance verification of the heating temperature, impact energy, and holding time output by the inverse mapping of the action is realized, and the final executable absolute process parameter instruction set is generated. By using the action space inverse mapping model, the normalized action vector is transformed into absolute parameter commands that meet the requirements of equipment dynamics boundary and process safety, so as to directly apply the reinforcement learning strategy output to the physical control level of the GK-5000 forging press. For example, on the GK-5000 device, the adjustable range of heating temperature is set to 80℃, the reference temperature to 1150℃, the adjustable range of impact energy per strike to 4kJ, the reference energy to 28kJ, the adjustable range of holding time to 1.5s, and the reference holding time to 8.0s. Assuming the normalized output motion vector is (0.25, -0.4, 0.1), the heating temperature setpoint is calculated as: 1150 + 80 × 0.25 = 1170℃; the impact energy per strike is calculated as: 28 + 4 × -0.4 = 26.4kJ; and the holding time is calculated as: 8.0 + 1.5 × 0.1 = 8.15s. During energy setting, the impact frequency is checked to ensure it does not exceed 16 spm, and during the pressure holding stage, the pressure is verified to ensure it does not exceed the safety limit of 50,000 kN. The final output is an absolute instruction set with a heating temperature of 1170℃, an impact energy of 26.4 kJ, and a pressure holding time of 8.15 s. After execution, the geometric accuracy stability of the forging and the welding quality are significantly improved under the condition of high fluctuation billet, while the energy consumption is controlled within the target range of process optimization.

[0025] Step S7: Execute the process parameter adjustment command defined by the three-dimensional continuous motion vector, control the forging equipment to operate in the updated process window, and simultaneously collect actual forming quality, measured energy consumption values, and mold stress feedback data to form a real environmental response result under closed-loop control. Specifically, this includes: S7.1: Based on the three-dimensional continuous motion vector output by the soft actor-critic algorithm framework, the heating temperature setpoint, the energy increment per strike, and the pressure holding time adjustment contained therein are analyzed as process parameter adjustment instructions to be executed. The motion vector has been limited by the tanh activation function and physical boundary verification is performed in combination with the number of strokes of 16 sppm and the nominal force of 50000 kN of the GK-5000 equipment to ensure that the instructions are within a safe and executable range. Based on the three-dimensional continuous motion vector output by the soft actor-critic algorithm, the parameter parsing method (input: normalized motion vector, dimension index = 1, 2, 3) is used to map each dimension to the control variables corresponding to the heating temperature setpoint, the energy increment per hit, and the pressure holding time adjustment. Furthermore, by using a boundary constraint detection algorithm (parameters: GK-5000 equipment stroke count 16 sppm, nominal force 50000 kN), the physical characteristics of the equipment are verified by the analyzed control variables, and the set of allowable range data for each process parameter is obtained. Furthermore, based on the tanh normalized inversion model (parameters: normalized action value ∈ [-1,1], preset process parameter adjustment amplitude, reference operating point), the normalized control variable is quickly converted into absolute process parameter command, and three execution data are generated: absolute setpoint of heating temperature, absolute increment of impact energy, and absolute adjustment of holding time. Furthermore, the safety margin calculation method is adopted (formula: ,in For the current action pressure command, (With a nominal force of 50,000 kN), it performs a safety assessment of pressure commands and generates safety enable flag data. To control the authority to issue subsequent instructions; By using boundary verification and safety margin assessment, the analysis results of the previous step are transformed into executable process parameter instructions that meet equipment dynamic constraints, thereby achieving safe and executable process window adjustments. For example, in a certain dynamic optimization loop, the three-dimensional continuous motion vector output by the soft actor-critic algorithm is [0.35, -0.15, 0.25]. This vector, after parameter analysis, corresponds to a relative change of 0.35 in the heating temperature setpoint, a relative change of -0.15 in the impact energy per strike, and a relative change of 0.25 in the holding time. Using a boundary constraint detection algorithm, combined with the adjustable ranges of the GK-5000 heating temperature [1150℃, 1250℃], impact energy [40kJ, 60kJ], and holding time [2.0s, 4.0s], an allowable range set is generated. Calculated using the tanh normalized inversion model: Heating temperature command = 1200℃ + 0.35 × 50℃ = 1217.5℃; Impact energy command = 50kJ + (-0.15) × 10kJ = 48.5kJ; Holding time command = 3.0s + 0.25 × 1.0s = 3.25s. Using the safety margin calculation method, assuming the current pressure command corresponds to 40,000kN, the margin coefficient is... When the value is below the preset safety threshold of 0.8, the safety enable flag is valid, and the command can be issued. The execution results show that after this adjustment, the deviation between the billet forming quality and the target is significantly reduced, the unit forming energy consumption is significantly optimized, the mold stress accumulation rate decreases, and the effect on improving multi-objective performance is significant. S7.2: The verified process parameter adjustment command is transmitted to the programmable logic controller (PLC) of the forging press, which drives the heating system to adjust the billet heating temperature to the target set value, controls the hydraulic servo system to dynamically adjust the striking stroke and speed according to the energy increment of each strike, and precisely sets the holding time through the time relay to realize the real-time setting of the process window of the GK-5000 forging press. S7.3: During the new process window of the equipment, the multi-channel data synchronous acquisition module is activated to acquire the actual forming quality data, the measured value of energy consumption per unit product and the stress feedback signal of the key parts of the mold during the forming process. The actual forming quality data is acquired by the online vision inspection system and the dimensional measuring instrument. The measured value of energy consumption is obtained by the smart meter recording the instantaneous power integral of the main drive motor. The mold stress feedback signal comes from the time-series voltage signal output by the embedded strain sensor array. S7.4: Perform geometric accuracy evaluation processing on the collected actual forming quality data, calculate the root mean square error of the key dimension deviation of the forging; normalize the measured value of energy consumption per unit product to generate the unit forming energy consumption index; perform frequency domain filtering and peak extraction on the mold stress feedback signal to generate the mold stress accumulation index. The three together constitute the real environment response result vector under closed-loop control, which is used to quantify the current strategy execution effect. S7.5: Upload the vector of real-world response results to the cloud optimization server as the basis for subsequent reverse fine-tuning of the priority configuration parameters in the discrimination threshold and weight mapping table of the lightweight convolutional neural network classifier, forming a data closed loop from policy execution to performance feedback, supporting the continuous evolution of the dual closed-loop self-evolution mechanism of environment recognition and reward adjustment.

[0026] Step S8: The deviation signal between the real environment response result and the expected target is converted into a gradient signal that can be used to optimize the classifier's discrimination threshold. Gradient backpropagation is performed to adjust the discrimination threshold of the output layer of the convolutional neural network classifier. The priority ratio among the forging forming quality score, unit product energy consumption cost, and mold stress accumulation index is adjusted to form a dual-closed-loop self-evolutionary mechanism of environmental recognition and reward adjustment, thereby improving the adaptability and global optimization stability of the multi-objective reinforcement learning strategy under dynamic working conditions. Specifically, this includes: S8.1: Obtain the real environmental response results collected after the process parameters are adjusted under closed-loop control. The real environmental response results include actual forming quality data, measured energy consumption per unit product and mold stress accumulation index feedback signal, which serve as the basis for evaluating the performance of the current strategy and are used to drive the online optimization of the classifier and weight mapping table. Based on the real environmental response result vector uploaded in step S7.5, three types of indicators are extracted as the input dataset for this step: actual forming quality data, measured energy consumption per unit product, and feedback signal of mold stress accumulation index.

[0027] By employing a data parsing method (parameters: forming quality detection data package, energy consumption detection record, mold stress sensing signal), various raw response results are decoded into independent structured numerical sequences, while maintaining a unique association index with the adjustment strategy.

[0028] Furthermore, by using a time-series synchronization mapping algorithm (parameters: policy execution timestamp, response result collection timestamp, and allowed synchronization deviation threshold), strict temporal alignment of the three types of response results is achieved, resulting in a multi-index response vector with a unified time benchmark.

[0029] Furthermore, a multi-scale feature extraction method (parameters: geometric accuracy error curve, energy consumption fluctuation curve, stress accumulation curve) is adopted to calculate the root mean square error of key dimensional deviations in forming quality data, extract the unit forming energy consumption index from energy consumption data, and statistically analyze the peak stress and cumulative stress increment of stress signals, thereby generating a set of performance quantitative indicators that are comparable and traceable.

[0030] Furthermore, an outlier removal algorithm (parameters: three times the absolute deviation threshold of the median, sliding window length) is used to clean up outliers in three types of performance indicators, so as to eliminate the interference of sensor instantaneous distortion or communication packet loss and obtain a stable and reliable response feature dataset.

[0031] Through the aforementioned data processing chain, the actual execution feedback results under closed-loop control are transformed into structured, robust, and usable real-world response indicators. These indicators serve as the driving source data for online optimization of the classifier's discrimination threshold and the priority parameters of the weight mapping table, thereby enabling precise input conditions for subsequent adaptive adjustment mechanisms.

[0032] S8.2: Based on the deviation between the actual environmental response result and the expected target, calculate the environmental state classification error index, and use the error index to adjust the gradient backpropagation of the output layer discrimination threshold of the lightweight convolutional neural network classifier to improve its classification accuracy and robustness under the current working conditions.

[0033] Based on the actual environmental response results obtained during the closed-loop control feedback phase and the preset expected target value, a deviation measurement algorithm is adopted (parameter: actual forming quality RMS deviation Q). real Target forming quality RMS deviation Q set Actual unit product energy consumption E real Energy consumption per unit of product E set Actual mold stress accumulation index S real Target mold stress accumulation index S set This allows for the calculation of the differences between each performance indicator and the expected target.

[0034] Furthermore, the mean squared normalized error calculation method (parameter: deviation array Δv of each performance index) is used. i (Total number of samples N), this unifies the differences in indicators with different dimensions into a dimensionless error space. The environmental state classification error index is calculated using the following formula:

[0035] in, The deviation after normalization of each indicator. This represents the total number of indicators.

[0036] Furthermore, the gradient sensitivity inversion method (parameter: classifier output layer activation vector y) is used. pred Target category probability distribution vector y target The learning rate η is used to calculate the gradient for adjusting the discrimination threshold of the output layer of a lightweight convolutional neural network classifier. Specifically, the gradient expression of the cross-entropy loss function is used:

[0037] in The gradient vector, To predict probabilities, Let be the target probability distribution.

[0038] Furthermore, the threshold update formula is determined by discrimination (parameter: current threshold vector θ). current (gradient vector g, learning rate η) to achieve adaptive threshold adjustment through iterative steps:

[0039] in For the classification threshold vector, For the adjusted threshold vector, θ current This is the current threshold vector.

[0040] By using the above-mentioned method of correcting the output layer parameters based on gradient backpropagation, the environmental state classification error calculated in the previous step is transformed into a gradient signal that can be used to optimize the classifier's discrimination threshold, thereby enhancing the accuracy and robustness of environmental state recognition under dynamic conditions.

[0041] For example, in one strategy loop, the actual forming quality RMS deviation collected in the closed-loop feedback phase is 0.28 mm, the target deviation is 0.20 mm, the actual unit energy consumption is 5.5 MJ, the target is 5.0 MJ, and the actual mold stress accumulation index is 1.35 × 10^4, the target is 1.20 × 10^4. 4 After dimensional normalization, Δv is obtained. 质量 =0.4, Δv 能耗 =0.1, Δv 模具应力 =0.125, N=3, substitute into the formula:

[0042] The calculated classification error index is E≈0.268. The predicted probability vector y output by the convolutional neural network classifier is... predGiven the range [0.65, 0.25, 0.10], the target probability distribution y target Given [1, 0, 0], the gradient vector is obtained as g = [-0.35, 0.25, 0.10]. With a learning rate η = 0.05, the discrimination threshold is updated as follows:

[0043] That is, the threshold corresponding to each category is fine-tuned according to the gradient components. This adjustment process effectively reduces the frequency of misclassification under highly fluctuating billet conditions, increases the classifier's probability of distinguishing similar inputs in the next cycle, and greatly enhances the stability of dynamic environment state discrimination.

[0044] S8.3: Perform correlation analysis on the changing trends of the updated classification accuracy and the multi-objective comprehensive reward value to generate a classifier performance gain factor, which serves as a guiding signal for the correction of the weight mapping table parameters, ensuring that the improvement in environmental recognition accuracy can be effectively transmitted to the priority configuration of the reward function.

[0045] Based on the classifier's discrimination threshold update result output by S8.2 and the real environment response result under closed-loop control, the Pearson correlation coefficient calculation method (parameters: classification accuracy time series, multi-objective comprehensive reward value time series) is adopted to realize the linear correlation quantification of the two in the entire closed-loop operation cycle.

[0046] Furthermore, by using a sliding window correlation analysis algorithm (parameters: window length = 50 sampling periods, step size = 1 sampling period), the local dynamic correlation between the classification accuracy change sequence and the multi-objective comprehensive reward value change sequence is estimated, and the sliding window correlation coefficient matrix is ​​obtained, reflecting the robustness and time-varying characteristics of the correlation in different time periods.

[0047] Furthermore, the synchronous change rate of the classification accuracy sequence and the reward value sequence is calculated using the differential fluctuation matching method (parameter: first-order differential operator Δ). Based on the change rate direction consistency test model (threshold: direction consistency ratio ≥ 0.6), the trend matching degree between the current environmental recognition performance change and the reward function gain change is judged, and the matching degree determination variable is obtained.

[0048] Furthermore, a weighted fusion calculation method is adopted (parameters: global Pearson correlation coefficient weight β1, sliding window correlation coefficient mean weight β2, trend matching ratio weight β3, and β1+β2+β3=1), to construct the calculation formula for the classifier performance gain factor G:

[0049] in, The global Pearson correlation coefficient. The mean of the correlation coefficients of the sliding window. Match the proportional variable to the trend of change.

[0050] Furthermore, by using a normalized mapping method (interval [0,1], nonlinear sigmoid function), the performance gain factor is... The mapping is used to obtain the guiding signal strength value that can be used for adaptive correction of the weight mapping table parameters, thereby realizing the effective transmission of the improvement of environmental recognition accuracy to the adjustment of the priority ratio of the reward function.

[0051] Through the above correlation analysis and weighted fusion calculation method, the coupling characteristics of changes in classification accuracy and changes in comprehensive reward are quantified into performance gain factors, realizing the closed loop from improving perception accuracy to adjusting the priority of strategy optimization.

[0052] For example, under the dynamic operating condition of the GK-5000 forging press running continuously for 500 control cycles, the global average classification accuracy was 0.92, and the global average multi-objective comprehensive reward value was 0.78. These values ​​were calculated using the Pearson correlation coefficient. The average correlation coefficient was obtained by using a sliding window with a window length of 50 and a step size of 1. Based on the first-order difference fluctuation matching method, the consistency ratio of the trend direction is calculated to be M=0.65. Setting β1=0.4, β2=0.35, and β3=0.25, substituting these values ​​into the performance gain factor formula:

[0053] The calculated value of G is 0.7815. After normalization mapping using the S-shaped function, a guiding signal strength value of 0.79 is generated. This value is used as the input for the S8.4 weight mapping table correction process, which enables the weight vector to automatically increase the forming quality weight under high-fluctuation billet conditions. This achieves adaptive optimization of the priority of process objectives and significantly improves the global optimization stability of the strategy in dynamic environments.

[0054] S8.4: Based on the comparison results between the performance gain factor and the historical optimization trajectory, the Bayesian optimization algorithm is used to iteratively update the normalized weight vector corresponding to each state category in the weight mapping table, and adjust the priority ratio among the forging forming quality score, unit product energy consumption cost and mold stress accumulation index.

[0055] Based on the comparison results between the classifier performance gain factor and the historical optimization trajectory, a Bayesian optimization algorithm is adopted (parameters: classifier performance gain factor g, historical optimization trajectory H, target weight vector W). tThe process involves iteratively updating the normalized weight vectors corresponding to each state category in the weight mapping table. Further, a Gaussian process regression model (parameters: kernel function type = radial basis function, prior hyperparameter set Θ) is used to estimate the posterior distribution of the objective function—the comprehensive performance of multi-objective reinforcement learning—and obtain the expected improvement value EI for each candidate weight vector. Further, the expected improvement criterion (parameter: EI threshold ε) is used to prioritize the sampling points of the weight vectors in the search space, retaining the candidate solution set C with an EI value greater than ε. Further, by maximizing the sampling function, the sampling center of the next round of weight vectors is dynamically determined, and candidate updated solutions are generated. Further, weight vector normalization ensures the comparability of weights between different optimization objectives by achieving a normalization constraint that the sum of all components is 1. Finally, a Bayesian optimization iterative update method is used to transform the performance gain information obtained in the previous step into the latest priority configuration corresponding to the environmental state category, enabling the adaptive evolution of the weight mapping table under dynamic conditions.

[0056] For example, in a certain optimization loop, for the "high-fluctuation billet" state category, the classifier performance gain factor is 0.25. Historical optimization trajectories show that the weights for forging forming quality score, energy cost, and die stress remain consistently at 0.55, 0.25, and 0.20, respectively. A Gaussian process regression model is initialized using a Bayesian optimization algorithm, with a radial basis function kernel selected. Hyperparameters are set as follows: length scale l = 1.5, signal variance σ² = 0.02, and noise variance σ² = 0.02. n =0.0001. Based on the current gain factor g=0.25 and the historical trajectory H, the expected improvement value distribution of the multi-objective integrated performance is calculated, and the EI values ​​of the five candidate weight vectors are 0.015, 0.042, 0.038, 0.025, and 0.047, respectively. Based on the threshold screening of ε=0.03, three vectors with EI greater than 0.03 are retained ([0.60,0.22,0.18], [0.58,0.24,0.18], and [0.57,0.23,0.20], respectively). According to the principle of the maximum value of the acquisition function, the third vector [0.57,0.23,0.20] is selected as the current iteration center and normalized correction is performed: the sum of each component is 1.00, and no scaling adjustment is required. The updated weight configuration is stored in the mapping table, increasing the priority of forming quality score from 0.55 to 0.57, decreasing energy consumption cost from 0.25 to 0.23, and keeping the mold stress accumulation index unchanged from 0.20. Execution verification shows that this update strategy significantly improves forming quality stability in subsequent strategy cycles, while maintaining energy consumption control and mold life indicators within the optimal fluctuation range, achieving a balanced improvement in multi-objective performance.

[0057] S8.5: The updated discrimination threshold and the optimized weight mapping table parameters are synchronously written into the online learning module to form a two-way closed-loop self-evolutionary structure from execution feedback to perception modeling and decision-making mechanism, thereby enhancing the system's long-term adaptability to non-steady-state disturbances such as billet fluctuations and equipment status changes.

[0058] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0059] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and rules of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for intelligent optimization of metal forging process parameters, characterized in that, Includes the following steps: S1: Acquire multi-source process data collected in real time during the forging process and record its corresponding time series labels; S2: Perform sliding window normalization on the multi-source process data to generate a standardized time-series feature vector, and input the standardized time-series feature vector into a pre-trained convolutional neural network classifier to determine the environmental state category; S3: Based on the discrimination results output by the convolutional neural network classifier, generate the current production environment status category label; S4: Query the preset weight mapping table according to the state category label to obtain the corresponding normalized weight vector; S5: Based on the normalized weight vector, the components in the multi-objective reward function are weighted and fused to generate a dynamic comprehensive reward value; S6: Input the sample tuple containing the dynamic comprehensive reward value into the soft actor-critic algorithm framework for training and convergence, and input the current production environment state category label into the converged soft actor-critic algorithm framework to generate a three-dimensional continuous action vector; S7: Execute the process parameter adjustment command defined by the three-dimensional continuous motion vector, control the forging equipment to run in the updated process window, and simultaneously collect the actual forming quality, measured energy consumption and mold stress feedback data to form the real environmental response result under closed-loop control; S8: The deviation signal between the actual environmental response result and the expected target is converted into a gradient signal that can be used to optimize the classifier's discrimination threshold. The gradient backpropagation is performed to adjust the discrimination threshold of the output layer of the convolutional neural network classifier, and the priority ratio among the forging forming quality score, unit product energy consumption cost and mold stress accumulation index is adjusted.

2. The intelligent optimization method for metal forging process parameters according to claim 1, characterized in that, The multi-source process data includes billet size deviation, material composition fluctuation, mold surface temperature gradient, and master cylinder pressure response delay signal.

3. The intelligent optimization method for metal forging process parameters according to claim 1, characterized in that, The input to the convolutional neural network classifier is a multi-source temporal feature vector after sliding window Z-score normalization. The batch size is 1-128, the temporal length is 10-60 seconds, the feature dimension is 4, the convolutional layer kernel size is 3×3, 4×4, 5×5, the stride is 1, and the activation function is ReLU. The output is processed by batch normalization and max pooling and then connected to a fully connected layer to obtain a deep spatial and temporal feature representation vector.

4. The intelligent optimization method for metal forging process parameters according to claim 1, characterized in that, Step S3 specifically includes: Based on the standardized temporal feature vector output in step S2, the original output response value of the convolutional neural network classifier is obtained. The original output response value is the unnormalized confidence score of each candidate environmental state category. The unnormalized confidence score is subjected to Softmax normalization to convert it into a posterior probability vector of the environment state in the form of a probability distribution. Based on the posterior probability vector of the environmental state, the maximum a posteriori probability criterion is used to perform state category decision, the category with the highest probability value is selected as the optimal state estimation result, and the corresponding discrete state category label is generated. The discrete state category label is encoded as a unique integer index value, and the integer index value corresponds one-to-one with the row number in the preset weight mapping table to construct an addressable mapping channel from the environmental identification result to the multi-objective optimization priority configuration. The integer-indexed state category label is output as a control enable signal for the dynamic weighting of the multi-objective reward function, so as to drive the weight mapping module to select the appropriate normalized weight vector according to the current actual working conditions.

5. The intelligent optimization method for metal forging process parameters according to claim 1, characterized in that, Step S4 specifically includes: The current production environment status category label output in step S3 is used as the index input variable; Construct and load a pre-defined weight mapping table, whose row index is the environmental state category label, and whose columns contain the initial weight coefficients of three optimization objectives: forging forming quality score, unit product energy consumption cost, and mold stress accumulation index. Based on the current production environment status category label, perform an exact match query in the weight mapping table, extract the original weight coefficient vector of the corresponding row, and perform L1 norm normalization on the original weight coefficient vector to obtain the normalized weight vector. Each component in the normalized weight vector is assigned to the corresponding sub-objective item in the multi-objective reward function, thus completing the mapping transformation from environmental state to multi-objective priority.

6. The intelligent optimization method for metal forging process parameters according to claim 5, characterized in that, The weight mapping table is constructed by combining domain expert experience and historical process trajectories, with each environmental state category corresponding to a set of reward weight allocations.

7. The intelligent optimization method for metal forging process parameters according to claim 5, characterized in that, Each component in the normalized weight vector is assigned to the corresponding sub-objective in the multi-objective reward function. The forging forming quality score receives the first weight component, the unit product energy consumption cost receives the second weight component, and the mold stress accumulation index receives the third weight component.

8. The intelligent optimization method for metal forging process parameters according to claim 1, characterized in that, The sample tuple in step S6 also includes the current production environment status category label, the process parameter action vector generated in the previous time step, and the environment status in the next time step.

9. The intelligent optimization method for metal forging process parameters according to claim 1, characterized in that, The three-dimensional continuous motion vector includes the heating temperature setpoint, the energy increment per strike, and the pressure holding time adjustment.

10. The intelligent optimization method for metal forging process parameters according to claim 1, characterized in that, The output of the policy network is limited by the tanh activation function and combined with the number of strokes of the GK-5000 device (16 spm) and the nominal force (50000 kN) as physical boundary constraints.

Images