Intelligent adaptive control system based on forging process

Abstract

The application relates to the field of forging control technology of phosphate fertilizer production equipment, and discloses an intelligent self-adaptive control system based on a forging process. The system obtains phosphate ore characteristic data through a spectral detection device, collects production working condition data through an online sensing device, calculates a forging fatigue life threshold through a multi-physical field coupling simulation engine; a digital twin platform rehearses a forging process combination, a multi-target deep reinforcement learning controller takes life matching working condition fluctuation and mechanical property adaptation raw material characteristics as targets, and adjusts forging equipment parameters; measured data is reversely input through a feedback link, forming a bidirectional self-adaptive closed-loop control. The system integrates multi-modal feature fusion, federal learning and other technologies, solves the problems of traditional forging and production process fragmentation and poor adaptability, and improves the service life of forgings and production stability.

Description

Technical Field

[0001] This invention relates to the field of forging control technology for phosphate fertilizer production equipment, and discloses an intelligent adaptive control system based on forging process. Background Technology

[0002] Currently, in the manufacturing of forged cylinders for high-pressure acid hydrolysis reactors for phosphate fertilizer, the control system only receives the dimensional design drawings and standard mechanical property indicators of the forgings. During the forging process, the system controls the forging equipment based on preset fixed forging temperatures, deformation amounts, and heat treatment curves. In subsequent phosphate fertilizer production, the production control system independently adjusts the reaction pressure, temperature, and material flow rate according to the acid hydrolysis process standards until the forgings reach their designed retirement standards and are then replaced. The forging control system and the phosphate fertilizer production control system operate independently, and their data are not interconnected.

[0003] The aforementioned control method results in a disconnect between the forging process and the phosphate fertilizer manufacturing process. The grade and impurity content of the phosphate ore to be processed are in a state of dynamic fluctuation, and the corrosiveness of the medium and the service load in the reactor change in real time accordingly. However, the forgings produced by the fixed forging process only possess static mechanical properties and fatigue life. The independently operating control system cannot adaptively adjust the forging actuator parameters of the forging based on the fluctuations in the characteristics of the phosphate ore raw materials and the dynamic operating conditions throughout the production cycle, causing a disconnect between the service life of the core equipment and the actual production requirements. Summary of the Invention

[0004] The purpose of this invention is to provide a solution that can effectively address the problems described in the background section.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0006] The intelligent adaptive control system based on forging process includes:

[0007] A spectral analysis device for the characteristics of phosphate rock raw materials is used to obtain data on the grade, impurity content, and mineral phase composition of the phosphate rock to be processed.

[0008] The phosphate fertilizer production full-condition online sensing device is used to collect data on acid hydrolysis reaction pressure, temperature, media corrosivity, and material throughput.

[0009] A multiphysics coupling simulation engine is used to calculate the service load, corrosion rate and fatigue life threshold of the cylinder forging of the high-pressure acid hydrolysis reactor for phosphate fertilizer based on the above data.

[0010] A digital twin platform for the entire forging process is used to predict the forging temperature, deformation, heat treatment parameters, and forging path combinations of forgings based on fatigue life thresholds.

[0011] A multi-objective deep reinforcement learning adaptive controller is used to adjust the actuator parameters of the forging equipment with the optimization objectives of matching the service life of forgings with the fluctuations of the working conditions throughout the entire production cycle of phosphate fertilizer and adapting the mechanical properties of forgings to the fluctuations of the characteristics of phosphate rock raw materials.

[0012] The measured data feedback link is used to input the measured mechanical properties and metallographic structure data of the forged parts into the phosphate fertilizer production process control system in reverse, forming a two-way adaptive closed-loop control.

[0013] Preferably, the process of acquiring mineral phase composition data by the phosphate rock raw material characteristic spectral detection device includes: using a multimodal deep feature fusion method to map the visible and short-wave infrared spectra and X-ray fluorescence spectra of phosphate rock to a high-dimensional feature space; capturing the complementary correlation between different spectral modes through a cross-modal attention mechanism to extract a joint feature vector characterizing the distortion of the phosphate rock crystal structure and the occurrence state of impurity elements; inputting the joint feature vector into a pre-constructed phosphate fertilizer acidolysis kinetic prediction network to output the reactivity index of phosphate rock at a specific acid concentration and temperature as a dynamic input parameter for subsequent calculation of corrosion rate.

[0014] Preferably, the process of calculating the fatigue life threshold using the multiphysics coupling simulation engine includes: extracting time-series data of temperature, stress, and flow fields of the reactor cylinder under historical phosphate fertilizer production conditions, and constructing a dynamic spatiotemporal graph with nodes as spatial locations and edges as physical quantity transfer relationships; using a spatiotemporal graph convolutional network to mine the nonlinear coupling features of multiphysics parameters in spatial topology; and adopting a transfer learning strategy based on multi-condition similarity measurement to transfer the weight parameters of the source domain spatiotemporal graph convolutional network to the target domain of the current phosphate mine, and outputting the fatigue life threshold of the forging under the current condition through fine-tuning.

[0015] Preferably, the process of pre-simulating forging path combinations in the digital twin platform for the entire forging process includes: establishing a Markov decision process in the digital twin space with forging temperature, deformation amount, and heat treatment parameters as the state space; using a generator in a generative adversarial network to randomly sample and generate candidate forging process paths, and a discriminator to distinguish between true and false candidate paths in combination with the boundary constraints of the high-pressure acid hydrolysis reaction of phosphate fertilizer; introducing Monte Carlo tree search to perform breadth-first expansion and depth simulation evaluation on the candidate paths output by the generator, and outputting the pool of process combinations with the highest comprehensive score for optimization.

[0016] Preferably, the process of adjusting actuator parameters by a multi-objective deep reinforcement learning adaptive controller includes: modeling the operating condition fluctuations of different phosphate fertilizer production lines as independent agents to construct a multi-agent reinforcement learning environment; designing a composite reward function that includes service life matching degree, mechanical performance adaptability degree, and phosphate fertilizer conversion rate stability; while the agents interact locally, a federated learning architecture is used to aggregate the reinforcement learning gradient information of each production line in the cloud, and while protecting the privacy of the operating condition data of each phosphate fertilizer production line, updating the global adaptive control strategy network, and then outputting the actuator parameter adjustment amount.

[0017] Preferably, the process of capturing complementary correlations through cross-modal attention mechanisms is constrained by a knowledge graph of phosphate fertilizer chemical mechanisms: a knowledge graph containing entities of phosphate rock mineral phases, impurity ions, acidolysis products, and corrosion side reactions is constructed, and stoichiometric relationships and thermodynamic reaction paths between entities are extracted as prior rules; the prior rules are transformed into graph embedding vectors and fused with joint feature vectors using tensors; a regularization penalty term based on graph embedding vectors is added to the calculation graph of attention weights to suppress spectral feature responses that violate the phosphate fertilizer acidolysis chemical mechanism and improve the physical interpretability of the reactivity index output.

[0018] Preferably, the training process of the spatiotemporal graph convolutional network integrates a physical information neural network: the partial differential equations describing the heat conduction and structural mechanics of the reactor cylinder are discretized and embedded as hard constraints into the hidden layer of the spatiotemporal graph convolutional network; during the backpropagation stage of the network, the residuals of the partial differential equations are calculated, and the L2 norm of the residuals is used as part of the unsupervised loss function; by jointly minimizing the loss of multiphysics prediction data and physical residuals, the spatiotemporal graph convolutional network is forced to strictly follow the laws of energy conservation and momentum conservation in the acidolysis reaction process of phosphate fertilizer during transfer learning fine-tuning.

[0019] Preferably, the discriminator introduces a contrastive learning mechanism and reinforcement learning feedback in the process of distinguishing the authenticity of candidate paths: positive sample forging paths that conform to the extreme boundary conditions of phosphate rock impurities and negative sample paths that deviate from the boundary conditions are extracted in the digital twin space; the distance between positive samples and negative samples in the feature space is shortened by contrastive learning, and a dynamic decision boundary is constructed; the authenticity probability of the path output by the discriminator is fed back as a reward signal to the simulation stage of Monte Carlo tree search, guiding the tree search strategy to focus on the process combination region with a high probability of success.

[0020] Preferably, in a multi-agent reinforcement learning environment, the policy update of each agent adopts Nash equilibrium solution and meta-learning algorithm: the policy network parameters of each agent are regarded as game participants, and the Nash equilibrium point across the phosphate fertilizer production line is solved by calculating the Jacobian matrix of the gradient ascent policy; after reaching the equilibrium point, the model-independent meta-learning algorithm in meta-learning is used to perform second-order gradient optimization on the virtual working condition distribution of various phosphate ore grade fluctuations, and the initial parameters of the global policy with high generalization ability are extracted, so that when faced with sudden changes in phosphate ore grade, the agents only need a small number of interactions to converge.

[0021] Preferably, the convergence process of the agent under a small number of interactions adopts a fusion of few-shot learning and cross-domain attention: for the phosphate rock grade that changes abruptly, a small amount of perception data from the initial stage of the current acidolysis reaction is extracted to construct a support set; the correlation matrix between the features of the support set and the key channel features in the initial parameters of the global strategy is calculated through the cross-domain attention mechanism; the initial parameters of the global strategy are fine-tuned at the channel level using the correlation matrix, and the fine-tuned strategy output is mapped to a coordinated scheduling instruction for the throughput and feeding rate of acidolysis reaction materials on the phosphate fertilizer production line, so as to realize the adaptive forging process and the synchronous matching of the phosphate fertilizer production cycle.

[0022] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0023] 1. This invention acquires phosphate rock characteristic data and operating condition data through a spectral detection device and an online sensing device, respectively. The fatigue life threshold is calculated using a multi-physics coupled simulation engine. The process combination is pre-simulated on a digital twin platform, and a multi-objective deep reinforcement learning controller adjusts the actuator parameters. The measured data is then fed back in reverse, constructing a bidirectional adaptive closed-loop control. This structure directly couples forging control variables with the dynamic boundary conditions of phosphate fertilizer production. It adaptively adjusts the forging actuator parameters based on changes in actual service load and corrosion rate, ensuring that the service life of the output forging matches the fluctuations in operating conditions throughout the production cycle, and that the mechanical properties are adapted to the fluctuations in the characteristics of the phosphate rock raw materials.

[0024] 2. By combining multimodal deep feature fusion and cross-modal attention mechanisms with a knowledge graph of phosphate fertilizer chemical mechanisms, graph embedding regularization is used to suppress spectral features that violate acidolysis chemical mechanisms, maintaining the physical interpretability of the reactivity index output. The residuals of partial differential equations are embedded as unsupervised loss functions into a spatiotemporal graph convolutional network, combined with transfer learning and meta-learning algorithms, enabling multiphysics simulations to output fatigue life thresholds with minimal interaction when facing sudden changes in phosphate ore grade. Generative adversarial networks and contrastive learning mechanisms are introduced into the digital twin platform, with the discrimination probability used as a reward signal to feed back to Monte Carlo tree search. Combined with Nash equilibrium solution and federated learning architecture, cross-domain collaboration and data privacy protection are achieved for optimizing forging process combinations across multiple production lines. Attached Figure Description

[0025] Figure 1 This is the overall flowchart of the intelligent adaptive control of the forging process of the present invention;

[0026] Figure 2 This is a flowchart of the process for obtaining phosphate rock raw material characteristic data according to the present invention;

[0027] Figure 3 This is a flowchart of the fatigue life threshold calculation for forgings according to the present invention;

[0028] Figure 4 This is a pre-simulation flow chart of the forging process combination of the present invention;

[0029] Figure 5 This is a flowchart of the actuator parameter adjustment process of the present invention;

[0030] Figure 6 This is a flowchart of the bidirectional adaptive closed-loop control of the present invention. Detailed Implementation

[0031] This specific embodiment discloses an intelligent adaptive control system based on forging process, applied to the full-process control of forging of the cylinder forging of phosphate fertilizer high-pressure acid hydrolysis reactor and the bidirectional collaborative control of phosphate fertilizer production process. The system includes a phosphate rock raw material characteristic spectrum detection device, an online sensing device for the entire phosphate fertilizer production process, a multi-physics field coupled simulation engine, a digital twin platform for the entire forging process, a multi-objective deep reinforcement learning adaptive controller, and a measured data feedback link.

[0032] Example 1: Please refer to the appendix Figure 1 and attached Figure 2 The process by which the phosphate rock raw material characteristic spectral detection device acquires data on the grade, impurity content, and mineral phase composition of the phosphate rock to be processed, and outputs the reactivity index, is as follows:

[0033] The visible and short-wave infrared spectrometer acquires diffuse reflectance spectral data of the phosphate rock to be treated in the 350nm-2500nm band, and the X-ray fluorescence spectrometer acquires characteristic fluorescence spectral data of Na, Mg, Al, Si, P, S, Cl, K, Ca, and Fe elements in the phosphate rock to be treated; the embedded processing unit sequentially performs baseline correction, noise removal, and normalization processing on the acquired raw spectral data, wherein the calculation formula for normalization is:

[0034]

[0035] In the formula, These are single-channel sampled values ​​from the original spectral data. This represents the minimum value of the spectral data for that channel. This represents the maximum value of the spectral data for that channel. This is the normalized spectral data.

[0036] A multimodal deep feature fusion method is employed, where normalized visible and short-wave infrared spectra and X-ray fluorescence spectra are input into two independent feature encoders. Each feature encoder uses a three-layer one-dimensional convolutional neural network structure with kernel sizes of 7, 5, and 3, a stride of 1, and the ReLU activation function. This maps the two spectral data to high-dimensional feature spaces, outputting a high-dimensional feature matrix of visible and short-wave infrared spectra. With X-ray fluorescence high-dimensional feature matrix ,in The number of feature channels, For feature dimensions.

[0037] By capturing complementary correlations between different spectral modes through a cross-modal attention mechanism, a cross-modal cross-attention module is constructed. The query vector of the cross-attention module... Depend on Linear mapping generation, key vector AND value vector Depend on Linear mapping is generated, and the weight matrices of the linear mapping are respectively... , , ,in The dimension of the key vector. The dimension is the value vector; the formula for calculating the attention weights is:

[0038]

[0039] In the formula, This is a regularization penalty term matrix generated based on a knowledge graph of phosphate fertilizer chemical mechanisms. This is a scaling factor used to prevent the gradient of the Softmax function from vanishing due to excessively large inner product values.

[0040] The computational process of the cross-modal attention mechanism is constrained by a knowledge graph of phosphate fertilizer chemical mechanisms. A knowledge graph of phosphate fertilizer chemical mechanisms is pre-constructed, and the entities in this knowledge graph include phosphate rock mineral phases, impurity ions, acidolysis products, and corrosion side reactions. The relationships between entities include stoichiometric relationships, thermodynamic reaction paths, and corrosion reaction coupling relationships. The TransE algorithm is used to perform graph embedding learning on the knowledge graph, outputting a low-dimensional graph embedding vector for each entity and relationship. , ,in, This refers to the graph embedding vector corresponding to the entity in the knowledge graph. This refers to the graph embedding vector corresponding to the relation in the knowledge graph. For the graph embedding dimension; concatenate the graph embedding vectors of all entities and relations to generate a global prior rule embedding matrix. ,in This represents the total number of entities and relations in the knowledge graph.

[0041] Embedding global prior rules into a matrix By mapping the fully connected layer to the same dimension as the attention weight matrix, a regularization penalty term matrix is ​​generated. The mapping formula is:

[0042]

[0043] In the formula, This is the weight matrix of the fully connected layer. This is the bias matrix; during the attention weight calculation process, This is used to apply a negative penalty to feature associations that violate chemical mechanisms, suppress their attention weights, and ensure that only spectral feature associations that conform to the chemical mechanism of phosphate fertilizer acidolysis are activated.

[0044] After optimization using the cross-modal attention mechanism and By concatenating the channels, a joint feature vector characterizing the distortion of the phosphate rock crystal structure and the occurrence state of impurity elements is extracted. The joint feature vector is input into a pre-constructed phosphate fertilizer acidolysis kinetic prediction network. This prediction network employs a four-layer fully connected neural network structure with 512, 256, 128, and 64 neurons in the hidden layers, respectively. The LeakyReLU activation function is used, and the Sigmoid function is used in the output layer. The output layer shows the reactivity index of phosphate rock under specific acid concentrations and temperatures. The reactivity index serves as a dynamic input parameter for the subsequent multiphysics coupling simulation engine to calculate the corrosion rate.

[0045] As a preferred embodiment, refer to the appendix. Figure 3 The process by which the multiphysics coupled simulation engine calculates the service load, corrosion rate, and fatigue life threshold of the cylindrical forging of the high-pressure acid lysis reactor for phosphate fertilizer is as follows:

[0046] The temperature, stress, and flow field time-series data of the reactor shell under historical phosphate fertilizer production conditions were extracted to construct a dynamic spatiotemporal graph with nodes as spatial locations and edges as physical quantity transfer relationships. ,in For time-series sampling time, for The set of nodes at time points This represents the total number of mesh nodes in the finite element model of the reactor vessel, with each node... The feature vector includes the node's... Physical quantities in four dimensions: temperature, equivalent stress, fluid pressure, and media corrosion rate at any given time; Let be the set of edges. The weights of the edges are determined by the physical quantity transfer relationships between nodes. For adjacent grid nodes, the edge weight is the weighted sum of the thermal conductivity coefficient and the structural stiffness coefficient. For non-adjacent nodes, the edge weight is the mapped value of the fluid convection heat transfer coefficient. The formula for calculating the edge weight is:

[0047]

[0048] In the formula, for Time Node With nodes The thermal conductivity between them The structural stiffness coefficient, The convective heat transfer coefficient is... , , These are weighting coefficients, and .

[0049] This paper utilizes a spatiotemporal graph convolutional network to mine the nonlinear coupling features of multiphysics parameters in spatial topology. The spatiotemporal graph convolutional network includes spatial graph convolutional layers and temporal convolutional layers. The spatial graph convolutional layers employ a graph convolution operator with Chebyshev polynomial approximation to extract coupling features in the spatial dimension. The temporal convolutional layers employ two layers of one-dimensional causal convolution to extract dynamic evolution features in the temporal dimension. The calculation formula for the spatial graph convolution is as follows:

[0050]

[0051] In the formula, The input node feature matrix, To output the feature matrix, Let the order be the Chebyshev polynomial. For learnable convolution kernel parameters, It is a k-th order Chebyshev polynomial. The normalized Laplace matrix, , For the graph Laplace matrix, for The largest eigenvalue, It is an identity matrix.

[0052] The training process of the spatiotemporal graph convolutional network integrates a physical information neural network, discretizing the partial differential equations describing the heat conduction and structural mechanics of the reactor vessel cylinder, and embedding them as hidden layers of the spatiotemporal graph convolutional network as hard constraints; wherein, the heat conduction partial differential equation is:

[0053]

[0054] In the formula, Density of the reactor vessel cylinder material. For isobaric specific heat capacity, For temperature field distribution, The thermal conductivity coefficient, The density of the volumetric heat source;

[0055] The partial differential equations of equilibrium in structural mechanics are as follows:

[0056]

[0057] In the formula, For stress tensor, It is a volume force vector.

[0058] The above system of partial differential equations is discretized using the finite difference method. During the backpropagation phase of the network, the residuals of the partial differential equations at each node are calculated. The residual is the difference between the calculated value on the left side and the true value on the right side of the discretized partial differential equation; the L2 norm of the residual is used as part of the unsupervised loss function, and the total loss function of the network is:

[0059]

[0060] In the formula, To ensure supervised loss, mean square error loss is used to calculate the error between the predicted and measured values ​​of multiphysics parameters. This is the physical residual loss; The balancing coefficient is used to adjust the weights of physical constraints; by jointly minimizing the total loss function, the spatiotemporal graph convolutional network is forced to strictly follow the laws of energy and momentum conservation in the acidolysis reaction of phosphate fertilizer during training and fine-tuning.

[0061] A transfer learning strategy based on multi-condition similarity measurement is adopted. A spatiotemporal graph convolutional network pre-trained under historical large-scale stable phosphate fertilizer production conditions is used as the source domain model, and the current working condition corresponding to the phosphate mine to be processed is used as the target domain. First, the similarity between the source domain working conditions and the target domain working conditions is calculated. The maximum mean difference measure is used for similarity, and the calculation formula is as follows:

[0062]

[0063] In the formula, For the source domain sample set, For the target domain sample set, For the mapping function of the reproducing kernel Hilbert space, , These represent the number of samples in the source domain and the target domain, respectively. Let be the Hilbert space norm.

[0064] The weights of the bottom feature extraction layer of the fixed source domain model are adjusted only for the top fatigue life prediction layer and some intermediate layers. The total loss function mentioned above is used for optimization during the fine-tuning process. The final output is the service load, corrosion rate and fatigue life threshold of the reactor cylinder forging under the current working conditions. The fatigue life threshold is the critical number of cyclic stresses of the forging under the current working conditions, which serves as the input boundary condition for the digital twin platform of the entire forging process.

[0065] As a preferred embodiment, refer to the appendix. Figure 4 The digital twin platform for the entire forging process constructs a digital twin of the entire forging process of the reactor cylinder forging. This digital twin includes a forging blank model, a forging equipment model, a heating furnace model, a heat treatment furnace model, and the physical field coupling relationships between these models. The process of predicting the forging temperature, deformation, heat treatment parameters, and forging path combination of the forging based on fatigue life thresholds is as follows:

[0066] A Markov decision process is established within a digital twin space, with forging temperature, deformation amount, and heat treatment parameters as the state space. This includes forging temperature, deformation amount, and heat treatment parameters, including quenching temperature, tempering temperature, holding time, and movement space. This includes the feed rate of the forging path, the number of forging passes, the forging direction, the adjustment of heating power, and the reward function. The state transition probability represents the degree of matching between the predicted fatigue life of the forging and the target fatigue life threshold. The physical simulation model of the digital twin determines the optimization objective of the Markov decision process, which is to maximize the cumulative reward, that is, the output forging process combination makes the fatigue life of the forging closest to the target fatigue life threshold.

[0067] A generator in a generative adversarial network (GAN) randomly samples and generates candidate forging process paths, while a discriminator, considering the boundary constraints of the high-pressure acidolysis reaction of phosphate fertilizer, distinguishes between genuine and false candidate paths. The GAN comprises a generator and a discriminator; the generator employs a 5-layer transposed convolutional neural network structure, with random Gaussian noise vectors as input. ,in For the noise dimension, the output is a sequence of candidate forging process paths. The process path sequence includes the forging temperature time-series curve, deformation distribution, heat treatment parameter curve, and forging path coordinate point sequence for the entire process. The discriminator adopts a 4-layer one-dimensional convolutional neural network structure, with the input being the candidate process path sequence and the output being the probability of whether the path meets the boundary constraints. .

[0068] The discriminator introduces a contrastive learning mechanism and reinforcement learning feedback into the process of distinguishing between true and false candidate paths. It extracts positive forging paths that meet the extreme boundary conditions of phosphate rock impurities and negative forging paths that deviate from these boundary conditions within the digital twin space. The positive samples are forging process paths that meet fatigue life requirements under high impurity, low grade, and strong corrosion conditions. The negative samples are forging process paths where the temperature exceeds the material phase transformation range, the deformation exceeds the rated load of the forging press, and the heat treatment parameters result in coarse grains. A momentum contrastive learning method is used to construct a dynamic query dictionary. A contrastive loss function is used to shorten the distance between positive samples and negative samples in the feature space and widen the distance between them. The contrastive loss function is:

[0069]

[0070] In the formula, To query the feature vector of a sample, The feature vector of the positive sample. The feature vector of the negative sample. For temperature coefficient, The number of negative samples is used; by minimizing the contrastive loss function, a dynamic decision boundary is constructed to improve the discriminator's accuracy in distinguishing the authenticity of process paths.

[0071] Monte Carlo tree search is introduced to perform breadth-first expansion and depth-first simulation evaluation of the candidate paths output by the generator. The Monte Carlo tree search includes four stages: selection, expansion, simulation, and backpropagation. In the selection stage, the upper confidence interval algorithm is used to select the optimal node. The formula for calculating the upper confidence interval value is:

[0072]

[0073] In the formula, For nodes The cumulative reward value, For nodes Number of visits, For nodes The parent node, To explore coefficients.

[0074] During the simulation phase, the path authenticity probability output by the discriminator is fed back as a reward signal to the simulation phase of the Monte Carlo tree search, i.e., the reward value for the simulation phase. The tree search strategy is guided to focus on the process combination region with a high probability of success; during the backpropagation phase, the reward value of the simulation phase is updated to all nodes on the path; after a preset number of iterative searches, the pool of the whole process combination with the highest cumulative reward value is output for optimization.

[0075] As a preferred embodiment, refer to the appendix. Figure 5The multi-objective deep reinforcement learning adaptive controller, with the optimization objectives of matching the service life of forgings to the fluctuations of the entire phosphate fertilizer production cycle and adapting the mechanical properties of forgings to the fluctuations of phosphate rock raw material characteristics, adjusts the actuator parameters of the forging equipment as follows:

[0076] The operating condition fluctuations of different phosphate fertilizer production lines are modeled as independent agents, and a multi-agent reinforcement learning environment is constructed, with the number of agents matching the number of phosphate fertilizer production lines. Each agent is deployed at the edge control node of its corresponding phosphate fertilizer production line, and each agent's local observation space... This includes data on the characteristics of phosphate rock raw materials for the corresponding production line, acid leaching reaction conditions, target fatigue life thresholds for forgings, and current forging process parameters; the action space of each agent. This includes parameter adjustments for the actuators of forging equipment, specifically the pressure adjustments of the hydraulic servo system of the forging press, the temperature adjustments of the heating furnace, the feed and forging path adjustments of the robotic arm, and the temperature control curve adjustments of the heat treatment furnace; the environmental state space. It is the joint set of the local observation spaces of all intelligent agents.

[0077] The design incorporates a composite reward function that considers service life matching, mechanical property compatibility, and phosphate fertilizer conversion rate stability. The calculation formula is as follows:

[0078]

[0079] In the formula, , , These are weighting coefficients, and ;

[0080] As a reward for service life matching, The predicted fatigue life of the forging under the current process parameters. The target fatigue life threshold;

[0081] As a reward for mechanical performance fit, Predict a vector of mechanical properties for forgings, including yield strength, tensile strength, impact toughness, and hardness. Let the target mechanical property vector be... This is the scaling factor;

[0082] As a reward for the stability of phosphate fertilizer conversion rate, The coefficient of variation is the phosphate fertilizer conversion rate from acid hydrolysis, used to measure the degree of fluctuation in the conversion rate.

[0083] While each agent interacts locally, a federated learning architecture is used to aggregate reinforcement learning gradient information from each production line in the cloud. This updates the global adaptive control policy network while protecting the privacy of the operating data of each phosphate fertilizer production line. Each agent executes policy interactions in its local environment, collects local trajectory data, and calculates local policy gradients based on a near-end policy optimization algorithm. Each edge control node only uploads local gradient information to the cloud policy aggregation server, without uploading the original operating data. The cloud server uses a federated averaging algorithm to aggregate the local gradients of each agent and updates the global adaptive control policy network. The aggregation formula is as follows:

[0084]

[0085] In the formula, The total number of agents. For the first Number of local samples per agent The total number of samples across all agents. For the first The local policy network parameters updated for each agent The updated global policy network parameters are then distributed from the cloud server to each edge control node. Each agent updates its local policy network based on the global parameters, outputs the adjustment amount of the actuator parameters, and distributes it to the actuator of the forging equipment for execution.

[0086] In the multi-agent reinforcement learning environment, the policy updates of each agent employ Nash equilibrium solving and meta-learning algorithms. The policy network parameters of each agent are considered as game participants, constructing a non-cooperative game model. The payoff function for each agent is the aforementioned composite reward function. The Nash equilibrium point across the phosphate fertilizer production line is solved by calculating the Jacobian matrix of the gradient ascent policy. The Nash equilibrium point satisfies the following condition for any agent: their strategy satisfy ,in To exclude intelligent agents Equilibrium strategies of all other agents; Jacobian matrix The elements are ,in For the first The policy parameters of each agent are obtained by iteratively solving the fixed points of the Jacobian matrix to obtain the policy parameters corresponding to the Nash equilibrium point.

[0087] After reaching the equilibrium point, the model-independent meta-learning algorithm in meta-learning is used to perform second-order gradient optimization on virtual working conditions with varying phosphate ore grades, extracting global policy initial parameters with high generalization ability; the optimization objective of model-independent meta-learning is:

[0088]

[0089] In the formula, For virtual operating condition distribution, For a single virtual working condition, For global policy networks, These are the policy parameters after a single-step gradient update. For the inner loop learning rate, For working conditions The loss function is defined below; by minimizing the above objective function, the initial parameters of the global policy are obtained, so that the agent can converge with only a small number of interactions when faced with sudden changes in phosphate ore grade.

[0090] The convergence process of the intelligent agent with limited interactions employs a fusion of few-shot learning and cross-domain attention. For abrupt changes in phosphate rock grade, a small amount of sensory data from the initial stage of the current acidolysis reaction is extracted to construct a support set. ,in, To support the number of samples in the set; Let k be the vector of observed acidolysis reaction conditions. This refers to the control action vector under the corresponding operating condition. The reward value under the corresponding working condition; the correlation matrix between the support set features and the key channel features in the initial parameters of the global policy is calculated through a cross-domain attention mechanism; the query vector of the cross-domain attention mechanism is used to calculate the correlation matrix between the support set features and the key channel features in the initial parameters of the global policy. Key vectors are generated from support set feature maps. The correlation matrix is ​​generated by mapping the channel weights of the initial parameters of the global policy. The calculation formula is:

[0091]

[0092] In the formula, The dimension of the feature vector.

[0093] Using the correlation matrix The initial parameters of the global policy are fine-tuned using channel-level weighting. The fine-tuned policy parameters are as follows:

[0094]

[0095] In the formula, These are the initial parameters for the global policy. The fine-tuning coefficient is used to map the output of the fine-tuned strategy into a coordinated scheduling instruction for the throughput and feeding rate of acid hydrolysis reaction materials on the phosphate fertilizer production line, thereby achieving adaptive forging process and synchronous matching of phosphate fertilizer production cycle.

[0096] As a preferred embodiment, refer to the appendix. Figure 6 Before the phosphate rock raw materials enter the plant, the phosphate rock raw material characteristic spectral detection device collects the spectral data of the phosphate rock to be processed, and outputs the grade, impurity content, mineral phase composition data of the phosphate rock to be processed, as well as the reaction activity index under the corresponding acid leaching conditions; at the same time, the phosphate fertilizer production full-condition online sensing device collects the pressure, temperature, media corrosivity and material throughput time series data of the acid leaching reaction of the current production line in real time, and uploads them to the multi-physics field coupled simulation engine.

[0097] The multiphysics coupling simulation engine constructs a dynamic spatiotemporal graph based on the received phosphate rock characteristic data and operating condition data. By using a spatiotemporal graph convolutional network that integrates physical information constraints, it calculates the service load, corrosion rate, and fatigue life threshold of the reactor cylinder forging under the current operating conditions. The fatigue life threshold is then used as the target boundary condition for forging process optimization and sent to the digital twin platform for the entire forging process.

[0098] The digital twin platform for the entire forging process constructs a Markov decision process within the digital twin space, with fatigue life threshold as the optimization objective. It generates candidate forging process paths through a generative adversarial network, and combines a discriminator optimized by contrastive learning with Monte Carlo tree search to output a pool of process combinations with the highest comprehensive score, which is then distributed to a multi-objective deep reinforcement learning adaptive controller.

[0099] The multi-objective deep reinforcement learning adaptive controller aims to match the service life of forgings with the fluctuations in the working conditions of the entire phosphate fertilizer production cycle and to adapt the mechanical properties of forgings to the fluctuations in the characteristics of phosphate rock raw materials. Based on the process combination pool, it generates parameter adjustment amounts for the actuators of forging equipment through a global policy network optimized by multi-agent reinforcement learning and federated learning. These adjustments are then distributed to actuators such as forging presses, heating furnaces, heat treatment furnaces, and robotic arms to execute the corresponding forging processes.

[0100] After the forging is formed, the measured mechanical properties and metallographic structure data of the forging are collected by mechanical property testing equipment and metallographic structure testing equipment. The measured data is fed back into the phosphate fertilizer production process control system through the measured data feedback link, and simultaneously fed back into the multiphysics coupling simulation engine and the digital twin platform of the forging process. The multiphysics coupling simulation engine optimizes the parameters of the multiphysics coupling calculation model based on the measured data, the digital twin platform of the forging process corrects the simulation accuracy of the digital twin based on the measured data, and the multi-objective deep reinforcement learning adaptive controller updates the parameters of the strategy network based on the measured data, forming a two-way adaptive closed-loop control of the forging process and the phosphate fertilizer production process.

[0101] The above description is merely an optional embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention. For example, the phosphate rock raw material characteristic spectral detection device can be supplemented with a Raman spectrometer to obtain mineral phase detection data; the phosphate fertilizer production full-condition online sensing device can be supplemented with an acoustic emission sensor to obtain fatigue crack propagation data of the reactor cylinder; and the multiphysics coupling simulation engine can be deployed on a cloud server to achieve parallel computing across multiple production lines. Such variations all fall within the scope of protection of the present invention.

Claims

1. A smart adaptive control system based on forging process, characterized in that, include: A spectral analysis device for the characteristics of phosphate rock raw materials is used to obtain data on the grade, impurity content, and mineral phase composition of the phosphate rock to be processed. The phosphate fertilizer production full-condition online sensing device is used to collect data on acid hydrolysis reaction pressure, temperature, media corrosivity, and material throughput. A multiphysics coupling simulation engine is used to calculate the service load, corrosion rate and fatigue life threshold of the cylinder forging of the high-pressure acid hydrolysis reactor for phosphate fertilizer based on the above data. A digital twin platform for the entire forging process is used to predict the forging temperature, deformation, heat treatment parameters, and forging path combinations of forgings based on fatigue life thresholds. A multi-objective deep reinforcement learning adaptive controller is used to adjust the actuator parameters of the forging equipment with the optimization objectives of matching the service life of forgings with the fluctuations of the working conditions throughout the entire production cycle of phosphate fertilizer and adapting the mechanical properties of forgings to the fluctuations of the characteristics of phosphate rock raw materials. The measured data feedback link is used to input the measured mechanical properties and metallographic structure data of the forged parts into the phosphate fertilizer production process control system in reverse, forming a two-way adaptive closed-loop control.

2. The intelligent adaptive control system based on forging process according to claim 1, characterized in that, The process of acquiring mineral phase composition data by the phosphate rock raw material characteristic spectral detection device includes: using a multimodal deep feature fusion method to map the visible short-wave infrared spectrum and X-ray fluorescence spectrum of phosphate rock to a high-dimensional feature space respectively; By capturing complementary correlations between different spectral modes through a cross-modal attention mechanism, a joint feature vector characterizing the distortion of phosphate rock crystal structure and the occurrence state of impurity elements is extracted. The joint feature vector is input into the pre-constructed phosphate fertilizer acidolysis kinetic prediction network, and the output phosphate rock reactivity index under acid concentration and temperature is used as the dynamic input parameter for subsequent calculation of corrosion rate.

3. The intelligent adaptive control system based on forging process according to claim 1, characterized in that, The process of calculating the fatigue life threshold by the multiphysics coupling simulation engine includes: extracting the time series data of temperature field, stress field and flow field of the reactor cylinder under historical phosphate fertilizer production conditions, and constructing a dynamic spatiotemporal graph with nodes as spatial locations and edges as physical quantity transmission relationships; Spatiotemporal graph convolutional networks are used to mine the nonlinear coupling features of multi-physics parameters in spatial topology. A transfer learning strategy based on multi-condition similarity measurement is adopted to transfer the weight parameters of the source domain spatiotemporal graph convolutional network to the target domain of the current phosphate mine to be processed, and the fatigue life threshold of the forging under the current condition is output by fine-tuning.

4. The intelligent adaptive control system based on forging process according to claim 1, characterized in that, The process of pre-simulating forging path combinations on the digital twin platform for the entire forging process includes: A Markov decision process with forging temperature, deformation amount, and heat treatment parameters as the state space is established in the digital twin space. A generator in a generative adversarial network is used to randomly sample and generate candidate forging process paths. A discriminator combines the boundary constraints of the high-pressure acid hydrolysis reaction of phosphate fertilizer to distinguish between true and false candidate paths. Monte Carlo tree search is introduced to perform breadth-first expansion and depth simulation evaluation on the candidate paths output by the generator. The pool of full-process combination with the highest comprehensive score is output for optimization.

5. The intelligent adaptive control system based on forging process according to claim 1, characterized in that, The process of adjusting actuator parameters by a multi-objective deep reinforcement learning adaptive controller includes: modeling the operating condition fluctuations of different phosphate fertilizer production lines as independent agents and constructing a multi-agent reinforcement learning environment; designing a composite reward function that includes service life matching degree, mechanical performance adaptability degree, and phosphate fertilizer conversion rate stability; while the agents interact locally, a federated learning architecture is used to aggregate the reinforcement learning gradient information of each production line in the cloud, and while protecting the privacy of the operating condition data of each phosphate fertilizer production line, the global adaptive control strategy network is updated, thereby outputting the actuator parameter adjustment amount.

6. The intelligent adaptive control system based on forging process according to claim 2, characterized in that, The process of capturing complementary associations through cross-modal attention mechanisms is constrained by the knowledge graph of phosphate fertilizer chemical mechanisms: constructing a knowledge graph containing entities of phosphate rock mineral phases, impurity ions, acidolysis products and corrosion side reactions, and extracting stoichiometric relationships and thermodynamic reaction paths between entities as prior rules; The prior rules are transformed into graph embedding vectors and fused with the joint feature vectors using tensors. A regularization penalty term based on the graph embedding vectors is added to the computation graph of the attention weights to suppress the spectral feature responses that violate the chemical mechanism of phosphate fertilizer acidolysis.

7. The intelligent adaptive control system based on forging process according to claim 3, characterized in that, The training process of the spatiotemporal graph convolutional network integrates physical information neural networks: the partial differential equations describing the heat conduction and structural mechanics of the reactor cylinder are discretized and embedded as hard constraints into the hidden layer of the spatiotemporal graph convolutional network. During the backpropagation phase of the network, the residuals of the partial differential equations are calculated, and the L2 norm of the residuals is used as part of the unsupervised loss function. By jointly minimizing the loss of multiphysics prediction data and physical residuals, the spatiotemporal graph convolutional network is forced to follow the laws of energy conservation and momentum conservation in the phosphate fertilizer acidolysis reaction process during transfer learning fine-tuning.

8. The intelligent adaptive control system based on forging process according to claim 4, characterized in that, The discriminator introduces a contrastive learning mechanism and reinforcement learning feedback to distinguish between true and false candidate paths: positive sample forging paths that conform to the extreme boundary conditions of phosphate rock impurities and negative sample paths that deviate from the boundary conditions are extracted in the digital twin space. By contrastive learning, the distance between positive samples and negative samples in the feature space is narrowed, and a dynamic decision boundary is constructed. The path authenticity probability output by the discriminator is fed back as a reward signal to the simulation stage of the Monte Carlo tree search, guiding the tree search strategy to focus on the process combination region with a high probability of success.

9. The intelligent adaptive control system based on forging process according to claim 5, characterized in that, In a multi-agent reinforcement learning environment, the policy updates of each agent are achieved using Nash equilibrium solving and meta-learning algorithms. The policy network parameters of each agent are regarded as game participants, and the Nash equilibrium point across the phosphate fertilizer production line is solved by calculating the Jacobian matrix of the gradient ascent policy. After reaching the equilibrium point, the model-independent meta-learning algorithm in meta-learning is used to perform second-order gradient optimization on virtual working conditions with various phosphate ore grade fluctuations. The initial parameters of the global policy with high generalization ability are extracted, so that when faced with sudden changes in phosphate ore grade, the agent can converge with only a small number of interactions.

10. The intelligent adaptive control system based on forging process according to claim 9, characterized in that, The convergence process of the agent with a limited number of interactions employs a fusion of few-shot learning and cross-domain attention: For phosphate rock grades that change abruptly, a support set is constructed by extracting a small amount of sensing data from the initial stage of the current acidolysis reaction. The correlation matrix between support set features and key channel features in the initial parameters of the global policy is calculated using a cross-domain attention mechanism. The initial parameters of the global strategy are fine-tuned at the channel level using the correlation matrix, and the fine-tuned strategy output is mapped to a coordinated scheduling instruction for the throughput and feeding rate of acid hydrolysis materials on the phosphate fertilizer production line.

Images