A method and system for dynamically scheduling resources for tower production

Abstract

The present application relates to the technical field of iron tower production, and proposes an iron tower production resource dynamic scheduling method and system. The method first fuses multi-source heterogeneous data, constructs a multi-modal graph neural network incorporating material mechanics and thermodynamics prior constraints, and forms an evaluation atlas defining a safety boundary. Further, with the objectives of maximizing material utilization and minimizing comprehensive energy consumption, a flexible process path is autonomously solved using multi-agent reinforcement learning to generate a preliminary scheduling strategy. During execution, millisecond-level real-time safety monitoring and emergency response are performed by an edge lightweight graph network, and abnormal events trigger incremental learning and global rescheduling of a cloud hybrid expert large model. The present application eliminates dangerous instructions from a physical mechanism, and achieves global optimization of resources and energy consumption under the premise of ensuring production safety.

Description

Technical Field

[0001] This invention relates to the field of iron tower production technology, and in particular to a method and system for dynamic scheduling of iron tower production resources. Background Technology

[0002] With the rapid development of ultra-high voltage (UHV) transmission line construction, the manufacturing of transmission towers is gradually shifting towards a flexible production model characterized by diverse varieties, small batches, and customization. The actual production of towers involves numerous complex processes, including plate cutting, hole making, assembly, and welding, with a large number of equipment and frequent process disturbances. Traditional tower production scheduling relies heavily on manual experience or a single centralized information system, making it difficult to respond to dynamic changes on-site in real time. This often leads to problems such as high equipment vacancy rates, low material utilization, and excessive energy consumption. Therefore, how to introduce artificial intelligence algorithms and combine them with a cloud-edge collaborative architecture to achieve dynamic and optimal allocation of tower production resources has become a pressing technical challenge in this field.

[0003] An existing invention patent application with patent number CN120949728A discloses a method and system for intelligent manufacturing workshop management based on edge-cloud collaboration. By processing abnormal states and comprehensively considering real-time equipment stability, aging degree, and failure frequency to obtain an equipment status performance index, the system categorizes equipment into different status levels and matches corresponding reserve coefficients. Production speed is scientifically increased starting with equipment in excellent condition, compensating for delays while ensuring equipment safety, and achieving a balance between production efficiency and equipment lifespan. Simultaneously, through data-driven scheduling strategies, production resource allocation is optimized to ensure timely delivery of production tasks.

[0004] As with related technologies, such methods typically treat production data as pure numerical values ​​for black-box neural network training, failing to embed physical laws such as the constitutive equations of material mechanics and thermodynamics during tower manufacturing into the model. This leads to scheduling strategies that easily violate engineering common sense, causing irreversible damage to equipment or materials. Furthermore, existing solutions primarily rely on the cloud for anomaly analysis and decision-making. Faced with sudden equipment vibration exceeding limits or rapid temperature increases during tower production, the cloud's transmission and computational latency cannot meet the millisecond-level emergency shutdown or derating protection requirements. When unknown anomalies occur on-site, existing technologies typically employ conventional full-scale fine-tuning or directly discard the anomaly. This not only incurs huge communication overhead and data latency but also easily leads to catastrophic forgetting of large models. Additionally, their scheduling objectives are often singular, failing to find the optimal balance between material utilization and overall energy consumption. Summary of the Invention

[0005] In view of this, the present invention proposes a dynamic scheduling method and system for tower production resources that combines physical hard constraints, millisecond response and dual-objective optimization, in order to solve the problem that existing solutions are difficult to coordinately optimize material utilization and comprehensive energy consumption and respond quickly to dynamic disturbances while ensuring production safety, due to the lack of physical constraints, slow edge response and single scheduling objective.

[0006] The technical solution of this invention is implemented as follows: On the one hand, the present invention provides a method for dynamic scheduling of tower production resources, comprising the following steps: Acquire multi-source heterogeneous data, including process parameters and on-site sensor data, corresponding to the entire process of iron tower production; Using equipment, molds, and sheet metal to be processed in the tower production workshop as graph nodes, multi-source heterogeneous data as node features, and process flow relationships and physical contact relationships as edges, a multimodal graph neural network is initialized and constructed. The material mechanics constitutive equation and thermodynamic laws are transformed into penalty terms in the loss function of the multimodal graph neural network for training, resulting in a multimodal safety evaluation graph that defines the safety boundary. Multi-source heterogeneous data is mapped into a high-dimensional state space, and a reward function is constructed with the objectives of maximizing material utilization and minimizing comprehensive energy consumption. Multi-agent reinforcement learning based on multi-agent deep deterministic policy gradient algorithm is used to solve the flexible process path, generate a preliminary scheduling strategy and execute it. During the execution of the initial scheduling strategy, a lightweight graph isomorphic network pre-deployed at the edge is invoked. First, the deviation between the field sensor data and the multimodal safety evaluation map is calculated in real time. Then, the local sensitivity of the deviation relative to the process parameters is calculated. When the local sensitivity exceeds the preset threshold, a millisecond-level safety shutdown or derated operation is triggered at the edge. The abnormal event is traced back to the cloud to trigger incremental fine-tuning and global rescheduling of the hybrid expert big model.

[0007] Based on the above technical solutions, preferably, before calling the lightweight graph isomorphic network pre-deployed at the edge, the field sensor data is preprocessed. The specific steps include: Fast Fourier transform and wavelet packet decomposition were performed on the vibration, temperature and acoustic signature time series data collected on site to extract frequency domain energy features and time domain statistical features. The extracted frequency domain energy features and time domain statistical features are used to construct an adjacency matrix, which is then input into a graph isomorphic network for real-time topology feature updates.

[0008] Based on the above technical solutions, the preferred method for mapping multi-source heterogeneous data into a high-dimensional state space includes: using one-hot encoding to discretize and characterize the type of equipment and the specifications of the mold. Minimum-maximum normalization is used to continuously scale the dimensions and material hardness of the sheet material to be processed. The discretized representation is concatenated with the vector scaled from continuous values ​​to form the environment state input for the multi-agent deep deterministic policy gradient algorithm.

[0009] Based on the above technical solutions, the preferred method involves using a lightweight graph isomorphic network pre-deployed at the edge to calculate in real time the deviation between the on-site sensor data and the multimodal safety evaluation map, as well as the local sensitivity. Specifically, this includes: Extracting topological feature vectors from field sensor data using graph isomorphic networks; Calculate the deviation between the topological feature vector and the feature vector of the corresponding node in the safety evaluation map, and calculate the local sensitivity of the deviation relative to the process parameters; If the local sensitivity exceeds the preset threshold, it is determined that the system is about to undergo irreversible physical deformation or equipment damage, and the edge end immediately executes the preset degraded operation command.

[0010] Based on the above technical solutions, the preferred hybrid expert large model includes multiple expert networks and a gated network; The gated network is used to receive the feature vector of abnormal events traced back from the edge, calculate the matching degree between the feature vector of abnormal events and the output feature vector of each expert network through the Softmax function, and define the calculation result as the attention weight; The top K expert networks with the highest matching degree are selected for activation. The outputs of the activated expert networks are weighted and summed according to the attention weights to obtain the final global rescheduling policy. The parameters of the inactive expert networks remain unchanged.

[0011] Based on the above technical solutions, the preferred incremental fine-tuning also includes a model parameter solidification step: When the cloud receives abnormal data traced back from the edge, it calculates the cosine similarity between the abnormal data and the historical fault feature database. If the similarity is lower than the preset threshold, it is determined to be an unknown fault type. At this time, the backbone parameters of the hybrid expert model are frozen, new fully connected layers are added only to the activated expert network modules, and the parameters are updated using the contrastive learning loss function.

[0012] Based on the above technical solutions, preferably, when updating parameters, an elastic weight solidification regularization method is used to constrain the parameters of the activated expert network.

[0013] Based on the above technical solutions, the preferred reward function configuration is as follows: By setting the total energy budget of the system, the energy consumption of equipment startup and operation is mapped to cost consumption items, the area of ​​the board to be processed is mapped to effective output items, and a nonlinear reward function is constructed by combining process constraint penalty items. By maximizing the cumulative expected return of the reward function, the multi-agent system iteratively optimizes the policy network parameters and obtains a balanced scheduling strategy that takes into account both material utilization and overall energy consumption.

[0014] Based on the above technical solutions, the preferred method, after tracing abnormal events back to the cloud to trigger incremental fine-tuning and global rescheduling of the hybrid expert large model, also includes: Based on the global rescheduling strategy, the operating status of equipment and mold configuration in the workshop are dynamically adjusted to achieve a dynamic balance of production resources.

[0015] On the other hand, the present invention provides a dynamic scheduling system for tower production resources, used to implement the above-mentioned dynamic scheduling method for tower production resources, comprising: The data acquisition layer is used to collect multi-source heterogeneous data and on-site sensor data throughout the entire tower production process. The graph construction layer is used to perform the steps of constructing a multimodal security evaluation graph; The cloud-based scheduling layer is equipped with a hybrid expert model and a multi-agent reinforcement learning algorithm to perform the steps of generating an initial scheduling strategy and global rescheduling. The edge control layer, deployed with a lightweight graph isomorphic network, performs real-time calculations of local sensitivity and triggers millisecond-level protection steps. The edge control layer and the cloud scheduling layer interact with each other via industrial Ethernet.

[0016] The method and system for dynamic scheduling of tower production resources of the present invention have the following advantages over the prior art: (1) By deeply coupling the constitutive equations of mechanics of materials with the laws of thermodynamics as prior constraints for model training, the scheduling model is endowed with an inherent logic of physical mechanism from the bottom layer, fundamentally eliminating dangerous instructions that violate engineering common sense. At the same time, a dual-objective reward function with material utilization rate and comprehensive energy consumption as the core is constructed through multi-agent reinforcement learning, guiding the system to autonomously evolve the optimal production rhythm within the safety boundary. Combined with the collaborative mechanism of lightweight GIN real-time protection at the edge and global overall planning of the large model in the cloud, it not only ensures millisecond-level emergency response in the event of sudden danger, but also realizes efficient allocation of macro-resources in the workshop.

[0017] (2) By introducing Fast Fourier Transform and Wavelet Packet Decomposition techniques at the edge, noise in the sensor signal can be accurately removed, and highly representative frequency domain energy and time domain statistical features can be extracted. On this basis, a lightweight graph isomorphic network is used to calculate the deviation between the topological features and the security spectrum and the local sensitivity in real time, which can accurately detect anomalies in the early stage of irreversible deformation or damage to the equipment. Once the threshold is exceeded, the edge can immediately and seamlessly switch to degraded operation or safe shutdown, which can effectively curb the occurrence of accidents and greatly reduce the failure rate of major equipment.

[0018] (3) This invention abandons the traditional inefficient full-scale fine-tuning and instead adopts a hybrid expert large model architecture. By dynamically activating the best-matching expert module through the attention mechanism of the gating network, and combining elastic weight solidification regularization with contrastive learning loss function, the model avoids catastrophic forgetting while absorbing new knowledge. This mechanism of learning when encountering new information and applying what it has learned enables the system to automatically generate highly adaptable global rescheduling strategies for unknown anomalies, realizing the self-evolution of the manufacturing system throughout its entire life cycle.

[0019] (4) This invention integrates equipment startup energy consumption, effective sheet metal output, and process constraints into a nonlinear reward function, guiding multiple agents to find the optimal balance between energy consumption and material utilization through long-term trial and error and iteration, thus completely solving the pain point of neglecting one aspect while addressing another in traditional scheduling. On this basis, by dynamically adjusting the operating status of workshop equipment and mold configuration, the rescheduling strategy issued from the cloud is further digested. This dynamic adaptation capability combining hardware and software ensures that the tower production system can quickly rebuild dynamic balance after encountering disturbances, providing strong technical support for the continuous and stable operation of intelligent manufacturing. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a flowchart of the dynamic scheduling method for tower production resources according to the present invention; Figure 2 This is an architecture diagram of the dynamic scheduling system for tower production resources of the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0023] In the description of the embodiments of the present invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "connected" and "linked" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium. Those skilled in the art can understand the specific meaning of the above terms in the embodiments of the present invention based on the specific circumstances.

[0024] In the description of the embodiments of the present invention, it should be noted that the terms "center", "longitudinal", "lateral", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", and "outer" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing the embodiments of the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the embodiments of the present invention.

[0025] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0026] Embodiments of the present invention are described in detail below, examples of which are illustrated 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 intended to explain the present invention, and should not be construed as limiting the present invention.

[0027] The following disclosure provides numerous different embodiments or examples for implementing various structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. 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. Additionally, examples of various specific processes and materials are provided in this invention; however, those skilled in the art will recognize the applicability of other processes and / or the use of other materials.

[0028] like Figure 1 As shown, the method for dynamic scheduling of tower production resources of the present invention includes: Step S1: Obtain multi-source heterogeneous data and on-site sensor data corresponding to the entire tower production process, including process parameters.

[0029] In this step, multi-source heterogeneous data is collected in real time through IoT sensors deployed in the tower production workshop and workshop environmental monitoring units. This multi-source heterogeneous data includes, but is not limited to: equipment operating status such as on / off, no-load, and full-load; process parameters such as cutting speed and welding current; environmental data such as workshop temperature and humidity; and order information. Simultaneously, static data such as the specifications, material hardness, and order batch of the sheet metal to be processed are obtained from the product data management system of the manufacturing execution system. Furthermore, high-precision vibration sensors, temperature sensors, and acoustic sensors are deployed on-site to collect time-series data at high frequency, such as 2000 times per second, providing a data foundation for subsequent real-time physical condition monitoring and safety early warning.

[0030] Step S2: Using the equipment, molds, and plates to be processed in the tower production workshop as graph nodes, multi-source heterogeneous data as node features, and process flow relationships and physical contact relationships as edges, a multimodal graph neural network is initialized and constructed. The material mechanics constitutive equation and thermodynamic laws are transformed into penalty terms in the loss function of the multimodal graph neural network for training, resulting in a multimodal safety evaluation graph that defines the safety boundary.

[0031] The specific implementation process for this step is as follows: First, the physical entities in the workshop are abstracted into a graph structure: for example, equipment such as laser cutting machines and radial drilling machines, each mold, and each piece of sheet material to be processed are all treated as nodes in the graph.

[0032] The node feature vector is formed by standardizing the corresponding multi-source heterogeneous data collected in step S1. The process flow relationship between entities, such as the next process of sheet A entering equipment B, and the physical contact relationship, such as mold C being installed on equipment D, are abstracted as edges connecting these nodes, thereby constructing a dynamic multimodal production graph.

[0033] Based on this graph, a three-layer Graph Attention Network (GAT) is initialized as a multimodal graph neural network. Its message passing and feature aggregation processes are as follows: For any node in the graph In the Output features in layered networks By itself and its neighboring nodes The characteristics of these elements determine the calculation formula as follows: ; in, Represents a node In the The feature vector of the layer; Represents a node neighboring nodes In the The feature vector of the layer; Indicates the first Layer-learnable weight matrix; This represents a non-linear activation function, and in this embodiment, the ReLU function is preferred. Represents a node For nodes The attention coefficient, calculated from the attention mechanism, reflects the differences in the importance of different neighbors; Represents a node The set of neighboring nodes.

[0034] During model training, in addition to minimizing the standard loss of the main task... In addition, the core innovation lies in embedding physical laws as hard constraints to introduce physical mechanism penalty terms.

[0035] against Those skilled in the art can configure it as a conventional loss function, such as map reconstruction error, cross-entropy loss in classification tasks, or mean square error in regression tasks, depending on the specific application scenario of iron tower production.

[0036] Define the total loss function as follows: = + ; in, The main task loss function is used to drive the model to learn the inherent patterns in the production data; This is a penalty term in materials mechanics. For example, according to the Von Mises yield criterion, a penalty is imposed when the predicted equivalent stress exceeds the material's yield strength; This is a thermodynamic penalty term. Based on the law of conservation of energy, it ensures a balance between heat generation and dissipation in the equipment, preventing overheating damage. This represents the L2 regularization term to prevent the model from overfitting. , , These represent the balance weight coefficients for the mechanical penalty term, the thermal penalty term, and the regularization term, respectively. In a preferred embodiment of the invention, these are set as follows: =0.4, =0.3, =0.3.

[0037] Specifically, The specific calculation formula is as follows: ; in, and These are the sets of all weight parameters and bias parameters in the model, respectively. and These represent the i-th specific weight value and bias value in the set, respectively. This is the index number of the model parameter.

[0038] The network was trained using historical production data and corresponding sensor data, with an optimal training epoch count of 500. The optimizer was Adam, and the initial learning rate was set to 1. Once trained, the network becomes a multimodal safety assessment map capable of accurately evaluating the system's distance from the physical safety boundary under different production conditions. This map accurately reflects the theoretical safety limits of equipment and materials under different operating conditions.

[0039] Step S3: Map the multi-source heterogeneous data into a high-dimensional state space, construct a reward function with the objectives of maximizing material utilization and minimizing overall energy consumption, and use multi-agent reinforcement learning based on a multi-agent deep deterministic policy gradient algorithm to solve for the flexible process path, generate a preliminary scheduling strategy, and execute it; the specific implementation is as follows: As a further optimization of step S3, the step of mapping multi-source heterogeneous data into a high-dimensional state space specifically includes: First, the data mapping process is as follows: Discrete features such as the type of production equipment and the specifications of the mold are represented using one-hot encoding to generate sparse vectors. Then, continuous features such as the size and material hardness of the sheet metal to be processed are scaled using min-max normalization, mapping them to the [0,1] interval. The formula is as follows: ; in, This represents the original continuous characteristic value, such as the original value of the plate thickness; and These represent the minimum and maximum values ​​of the feature in the entire dataset, respectively. This represents the normalized eigenvalues. Finally, the discretized representation vector is concatenated with the vector scaled from the continuous values ​​to form the high-dimensional state vector. , which serves as the environment state input for the multi-agent deep deterministic policy gradient algorithm.

[0040] Secondly, construct a bi-objective nonlinear reward function. Guide the agent to learn the optimal scheduling strategy: = - - ; in, Indicates the action Below is the actual effective area of ​​the board material that is utilized; For the total area of ​​the boards used, the formula is related to... The ratio is the material utilization rate; To indicate the execution of an action Then, the sum of the startup energy consumption and operating energy consumption generated by all related equipment; The total system energy budget set for the current production task, where and The ratio represents the overall energy consumption level. As an indicator function, when a scheduling action is performed. Its value is 1 when any process constraint is violated, and 0 otherwise. This item is a process constraint penalty item. , , and are weighting coefficients for material utilization rate, overall energy consumption, and constraint penalty, respectively. In a preferred embodiment of the present invention, in order to achieve a balance between improving material utilization rate and reducing energy consumption, and to strictly ensure feasibility, the following settings are established: =0.6, =0.3; =0.1.

[0041] This nonlinear reward function This constitutes the core optimization objective of the multi-agent deep deterministic policy gradient algorithm. Within this algorithm, each agent (representing a schedulable production resource unit, such as a machine tool) has an independent Actor network and Critic network. The agent's objective is to maximize the cumulative discounted reward. ; in, Representation Strategy That is, the objective function that needs to be maximized; Indicates about trajectory Mathematical expectation, trajectory In strategy The state-action sequence is obtained by sampling from the environment; Indicates by parameters A parameterized policy function that is based on the current state. Output status The probability distribution; The time step index representing the decision point; This represents the total number of time steps in a scheduling task cycle or a training round. This represents the discount factor, with a value range of [0,1], used to weigh the importance of current rewards against future rewards. In this embodiment, it is set... =0.99 indicates that the algorithm focuses more on long-term gains. That is, as defined above, at time step In state And perform the action The instant reward obtained afterward.

[0042] The MADDPG algorithm was explored and iterated tens of thousands of times in a simulation environment to continuously optimize the strategy parameters. To maximize Ultimately, a strategy network capable of generating preliminary scheduling strategies is trained. This strategy achieves a Pareto optimal balance between material utilization and overall energy consumption while satisfying all process constraints. The workshop production system will execute this strategy.

[0043] Step S4: During the execution of the initial scheduling strategy, the lightweight graph isomorphic network pre-deployed at the edge is invoked. First, the deviation between the field sensor data and the multimodal safety evaluation map is calculated in real time. Then, the local sensitivity of the deviation relative to the process parameters is calculated. When the local sensitivity exceeds the preset threshold, the edge is triggered to perform a millisecond-level safety shutdown or derated operation. The abnormal event is traced back to the cloud to trigger the incremental fine-tuning and global rescheduling of the hybrid expert large model.

[0044] First, during the execution of the initial scheduling strategy, the edge control layer continuously preprocesses sensor data: performing Fast Fourier Transform (FFT) and wavelet packet decomposition on the time-series data of vibration, temperature, and acoustic signatures collected on-site. The FFT is used to extract the frequency domain energy features of the signal. Wavelet packet decomposition can provide more refined time-frequency analysis, and its decomposition process can be represented as: ; in, The original discrete-time signal; For discrete wavelet packet basis functions, derived from the mother wavelet through scaling Peaceful relocation get; For the signal at scale ,frequency band The wavelet packet coefficients are calculated. The energy of each frequency band coefficient is then determined. It can extract time-domain statistical features such as energy entropy and variance that characterize early faults such as bearing wear and tool micro-breakage. These frequency-domain and time-domain features are combined with equipment ID, timestamp, etc. to construct a feature vector, which is then organized into a topology graph that characterizes the current real-time status of the workshop and input into a lightweight graph isomorphic network pre-deployed on an edge computing device.

[0045] Next, the following computation process is performed: the lightweight graph isomorphic network extracts the topological feature vector of the current state in real time. Calculate the vector and its corresponding standard state node features in the multimodal security evaluation map trained in step S2. The deviation was calculated, and the local sensitivity was further calculated. : ; in, The L2 norm of a vector; This represents a vector of key process parameters, such as spindle speed and feed rate. Sensitivity. This reflects the rate at which a small change in process parameters causes the system to deviate from a safe state. If Exceeding the preset threshold In this embodiment, the following is set If the value is 0.85, it is determined that the system is about to undergo irreversible physical deformation or equipment damage. At this time, the edge control layer will bypass the cloud and issue a command to the device controller for safe shutdown or derated operation within 5 milliseconds.

[0046] At the same time, the feature vector of the abnormal event, the process parameters at the trigger time, and the processing result will be encapsulated and traced back to the cloud scheduling layer via industrial Ethernet.

[0047] Upon receiving an abnormal event, the cloud-based scheduling layer initiates the incremental fine-tuning and global rescheduling process of the hybrid expert large model: Gating and Expert Selection: The hybrid expert large model comprises one gating network and N expert networks. The gating network receives anomalous feature vectors. The matching degree between it and each expert network is calculated using the Softmax function: ; in, and The weights and biases of the gated network; Operation with the largest retained value One component is set, and the values ​​of all other components are set to negative infinity. In this embodiment, it is set to... =2. Select the top option with the highest attention weight. An expert network is activated and participates in the calculation of the current sample.

[0048] Unknown Fault Identification and Parameter Consolidation: Calculating Anomaly Characteristics Compared with all features in the historical fault feature library Cosine similarity: ; in, This is the feature vector obtained after feature extraction of the current device state; Representing the first in the historical database Each sample's feature vector corresponds to a known fault mode.

[0049] If the maximum similarity is lower than the preset threshold τ (τ=0.85 in this embodiment), it is determined to be an unknown fault type. At this time, the backbone parameters of the hybrid expert model are frozen, only those activated in step 1 are considered. A small adaptation layer is added to each expert network. Then, a contrastive learning loss function is used. By combining the elastic weight solidification regularization method, only the parameters of the newly added adaptation layer are incrementally fine-tuned, enabling the model to learn new failure modes while effectively protecting knowledge from old tasks and avoiding catastrophic forgetting.

[0050] Incremental fine-tuning: using a contrastive learning loss function The elastic weight solidification regularization method is used to fine-tune the newly added adaptation layer to learn new failure modes while avoiding forgetting old knowledge.

[0051] The contrastive learning loss function adopts the InfoNCE form: ; in, Features of the current abnormal samples To generate positive sample features through data augmentation, Features of all samples (including negative samples) within a batch. (·) represents the cosine similarity function. The temperature parameter is set to 0.07, and M is the batch size.

[0052] The elastic weight solidification regularization term is added to the total loss to constrain key parameters from deviating from their old values. ; in, To solidify the regularization coefficients for the elastic weights, this embodiment sets... =1000; The first one on the diagonal of the Fisher information matrix The importance of each parameter; This is the value of the parameter before fine-tuning.

[0053] Generate and execute a global rescheduling strategy: The fine-tuned hybrid expert model integrates the current global production status and new fault knowledge to generate a global rescheduling strategy. This strategy dynamically balances workshop resources by adjusting equipment operation sequences, replacing molds, and reallocating tasks, and is then sent to the edge control layer for execution, thereby enabling the system to quickly restore efficient and safe operation after anomalies are resolved.

[0054] Based on the above methods, such as Figure 2 As shown, the present invention also provides a dynamic scheduling system for tower production resources, the collaborative workflow of which is as follows: Data acquisition layer: Composed of various sensors deployed in the workshop and MES / PDM system interfaces, it is used to collect multi-source heterogeneous data and field sensor data, and synchronously upload the data to the map construction layer and edge control layer.

[0055] The graph construction layer is typically deployed on an enterprise private cloud or a high-performance server. After receiving data, it executes step S2 to build and continuously update the multimodal security evaluation graph, and then sends the graph model parameters to the edge control layer.

[0056] Cloud-based scheduling layer: Deployed on a central cloud server, it includes a hybrid expert model and a multi-agent reinforcement learning algorithm module. After receiving data, it executes step S3 to generate a preliminary scheduling policy and distributes it to the workshop. Simultaneously, it receives abnormal events from the edge control layer, triggering the hybrid expert model to perform incremental fine-tuning and generate a new global rescheduling policy.

[0057] Edge control layer: Composed of edge computing devices deployed on-site in the workshop, such as industrial control computers, which load a lightweight security graph model from the graph construction layer. It executes step S4 in real time, performing millisecond-level risk assessment and emergency response, and traces the characteristics of abnormal events back to the cloud scheduling layer via industrial Ethernet.

[0058] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for dynamic scheduling of tower production resources, characterized in that, Includes the following steps: Acquire multi-source heterogeneous data, including process parameters and on-site sensor data, corresponding to the entire process of iron tower production; Using equipment, molds, and sheet metal to be processed in the tower production workshop as graph nodes, and multi-source heterogeneous data as node features, combined with process flow relationships and physical contact relationships as edges, a multimodal graph neural network is initialized and constructed. The material mechanics constitutive equation and thermodynamic laws are transformed into penalty terms in the loss function of the multimodal graph neural network for training, resulting in a multimodal safety evaluation graph that defines the safety boundary. The multi-source heterogeneous data is mapped into a high-dimensional state space, and a reward function is constructed with the objectives of maximizing material utilization and minimizing comprehensive energy consumption. The flexible process path is solved by multi-agent reinforcement learning based on the multi-agent deep deterministic policy gradient algorithm, and a preliminary scheduling strategy is generated and executed. During the execution of the initial scheduling strategy, a lightweight graph isomorphic network pre-deployed at the edge is invoked. First, the deviation between the field sensor data and the multimodal safety evaluation map is calculated in real time. Then, the local sensitivity of the deviation relative to the process parameters is calculated. When the local sensitivity exceeds a preset threshold, a millisecond-level safety shutdown or derating operation is triggered at the edge. The abnormal event is traced back to the cloud to trigger incremental fine-tuning and global rescheduling of the hybrid expert large model.

2. The method for dynamic scheduling of tower production resources as described in claim 1, characterized in that, Before invoking the lightweight graph isomorphic network pre-deployed at the edge, the field sensor data is preprocessed, specifically including the following steps: Fast Fourier transform and wavelet packet decomposition were performed on the vibration, temperature and acoustic signature time series data collected on site to extract frequency domain energy features and time domain statistical features. The extracted frequency domain energy features and time domain statistical features are used to construct an adjacency matrix, which is then input into the graph isomorphic network for real-time topology feature updates.

3. The method for dynamic scheduling of tower production resources as described in claim 1, characterized in that, Mapping the multi-source heterogeneous data into a high-dimensional state space includes: The type of equipment and the specifications of the mold are discretized using unique thermal coding; Minimum-maximum normalization is used to continuously scale the dimensions and material hardness of the sheet material to be processed. The discretized representation is concatenated with the vector scaled by the continuous value to form the environment state input of the multi-agent deep deterministic policy gradient algorithm.

4. The method for dynamic scheduling of tower production resources as described in claim 1, characterized in that, The process of calling a lightweight graph isomorphic network pre-deployed at the edge to calculate in real time the deviation and local sensitivity between the field sensor data and the multimodal safety assessment map specifically includes: The topological feature vector of the field sensor data is extracted using a graph isomorphic network. Calculate the deviation between the topological feature vector and the feature vector of the corresponding node in the safety evaluation map, and calculate the local sensitivity of the deviation relative to the process parameters; If the local sensitivity exceeds a preset threshold, it is determined that the system is about to undergo irreversible physical deformation or equipment damage, and the edge terminal immediately executes a preset degraded operation command.

5. The method for dynamic scheduling of tower production resources as described in claim 1, characterized in that: The hybrid expert large model includes multiple expert networks and a gated network; The gated network is used to receive the abnormal event feature vectors traced back from the edge, calculate the matching degree between the abnormal event feature vectors and the output feature vectors of each expert network through the Softmax function, and define the calculation result as the attention weight. The top K expert networks with the highest matching degree are selected for activation. The outputs of the activated expert networks are weighted and summed according to the attention weights to obtain the final global rescheduling policy. The parameters of the inactive expert networks remain unchanged.

6. The method for dynamic scheduling of tower production resources as described in claim 5, characterized in that, The incremental fine-tuning also includes a model parameter solidification step: When the cloud receives abnormal data traced back from the edge, it calculates the cosine similarity between the abnormal data and the historical fault feature database. If the similarity is lower than a preset threshold, it is determined to be an unknown fault type. At this time, the backbone parameters of the hybrid expert model are frozen, new fully connected layers are added only to the activated expert network modules, and the parameters are updated using the contrastive learning loss function.

7. The method for dynamic scheduling of tower production resources as described in claim 6, characterized in that: When updating parameters, the flexible weight solidification regularization method is used to constrain the parameters of the activated expert network.

8. The method for dynamic scheduling of tower production resources as described in claim 1, characterized in that, The reward function is specifically configured as follows: By setting the total energy budget of the system, the energy consumption of equipment startup and operation is mapped to cost consumption items, the area of ​​the board to be processed is mapped to effective output items, and a nonlinear reward function is constructed by combining process constraint penalty items. By maximizing the cumulative expected return of the reward function, the multi-agent system iteratively optimizes the policy network parameters and solves for a balanced scheduling strategy that takes into account both material utilization and overall energy consumption.

9. The method for dynamic scheduling of tower production resources as described in claim 1, characterized in that: After tracing the abnormal event back to the cloud to trigger incremental fine-tuning and global rescheduling of the hybrid expert large model, it also includes: Based on the global rescheduling strategy, the operating status of equipment and mold configuration in the workshop are dynamically adjusted to achieve a dynamic balance of production resources.

10. A dynamic scheduling system for tower production resources, used to implement the dynamic scheduling method for tower production resources as described in any one of claims 1 to 9, characterized in that, include: The data acquisition layer is used to collect multi-source heterogeneous data and on-site sensor data throughout the entire tower production process. The graph construction layer is used to perform the steps of constructing a multimodal security evaluation graph; The cloud-based scheduling layer is equipped with a hybrid expert model and a multi-agent reinforcement learning algorithm to perform the steps of generating an initial scheduling strategy and global rescheduling. The edge control layer, deployed with a lightweight graph isomorphic network, performs real-time calculations of local sensitivity and triggers millisecond-level protection steps. The edge control layer and the cloud scheduling layer interact with each other via industrial Ethernet.

Images