Dynamic scheduling and decision optimization system for machine tool feeding and discharging

Abstract

The application discloses a machine tool feeding and discharging dynamic scheduling and decision optimization system and belongs to the technical field of production scheduling management, and comprises the following modules: a solution set fingerprint construction module, which is used for constructing a solution set fingerprint for the obtained Pareto solution set; a path prediction model training module, which is used for training a solution set structure path prediction model based on a historical solution set fingerprint sequence; a scheduling anchor solution determination module, which is used for determining a scheduling anchor solution of a current period according to a current Pareto solution set and a future solution set form output by the prediction model; and an anti-disturbance test module, which is used for forming a resource allocation cluster based on the scheduling anchor solution. The solution set fingerprint can be constructed, and the solution set structure path prediction model can be trained, so that the dynamic scheduling is converted from continuous static optimization to smooth evolution decision, resource waste caused by frequent recalculation of a traditional algorithm due to disturbance is avoided, production line oscillation and stop caused by invalidation of an old scheme or waiting during a recalculation empty window period is eliminated, and the continuity and stability of scheduling in a dynamic environment are improved.

Description

Technical Field

[0001] This invention relates to the field of production scheduling and management technology, and more specifically, to a dynamic scheduling and decision optimization system for machine tool loading and unloading. Background Technology

[0002] In traditional production scheduling, multi-objective optimization algorithms, such as multi-objective genetic algorithms and multi-objective particle swarm optimization, are widely used to find the optimal production scheduling scheme, especially in multi-objective and multi-constraint environments. However, in dynamic production environments, the application of these algorithms has certain limitations, mainly reflected in their insufficient adaptability and stability in the face of real-time changes.

[0003] First, traditional multi-objective optimization algorithms attempt to find a compromise between different objectives by calculating the Pareto optimal solution set in a static production environment. These algorithms can solve resource allocation problems in production scheduling to a certain extent, such as production time, machine load, and workpiece processing sequence. However, disturbances in a dynamic production environment, such as equipment failure, process delays, and emergency orders, often cause the original optimal solution set to become invalid. At this time, the algorithm needs to recalculate a new scheduling scheme, which often leads to an increase in system uncertainty. The recalculation process can cause the production line to enter a window period, during which either an invalid scheduling scheme is used or the system is stuck in a waiting state, causing production to stop or equipment to idle, thus wasting a lot of computing resources.

[0004] The deeper problem is that existing algorithms treat dynamic scheduling problems as a series of independent static optimization problems, ignoring the continuity and stability over time. Each optimization calculation often focuses only on the Pareto optimal solution at a certain instant, lacking consideration of historical scheduling states and future disturbance trends. This one-sidedness leads to discontinuity and drastic fluctuations in scheduling schemes over time, making it impossible to ensure that the production system maintains stable and efficient operation in a changing production environment.

[0005] For example, in a machining workshop, when equipment failure or urgent orders occur, traditional multi-objective optimization algorithms may quickly adjust the scheduling scheme in order to reach a new optimal solution. However, because the correlation between scheduling decisions is ignored, frequent scheduling changes often lead to the rearrangement of processes and frequent switching of personnel and equipment, which in turn causes problems such as production stoppages or equipment idleness. Ultimately, this overly sensitive adjustment mechanism not only fails to improve production efficiency, but may also lead to the waste of production resources. Summary of the Invention

[0006] To address the problems existing in the prior art, the purpose of this invention is to provide a dynamic scheduling and decision optimization system for machine tool loading and unloading. This system can transform dynamic scheduling from continuous static optimization to smooth evolutionary decision-making by constructing solution set fingerprints and training a solution set structure path prediction model, thus avoiding the resource waste caused by frequent recalculations due to disturbances in traditional algorithms.

[0007] To solve the above problems, the present invention adopts the following technical solution:

[0008] A dynamic scheduling and decision optimization system for machine tool loading and unloading includes:

[0009] The solution set fingerprint construction module is used to construct a solution set fingerprint for the obtained Pareto solution set.

[0010] The path prediction model training module is used to train a path prediction model for solution set structure based on historical solution set fingerprint sequences.

[0011] The scheduling anchor solution determination module is used to determine the scheduling anchor solution for the current period based on the current Pareto solution set and the future solution set output by the prediction model.

[0012] The disturbance immunity test module is used to form resource allocation clusters based on scheduling anchor solutions and to perform disturbance immunity tests on the resource allocation clusters to obtain their disturbance immunity margin.

[0013] The dynamic disturbance adjustment module is used to determine whether the disturbance exceeds the disturbance immunity margin when it occurs. If it does not exceed the margin, adjustments are made within the resource allocation cluster. If it exceeds the margin, the scheduling anchor solution determination module is triggered to run again to determine a new scheduling anchor solution.

[0014] The feedback learning update module is used to update the solution set fingerprint database and the solution set structure path prediction model based on actual scheduling feedback.

[0015] Furthermore, the deset fingerprint construction module is also used for:

[0016] Calculate the gradient direction of each solution in the Pareto solution set, and map the Pareto solution set to a low-dimensional tradeoff feature space according to the preset gradient direction similarity criterion.

[0017] In the weighted feature space, dense kernels and sparse paths connecting dense kernels in the solution point cloud are identified, and the topological skeleton is extracted.

[0018] Dense kernels and sparse paths are encoded into symbol sets and assigned parameters to generate structural blueprints;

[0019] The geometric parameters in the structural blueprint are normalized based on the dynamic context of the production system, and the solution set fingerprint is generated.

[0020] Furthermore, the path prediction model training module is also used for:

[0021] Based on the core symbol set and the relation symbol set, the historical solution fingerprint sequence is split into kernel trajectory and relation strength trajectory;

[0022] For the kernel trajectory, the local tangent space and local manifold are learned within a local window at each time point to obtain the manifold parameters;

[0023] Candidate states for each core symbol are generated based on manifold parameters, and candidate states are competitively reorganized according to the connection relationships represented by the relation symbol set to synthesize candidate future structural blueprints.

[0024] The mechanisms of trajectory decomposition, manifold learning, and competitive recombination are encapsulated into a trainable and executable solution set structure path prediction model.

[0025] Furthermore, the scheduling anchor solution determination module is also used for:

[0026] Calculate the distance between each solution in the current Pareto solution set and each structural symbol in the predicted future solution set in the weighted feature space, and generate a bidirectional soft correlation matrix;

[0027] Based on the soft correlation matrix, a multi-step virtual deduction is performed on each solution in the current Pareto solution set to generate virtual paths leading to different structural symbols in the predicted future solution set form.

[0028] Based on the changes in the weights of each virtual path, the path convergence entropy is calculated; and parameter perturbation is simulated to test the ability of each path to maintain its evolution direction, thereby selecting high-quality virtual paths.

[0029] A starting solution is selected from high-quality virtual paths, and a decision instruction set containing a specific scheduling scheme, backup solutions, and expected adjustment direction instructions is generated around the starting solution as the scheduling anchor solution for the current cycle.

[0030] Furthermore, the disturbance rejection test module is also used for:

[0031] The decision instruction set is analyzed, the resource and process binding relationship common to all scheduling schemes is identified, resource and process cells are formed, and the cells are connected into a scheduling backbone network according to the process constraints.

[0032] Calculate the multidimensional affinity between flexible processes and cells in the scheduling backbone network, and adsorb flexible processes to corresponding cells based on the multidimensional affinity to generate tree-like resource allocation clusters;

[0033] Virtual resource erosion is performed on the tree-like resource allocation cluster to simulate the propagation process upstream and to branches after the selected cell fails, generating a failure propagation chain.

[0034] Furthermore, the disturbance rejection test module is also used for:

[0035] Analyze the failure propagation chain, extract key nodes, average propagation depth and structural entropy change index, and calculate the structural robustness score and collapse threshold based on the index;

[0036] By combining structural robustness scores, collapse thresholds, and real-time load rates, a disturbance tolerance margin is generated, and key nodes are marked on a tree-like resource allocation cluster to form a structural vulnerability map.

[0037] Furthermore, the dynamic disturbance adjustment module is also used for:

[0038] The disturbance events are analyzed in a multidimensional quantitative manner and mapped to a structural vulnerability map to identify the key affected nodes and load impacts.

[0039] Starting from critical nodes, stress distribution is calculated based on propagation patterns, and cascading collapses are predicted based on collapse thresholds.

[0040] Furthermore, the dynamic disturbance adjustment module is also used for:

[0041] If no cascading collapse is triggered, a reverse resolution operation is performed based on the stress distribution to reconstruct the resource allocation cluster;

[0042] If a chain reaction of crashes is triggered, the quantitative analysis results of the disturbance event, the corresponding failure propagation chain, and the information of the subgraph about to break identified from the tree-like resource allocation cluster will be encapsulated to generate a rescheduling request.

[0043] Furthermore, the feedback learning update module is also used for:

[0044] Align the actual production state sequence with the structural blueprint of the predicted future solution set form to generate actual symbol trajectories, and compare them with the preferred virtual path to cut out verified effective segments and empirical deviation segments.

[0045] The deviation patterns of empirical deviation segments are analyzed, and the deviation patterns are correlated with recorded disturbance events and local stress distributions to establish a causal hypothesis chain between disturbance type, impact spectrum and deviation patterns.

[0046] Furthermore, the feedback learning update module is also used for:

[0047] The confidence level of the fingerprints involved is enhanced based on the verified valid fragments, and the parameters of the fingerprints are fine-tuned or new candidate fingerprints are generated based on empirical bias fragments and causal hypothesis chains to update the solution set fingerprint library.

[0048] Based on the causal hypothesis chain, the parameters and rules of the competitive recombination mechanism and the local manifold learning mechanism in the solution set structure path prediction model are adjusted in a targeted manner.

[0049] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0050] (1) This scheme transforms dynamic scheduling from continuous static optimization to smooth evolution decision-making by constructing solution set fingerprints and training solution set structure path prediction model. This avoids the waste of resources caused by frequent recalculation due to disturbances in traditional algorithms. At the same time, it eliminates the production line oscillation and stoppage caused by the failure or waiting of old schemes during the recalculation window, and improves the continuity and stability of scheduling in dynamic environment.

[0051] (2) This solution performs disturbance resistance testing on the resource allocation cluster and generates disturbance resistance margin. When a disturbance occurs, the scope of impact can be accurately determined: small disturbances are directly resolved by local adjustment within the cluster, while large disturbances trigger rescheduling. This avoids the inefficiency of excessive recalculation and can quickly respond to anomalies, reducing the impact of equipment failures and emergency order insertions on the production rhythm and enhancing the disturbance resistance of the production line.

[0052] (3) This scheme splits effective and deviation segments by actual scheduling feedback, associates disturbance and deviation patterns and updates the solution fingerprint library and prediction model, so that the fingerprint library continuously fits the solution characteristics of actual production, and the manifold learning and competitive reorganization mechanism of the prediction model are more suitable for real scenarios. After long-term iteration, it can gradually improve the accuracy and adaptability of the scheduling scheme and realize the self-optimization of decision-making.

[0053] (4) This scheme identifies the backbone network of scheduling and absorbs flexible processes to form a tree-like resource allocation cluster. It not only preserves the stability of core resources and process binding, but also improves resource utilization by utilizing the multi-dimensional affinity allocation of flexible processes. At the same time, the adjustability of flexible processes provides flexible space for scheduling, balances production stability and resource efficiency, and adapts to the scheduling needs of different production goals. Attached Figure Description

[0054] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0055] Figure 1 This is a flowchart illustrating the relationships between the various modules in the machine tool loading and unloading dynamic scheduling and decision optimization system of this invention. Detailed Implementation

[0056] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0057] Please see Figure 1 A dynamic scheduling and decision optimization system for machine tool loading and unloading, including:

[0058] The solution set fingerprint construction module is used to construct a solution set fingerprint for the obtained Pareto solution set.

[0059] The path prediction model training module is used to train a path prediction model for solution set structure based on historical solution set fingerprint sequences.

[0060] The scheduling anchor solution determination module is used to determine the scheduling anchor solution for the current period based on the current Pareto solution set and the future solution set output by the prediction model.

[0061] The disturbance immunity test module is used to form resource allocation clusters based on scheduling anchor solutions and to perform disturbance immunity tests on the resource allocation clusters to obtain their disturbance immunity margin.

[0062] The dynamic disturbance adjustment module is used to determine whether the disturbance exceeds the disturbance immunity margin when it occurs. If it does not exceed the margin, adjustments are made within the resource allocation cluster. If it exceeds the margin, the scheduling anchor solution determination module is triggered to run again to determine a new scheduling anchor solution.

[0063] The feedback learning update module is used to update the solution set fingerprint database and the solution set structure path prediction model based on actual scheduling feedback.

[0064] The fingerprint construction module also performs the following steps:

[0065] Step 1: Calculate the gradient direction of each solution in the Pareto solution set, and map the Pareto solution set to a low-dimensional tradeoff feature space according to a preset gradient direction similarity criterion. The specific operations are as follows:

[0066] Each solution in the Pareto solution set corresponds to a set of objective function values ​​for a multi-objective optimization problem of machine tool loading and unloading scheduling, such as processing efficiency, resource utilization, and cost. First, the gradient direction of each solution is calculated: a high-dimensional objective space is constructed with each objective function as the dimension. By solving the partial derivative matrix of each solution with respect to neighboring solutions, the rate of change and direction of the solution in each objective dimension are extracted and integrated to obtain the gradient direction vector of the solution. Then, the gradient directions of all solutions are correlated according to a preset gradient direction similarity criterion. This criterion is constructed based on similarity measures such as the cosine of the gradient vector angle and Euclidean distance. Its logic is to classify solutions with gradient direction angles less than a preset threshold and Euclidean distances that meet the similarity conditions into the same category. Through this criterion, the Pareto solution set in the high-dimensional objective space is projected to a low-dimensional trade-off feature space. This mapping process is not a simple dimension reduction, but rather, by preserving the multi-objective trade-off relationship represented by the gradient direction, the low-dimensional trade-off feature space can accurately reflect the relative relationship of each solution in the objective dimension.

[0067] Step 2: In the weighted feature space, identify the dense kernels and sparse paths connecting the dense kernels in the solution point cloud, and extract the topological skeleton. The specific operations are as follows:

[0068] In the reduced-dimensionality tradeoff feature space, all solutions are distributed in the form of point clouds. First, the local density of each solution point is calculated using a density clustering algorithm. The calculation of local density is based on the neighborhood range of the solution point and the number of solution points within the neighborhood. Continuous solution point regions with local density higher than a preset threshold and the number of solution points in the neighborhood meeting the minimum requirement are identified as dense kernels. These dense kernels correspond to regions where the solution distribution is concentrated and the overall optimization effect is better in multi-objective tradeoffs. For the connection between different dense kernels, solution point sequences with local density lower than the threshold but capable of connecting different dense kernels are selected and defined as sparse paths. These paths reflect the transition characteristics and connection logic between different core tradeoff regions. Based on the identified dense kernels and sparse paths, a topological skeleton extraction algorithm is used to remove redundant solution point information from the point cloud, retaining only the topological structure that can completely represent the spatial distribution of the solution set, i.e., the topological skeleton. This skeleton contains both the spatial location and distribution characteristics of the dense kernels and the connection relationships of the sparse paths.

[0069] Step 3: Encode the dense kernel and sparse path into a symbol set and assign parameters to generate the structural blueprint. The specific operations are as follows:

[0070] For each extracted dense kernel, a unique core symbol is assigned based on its features such as center coordinates, volume, local density, and number of solution points in the tradeoff feature space. At the same time, each core symbol is assigned a corresponding quantization parameter, which covers key physical quantities such as the spatial location coordinates of the dense kernel, local density value, and multi-objective tradeoff preference weight. For each sparse path, a unique relation symbol is assigned based on its starting and ending dense kernels, path length, gradient change rate of solution points on the path, and connection strength, and a quantization parameter characterizing the transition characteristics of the path is assigned simultaneously.

[0071] After the allocation and parameter assignment of core symbols and relational symbols are completed, based on the connection relationships of the topological skeleton itself, all core symbols, relational symbols, and their corresponding quantization parameters are structurally integrated. This integration process is based on the physical structural characteristics of the topological skeleton and follows the logic of unique feature encoding and relational feature association encoding: First, the features of each dense kernel are independently encoded to ensure that each dense kernel corresponds to a unique core symbol and parameter combination; then, for each sparse path, according to the dense kernel identifier it connects to, the relational symbols and parameters corresponding to the path are associated and bound to the corresponding core symbols. Through this encoding logic, it can be ensured that each different topological structure corresponds to a unique set of symbols. The final symbolized set is the structural blueprint. The structural blueprint fully replicates the topological structure of the Pareto solution set through the association relationships between symbols and the quantization parameter characteristics corresponding to each symbol, so that the dynamic evolution law of the solution set can be mathematically described and analyzed through the changes in the symbol sequence and the adjustment of the parameters corresponding to each symbol.

[0072] Step 4: Normalize the geometric parameters in the structural blueprint based on the dynamic context of the production system, and encapsulate and generate a solution set fingerprint. The specific operations are as follows:

[0073] The dynamic operating environment of the production system includes real-time changing environmental parameters such as equipment load rate, process priority, material supply status, and disturbance frequency. These parameters directly affect the practical application value and evolution trend of the Pareto solution set. First, various geometric parameters in the structural blueprint are normalized. These parameters include dense kernel center coordinates, sparse path lengths, and the local density corresponding to core symbols. The normalization process can be achieved using the following formula: ;

[0074] This formula is derived by combining the weights of the production dynamic context. First, the original parameter x is mapped to the base interval through a linear transformation. Then, dynamic context weights and baseline offsets are introduced to make the normalization result adapt to the characteristics of the current production scenario and avoid the decrease in parameter comparability due to scenario differences. Here, x' represents the normalized geometric parameter value; x represents the original geometric parameter value in the structural blueprint. and These are the minimum and maximum values ​​of this type of parameter in historical samples, respectively; The dynamic context weights are calculated by weighting parameters such as equipment load rate and process priority. The baseline offset is set to 0.1 to ensure that the parameter values ​​are non-zero and evenly distributed. After normalizing all parameters, the symbol set of the structural blueprint, the normalized parameter set, and the key feature parameters of the dynamic context are structurally encapsulated to form a solution set fingerprint. This fingerprint retains the core features of the solution set topology and incorporates the dynamic environment information of the production system. It has scene adaptability and unique identification, and can be used as a historical data sample for training the solution set structure path prediction model. It also provides a standardized reference for comparative analysis of solution sets at different periods.

[0075] In a preferred embodiment of the present invention, the path prediction model training module is further configured to perform the following steps:

[0076] Step 21: Based on the core symbol set and the relation symbol set, the historical solution fingerprint sequence is split into kernel trajectory and relation strength trajectory. The specific operation is as follows:

[0077] The historical solution set fingerprint sequence is a sequence formed by arranging the solution set fingerprints generated in each production cycle in chronological order. Each solution set fingerprint contains a core symbol set representing the dense kernel features and a relation symbol set representing the sparse path connection relationships. Based on the attribute differences between these two types of symbol sets, the sequence is split: for the core symbol set, the quantization parameters corresponding to all core symbols in each cycle are extracted in chronological order to form a sequence that reflects the characteristics of the dense kernel changing over time, i.e., the kernel trajectory; for the relation symbol set, the connection strength parameters corresponding to each relation symbol in each cycle are extracted and arranged in chronological order to form a relation strength trajectory. Through this split, the temporal evolution of the solution set structure is decomposed into two dimensions: the evolution of the core structure and the evolution of the connection strength.

[0078] Step 22: For the kernel trajectory, learn the local tangent space and local manifold within a local window at each time point to obtain the manifold parameters. The specific operations are as follows:

[0079] For the core trajectory, a local window is first defined for each time point. This window includes the core trajectory segments of the current time point and several adjacent production cycles before and after it. Its length is determined based on the temporal resolution and evolution rate of the historical sequence to cover the complete change trend within the local range. Within each local window, the basis vector of the local tangent space is obtained by analyzing the variation law of the core symbol quantization parameters in the core trajectory. This basis vector is used to describe the local linear approximation characteristics of the core trajectory within this local range. At the same time, the dimension, curvature, and embedding mapping relationship of the low-dimensional local manifold corresponding to the core trajectory in the high-dimensional feature space are determined through embedding analysis. The set of parameters describing the characteristics of the local tangent space and the local manifold are the manifold parameters. These manifold parameters accurately characterize the evolution law of the core trajectory within the local time range.

[0080] Step 23: Generate candidate states for each core symbol based on manifold parameters, and competitively reorganize the candidate states according to the connection relationships represented by the relation symbol set to synthesize candidate future structural blueprints. The specific operations are as follows:

[0081] First, using the manifold parameters obtained in step 22 and the historical evolution trend of the core trajectory, the possible quantization parameter values ​​of each core symbol at future time points are deduced, resulting in multiple potential parameter combinations for each core symbol, i.e., candidate states of the core symbol. Then, based on the connection relationship constraints between core symbols represented by the relation symbol set, the candidate states of each core symbol are combined and matched. In this process, a competitive recombination mechanism is introduced, retaining only the candidate state combinations that conform to the connection relationship constraints and whose parameter combinations are more closely aligned with the historical evolution law. Combinations that deviate from the constraints or evolution law are eliminated. Through this competitive recombination process, a symbolic set that can reflect the future solution set topology is finally formed, i.e., the future structural blueprint candidate.

[0082] Step 24 encapsulates the trajectory decomposition, manifold learning, and competitive recombination mechanisms into a trainable and executable solution set structure path prediction model. The specific operations are as follows:

[0083] The trajectory splitting logic in step 21 is encapsulated into the model's data preprocessing module, responsible for converting the input historical solution set fingerprint sequence into kernel trajectories and relation strength trajectories. The local tangent space and local manifold learning logic in step 22 is encapsulated into the model's feature learning module, which automatically extracts the evolution law of kernel trajectories and outputs manifold parameters. The candidate state generation and competitive recombination logic in step 23 is encapsulated into the model's prediction generation module, which outputs future structural blueprint candidates based on manifold parameters and connectivity constraints. At the same time, adjustable training parameters are set for these modules, enabling the model to iteratively train using historical solution set fingerprint sequences, continuously optimizing the running accuracy of each module, and finally forming a solution set structure path prediction model with both training and execution capabilities. This model can receive historical solution set fingerprint sequence input and output future structural blueprint candidates.

[0084] In a preferred embodiment of the present invention, the scheduling anchoring solution determination module is further configured to perform the following steps:

[0085] Step 31: Calculate the distance between each solution in the current Pareto solution set and each structural symbol in the predicted future solution set in the weighted feature space, and generate a bidirectional soft correlation matrix. The specific operation is as follows:

[0086] Each solution in the current Pareto solution set is mapped to the previously constructed low-dimensional tradeoff feature space, and its spatial location can represent the core relationship of multi-objective tradeoffs. The future solution set output by the prediction model has its structural symbols corresponding to the core topological units of the future solution set, which also have corresponding representation positions in this tradeoff feature space. Based on this spatial carrier, the spatial distance between each solution in the current Pareto solution set and each structural symbol in the future solution set is calculated. The magnitude of this distance directly reflects the degree of matching between the two in the multi-objective tradeoff dimension. Subsequently, these distance values ​​are converted into correlation degrees. The smaller the distance, the higher the correlation degree. According to the correspondence between the current solution and the future structural symbols, all correlation degree values ​​are integrated to form a matrix, namely a two-way soft correlation matrix. This matrix does not rigidly bind the current solution and the future structural symbols, but reflects the potential correspondence between the two through the correlation degree values.

[0087] Step 32: Based on the soft correlation matrix, perform multi-step virtual deduction on each solution in the current Pareto solution set to generate virtual paths leading to different structural symbols in the predicted future solution set form. The specific operations are as follows:

[0088] Based on the soft correlation matrix generated in step 31, for each solution in the current Pareto solution set, a multi-step virtual evolutionary deduction is performed according to the preset time step corresponding to the subsequent continuous production cycle. In each step of the deduction, the evolutionary tendency of the solution towards the corresponding future structural symbol is determined according to the correlation degree value. Combined with the evolutionary rules of the historical solution set structure, its position in the tradeoff feature space and its correlation with the future structural symbol are adjusted. Through this continuous multi-step deduction, each current solution will form an evolutionary sequence that starts from its current position and gradually approaches a certain structural symbol in the future solution set form, that is, a virtual path leading to that structural symbol. Different current solutions will correspond to different virtual paths, and each path represents a potential possibility for the current solution to evolve into the future solution set form.

[0089] Step 33: Based on the changes in the attribution weights of each virtual path, calculate the path convergence entropy; and simulate applying parameter perturbations to test the ability of each path to maintain its evolution direction, thereby selecting high-quality virtual paths. The specific operations are as follows:

[0090] In the multi-step derivation of each virtual path, the weight assigned to the future structural symbol by the current solution changes with the time step. Based on the distribution characteristics of this weight change, the path convergence entropy is calculated. This entropy value measures the concentration of the weight change; the lower the entropy value, the more stable the path is towards a certain future structural symbol, and the better the convergence. Subsequently, a parameter perturbation is applied to each virtual path, i.e., a small adjustment is made to some of the quantization parameters of the current solution in the path, to test whether the path can maintain its original evolution direction. If the path still tends towards the target future structural symbol after the perturbation, its anti-perturbation ability is strong. Combining the magnitude of the path convergence entropy and the results of the anti-perturbation test, virtual paths with good convergence and strong anti-perturbation ability are selected, i.e., high-quality virtual paths.

[0091] Step 34: Select a starting solution from the high-quality virtual paths, and generate a decision instruction set around the starting solution that includes a specific scheduling scheme, backup solutions, and expected adjustment direction meta-instructions, as the scheduling anchor solution for the current cycle. The specific operations are as follows:

[0092] From the high-quality virtual paths in step 33, select the starting solution corresponding to the path. This solution is the unit in the current Pareto solution set that can stably evolve into the future solution set form. Based on this starting solution, firstly, formulate a specific machine tool loading and unloading scheduling plan for the current production cycle, clarifying the allocation logic of equipment, materials, and processes. At the same time, based on the neighborhood solutions of the starting solution in the trade-off feature space, determine several backup solutions to cope with small fluctuations in production. In addition, combined with the evolution direction of the virtual path, extract the subsequent adjustment direction meta-instructions of the scheduling plan, clarify the adjustment dimensions and trends when the state changes, and structurally integrate the specific scheduling plan, backup solutions, and expected adjustment direction meta-instructions to form a decision instruction cluster. This cluster is the scheduling anchor solution for the current cycle, which can guide the current production while maintaining consistency with future scheduling decisions.

[0093] In a preferred embodiment of the present invention, the disturbance rejection test module is further configured to perform the following steps:

[0094] Step 41: Parse the decision instruction cluster, identify the resource and process binding relationships common to all scheduling schemes, form resource and process cells, and connect the cells into a scheduling backbone network according to process constraints. The specific operations are as follows:

[0095] The decision instruction set contains the specific scheduling scheme, backup solution, and adjustment direction meta-instructions for the current cycle. First, the decision instruction set is structured and parsed to filter out all scheduling schemes, including the main scheme and backup solutions. These schemes all have inseparable resource and process binding relationships. These relationships are the foundation of the scheduling scheme and will not change due to backup solution switching or minor adjustments. Each such binding relationship is encapsulated into an independent resource and process cell. Each cell corresponds to a specific set of fixed matching relationships between equipment, materials, and other resources and corresponding processes. Subsequently, based on the established process constraints in the production process, such as the execution order of processes, parallel or serial constraints, and resource mutual exclusion constraints, all resource and process cells are connected according to the constraint logic to form a scheduling backbone network with hierarchical and dependent characteristics. This network constitutes the skeleton of subsequent resource allocation, ensuring that resource allocation meets the basic constraints of the production process.

[0096] Step 42: Calculate the multidimensional affinity between flexible processes and cells in the scheduling backbone network. Based on the multidimensional affinity, absorb the flexible processes into the corresponding cells to generate a tree-like resource allocation cluster. The specific operations are as follows:

[0097] Flexible processes refer to processes that are not strongly bound to specific resources and can adapt to multiple resource process cells. Their allocation flexibility is key to dealing with production disturbances. First, the multidimensional affinity between flexible processes and each resource process cell in the scheduling backbone network is calculated. This affinity is quantitatively evaluated from multiple dimensions, including process matching degree (the process adaptability between the flexible process and the processes within the cell), resource utilization rate (the efficiency of resource utilization after the flexible process is allocated to the cell), and time adaptability (the impact of the flexible process on the overall process cycle after being integrated into the cell). The dimension weights are dynamically adjusted according to production objectives, such as efficiency priority and cost priority. Based on the calculated multidimensional affinity values, each flexible process is adsorbed onto the resource process cell with the highest affinity. With the scheduling backbone network as the core trunk and the adsorbed flexible processes as branches and leaves, a tree-like resource allocation cluster with clear hierarchy and covering all scheduling processes and resources is finally formed. This structure retains the stability of the core scheduling logic while providing adjustment space brought by flexible processes.

[0098] Step 43: Perform virtual resource erosion on the tree-like resource allocation cluster, simulating the propagation process upstream and to branches and leaves after the selected cell fails, generating a failure propagation chain. The specific operations are as follows:

[0099] A virtual resource erosion operation is performed on the generated tree-like resource allocation cluster. This operation simulates the cell failure scenarios that may occur in actual production, such as equipment failure and resource shortage. First, the resource process cell to be tested is selected, usually the core cell in the backbone network is selected first, and it is marked as a failure state. Then, the propagation process of the failure state is tracked, and it is propagated upstream to the upstream resource process cells that depend on the output of the cell, and propagated to the branches and leaves to all flexible processes attached to the cell. The failure trigger time, impact degree and propagation path of each affected node are recorded. This propagation sequence and impact characteristics starting from the failure cell and covering all affected nodes are integrated into a failure propagation chain. This chain fully reflects the diffusion law of the failure impact within the tree-like resource allocation cluster after a single cell fails.

[0100] Step 44: Analyze the failure propagation chain, extract key nodes, average propagation depth, and structural entropy change indicators. Calculate the structural robustness score and collapse threshold based on these indicators. The specific operations are as follows:

[0101] First, core characteristic indicators are extracted from the failure propagation chain: Key nodes are resource process cells that play a pivotal role in the propagation process and trigger a wide-ranging chain effect after failure; their identification is based on the node's connectivity in the propagation chain and the number of nodes affected. Average propagation depth is the average level at which the failure impact spreads outward from the failed cell in the tree structure, reflecting the vertical penetration degree of the failure impact. The structural entropy change indicator is the change in structural disorder of the resource allocation cluster after the failure; a larger entropy change indicates a higher degree of structural disorder. Based on these indicators, the structural robustness score is calculated using the following formula: ;

[0102] This formula is abstracted from the structural stability logic of production systems. It uses average propagation depth and structural entropy change to reflect the negative impact of failure, and the proportion of critical nodes to reflect structural vulnerability. These are integrated into a comprehensive robustness score through weighted coefficients. This is the structural robustness score, ranging from 0 to 1, with higher values ​​indicating stronger robustness. For average propagation depth, This represents the highest level of the tree structure. It is an index of structural entropy change; This is the theoretical maximum value of entropy change; Number of key nodes; This represents the total number of nodes; , , The weighting coefficients are set based on the importance of different indicators in the production scenario, and satisfy the following conditions: For example, if a production scenario prioritizes scheduling stability, such as precision parts processing where the continuity of process steps is crucial, then the structural entropy change index, which reflects structural disorder, would have a higher weight and could be set as follows: Corresponding to average propagation depth, Corresponding to structural entropy change, This corresponds to the proportion of key nodes; if the production scenario prioritizes efficiency, such as mass production of standardized products, focusing on resource utilization efficiency and control of the scope of disturbance impact, then the weight of the average propagation depth, which reflects the degree of failure penetration, will be increased and can be set to... , , Based on the critical characteristics such as the proportion of failed nodes and the propagation depth when the resource allocation cluster loses its scheduling function in the failure propagation chain, a collapse threshold is determined. This threshold is the maximum failure impact boundary for the resource allocation cluster to maintain normal scheduling function.

[0103] Step 45: Combining structural robustness score, collapse threshold, and real-time load rate, generate disturbance tolerance margin, and mark key nodes on the tree-like resource allocation cluster to form a structural vulnerability map. The specific operations are as follows:

[0104] Disturbance resilience margin is an indicator that measures the maximum degree of disturbance that a resource allocation cluster can withstand. Its generation process requires combining multiple dimensions of parameters: the structural robustness score reflects the inherent failure resistance of the cluster, the collapse threshold reflects the critical boundary for maintaining the cluster's function, and the real-time load rate corrects the above two indicators. The higher the real-time load rate, the lower the resource redundancy and the weaker the disturbance resilience. Therefore, the robustness score and the collapse threshold need to be negatively corrected. The real-time load rate can be obtained by the ratio of the actual load of the current resource process cell to the rated load. By integrating the correlation between the corrected structural robustness score, the collapse threshold, and the real-time load rate, the disturbance resilience margin is finally generated. This margin value directly represents the upper limit of the disturbance intensity that the resource allocation cluster can withstand under the current production state. At the same time, the key nodes identified in step 44 are accurately marked at the corresponding positions in the tree-like resource allocation cluster, and the failure impact range and propagation characteristics of each key node are marked to form a structural vulnerability map. This map intuitively presents the weak links in the resource allocation cluster.

[0105] In a preferred embodiment of the present invention, the dynamic disturbance adjustment module is further configured to perform the following steps:

[0106] Step 51: Perform multi-dimensional quantitative analysis on the disturbance event and map it to the structural vulnerability map to identify the key affected nodes and load impacts. The specific operations are as follows:

[0107] Disturbance events encompass production anomalies such as sudden equipment failures, emergency order insertions, and process delays. First, a multi-dimensional quantitative analysis of the disturbance event is performed: the actual characteristics of the disturbance are transformed into standardized quantitative parameters based on the disturbance type (resource-based, process-based, order-based); the disturbance intensity (failure duration, number of inserted processes, and delay duration); and the scope of impact (the number of initial processes involved). These quantitative parameters are then mapped to a previously generated structural vulnerability map. Using the distribution information of key nodes already marked in the map, the key nodes directly affected by the disturbance event are matched. Simultaneously, the load impact on these nodes due to the disturbance is calculated. This impact represents the increase in node resource load or process execution pressure caused by the disturbance, and its value is determined comprehensively based on the disturbance intensity and the node's original load state. Through this process, the abstract impact of the disturbance event is transformed into the quantitative effect of specific nodes within the resource allocation cluster.

[0108] Step 52: Starting from the critical node, calculate the stress distribution according to the propagation mode, and predict cascading collapses based on the collapse threshold. The specific operations are as follows:

[0109] Starting with the critical nodes affected in step 51, and based on the failure propagation patterns obtained during the previous virtual resource erosion process, i.e. the dependencies and impact transmission logic between nodes, the stress distribution of the disturbance impact within the tree-like resource allocation cluster is calculated. This stress is a quantitative value of the failure risk borne by the node due to the disturbance impact and transmission effect. Its calculation requires combining the connection strength between nodes, resource redundancy, and the transmission coefficient of the disturbance load impact to ultimately form a stress distribution state covering the affected area and potential associated nodes. Subsequently, the stress value of each node is compared with the collapse threshold obtained in step 44. If the stress value of a node is close to or exceeds its collapse threshold, and the failure of that node will cause the stress of adjacent nodes to continue to rise through the propagation pattern, it can be predicted that the disturbance will trigger a chain collapse of the resource allocation cluster; otherwise, it is determined that the disturbance impact is within a controllable range.

[0110] Step 53: If no cascading collapse is triggered, perform a reverse resolution operation based on the stress distribution to reconstruct the resource allocation cluster. The specific operation is as follows:

[0111] When step 52 predicts that the disturbance will not trigger a chain reaction of collapse, a reverse resolution operation is performed based on the calculated stress distribution. This operation starts from the node with the highest stress value and traces its dependency relationship in the resource allocation cluster in reverse. It prioritizes adjusting the flexible processes associated with the node, such as transferring some flexible processes from the high-stress node to other cells with suitable affinity and lower stress, or adjusting the execution order of flexible processes to distribute the node load. At the same time, it fine-tunes the resource allocation ratio between the adjustment points and improves the resource redundancy of the high-stress node. Through this series of local adjustments, the stress value of each node is gradually reduced to a safe range. Finally, a tree-like resource allocation cluster adapted to the current disturbance state is reconstructed. This cluster retains the core structure of the original scheduling anchor solution and eliminates the load impact caused by the disturbance.

[0112] Step 54: If a cascading failure is to be triggered, the quantitative analysis results of the disturbance event, the corresponding failure propagation chain, and the information of the subgraph about to break identified from the tree-like resource allocation cluster are encapsulated to generate a rescheduling request. The specific operations are as follows:

[0113] When step 52 predicts that the disturbance will trigger a chain reaction of collapses, it indicates that the current resource allocation cluster can no longer maintain normal function through local adjustments. At this time, it is necessary to quantify and analyze the disturbance event, the failure propagation chain under the corresponding disturbance scenario (i.e., the impact propagation path after the failure of such nodes obtained from previous simulations), and the subgraph information about to break from the tree-like resource allocation cluster (i.e., the local structure that is about to lose function due to stress exceeding the threshold), and then encapsulate them in a structured manner. This encapsulated information fully contains the impact characteristics of the disturbance, the failure risk path, and the fragile structure of the current resource allocation. After integrating them into a standardized rescheduling request, the scheduling anchor solution determination module is triggered to run again, so that the module can recalculate the scheduling anchor solution for the current period based on this information to adapt to the production state after the disturbance.

[0114] In a preferred embodiment of the present invention, the feedback learning update module is further configured to perform the following steps:

[0115] Step 61: Align the actual production state sequence with the structural blueprint of the predicted future solution set form to generate the actual symbol trajectory, and compare it with the preferred virtual path to cut out the verified valid segments and empirical deviation segments. The specific operations are as follows:

[0116] The actual production state sequence is a continuous sequence formed by the solution set structure, resource allocation status, and operating parameters in each production cycle during the actual scheduling and execution process, arranged in chronological order. The structural blueprint for predicting the future solution set form is a symbolic topological description output by the solution set structure path prediction model. First, the actual production state sequence and the structural blueprint are aligned along the time dimension according to the production cycle. Based on the symbol encoding rules of the structural blueprint, the actual production state is transformed into a corresponding symbolic sequence, i.e., the actual symbolic trajectory, so that the actual production state and the predicted symbolic structure have a comparable carrier. Then, the actual symbolic trajectory is compared with the previously determined preferred virtual path on a cycle-by-cycle feature comparison. The part of the actual symbolic trajectory that completely matches the features of the preferred virtual path is identified and defined as a verified valid segment. Such segments can prove the actual adaptability of the corresponding prediction logic. At the same time, the parts where the actual trajectory deviates from the virtual path are separated as empirical deviation segments.

[0117] Step 62: Analyze the deviation patterns of the empirical deviation segments and correlate the deviation patterns with the recorded disturbance events and local stress distribution to establish a causal hypothesis chain between disturbance type, impact spectrum, and deviation patterns. The specific operations are as follows:

[0118] First, a structured analysis is performed on the empirical deviation segments to extract their deviation patterns. For example, the quantification parameters of core symbols may deviate from the preset range, the connection strength of relational symbols may not match the connection characteristics in actual production, or the evolution rate of symbol trajectories may differ from reality. This clarifies the type, degree, and distribution characteristics of the deviation in the symbol trajectory. Subsequently, disturbance events recorded within the corresponding production cycle of the deviation segment, such as sudden equipment failures, emergency order insertions, and process delays, as well as the local stress distribution data of the resource allocation cluster at that time, are retrieved. The deviation patterns are then correlated with these data to determine whether a specific type of disturbance will trigger a corresponding type of deviation, and whether the concentration of local stress distribution is related to the severity of the deviation. Based on the results of these correlation analyses, a disturbance type, the corresponding impact spectrum of the disturbance, i.e., the influence range and intensity distribution of the disturbance within the resource allocation cluster, and the causal hypothesis chain between the deviation patterns are established. This chain provides a clear causal indication for subsequent adjustments to the solution set fingerprint and prediction model.

[0119] Step 63: Based on the verified valid segments, the confidence of the fingerprints involved is enhanced, and the parameters of the fingerprints are fine-tuned or new candidate fingerprints are generated based on empirical bias segments and causal hypothesis chains to update the solution set fingerprint database. The specific operations are as follows:

[0120] For verified valid segments, their corresponding solution set fingerprints can accurately match the solution set structure features in actual production. Therefore, based on these segments, the confidence of the relevant solution set fingerprints is strengthened: specifically, the weight coefficient of the fingerprint in the solution set fingerprint library is increased, the priority of its selection in subsequent model training is increased, and its influence on the model prediction logic is strengthened. For empirical deviation segments, combined with the causal hypothesis chain established in step 62, the source of the deviation of the fingerprint parameter is analyzed: if the deviation is due to a small deviation of the fingerprint parameter, the corresponding quantization parameter of the solution set fingerprint is finely adjusted to better fit the structural features in actual production; if the deviation exceeds the parameter adjustment range of the existing fingerprint, or corresponds to a completely new solution set structure, a new candidate fingerprint is generated and added to the solution set fingerprint library. Through this process, the accuracy and coverage of the solution set fingerprint library are improved simultaneously.

[0121] Step 64: Based on the causal hypothesis chain, the parameters and rules of the competitive recombination mechanism and the local manifold learning mechanism in the solution set structure path prediction model are adjusted in a targeted manner. The specific operations are as follows:

[0122] The competitive recombination mechanism and the local manifold learning mechanism in the solution set structure path prediction model are the core modules that determine the prediction effect of the future solution set shape. Based on the causal hypothesis chain established in step 62, the parameters and rules of these two mechanisms are adjusted in a targeted manner: for example, if the causal hypothesis indicates that a certain type of disturbance will cause the window length of the local manifold learning to be insufficient and unable to fully capture the local evolution trend of the solution set structure, then the window length parameter of the local manifold learning is adjusted; if the candidate state generated by the competitive recombination mechanism under a specific disturbance scenario deviates significantly from the actual solution set structure, then the constraint rules in the recombination process are optimized, and the weight factor of the corresponding disturbance type is increased to make the recombination result more consistent with the evolution logic of the actual solution set. Through these targeted adjustments, the prediction model can better adapt to various disturbance scenarios in actual production and improve the accuracy and reliability of future solution set shape prediction.

[0123] The above description is merely a preferred embodiment of the present invention; however, the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and its improved concepts, should be covered within the scope of protection of the present invention.

Claims

1. A dynamic scheduling and decision optimization system for machine tool loading and unloading, characterized in that, include: The solution set fingerprint construction module is used to construct solution set fingerprints for the obtained Pareto solution set. It further includes: calculating the gradient direction of each solution in the Pareto solution set and mapping the Pareto solution set to a low-dimensional tradeoff feature space according to a preset gradient direction similarity criterion; identifying dense kernels and sparse paths connecting the dense kernels in the solution point cloud within the tradeoff feature space, and extracting the topological skeleton; encoding the dense kernels and sparse paths into a symbol set and assigning parameters to generate a structural blueprint; normalizing the geometric parameters in the structural blueprint based on the dynamic context of the production system, and encapsulating them to generate the solution set fingerprint. The path prediction model training module is used to train a solution set structure path prediction model based on historical solution set fingerprint sequences. It further includes: splitting the historical solution set fingerprint sequence into kernel trajectories and relation strength trajectories based on core symbol sets and relation symbol sets; learning the local tangent space and local manifold within local windows at each time point for the kernel trajectory to obtain manifold parameters; generating candidate states for each core symbol based on the manifold parameters, and competitively recombining the candidate states according to the connection relationships represented by the relation symbol set to synthesize future structure blueprint candidates; and encapsulating the mechanisms of trajectory splitting, manifold learning, and competitive recombining into a trainable and executable solution set structure path prediction model. The scheduling anchor solution determination module is used to determine the scheduling anchor solution for the current period based on the current Pareto solution set and the future solution set output by the prediction model. It also includes: calculating the distance between each solution in the current Pareto solution set and each structural symbol in the predicted future solution set in the weighted feature space, and generating a bidirectional soft correlation matrix; based on the soft correlation matrix, performing multi-step virtual deduction on each solution in the current Pareto solution set to generate virtual paths leading to different structural symbols in the predicted future solution set; calculating the path convergence entropy based on the changes in the attribution weights of each virtual path; simulating the application of parameter perturbations to test the ability of each path to maintain its evolution direction, thereby selecting high-quality virtual paths; selecting a starting solution from the high-quality virtual paths, and generating a decision instruction cluster around the starting solution containing specific scheduling schemes, backup solutions, and expected adjustment direction meta-instructions, as the scheduling anchor solution for the current period. The disturbance resilience testing module is used to form resource allocation clusters based on scheduling anchor solutions and to perform disturbance resilience testing on the resource allocation clusters to obtain their disturbance resilience margin. It also includes: analyzing the failure propagation chain, extracting key nodes, average propagation depth and structural entropy change indicators, and calculating structural robustness scores and collapse thresholds based on the indicators; combining the structural robustness scores, collapse thresholds and real-time load rates to generate disturbance resilience margins, and marking key nodes on the tree-like resource allocation clusters to form a structural vulnerability map. The dynamic disturbance adjustment module is used to determine whether the disturbance exceeds the disturbance immunity margin when it occurs. If it does not exceed the margin, adjustments are made within the resource allocation cluster. If it exceeds the margin, the scheduling anchor solution determination module is triggered to run again to determine a new scheduling anchor solution. The feedback learning update module is used to update the solution set fingerprint database and the solution set structure path prediction model based on actual scheduling feedback.

2. The machine tool loading and unloading dynamic scheduling and decision optimization system according to claim 1, characterized in that, The dynamic disturbance adjustment module is also used for: The disturbance events are analyzed in a multidimensional quantitative manner and mapped to a structural vulnerability map to identify the key affected nodes and load impacts. Starting from critical nodes, stress distribution is calculated based on propagation patterns, and cascading collapses are predicted based on collapse thresholds.

3. The machine tool loading and unloading dynamic scheduling and decision optimization system according to claim 2, characterized in that, The dynamic disturbance adjustment module is also used for: If no cascading collapse is triggered, a reverse resolution operation is performed based on the stress distribution to reconstruct the resource allocation cluster; If a chain reaction of crashes is triggered, the quantitative analysis results of the disturbance event, the corresponding failure propagation chain, and the information of the subgraph about to break identified from the tree-like resource allocation cluster will be encapsulated to generate a rescheduling request.

4. The machine tool loading and unloading dynamic scheduling and decision optimization system according to claim 1, characterized in that, The feedback learning update module is also used for: Align the actual production state sequence with the structural blueprint of the predicted future solution set form to generate actual symbol trajectories, and compare them with the preferred virtual path to cut out verified effective segments and empirical deviation segments. The deviation patterns of empirical deviation segments are analyzed, and the deviation patterns are correlated with recorded disturbance events and local stress distributions to establish a causal hypothesis chain between disturbance type, impact spectrum and deviation patterns.

5. The machine tool loading and unloading dynamic scheduling and decision optimization system according to claim 4, characterized in that, The feedback learning update module is also used for: The confidence level of the fingerprints involved is enhanced based on the verified valid fragments, and the parameters of the fingerprints are fine-tuned or new candidate fingerprints are generated based on empirical bias fragments and causal hypothesis chains to update the solution set fingerprint library. Based on the causal hypothesis chain, the parameters and rules of the competitive recombination mechanism and the local manifold learning mechanism in the solution set structure path prediction model are adjusted in a targeted manner.

Images