Severity perception layering system oriented to dynamic flexible job shop and scheduling method of severity perception layering system

Abstract

The invention belongs to the technical field of workshop scheduling, and particularly relates to a severity perception layering system facing a dynamic flexible job workshop and a scheduling method thereof. The system comprises a planning layer, a strategy layer and a meta-control layer. The planning layer is responsible for generating a high-quality initial scheduling scheme and constructing an elite knowledge base by adopting a self-adaptive genetic algorithm Q-AGA guided by Q learning; the strategy layer comprises a lightweight rapid repair strategy and a heavy optimization strategy; the meta-control layer is responsible for solving an algorithm selection problem on line: when an uncertain event occurs, selecting a rescheduling strategy most suitable for a current scene from a strategy library; the meta-control layer comprises a severity index feature engineering module, a linear decision model module and a threshold decision and calibration module; according to the method, the optimization efficiency and the system performance of the severity perception layering system are improved, and the requirements of a modern industrial automatic dispatching system for the real-time performance, the stability and the accuracy are met.

Description

Technical Field

[0001] This invention belongs to the technical field of workshop scheduling, and more specifically, relates to a severity perception hierarchical system and scheduling method for dynamic flexible work workshops. Background Technology

[0002] The Flexible Job Shop Scheduling (FJSP) problem is a core technological bottleneck in intelligent manufacturing during the Industry 5.0 era. In real-world manufacturing environments, companies typically experience multiple production disruptions each week, including equipment failures, urgent orders, and raw material shortages. These disruptions cause statically optimized solutions to deviate significantly from expectations during actual execution. Traditional static scheduling methods cannot handle this dynamism, while existing dynamic rescheduling methods face a fundamental "quality-efficiency-stability dilemma": high-quality global optimization algorithms are computationally expensive and significantly disrupt the original plan; while rapid repair rules ensure efficiency and stability, the solution quality deteriorates drastically under severe failures.

[0003] The core challenges of the Dynamic Flexible Job Shop Scheduling (DFJSP) problem mainly include three progressively layered technical issues: How to generate a high-quality, robust initial scheduling scheme? How to design a rescheduling algorithm that can efficiently cope with severe failures? How to construct an intelligent meta-control layer to achieve adaptive strategy selection under different disturbances? Most existing technologies focus on single-level improvements, lacking integrated solutions for multi-level problems, and even more so, lacking formal disturbance evaluation and algorithm selection mechanisms. Summary of the Invention

[0004] The present invention aims to overcome at least one of the defects of the prior art and provide a severity perception hierarchical system for dynamic flexible work workshops.

[0005] This invention also discloses a method for scheduling workshops using the aforementioned system. The complex dynamic scheduling problem is modeled as an "online algorithm selection problem." A three-layer collaborative architecture is constructed using a "hierarchical meta-control" design concept: a planning layer employing a Q-learning-guided adaptive genetic algorithm (Q-AGA) to model parameter control as a Markov decision process; a policy layer using an event-driven elite re-evolutionary algorithm (ERRE) to achieve knowledge inheritance based on the solution structure robustness assumption; and a meta-control layer using a lightweight surrogate function based on the severity index (SI) to achieve real-time policy selection. This improves the optimization efficiency and system performance of the severity-aware hierarchical system, meeting the real-time, stability, and accuracy requirements of modern industrial automation scheduling systems.

[0006] The detailed technical solution of this invention is as follows: A severity perception hierarchical system for dynamic flexible workshops, comprising a planning layer, a strategy layer, and a meta-control layer, is formally defined as follows: In this context, G represents the planning layer; P represents the strategy layer; M represents the meta-control layer; and K represents the knowledge base, which is a shared knowledge base that stores elite solution structures. The planning layer G is responsible for generating a high-quality initial scheduling scheme using the Q-learning-guided adaptive genetic algorithm Q-AGA. And building an elite knowledge base K; The strategy layer P is a reactive strategy library, including lightweight and fast repair strategies. and heavyweight optimization strategies The lightweight and fast repair strategy This refers to the right-shift rescheduling strategy, representing a traditional heuristic strategy that pursues the highest stability and speed; the aforementioned heavyweight optimization strategy This refers to a rescheduling strategy based on knowledge reuse; The meta-control layer M is the decision-making core of the system, responsible for solving the algorithm selection problem online: when uncertain events (such as machine failures) occur, it quickly and online retrieves algorithms from the policy library with extremely low computational cost. The most suitable rescheduling strategy for the current scenario is selected; the meta-control layer includes a feature engineering module for severity index, a linear decision model module, and a threshold decision and calibration module. The feature engineering module of the severity index is responsible for constructing various features used to calculate the severity index SI, forming a three-dimensional feature vector; The linear decision model module is used to fuse the three-dimensional feature vectors extracted by the feature engineering module of the severity index into a single scalar value that can be used for decision-making. The threshold decision and calibration module is used to select strategies and calibrate thresholds to determine the optimal threshold.

[0007] According to a preferred embodiment of the present invention, the three-dimensional feature vector is: , by characteristics Time pressure, characteristics Structural disturbances and characteristics Key components of resources; The features Time pressure is as shown in formula (1): (1) In equation (1), Indicates the fault recovery time. Indicates the fault start time. Indicates workpiece The One process, This represents the set of processes that are directly affected. Indicates workpiece The Processing time of each step; molecules The denominator represents the duration of the failure, indicating the time window available for production recovery; the shorter this time, the more urgent the situation. This represents the average expected processing time of those processes directly affected by the malfunctioning machine, and serves as a benchmark for measuring the recovery workload. The features The structural disturbance is as shown in formula (2): (2) In equation (2), This represents the set of all affected processes. This indicates the preceding closure of the process that is directly affected; The features Resource criticality is as shown in formula (3): (3) In equation (3), Indicates a faulty machine. Indicates a faulty machine Total planned workload Represents the set of all machines. This represents the average workload across all machines.

[0008] According to a preferred embodiment of the present invention, the step of fusing the three-dimensional feature vectors extracted by the feature engineering module of the severity index into a single, decision-oriented scalar value SI specifically includes using the three-dimensional feature vectors... Define a linear model: (4) In equation (4), Represents the weight vector, weight vector Determined using the Analytic Hierarchy Process (AHP).

[0009] According to a preferred embodiment of the present invention, the strategy selection through a threshold means defining a decision rule as shown in formula (5): (5) In equation (5), Threshold; Determining the optimal threshold through threshold calibration refers to setting an optimization target. As in formula (6): (6) In equation (6), It is an asymmetric loss function. Indicates the optimal strategy label. Indicates the fault scenario The three-dimensional feature vectors below; The detailed threshold calibration process is as follows: S1. Data preparation: For each fault scenario in the dataset... The following three pieces of information need to be prepared: ① Three-dimensional feature vector ; ② Severity index ③ Optimal strategy label For fault scenarios Execute offline respectively and Two strategies are employed to obtain rescheduling schemes and record performance metrics; the strategy with the better performance metrics is selected as the optimal strategy; through this process, each fault scenario is centralized in the dataset. They all got one Data pairs; S2. Quantization definition of asymmetric loss function The loss function receives two inputs: the policy predicted by the model and the actual optimal policy, and then outputs a cost value, as shown in formula (7): (7) In equation (7), This refers to the cost of false alarms; This refers to the cost of underreporting, the ratio. This reflects the degree of aversion to the risk of underreporting; S3. Search and Optimization: Based on the prepared data and loss function, a grid search method is used to find the optimal threshold. .

[0010] A scheduling method for dynamic flexible workshops, the method comprising: an offline preparation phase and an online response phase; The offline preparation phase includes: firstly, using the Q-learning-guided adaptive genetic algorithm Q-AGA to generate an initial scheduling scheme. Simultaneously, a knowledge base K containing elite solution structures is constructed; the Q-learning-guided adaptive genetic algorithm Q-AGA models the parameter control of the genetic algorithm GA as a Markov decision process, and uses the Q-learning reinforcement learning algorithm to achieve intelligent parameter adaptation. Then, the meta-control layer uses the Analytic Hierarchy Process (AHP) to complete the parameter calibration and perform threshold calibration to obtain the optimal threshold. The online response phase includes: firstly, fault detection, real-time monitoring of production disturbances (such as equipment failures, urgent orders, etc.), and recording of fault events. ,in: Indicates a faulty machine. Indicates the fault start time. Indicates the fault recovery time; Then, based on time pressure, structural disturbances, and resource criticality, a severity index (SI) is obtained through a linear decision model. Strategy selection is then based on the SI, comparing it with the optimal threshold: if SI < optimal threshold, a lightweight, rapid repair strategy is selected. If SI Optimal threshold, select heavyweight optimization strategy ; Finally, the status is updated after the solution is generated.

[0011] According to a preferred embodiment of the present invention, the Markov decision process is modeled as follows: The parameter control problem of Q-AGA is modeled as a Markov decision process quadruple. Where S represents the state space, A represents the action space, and R represents the reward function. Represents the discount factor; more specifically, the state space in the th... At that time, the environmental state observed by the Agent It is a two-dimensional vector, including the evolutionary stage and the degree of stagnation, with a total of 9 discrete states, formally represented as formula (8): (8) In equation (8), It refers to the stage of evolution. This refers to the degree of stagnation; The action space is designed with N predefined discrete crossover probabilities. With the probability of mutation Parameter combination ; The reward function is as shown in formula (9): (9) In equation (9), This refers to the optimal solution of the previous generation. It refers to the optimal solution of the new generation.

[0012] According to a preferred embodiment of the present invention, the workflow of the Q-learning-guided adaptive genetic algorithm Q-AGA is as follows: S1. Use application examples of flexible work workshops as algorithm input; S2. Use the MSOS hierarchical encoding method to encode the chromosome of the instance; S3. Use the GLR strategy for population initialization; S4. Begin the evolutionary operation generation by generation until the maximum number of iterations designed by the algorithm is reached. Steps S5-S8 are the specific steps of one evolution. S5. State awareness, calculating the current evolutionary stage. and degree of stagnation Combined into the current environmental state ; S6. Action selection: Generate a random number rand between (0,1). If rand is less than a preset threshold... Randomly select a group from the action space As an action If rand is greater than or equal to the preset threshold Then look up the current state in the Q table. For the corresponding row, select the action with the largest Q value in that row as... ,Right now ; S7. Perform the GA operation and select the action. Apply to the current population Perform standard genetic algorithm operations: selection, crossover, and mutation to generate the next generation population. ; S8. Reward calculation and state transition, evaluating the new population. To obtain the target value of its optimal solution The optimal value of the previous generation Compare the results and calculate the immediate reward based on the reward function. Then the system transitions to the next state. ; S9. Update the Q table. After S8, a complete transfer sample is obtained. Update the Q-table according to the core update formula of Q-learning. The value; S10. Determine if the current generation is the last generation. If so, save the initial scheduling scheme and the elite population.

[0013] According to a preferred embodiment of the present invention, the lightweight and rapid repair strategy This refers to the right-shift rescheduling strategy, the heavyweight optimization strategy. This refers to a rescheduling strategy based on knowledge reuse; The knowledge reuse-based rescheduling strategy The event-driven elite re-evolution algorithm ERRE is adopted. The detailed steps of the ERRE algorithm are as follows: S1. Input parameters: ① Fault events including faulty machine, fault start time and end time, ② Current system status, ③ Knowledge base storing historical elite solutions, ④ Genetic algorithm parameter settings; S2. First stage: Identify the affected domain, specifically including: ① Recording fault information, ② Identifying interrupted operations, ③ Identifying planned operations, ④ Identifying continuous operations, and ⑤ Calculating the set of affected operations. S3, Second Phase: Constructing a world snapshot, specifically including: ① Determining fixed operations, ② Handling residual processing time, ③ Constructing a rescheduling context; S4, Third Stage: Call the rescheduling engine, which includes: ① Obtain the elite population as the current population, ② Repair the elite chromosomes, ③ Initialize the optimal solution, ④ Iterative optimization, with each iteration consisting of steps ⑤-⑨ under step S4: ⑤ Fitness evaluation, ⑥ Update the optimal solution, ⑦ Select elites, ⑧ Generate new individuals, ⑨ Update the population, ⑩ Obtain the optimal solution after all iterations are completed, i.e., obtain the new rescheduling scheme; S5, Fourth Stage: Integrating the new scheduling scheme, specifically including: ① Cleaning the event queue, ② Adding repair completion events, ③ Using context-aware reactive decoding to generate detailed scheduling, the specific decoding process is ④-⑦ under S5, ④ Resetting internal state, extracting fault parameters, setting global constraints, ⑤ Loading context: handling fixed operations, adding fault constraints, ⑥ Decoding new operations: parsing chromosomes, scheduling jobs in sequence, ⑦ Calculating maximum completion time, ⑧ Updating the event queue, ⑨ Returning to the updated status.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) This invention is the first to model dynamic rescheduling as an online algorithm selection problem and proposes a formal hierarchical meta-control paradigm, breaking through the limitations of the traditional "one-size-fits-all" approach. The three-layer architecture collaborative design realizes a complete closed loop from offline planning to online decision-making: Q-AGA's reinforcement learning-driven parameter control, ERRE's knowledge inheritance mechanism, and SI's lightweight decision-making constitute an intelligent system that guarantees both performance and real-time performance.

[0015] (2) The Q-AGA algorithm proposed in this invention introduces reinforcement learning to dynamically and adaptively adjust the parameters of the meta-control layer. By designing a compact state space representation method and a hierarchical action space structure, it effectively alleviates the curse of dimensionality and significantly improves search efficiency and policy generalization ability.

[0016] (3) The ERRE algorithm proposed in this invention adopts the strategy of "inheriting the skeleton and optimizing resources" and only reconstructs the machine selection process. It is significantly different from the extreme paths of "complete rescheduling" or "pure rule repair" in the present invention, and achieves an excellent balance of quality, efficiency and stability. Attached Figure Description

[0017] Figure 1 This is a flowchart of the method for scheduling workshops using the SAHF system as described in this invention.

[0018] Figure 2 This is a flowchart of the Q-AGA algorithm described in this invention.

[0019] Figure 3This is the flowchart of the ERRE algorithm described in the present invention.

[0020] Figure 4 This is the performance comparison chart of the average Cmax under different faults in Embodiment 3 of the present invention.

[0021] Figure 5 This is the box plot of the discrimination degree of the SI index of RSR and ERRE in Embodiment 4 of the present invention.

[0022] Figure 6 This is the performance comparison chart of the AGA and Q-AGA algorithms in Embodiment 5 of the present invention.

[0023] Figure 7 This is the comparison chart of the convergence curves of AGA and Q-AGA on the box plot of the discrimination degree of the SI index of RSR and ERRE in the Mk15 instance in Embodiment 5 of the present invention. Detailed implementation manners

[0024] The following further describes the present disclosure in conjunction with the drawings and embodiments.

[0025] Embodiment 1 Refer (It seems there is something missing here. The text is incomplete. But I'll translate as it is for now.) Figure 1 , this embodiment provides a severity-aware hierarchical system for a dynamic flexible job shop, including a planning layer, a policy layer, and a meta-control layer, which is formally defined as: , where G represents the planning layer; P represents the policy layer; M represents the meta-control layer; K represents the knowledge base, which is a shared knowledge base storing elite solution structures; The core function of the SAHF is the mapping function , and the execution process is divided into two stages: (1) intelligent policy selection, which is expressed by the formula: ; (2) policy execution, which is expressed by the formula: ; The planning layer G is responsible for generating a high-quality initial scheduling plan and constructing an elite knowledge base K using the adaptive genetic algorithm Q-AGA guided by Q learning; The policy layer P is a reactive policy library, including a lightweight fast repair policy and a heavyweight optimization policy ; The lightweight fast repair policy refers to the right-shift rescheduling policy, representing a traditional heuristic policy that pursues the highest stability and speed; The heavyweight optimization policy refers to the rescheduling policy based on knowledge reuse; The meta-control layer M is the decision-making core of the system, responsible for online solving the algorithm selection problem: when an uncertain event (such as a machine failure) occurs, it can quickly and online select from the policy library at a very low computational cost The most suitable rescheduling strategy for the current scenario is selected; the meta-control layer includes a feature engineering module for severity index, a linear decision model module, and a threshold decision and calibration module. The feature engineering module of the severity index is responsible for constructing the various features (indicators) used to calculate the severity index SI. More specifically, based on principal component analysis and correlation analysis, this invention identifies the three most critical dimensions affecting the selection of rescheduling strategies, including features. Time pressure, characteristics Structural disturbances and characteristics Resource criticality constitutes a three-dimensional feature vector. The features Time pressure is as shown in formula (1): (1) In equation (1), Indicates the fault recovery time. Indicates the fault start time. Indicates workpiece The One process, This represents the set of processes that are directly affected. Indicates workpiece The Processing time of each step; molecules The denominator represents the duration of the failure, indicating the time window available for production recovery; the shorter this time, the more urgent the situation. This represents the average expected processing time of those processes directly affected by the malfunctioning machine, and serves as a benchmark for measuring the recovery workload. The features The structural disturbance is as shown in formula (2): (2) In equation (2), This represents the set of all affected processes. This indicates the preceding closure of the process that is directly affected; The features Resource criticality is as shown in formula (3): (3) In equation (3), Indicates a faulty machine. Indicates a faulty machine Total planned workload Represents the set of all machines. This represents the average workload across all machines; The linear decision model module is used to process the three-dimensional feature vector extracted by the feature engineering module of the severity index. The values ​​are integrated into a single, decision-making scalar value. In online decision-making scenarios, the computational speed of the model is crucial. Therefore, to avoid high online inference overhead, this invention prefers a linear model rather than a complex nonlinear model (such as a neural network or support vector machine). The linear model is as shown in formula (4): (4) In equation (4), Represents the weight vector, weight vector Determined using the analytic hierarchy process; The weight vector The determination of the weights is based on the Analytic Hierarchy Process (AHP). AHP allows the systematic and quantifiable integration of domain experts' experiential knowledge into the model. Experts can make pairwise comparisons based on their judgments of the relative importance of "time pressure," "structural disturbance," and "resource criticality," and AHP then uses mathematical methods to calculate the implicit weight vector. This allows the model's weights to be not only data-driven but also knowledge-driven, enhancing the model's robustness and credibility.

[0026] The threshold decision and calibration module mainly includes threshold decision function and threshold calibration function: The decision rule for the threshold decision function is as shown in formula (5): (5) In equation (5), Threshold; The threshold calibration function is completed offline, and its essence is a data-driven supervised learning and optimization process; in this threshold calibration process, the present invention sets an optimization objective. As in formula (6): (6) In equation (6), It is an asymmetric loss function. Indicates the optimal strategy label. Indicates the fault scenario The three-dimensional feature vectors below; The detailed threshold calibration process is as follows: S1. Data preparation: For each fault scenario in the dataset (referred to as a sample)... The following three pieces of information need to be prepared: ① Feature vector For this fault scenario The corresponding three-dimensional feature vector is calculated based on its specific parameters (faulty machine, start and end time, etc.). ; ② Severity index Using the weight vectors already determined by methods such as AHP The severity index of the scenario was calculated. ③ Optimal strategy label For this fault scenario Two strategies are executed offline: First, a lightweight strategy is run once. This yields a rescheduling scheme, and its performance metrics (e.g., maximum completion time) are recorded. Then, run the heavyweight strategy once. This yields another rescheduling scheme, and its performance metrics (e.g., maximum completion time) are recorded. Finally, compare the two results; if The performance is better than or close to Then the "optimal strategy" for this scenario is marked as Conversely, if If a significantly better result is obtained, then the optimal strategy is marked as... Through this process, each sample in the dataset They all have one Data pairs; S2. Quantization definition of asymmetric loss function The loss function takes two inputs: the policy predicted by the model and the actual optimal policy, and then outputs a cost (loss), as shown in formula (7): (7) In equation (7), This refers to the cost of false alarms, treating minor faults as major ones. It's a relatively small penalty, for example... ; This refers to the cost of underreporting, treating severe faults as minor ones, which is far greater than... The penalty value, for example or This ratio ( This directly reflects the degree of aversion to the risk of underreporting; S3. Search and Optimization: Based on data and loss function, this invention employs the most direct grid search method to find the optimal threshold. Specifically, this includes: ① Determining the search scope. First, iterate through the entire prepared dataset and find the minimum severity index of all samples. and maximum value Optimal threshold It must fall within this range; ② Create candidate thresholds, in Within a certain range, a series of discrete candidate thresholds are generated. For example, a value can be taken at very small intervals (such as 0.01) to form a set. ③ Iterate through and evaluate each candidate threshold. For each candidate threshold... Perform the following operations: a. Initialize a total loss Total_Loss of 0; b. Iterate through all samples in the dataset. Based on the current candidate threshold and samples Make a prediction, if ,So ;if ,So Obtain the true optimal policy label for this sample. ,use Calculate the loss for this sample: Add this loss to the total loss: c. After iterating through all data samples, Total_Loss becomes the candidate threshold. The total cost across the entire dataset; ④ Finding the optimal threshold, after evaluating all candidate thresholds, obtaining each The candidate threshold that minimizes the total loss is the optimal threshold we are looking for. .

[0027] Example 2 This embodiment provides a method for scheduling a workshop using the above system, the method including: an offline preparation phase and an online response phase; The offline preparation phase includes: firstly, using a Q-learning-guided adaptive genetic algorithm (Q-AGA) at the planning layer to generate an initial scheduling scheme. Simultaneously, a knowledge base K containing elite solution structures is constructed; Then, the meta-control layer uses the Analytic Hierarchy Process (AHP) to complete the parameter calibration and perform threshold calibration to obtain the optimal threshold. The online response phase includes: firstly, fault detection, real-time monitoring of production disturbances (such as equipment failures, urgent orders, etc.), and recording of fault events. ,in: Indicates a faulty machine. Indicates the fault start time. Indicates the fault recovery time; Then, based on time pressure, structural disturbances, and resource criticality, a severity index (SI) is obtained through a linear decision model. Strategy selection is then performed based on the SI, and the SI is compared with the optimal threshold: if SI... Optimal threshold, select lightweight and fast repair strategy If SI Optimal threshold, select heavyweight optimization strategy ; Finally, the status is updated after the solution is generated.

[0028] The Q-learning-guided adaptive genetic algorithm Q-AGA models the parameter control of the genetic algorithm GA as a Markov decision process and uses the Q-learning reinforcement learning algorithm to achieve intelligent parameter adaptation. The Markov decision process is formally modeled as follows: The parameter control problem of Q-AGA is modeled as a Markov decision process quadruple. Where S represents the state space, A represents the action space, and R represents the reward function. Represents the discount factor; more specifically, the state space in the th... At that time, the environmental state observed by the Agent It is a two-dimensional vector, including the evolutionary stage and the degree of stagnation, with a total of 9 discrete states, formally represented as formula (8): (8) In equation (8), It refers to the stage of evolution. This refers to the degree of stagnation; The action space is designed with N predefined discrete crossover probabilities. With the probability of mutation Parameter combination For example, N=6; specifically, this N parameter combination conforms to the following strategy: from high exploration to high utilization, such as: ① parameter combination (0.9, 0.01), the initial exploration stage with high crossover and very low mutation; ② parameter combination (0.9, 0.05), the balanced strategy stage; ③ parameter combination (0.8, 0.05), the mid-stage exploration stage with medium crossover and medium mutation; ④ parameter combination (0.7, 0.1), the late-stage escape stage with low crossover and high mutation; ⑤ parameter combination (0.6, 0.2), the severe stagnation stage with very low crossover and very high mutation; ⑥ parameter combination (0.95, 0.005), the super exploration stage; The reward function is as shown in formula (9): (9) In equation (9), This refers to the optimal solution of the previous generation. It refers to the optimal solution of the new generation.

[0029] The workflow of the Q-AGA algorithm is as follows: Figure 2 As shown, the specific description is as follows: S1. First, use application examples of flexible work workshops as algorithm input; S2. Chromosome encoding is performed on the instance. Specifically, the MSOS hierarchical encoding method is adopted. The specific execution steps of the MSOS encoding method are not within the scope of protection of this invention. They are encoding methods commonly used in the field and will not be described in detail here. S3. Population initialization is performed using the GLR strategy. The GLR strategy is a commonly used method in the field and is not within the scope of protection of this invention. It will not be described in detail here. S4. Begin the evolutionary operation generation by generation until the maximum number of iterations designed by the algorithm is reached. Steps S5-S8 are the specific steps of one evolution. S5. State awareness, calculating the current evolutionary stage. and degree of stagnation Combined into the current environmental state ; S6. Action selection: Generate a random number rand between (0,1). If rand is less than a preset threshold... Randomly select a group from the action space As an action If rand is greater than or equal to the preset threshold Then look up the current state in the Q table. For the corresponding row, select the action with the largest Q value in that row as... ,Right now ; S7. Perform the GA operation and select the action. Apply to the current population Perform standard genetic algorithm operations: selection, crossover (using...) ), mutation (using) ), generating the next generation of population ; S8. Reward calculation and state transition, evaluating the new population. To obtain the target value of its optimal solution The optimal value of the previous generation Compare the results and calculate the immediate reward based on the reward function. Logically, the system has transitioned to the next state. (This state will be formally calculated at the start of the next cycle); S9. Update the Q table. After S8, a complete transfer sample is obtained. Update the Q-table according to the core update formula of Q-learning. The value; S10. Determine if the current generation is the last generation. If so, save the initial scheduling scheme and the elite population.

[0030] The lightweight and fast repair strategy This refers to the right-shift rescheduling strategy, representing a traditional heuristic strategy that pursues the highest stability and speed; the aforementioned heavyweight optimization strategy This refers to a rescheduling strategy based on knowledge reuse; The knowledge reuse-based rescheduling strategy The event-driven elite re-evolution algorithm ERRE is employed. ERRE is designed based on a "solution structure robustness assumption," which states that "in high-quality scheduling solutions, the operation sequencing (OS) component is more structurally robust than the machine selection (MS) component; specifically, the priority relationships between workpieces determined by the OS remain relatively stable under local machine failures, while the MS component is more directly dependent on the machine state and more sensitive to failures." Based on this assumption, ERRE adopts a "reinherit skeleton, optimize resources" strategy: it retains the OS component of the elite solution as the optimization skeleton, and only re-optimizes the MS component to adapt to resource constraints after failures; for example... Figure 3 As shown, the detailed algorithm steps are as follows: S1. Input parameters: ① Fault events including faulty machine, fault start time and end time, ② Current system status, ③ Knowledge base storing historical elite solutions, ④ Genetic algorithm parameter settings; S2. First stage: Identify the affected domain, specifically including: ① Recording fault information, ② Identifying interrupted operations, ③ Identifying planned operations, ④ Identifying continuous operations, and ⑤ Calculating the set of affected operations. S3, Second Phase: Constructing a world snapshot, specifically including: ① Determining fixed operations, ② Handling residual processing time, ③ Constructing a rescheduling context; S4, Third Stage: Call the rescheduling engine, which includes: ① Obtain the elite population as the current population, ② Repair the elite chromosomes, ③ Initialize the optimal solution, ④ Iterative optimization, with each iteration consisting of steps ⑤-⑨ under step S4: ⑤ Fitness evaluation, ⑥ Update the optimal solution, ⑦ Select elites, ⑧ Generate new individuals, ⑨ Update the population, ⑩ Obtain the optimal solution after all iterations are completed, i.e., obtain the new rescheduling scheme; S5, Fourth Stage: Integrating the new scheduling scheme, specifically including: ① Cleaning the event queue, ② Adding repair completion events, ③ Using context-aware reactive decoding to generate detailed scheduling, the specific decoding process is ④-⑦ under S5, ④ Resetting internal state, extracting fault parameters, setting global constraints, ⑤ Loading context: handling fixed operations, adding fault constraints, ⑥ Decoding new operations: parsing chromosomes, scheduling jobs in sequence, ⑦ Calculating maximum completion time, ⑧ Updating the event queue, ⑨ Returning to the updated status.

[0031] Example 3 This embodiment compares the performance of the SAHF system described in this invention with the maximum makespan of pure RSR and pure ERRE under a fixed strategy, as detailed below: This embodiment extracts 35 test instances of different sizes from four classic FJSP benchmark datasets and compares the makespan performance of SAHF with fixed-strategy pure RSR and pure ERRE under three fault scenarios: mild, moderate and severe. All results are the statistical mean of 20 runs. Experimental results are as follows Figure 4 As shown, under mild faults, RSR has a significant advantage (440.11 vs 459.53, an improvement of 4.41%), under moderate faults, the two strategies are comparable (ERRE is slightly better by 1.67%), and under severe faults, ERRE is significantly superior (523.49 vs 617.98, an improvement of 15.29%). SAHF achieves intelligent selection by comparing the SI value with a threshold τ*=0.75 (this value is based on offline experimental calibration): RSR is selected when SI<0.75, and ERRE is selected when SI≥0.75; Figure 4 The experiments shown demonstrate that the optimal strategy is selected in all scenarios, avoiding the performance loss of a single strategy.

[0032] Example 4 This embodiment aims to further verify the SAHF intelligent decision-making mechanism by evaluating the performance of the meta-controller in 1000 fault scenario tests. Details are as follows: This embodiment designs a controlled disturbance generation model, controlling three types of fault scenarios—mild, moderate, and severe—based on machine load category (high / medium / low load) and fault duration ([0.5-1.5] average processing time / [2.0-4.0] average processing time / [5.0-8.0] average processing time). Based on the disturbance generation model, 1000 fault scenarios are preset to deeply evaluate the performance of the meta-control layer. Experimental results show a decision accuracy of 89.5%, an RSR recall of 98%, and an ERRE precision of 100%. Figure 5 As shown, the Severity Index (SI) has a strong ability to distinguish the severity of faults: the SI values ​​for minor faults are concentrated in the range of 0.4-0.6 (μ=0.483±0.029), while the SI values ​​for severe faults are distributed in the range of 1.0-1.7 (μ=1.398±0.346), and the Cohen's d effect size reaches 2.716, proving that the SI function can effectively distinguish fault scenarios of different severity levels.

[0033] Example 5 This embodiment verifies the effectiveness of the Q-AGA initial scheduling method described in this invention on a selected benchmark dataset, as detailed below: like Figure 6As shown, the performance of the standard Adaptive Genetic Algorithm (AGA) and the Q-AGA algorithm described in this invention are compared. The experiment sets up Q-AGA and AGA to share the same encoding and genetic operators, differing only in parameter control strategies. Six representative instances were selected, covering different scales from medium to large. The experimental results are as follows: Figure 6 As shown, Q-AGA outperforms AGA on all six instances, with the most significant improvements observed in Mk04 and Mk06 instances, reaching 6.06% and 5.71% respectively; it also performs exceptionally well on large-scale instances of 50×60 and 100×60. To ensure the reliability of the experimental results, all instances were run 20 times independently. Appropriate statistical tests (t-test or Wilcoxon test) were selected based on the data distribution characteristics. The test results showed that Q-AGA achieved statistically significant improvements (p<0.05) on all 6 instances, with an average performance improvement of approximately 3.4%. like Figure 7 The results shown are the convergence analysis results of the two algorithms evaluated on the Mk15 instance. AGA gets stuck in a local optimum after about 65 generations, while Q-AGA maintains its continuous optimization ability until about 235 generations through intelligent parameter adjustment of RL, and finally obtains a better solution. In summary, the experimental results of this embodiment are as follows: Q-AGA achieves performance improvements on all six test instances, with improvements ranging from 0.69% to 6.06%, and all improvements are statistically significant. The convergence curves verify that Q-AGA effectively avoids premature convergence through intelligent parameter control via RL, providing a high-quality initial scheme and knowledge base for the framework.

[0034] Example 6 This embodiment verifies the effectiveness of the ERRE rescheduling method described in this invention on a selected benchmark dataset, as detailed below: To verify the superiority of the ERRE algorithm described in this invention under moderate and severe faults, the performance of ERRE, Complete Rescheduling (CRS), and Right Shift Repair (RSR) will be compared under moderate and severe faults. Tables 1 and 2 show the performance comparison of RSR, CRS, and ERRE under moderate and severe faults, respectively, where bold indicates the optimal value and black dots indicate the suboptimal value. As can be seen from Tables 1 and 2, in the six application instances of different scales evaluated, ERRE demonstrates the best or suboptimal performance in terms of average Cmax (minimum maximum completion time). Results: Among the suboptimal examples, the best performance was achieved under the CRS strategy. However, the average runtime of the CRS strategy was 59 times (under moderate failure) and 56 times (under severe failure) of ERRE. Overall, ERRE is still superior. Although RSR has the best average time under moderate and severe failures, in practical applications, minimizing the maximum completion time is more important. Therefore, RSR is also inferior to ERRE. In summary, the performance differences of the three strategies under different scenarios verify that ERRE achieves an excellent balance in the "quality-efficiency-stability dilemma".

[0035] Table 1. Performance Comparison of Three Algorithms under Moderate Fault Conditions

[0036] Table 2 Performance Comparison of Three Algorithms under Severe Faults

[0037] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the technical solutions of the present invention, and are not intended to limit the specific implementation of the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the claims of the present invention should be included within the protection scope of the claims of the present invention.

Claims

1. A severity perception hierarchical system for dynamic flexible workshops, characterized in that, The system comprises a planning layer, a strategy layer, and a meta-control layer, formally defined as follows: In this context, G represents the planning layer; P represents the strategy layer; M represents the meta-control layer; and K represents the knowledge base, which is a shared knowledge base that stores elite solution structures. The planning layer G is responsible for generating a high-quality initial scheduling scheme using the Q-learning-guided adaptive genetic algorithm Q-AGA. And building an elite knowledge base K; The strategy layer P is a reactive strategy library, including lightweight and fast repair strategies. and heavyweight optimization strategies The lightweight and fast repair strategy This refers to the right-shift rescheduling strategy, representing a traditional heuristic strategy that pursues the highest stability and speed; the aforementioned heavyweight optimization strategy This refers to a rescheduling strategy based on knowledge reuse; The meta-control layer M is the decision-making core of the system, responsible for solving the algorithm selection problem online: when uncertain events occur, it selects algorithms from the policy library. The most suitable rescheduling strategy for the current scenario is selected; the meta-control layer includes a feature engineering module for severity index, a linear decision model module, and a threshold decision and calibration module. The feature engineering module of the severity index is responsible for constructing various features used to calculate the severity index SI, forming a three-dimensional feature vector; The linear decision model module is used to fuse the three-dimensional feature vectors extracted by the feature engineering module of the severity index into a single scalar value that can be used for decision-making. The threshold decision and calibration module is used to select strategies and calibrate thresholds to determine the optimal threshold.

2. The severity perception hierarchical system for dynamic flexible workshops according to claim 1, characterized in that, The three-dimensional feature vector is , by characteristics Time pressure, characteristics Structural disturbances and characteristics Key components of resources; The features Time pressure is as shown in formula (1): （1） In equation (1), Indicates the fault recovery time. Indicates the fault start time. Indicates workpiece The One process, This represents the set of processes that are directly affected. Indicates workpiece The Processing time of each step; molecules Indicates the duration of the fault, denominator This indicates the average expected processing time for those processes directly affected by the malfunctioning machine; The features The structural disturbance is as shown in formula (2): （2） In equation (2), This represents the set of all affected processes. This indicates the preceding closure of the process that is directly affected; The features Resource criticality is as shown in formula (3): （3） In equation (3), Indicates a faulty machine. Indicates a faulty machine Total planned workload Represents the set of all machines. This represents the average workload across all machines.

3. A severity perception hierarchical system for dynamic flexible workshops according to claim 1, characterized in that, The process of fusing the three-dimensional feature vectors extracted by the feature engineering module of the severity index into a single, decision-making scalar value SI specifically includes using the three-dimensional feature vectors... Define a linear model: （4） In equation (4), Represents the weight vector, weight vector Determined using the Analytic Hierarchy Process (AHP).

4. A severity perception hierarchical system for dynamic flexible workshops according to claim 1, characterized in that, The strategy selection based on the threshold refers to defining the decision rule as shown in formula (5): （5） In equation (5), Threshold; Determining the optimal threshold through threshold calibration refers to setting an optimization target. As in formula (6): （6） In equation (6), It is an asymmetric loss function. Indicates the optimal strategy label. Indicates the fault scenario The three-dimensional feature vectors below; The detailed threshold calibration process is as follows: S1. Data preparation: For each fault scenario in the dataset... The following three pieces of information need to be prepared: ① Three-dimensional feature vector ; ② Severity Index ③ Optimal strategy label For fault scenarios Execute offline respectively and Two strategies were used to obtain rescheduling schemes and record performance metrics; The strategy with better performance metrics is selected as the optimal strategy; through this process, each fault scenario in the dataset is analyzed. They all got one Data pairs; S2. Quantization definition of asymmetric loss function The loss function receives two inputs: the policy predicted by the model and the actual optimal policy, and then outputs a cost value, as shown in formula (7): （7） In equation (7), This refers to the cost of false alarms; This refers to the cost of underreporting, the ratio. This reflects the degree of aversion to the risk of underreporting; S3. Search and Optimization: Based on the prepared data and loss function, a grid search method is used to find the optimal threshold. .

5. A method for scheduling a workshop using the system described in claims 1-4, characterized in that, The method includes: an offline preparation phase and an online response phase; The offline preparation phase includes: firstly, using the Q-learning-guided adaptive genetic algorithm Q-AGA to generate an initial scheduling scheme. Simultaneously, a knowledge base K containing elite solution structures is constructed; the Q-learning-guided adaptive genetic algorithm Q-AGA models the parameter control of the genetic algorithm GA as a Markov decision process, and uses the Q-learning reinforcement learning algorithm to achieve intelligent parameter adaptation. Then, the meta-control layer uses the Analytic Hierarchy Process (AHP) to complete parameter calibration and threshold calibration to obtain the optimal threshold. The online response phase includes: firstly, fault detection, real-time monitoring of production disturbances, and recording of fault events. ,in: Indicates a faulty machine. Indicates the fault start time. Indicates the fault recovery time; Then, based on time pressure, structural disturbances, and resource criticality, a severity index (SI) is obtained through a linear decision model. Strategy selection is then based on the SI, comparing it with the optimal threshold: if SI < optimal threshold, a lightweight, rapid repair strategy is selected. If SI Optimal threshold, select heavyweight optimization strategy ; Finally, the status is updated after the solution is generated.

6. The method for scheduling workshops according to claim 5, characterized in that, The Markov decision process is modeled as follows: The parameter control problem of Q-AGA is modeled as a Markov decision process quadruple. Where S represents the state space, A represents the action space, and R represents the reward function. Represents the discount factor; more specifically, the state space in the th... At that time, the environmental state observed by the Agent It is a two-dimensional vector, including the evolutionary stage and the degree of stagnation, with a total of 9 discrete states, formally represented as formula (8): （8） In equation (8), It refers to the stage of evolution. This refers to the degree of stagnation; The action space is designed with N predefined discrete crossover probabilities. With the probability of mutation Parameter combination ; The reward function is as shown in formula (9): （9） In equation (9), This refers to the optimal solution of the previous generation. It refers to the optimal solution of the new generation.

7. The method for scheduling workshops according to claim 6, characterized in that, The workflow of the Q-learning-guided adaptive genetic algorithm Q-AGA is as follows: S1. Use application examples of flexible work workshops as algorithm input; S2. Use the MSOS hierarchical encoding method to encode the chromosome of the instance; S3. Use the GLR strategy for population initialization; S4. Begin the evolutionary operation generation by generation until the maximum number of iterations designed by the algorithm is reached. Steps S5-S8 are the specific steps of one evolution. S5. State awareness, calculating the current evolutionary stage. and degree of stagnation Combined into the current environmental state ; S6. Action selection: Generate a random number rand between (0,1). If rand is less than a preset threshold... Randomly select a group from the action space As an action If rand is greater than or equal to the preset threshold Then look up the current state in the Q table. For the corresponding row, select the action with the largest Q value in that row as... ,Right now ; S7. Perform the GA operation and select the action. Apply to the current population Perform standard genetic algorithm operations: selection, crossover, and mutation to generate the next generation population. ; S8. Reward calculation and state transition, evaluating the new population. To obtain the target value of its optimal solution The optimal value of the previous generation Compare the results and calculate the immediate reward based on the reward function. Then the system transitions to the next state. ; S9. Update the Q table. After S8, a complete transfer sample is obtained. Update the Q-table according to the core update formula of Q-learning. The value; S10. Determine if the current generation is the last generation. If so, save the initial scheduling scheme and the elite population.

8. The method for scheduling workshops according to claim 5, characterized in that, The lightweight and fast repair strategy This refers to the right-shift rescheduling strategy, the heavyweight optimization strategy. This refers to a rescheduling strategy based on knowledge reuse; The knowledge reuse-based rescheduling strategy The event-driven elite re-evolution algorithm ERRE is adopted. The detailed steps of the ERRE algorithm are as follows: S1. Input parameters: ① Fault events including faulty machine, fault start time and end time, ② Current system status, ③ Knowledge base storing historical elite solutions, ④ Genetic algorithm parameter settings; S2. First stage: Identify the affected domain, specifically including: ① Recording fault information, ② Identifying interrupted operations, ③ Identifying planned operations, ④ Identifying continuous operations, and ⑤ Calculating the set of affected operations. S3, Second Phase: Constructing a world snapshot, specifically including: ① Determining fixed operations, ② Handling residual processing time, ③ Constructing a rescheduling context; S4, Third Stage: Call the rescheduling engine, which includes: ① Obtain the elite population as the current population, ② Repair the elite chromosomes, ③ Initialize the optimal solution, ④ Iterative optimization, with each iteration consisting of steps ⑤-⑨ under step S4: ⑤ Fitness evaluation, ⑥ Update the optimal solution, ⑦ Select elites, ⑧ Generate new individuals, ⑨ Update the population, ⑩ Obtain the optimal solution after all iterations are completed, i.e., obtain the new rescheduling scheme; S5, Fourth Stage: Integrating the new scheduling scheme, specifically including: ① Cleaning the event queue, ② Adding repair completion events, ③ Using context-aware reactive decoding to generate detailed scheduling, the specific decoding process is ④-⑦ under S5, ④ Resetting internal state, extracting fault parameters, setting global constraints, ⑤ Loading context: handling fixed operations, adding fault constraints, ⑥ Decoding new operations: parsing chromosomes, scheduling jobs in sequence, ⑦ Calculating maximum completion time, ⑧ Updating the event queue, ⑨ Returning to the updated status.

Images