Workshop disturbance level sensing method based on double-layer Bayesian network

Abstract

A workshop disturbance level sensing method based on a double-layer Bayesian network comprises the following steps: constructing a virtual simulation model of a flexible job workshop based on a digital twin technology, simulating an uncertain disturbance event in a production process, and collecting production information data; defining key factors for quantifying the influence of disturbance on the production system; constructing a double-layer Bayesian network model for disturbance perception; carrying out conditional probability distribution on each node of the constructed double-layer Bayesian network model by utilizing production information data collected by a virtual simulation model of the flexible job shop; and based on the constructed double-layer Bayesian network model, calculating a comprehensive index W representing the overall disturbance level of the production system, and when the W value exceeds a preset threshold value, triggering a rescheduling decision. According to the method, the uncertainty disturbance in the production process can be quantitatively perceived, the comprehensive influence degree of the uncertainty disturbance on a production system is accurately evaluated, and a rescheduling decision is timely and reasonably triggered.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent manufacturing and production scheduling, and particularly relates to a workshop disturbance level perception method based on a double-layer Bayesian network. BACKGROUND

[0002] The production system of a discrete manufacturing enterprise usually has characteristics such as high complexity, strong dynamics and multiple uncertainties. The efficiency and stability of the production system depend largely on the scheduling scheme adopted. However, in actual production processes, uncertainty disturbance events such as changes in processing time, machine failures, urgent orders and rework of workpieces frequently occur, making it difficult to stably execute the initial scheduling scheme. When and how to trigger rescheduling according to the disturbance situation is a key problem faced by modern workshops.

[0003] Traditional rescheduling strategies mainly include periodic rescheduling, event-driven rescheduling and hybrid rescheduling. These methods can cope with disturbances to some extent, but often lack accurate quantitative perception of the disturbance level, making it difficult to judge the severity of the disturbance event and its global impact on the production system, which may lead to unnecessary rescheduling (increasing the burden on the system) or untimely rescheduling (resulting in production delays).

[0004] Digital twin technology can map the running state of a physical workshop in real time through virtual modeling, providing a data basis for perceiving and analyzing disturbances. Bayesian networks, as a kind of probabilistic graphical model, are good at handling uncertainty reasoning problems. Existing research has attempted to use Bayesian networks for workshop scheduling and disturbance management, but most methods are either too macroscopic, lacking details at the process level, or too microscopic, making it difficult to assess the propagation and comprehensive impact of disturbances at the system level. Therefore, a method is needed that can balance the microscopic process impact and the macroscopic system impact and quantitatively perceive disturbances to provide accurate and reliable basis for rescheduling decisions. SUMMARY

[0005] The present application proposes a workshop disturbance level perception method based on a double-layer Bayesian network, which can quantitatively perceive uncertainty disturbances in the production process, accurately assess the comprehensive impact of the disturbances on the production system, and timely and reasonably trigger rescheduling decisions.

[0006] The technical scheme adopted by the present application is as follows:

[0007] The workshop disturbance level perception method based on a double-layer Bayesian network comprises the following steps:

[0008] Step 1: Based on digital twin technology, a virtual simulation model of a flexible job shop is constructed to simulate uncertainty disturbance events in the production process and collect production information data.

[0009] Step 2: Define the key factors used to quantify the impact of disturbances on the production system;

[0010] Step 3: Construct a two-layer Bayesian network model for perturbation sensing;

[0011] Step 4: Using the production information data collected from the virtual simulation model of the flexible work workshop in Step 1, conditional probability assignment is performed on each node of the two-layer Bayesian network model constructed in Step 3.

[0012] Step 5: Based on the two-layer Bayesian network model constructed in Step 3, calculate the comprehensive index W that characterizes the overall disturbance level of the production system. When the value of W exceeds the preset threshold, trigger a rescheduling decision.

[0013] In step 1, based on the scheduling plan to be implemented during actual production, a virtual simulation model of the flexible workshop is constructed using digital twin technology, such as software like PlantSimulation. Based on historical production data, parameters such as machine failure rate and workpiece defect rate are set in the model to simulate uncertainties and disturbances such as processing time fluctuations, machine failures, reprocessing, and the insertion of emergency orders.

[0014] Run a virtual simulation model of the flexible workshop to collect detailed production information data such as the actual start time, actual end time, and equipment failures of each process.

[0015] In step 2, five key factors are defined to quantify the impact of disturbances on the production system: process reprocessing R i，j Machine malfunction B i，j Processing time delay D i，j Processing time fluctuation F i，j Processing timeout O i，j Among them, processing time exceeded O i，j It is the ultimate measure of the impact of disturbances on the process level.

[0016] In step 3, the two-layer Bayesian network model includes:

[0017] Process-level Bayesian network: used to characterize the probability of a single process experiencing processing timeout under the influence of reprocessing, machine failure, and processing fluctuation factors;

[0018] System-level Bayesian network: It is derived from the single-code network graph based on the initial scheduling scheme. The network nodes represent the timeout events of the process and the directed edges represent the temporal constraints between processes, which are used to characterize the transmission and impact of disturbances between processes.

[0019] The output O of the process layer Bayesian network i，j As input to the system-level Bayesian network, the two networks are coupled together.

[0020] wherein:

[0021] 3.1: Process layer Bayesian network construction includes:

[0022] For each process, a Bayesian network substructure is established with R i，j , B i，j , F i，j as parent nodes, D i，j as intermediate nodes, and O i，j as child nodes, representing the probabilistic relationship of the process being affected by disturbance and overtime.

[0023] 3.2: System layer Bayesian network construction includes:

[0024] According to the Gantt chart of the initial scheduling scheme, a single code network diagram is generated, establishing the time logic relationship between processes, providing a structural basis for subsequent probabilistic reasoning and disturbance quantification, as shown in (a) to (b) of Figure 4 . Then the nodes in the single code network diagram are mapped to the opportunity nodes representing "the process processing overtime O i，j ", realizing the conversion from static scheduling diagram to dynamic disturbance perception model, so that the system can perform probabilistic reasoning, disturbance quantification and intelligent decision-making in an uncertain environment, as shown in (b) to (c) of Figure 4 .

[0025] The time sequence constraint relationship between nodes (process constraint, machine constraint) is mapped to the directed edge in the Bayesian network, i.e. Figure 4 the arrow line in the Bayesian network in (c), thereby forming a system layer Bayesian network representing the transmission of disturbance between processes.

[0026] 3.3: Network coupling:

[0027] The output node O i，j of the process layer Bayesian network of each process is connected to the corresponding node in the system layer Bayesian network, forming a complete and coupled two-layer Bayesian network model.

[0028] In step 4, the nodes are assigned conditional probabilities (network parameter learning):

[0029] The condition probability table (CPT) is the quantitative core of the double-layer Bayesian network, which defines the probability of the occurrence of a child node (such as "the current process overtime") given the state of the parent node (such as "machine failure" and "previous process overtime"). Through the condition probability table CPT, the model converts the disturbance transmission relationship between the process layer and the system layer into calculable probability reasoning, and finally realizes the quantification of the system disturbance level, providing a basis for rescheduling decisions.

[0030] For the processing time fluctuation F i，j The improved K-means clustering method is used to reasonably divide the continuous variable of processing time fluctuation into three discrete levels with clear physical meaning (low, medium, and high), and then based on the statistical frequency of simulation data, the probability of each level is assigned, so that the continuous uncertainty is successfully integrated into the inference framework of discrete Bayesian network.

[0031] In step 4, the processing time fluctuation F i，j The improved K-means clustering method is used to reasonably divide the continuous variable of processing time fluctuation into three discrete levels with clear physical meaning (low, medium, and high), and then based on the statistical frequency of simulation data, the probability of each level is assigned, so that the continuous uncertainty is successfully integrated into the inference framework of discrete Bayesian network. i，j，k For each processing record, the fluctuation value ΔT is calculated, and the calculation formula is:

[0032] ΔT = T' i，j，k -T i，j，k ;

[0033] According to the domain knowledge (negative fluctuation or zero fluctuation has the same positive significance), all data are first divided into two groups Group A (ΔT≤0) and Group B (ΔT≥0), and Group A is directly marked as "low fluctuation" corresponding to cluster 1. For Group B, the K-means clustering algorithm is applied to divide it into two clusters, corresponding to cluster 2 and cluster 3. The objective function of K-means is to minimize the within-cluster sum of squares (WCSS):

[0034]

[0035] Where J is the objective function of K-means, K is the number of clusters, Ci For the i-th cluster, x j For the sample point belonging to the j-th class, ||x j -μ i || 2 For the Euclidean distance square between the sample point and the cluster center it belongs to, μ i For the i-th cluster center, calculated as:

[0036]

[0037] The cluster boundary is [min(C i ), max(C i ] where C i is the i-th cluster; after clustering, three clusters with clear physical meaning are obtained: low volatility: Group A; medium volatility: Group B with smaller center value; high volatility: Group B with larger center value) and its parameters, based on the frequency of each level appearing in the simulation data, assign the prior probability:

[0038] ① Count:

[0039] N low = the number of data in Group A;

[0040] N mid = the number of data in cluster 2;

[0041] N high = the number of data in cluster 3;

[0042] N total = N low + N mid + N high ;

[0043] ② Probability calculation:

[0044]

[0045] In step 5, disturbance quantification perception and rescheduling decision:

[0046] When a certain disturbance event (such as machine failure) is monitored in real time or predicted through simulation, it is input as evidence into the double-layer Bayesian network model, and the network inference is calculated:

[0047] a: processing time disturbance R1: based on the expected overtime calculated by network inference;

[0048] b: number of affected processes disturbance R2: based on the system layer Bayesian network structure, determine the number of affected subsequent processes.

[0049] The two are combined to obtain a system disturbance comprehensive index W, a threshold W0 is set, when W>W0, it is determined that the system is disturbed to a high degree, and a rescheduling process is triggered, otherwise, the system can continue to operate according to the original plan.

[0050] In step 5, the calculation formula of the comprehensive index W is:

[0051]

[0052] Wherein, R1 is a normalized processing time disturbance, R2 is a normalized number of affected process disturbances; the closer the value of W is to 1, the higher the degree of system disturbance.

[0053] The calculation of the processing time disturbance R1 is based on the expected overtime E(Δt) inferred by the Bayesian network, and the calculation formula is:

[0054]

[0055] The calculation of the number of affected process disturbances R2 is based on the ratio of the number of subsequent processes N affected by the current process overtime to the maximum possible number of affected processes N max The calculation formula is:

[0056]

[0057] The method for perceiving the disturbance level of a workshop based on a double-layer Bayesian network has the following beneficial effects:

[0058] 1) The method combines digital twinning with Bayesian networks, uses digital twinning to provide real-time / simulation data, uses Bayesian networks to perform uncertainty reasoning, and provides a data basis and theoretical model for disturbance perception.

[0059] 2) The double-layer Bayesian network structure proposed by the method takes into account the detailed disturbance factors of the process layer and the disturbance propagation relationship of the system layer, realizes comprehensive disturbance perception from “point” (process) to “line” (production line), and can more accurately reflect the real impact of the disturbance.

[0060] 3) The method defines a quantitative system disturbance level comprehensive index W, converts the abstract disturbance degree into a specific numerical value, provides a clear and objective basis for rescheduling decision-making, and improves the timeliness and scientificity of the decision-making.

[0061] 4) The method has universality and can be applied to dynamic scheduling scenarios of various discrete manufacturing workshops, effectively improving the anti-disturbance ability and operation efficiency of the production system. BRIEF DESCRIPTION OF DRAWINGS

[0062] The application will be further described below in combination with the drawings and examples:

[0063] Figure 1 This is a flowchart illustrating the overall process of the workshop disturbance sensing and rescheduling method based on digital twins and two-layer Bayesian networks in this invention.

[0064] Figure 2 This is a simulation workshop model diagram created using Plant-Simulation software in an embodiment of the present invention.

[0065] Figure 3 This is a schematic diagram of the Bayesian network structure of the process layer in an embodiment of the present invention.

[0066] Figure 4 This is a schematic diagram illustrating the process of converting a Gantt chart into a single-symbol network graph and then into a system-layer Bayesian network in an embodiment of the present invention.

[0067] Figure 5 This is a partial schematic diagram of the two-layer Bayesian network coupling structure in an embodiment of the present invention.

[0068] Figure 6 This is a schematic diagram of the discretization results of processing time fluctuation data (improved K-means clustering) in an embodiment of the present invention.

[0069] Figure 7 This is a schematic diagram illustrating the disturbance level W of the system under three different disturbance scenarios in an embodiment of the present invention. Detailed Implementation

[0070] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0071] like Figure 1 As shown in the figure, a workshop disturbance level sensing method based on a two-layer Bayesian network in this embodiment specifically includes the following steps:

[0072] Step S101, Virtual Modeling and Data Acquisition: (e.g.) Figure 2 As shown, a digital twin model of a flexible job shop (FJSP) with 6 workpieces, 6 processes, and 10 machines in a factory was constructed using PlantSimulation software. The machine failure rate was set to 5%, the workpiece defect rate to 4%, and the process processing time to follow a triangular distribution. The model ran for 31 days, executing the same initial scheduling scheme each day, collecting data such as the scheduled / actual start / end times of all processes and machine failure records, totaling 1116 processing records.

[0073] Step S102, Definition of Disturbance Factors: This embodiment defines five key disturbance variables: R i，j (Reprocessing), B i，j (Machine malfunction), D i，j (Processing delay), F i，j (Processing fluctuations), O i，j(Processing overtime).

[0074] Step S103, double-layer Bayesian network construction:

[0075] As shown in Figure 3 , the process layer Bayesian network is constructed, showing how R i，j , B i，j inevitably lead to D i，j , and how D i，j and F i，j jointly affect O i，j .

[0076] As shown in Figure 4 , a single-digit network diagram is generated according to the initial scheduling Gantt chart, and it is converted into a system layer Bayesian network, with nodes representing O i，j .

[0077] As shown in Figure 5 , the process layer network is coupled with the system layer network through O i，j nodes to form a complete double-layer Bayesian network.

[0078] Step S104, node probability assignment: use the data generated in step S101 to assign probabilities.

[0079] The prior probabilities of R i，j , B i，j are assigned according to the settings of 4% and 5%.

[0080] The conditional probabilities of D i，j are assigned according to the logical relationship between R i，j , B i，j , and delay is bound to occur when there is a failure or rework.

[0081] The probabilities of F i，j are obtained by analyzing the processing time deviation data. Improved K-means clustering is used, as shown in Figure 6 , to discretize the deviation values into three categories: "low", "medium", and "high", and calculate the probabilities of each category: about 50%, 29.1%, and 20.9%.

[0082] The conditional probabilities of O i，j and its conditions in the system layer are calculated by analyzing the theoretical time intervals between processes and the possible time variation amounts to calculate the probabilities of overtime under different disturbance combinations, forming the conditional probability table CPT. The probabilities of process overtime under different disturbance combinations are calculated by enumerating all state combinations of disturbance factors, and based on whether the total time delay under each combination exceeds the theoretical time buffer between the process and the subsequent process, a deterministic judgment is made: if the delay exceeds the buffer, the overtime probability is 1, otherwise it is 0, thus forming the conditional probability table of the nodes in the system layer Bayesian network.

[0083] Step S105, disturbance perception and decision: three disturbance scenarios are set to be verified.

[0084] Scenario 1: P 1，1 Reprocessing occurs. It is reasoned that W1=0.675.

[0085] Scenario 2: Machine 3 fails at t=6 (affects P 2，2 ). It is reasoned that W2=0.795.

[0086] Scenario 3: An emergency order is inserted at t=23. It is reasoned that the two processes of this order cause W 3A =0.297 and W 3B =0.31, respectively.

[0087] As shown in Figure 7 , the W values under the three scenarios are shown. Assuming that a threshold W0=0.6 is set, then scenario 1 and scenario 2 need to trigger rescheduling, and scenario 3 does not need to reschedule or only needs to be adjusted locally.

Claims

1. A workshop disturbance level sensing method based on a two-layer Bayesian network, characterized in that... Includes the following steps: Step 1: Based on digital twin technology, construct a virtual simulation model of the flexible workshop to simulate uncertain disturbance events in the production process and collect production information data; Step 2: Define the key factors used to quantify the impact of disturbances on the production system; Step 3: Construct a two-layer Bayesian network model for perturbation sensing; Step 4: Using the production information data collected from the virtual simulation model of the flexible work workshop in Step 1, conditional probability assignment is performed on each node of the two-layer Bayesian network model constructed in Step 3. Step 5: Based on the two-layer Bayesian network model constructed in Step 3, calculate the comprehensive index W that characterizes the overall disturbance level of the production system. When the value of W exceeds the preset threshold, trigger a rescheduling decision.

2. The workshop disturbance level sensing method based on a two-layer Bayesian network according to claim 1, characterized in that: In step 1, based on the scheduling plan to be implemented during actual production, a virtual simulation model of the flexible workshop is constructed using digital twin technology. Based on historical actual production data, parameters such as machine failure rate and workpiece defect rate are set in the model to simulate the occurrence of uncertain disturbance events such as processing time fluctuations, machine failures, reprocessing, and emergency order insertion. Run a virtual simulation model of the flexible workshop to collect detailed production information data such as the actual start time, actual end time, and equipment failures of each process.

3. The workshop disturbance level sensing method based on a two-layer Bayesian network according to claim 2, characterized in that: In step 2, five key factors are defined to quantify the impact of disturbances on the production system: process reprocessing R i,j Machine malfunction B i，j Processing time delay D i，j Processing time fluctuation F i，j Processing timeout O i，j Among them, processing time exceeded O i，j It is the ultimate measure of the impact of disturbances on the process level.

4. The workshop disturbance level sensing method based on a two-layer Bayesian network according to claim 3, characterized in that: In step 3, the two-layer Bayesian network model includes: Process-level Bayesian network: used to characterize the probability of a single process experiencing processing timeout under the influence of reprocessing, machine failure, and processing fluctuation factors; System-level Bayesian network: It is derived from the single-code network graph based on the initial scheduling scheme. The network nodes represent the processing timeout events of the process, and the directed edges represent the temporal constraints of process 6, which are used to characterize the transmission and impact of disturbances in process 6. The output O of the process layer Bayesian network i，j As input to the system-level Bayesian network, the two networks are coupled together.

5. The workshop disturbance level sensing method based on a two-layer Bayesian network according to claim 4, characterized in that: The construction of the process layer Bayesian network includes: For each process, establish an R-based system. i，j B i，j F i，j As the parent node, D i，j As an intermediate node, O i，j The Bayesian network substructure for child nodes represents the probability relationship of timeout caused by disturbance in this process; The construction of system-level Bayesian networks includes: Based on the Gantt chart of the initial scheduling scheme, a single-code network graph was generated, establishing the temporal logical relationships between processes. This provided a structural foundation for subsequent probabilistic reasoning and perturbation quantification. Then, the nodes in the single-code network graph were mapped to represent "processing timeout O". i,j "Opportunity nodes" enable the transformation from a static scheduling graph to a dynamic disturbance perception model; The temporal constraints of node 6 are mapped to directed edges in a Bayesian network, thus forming a system-level Bayesian network to characterize the propagation of disturbances in process 6. Network coupling: The output nodes O of the Bayesian network at the process layer of each process are coupled together. i，j It is connected to the corresponding node in the system layer Bayesian network to form a complete and coupled two-layer Bayesian network model.

6. The workshop disturbance level sensing method based on a two-layer Bayesian network according to claim 5, characterized in that: In step 4, the nodes perform conditional probability assignment: Using the large amount of production information data generated by the virtual simulation in step 1, and combined with domain knowledge, the conditional probability table (CPT) of all nodes in the two-layer Bayesian network model is learned and assigned values. The conditional probability table (CPT) is the core of quantization in the two-layer Bayesian network, defining the probability of a child node occurring given the state of the parent node. Through the conditional probability table (CPT), the model transforms the disturbance propagation relationship between the process layer and the system layer into computable probabilistic inference, ultimately achieving the quantification of the system disturbance level.

7. The workshop disturbance level sensing method based on a two-layer Bayesian network according to claim 6, characterized in that: For processing time fluctuation F i，j For continuous variables, the improved K-means clustering method is used to discretize them and assign probabilities. Among them, the improved K-means clustering method, which combines domain knowledge, reasonably divides the continuous variable of processing time fluctuation into three discrete levels with clear physical meaning: low, medium and high. Then, based on the statistical frequency of simulation data, a probability is assigned to each level, thus successfully integrating continuous uncertainty into the inference framework of discrete Bayesian networks.

8. The workshop disturbance level sensing method based on a two-layer Bayesian network according to claim 7, characterized in that: In step 4, the processing time fluctuation F i，j The node probability allocation employs an improved K-means clustering method to discretize the deviation between actual and theoretical processing times, classifying them into three categories: "low fluctuation," "medium fluctuation," and "high fluctuation." The actual processing time T′ for all processes is collected by running the PlantSimulation simulation model. i，j，k For each processing record, calculate its fluctuation value ΔT using the following formula: ΔT=T′ i，j，k -T i，j，k ; Based on domain knowledge, all data are first divided into two main groups: Group A (ΔT≤0) and Group B (ΔT≥0). Group A is directly labeled as "low volatility," corresponding to cluster 1. K-means clustering is applied to Group B, dividing it into two clusters, corresponding to cluster 2 and cluster 3. The objective function of K-means is to minimize the within-cluster sum of squares (Within-ClusterSumofSquares.WCSS). Where J is the objective function of K-means, K is the number of clusters, and C is the number of clusters. i For the i-th cluster, x j Let ||x| be a sample point belonging to class j. j -μ i || 2 μ is the squared Euclidean distance between a sample point and the centroid of its cluster. i The center of the i-th cluster is calculated as follows: The cluster boundary is [min(C i ), max(C i ], where C i This is the i-th level cluster; after clustering, three clusters with clear physical meanings were obtained: low volatility (Group A); medium volatility (Group B with smaller center values); and high volatility (Group B with larger center values). Based on the frequency of each level in the simulation data, prior probabilities were assigned according to their parameters: ① Counting: N low =Number of data entries in GroupA; N mid = The number of data entries in cluster 2; N high = The number of data entries in cluster 3; N total =N low +N mid +N high ; ②Probability calculation:

9. The workshop disturbance level sensing method based on a two-layer Bayesian network according to claim 8, characterized in that: In step 5, the disturbance quantization sensing and rescheduling decision are as follows: When a disturbance event is detected in real time or predicted through simulation, it is input as evidence into the two-layer Bayesian network model, and the network inference is used for calculation: a: Processing time perturbation R1: Calculated based on the expected timeout time inferred from the network; b: Disturbance R2: Based on the system-level Bayesian network structure, determine the number of subsequent processes affected; By combining the two, we obtain the comprehensive disturbance index W of the system. We set a threshold W0. When W>W0, we determine that the system is highly disturbed and trigger the rescheduling process. Otherwise, the system can continue to operate as originally planned.

10. The workshop disturbance level sensing method based on a two-layer Bayesian network according to claim 9, characterized in that: In step 5, the formula for calculating the comprehensive index W is: Where R1 is the normalized processing time disturbance, and R2 is the normalized disturbance of the number of affected processes; the closer the W value is to 1, the higher the degree of disturbance to the system. The processing time disturbance R1 is calculated based on the expected timeout E(Δt) inferred from the Bayesian network, and the calculation formula is as follows: The calculation of the disturbance R² caused by the number of affected processes is based on the number N of subsequent processes affected by the timeout of the current process and the maximum possible number of processes N of the system. max The ratio is calculated using the following formula:

Images