Automatic wharf operation delay propagation prediction method fusing abnormal perception and transfer entropy analysis

Abstract

The invention discloses an automatic wharf operation delay propagation prediction method fusing anomaly perception and transfer entropy analysis, which comprises the following steps: on the basis of an operation time sequence of operation equipment, introducing anomaly characteristics of the operation equipment, constructing an enhanced time sequence set, and enhancing the response capability to sudden disturbance in a prediction process; a candidate condition variable set is selected through a recursive feature elimination algorithm, so that the calculation complexity and the over-fitting risk caused by a high-dimensional variable space are effectively reduced, and the problem of curse of dimensionality is relieved; by constructing the delay propagation atlas, the mutual relation between operation equipment can be indicated, and intuition is achieved; a multi-stage operation delay conduction model is constructed according to the transformation probability of the operation equipment in different states, so that the influence of the equipment in different states on the port transportation process can be reflected more accurately; by using the causal relationship between the operation devices in the delay propagation map, how to propagate between the operation devices during delay propagation can be more clearly understood.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of port automation, and in particular to an automated wharf operation delay propagation prediction method combining abnormal perception and transfer entropy analysis. BACKGROUND

[0002] As a key node of cargo transportation, the safety and stability of the port directly affect the efficiency of maritime transportation and the stability of the supply chain. In recent years, port congestion events have occurred frequently, increasing shipping costs and posing a great challenge to wharf operations. Port congestion leads to a backlog of ships, causing the utilization rate of the wharf yard to rise to nearly 90%, sharply increasing the workload of loading and unloading, storage, and transfer in the short term. This not only increases the pressure on equipment usage and scheduling difficulty, but also makes it more likely to cause operation delays due to time window conflicts, further exacerbating the port congestion situation. As an important part of the port, the automated wharf has multiple operation links, including quay crane (QC) loading and unloading, automated guided vehicle (AGV) transfer, yard crane (YC) operation in the yard, and container pickup and delivery on the land side, forming an integrated system that is closely connected and mutually influenced. When a link fails or scheduling conflicts occur, the delay will spread throughout the entire wharf during the operation process.

[0003] Currently, most scholars focus on solving the problem of wharf operation delay from the perspectives of plan scheduling and scheduling optimization. For example, Zeng Qingcheng et al. established a multi-AGV path optimization model considering congestion, effectively alleviating congestion in horizontal transportation and improving transportation efficiency. Ma Ningli et al. established a mixed integer programming model to minimize AGV completion time, and through optimizing the charging strategy, they effectively improved the operation efficiency of the AGV by shunting the congestion of the road network partition. As a complex system, the automated wharf has closely linked operation links, and single-link optimization cannot fully explore the delay problem in the system. Xu Boma et al. first applied complex network theory to wharf operations, taking device resources as nodes and operation tasks as hyper-edges to construct a multi-level operation hyper-network, and used characteristic path length and clustering coefficient as indicators to analyze the operation delay caused by uncertain events in the hyper-network. LI et al. abstracted the devices and relationships in the main functional areas of the wharf as nodes and edges to construct a multi-level operation risk entity network of the automated wharf, and identified the multi-level operation risk transmission chain group through the Gillespie algorithm to simulate the dynamics of risk propagation.

[0004] However, the multi-level operation process of the wharf is dynamic, and current research on multi-level operations mainly constructs networks with task or entity associations as edges, making it difficult to accurately depict the complex interaction between devices in the operation system. SUMMARY

[0005] In order to solve the problems that it is difficult to comprehensively and accurately depict the complex interaction relationship between the working devices in the port operation system and it is impossible to predict the propagation prediction of the working device delay under the complex relationship in the prior art, the application provides an automatic wharf operation delay propagation prediction method fusing abnormal perception and transfer entropy analysis, and solves the above problems.

[0006] The specific scheme is as follows:

[0007] An automatic wharf operation delay propagation prediction method fusing abnormal perception and transfer entropy analysis comprises the following steps.

[0008] S1: constructing a time series set: acquiring the operation time data of the working devices inside the wharf and constructing a first time series set; acquiring the abnormal index data of the working devices in the operation process; uniformly coding the abnormal index data to obtain a second time series set; and fusing the first time series set and the second time series set into a time series set;

[0009] S2: constructing a selected condition variable set: calculating the importance scores of the variables in the time series set through a recursive feature elimination algorithm, and selecting the variables and constructing a selected condition variable set according to the importance scores of the variables;

[0010] S3: determining the causal relationship between the working devices and constructing a delay propagation graph: calculating the transfer entropy between the working devices according to the selected condition variable set; determining the causal relationship between the working devices according to the calculation result of the transfer entropy; and constructing a delay propagation graph through the causal relationship between the working devices;

[0011] S4: constructing a multi-level operation delay propagation model: constructing a multi-level operation delay propagation model according to the transformation probability between the working devices in different states and an SEIR model;

[0012] S5: performing delay propagation prediction; the multi-level operation delay propagation model performs delay propagation prediction by using the causal relationship between the devices in the delay propagation graph.

[0013] Preferably, the method for constructing the time series set in step S1 is as follows:

[0014] A1: dividing the operation time data at a fixed time interval to obtain divided operation time data;

[0015] A2: calculating the average operation time of each working device in each time interval in the divided operation time data; constructing an average operation time sequence of each working device according to the average operation time; and constructing a first time series set through the average operation time sequence of each working device;

[0016] A3: obtaining abnormal index data of the internal operation equipment; the abnormal index types include but are not limited to: abnormal operation behavior of the operation equipment, operation equipment not operating according to the plan, path deviation of the operation equipment and non-planned start of the operation equipment;

[0017] A4: uniformly encoding the abnormal index data to obtain a second time series set consistent with the first time series variable set structure; the uniform encoding is to binarize, value, residual or frequency code the abnormal index data according to the characteristics of the abnormal index;

[0018] A5: fusing the first time series set and the second time series set to obtain a time series set.

[0019] Preferably, the step of fusing the first time series set and the second time series set in step S1 is:

[0020] B1: uniformly constructing the first time series set and the second time series set into a multi-order lag structure and aligning them in the same feature space;

[0021] B2: merging the aligned first time series set and the second time series set to obtain a time series set.

[0022] Preferably, the method for constructing the selection condition variable set in step S2 is:

[0023] C1: selecting any one variable in the first time series set as a target variable;

[0024] C2: selecting variables other than the target variable in the time series set as candidate variables; constructing a candidate variable set corresponding to the target variable through the candidate variables;

[0025] C3: performing importance scoring on each variable in the candidate variable set through a recursive feature elimination algorithm and the target variable to obtain an importance score corresponding to the variable; deleting the variable with the lowest importance score from the candidate variable set to form a new candidate variable set; judging whether the new candidate variable set meets a preset stopping condition;

[0026] C4: if not, repeating the variable deletion process of C3 using the new candidate variable set, and if so, stopping iteration and taking the new candidate variable set meeting the preset stopping condition as the selection condition variable set corresponding to the target variable;

[0027] C5: selecting the next variable in the first time series set as a target variable, repeating steps C2-C4 to obtain the selection condition variable set corresponding to each variable in the first time series set.

[0028] Preferably, the method for scoring the importance of each variable in the candidate variable set by the recursive feature elimination algorithm and the target variable is:

[0029] a1: training a regression model by the candidate variable set and the target variable to obtain a trained regression model;

[0030] a2: scoring the importance of each variable in the candidate variable set by the trained regression model, and taking the absolute value of the importance score to obtain the importance score corresponding to each variable; the importance score represents the contribution of the variable in the candidate variable set to the target variable.

[0031] Preferably, the preset stopping condition includes whether the number of variables in the new candidate variable set meets the preset lower limit of the dimension, and whether the performance of the regression model decreases by more than the preset threshold.

[0032] Preferably, the method for judging the causal relationship between the work devices in step S3 is:

[0033] D1: selecting a variable in the first time series set as a target variable;

[0034] D2: selecting a variable other than the target variable from the first time series set as a source variable; judging whether the source variable exists in the selected condition variable set corresponding to the target variable, if it exists, deleting the source variable in the selected condition variable set to obtain a condition variable set; if it does not exist, taking the selected condition variable set as the condition variable set;

[0035] D3: taking the historical information of the target variable and the historical information of all variables in the condition variable set as conditions, calculating the conditional mutual information between the current information of the target variable and the historical information of the source variable to obtain the transfer entropy between the work devices corresponding to the target variable and the source variable;

[0036] D4: when the value of the transfer entropy is greater than 0, there is a causal relationship between the work devices corresponding to the target variable and the source variable; otherwise, there is no causal relationship between the work devices corresponding to the target variable and the source variable;

[0037] D5: reselecting the source variable from the first time series set; repeating steps D2-D4 to obtain the causal relationship between the work device corresponding to the target variable and all other work devices;

[0038] D6: reselecting the target variable from the first time series set, and repeating steps D2-D5 to obtain the causal relationship between all work devices.

[0039] Preferably, the delay propagation graph constructed in step S3 comprises nodes, links between nodes and arrows between links; the nodes represent job devices, the links between nodes indicate that there is a causal relationship between two nodes, and the arrows between links indicate the direction of influence between nodes.

[0040] Preferably, the conversion probability of the job devices in different states in step S4 comprises:

[0041] When the normal device contacts the delay transmission device, the normal job state node is converted into the delay node with a probability of a;

[0042] The delay device is converted into the delay transmission device with a probability of b, and the delay device is converted into the recovery device with a probability of q;

[0043] The delay transmission device is converted into the recovery device with a probability of g, and the delay transmission device is withdrawn from the job process with a probability of p;

[0044] The immune state device is converted into the susceptible state device with a probability of d.

[0045] Preferably, the multi-stage job delay transmission model in step S4 is based on an SEIR model, and the formula is:

[0046]

[0047] wherein, at time t, the normal job device with in-degree and out-degree of m and n The probability of being pointed to by the delay transmission device is P((i,j)|(m,n)) is defined as a connection probability, that is, the probability of connecting the node with in-degree and out-degree of i, j to the node with in-degree and out-degree of m, n; N(i,j|m,n,t) represents the number of devices connected to the node with in-degree and out-degree of m, n in normal operation at time t; N(i,j|m,n,d,t) represents the number of devices connected to the node with in-degree and out-degree of m, n in a delay state at time t; N(i,j|m,n,d,t) represents the number of devices connected to the node with in-degree and out-degree of m, n in a delay state at time t; N(i,j|m,n,r,t) represents the number of devices connected to the node with in-degree and out-degree of m, n in recovery at time t; m represents the in-degree of the screened node; n represents the out-degree of the screened node; N represents the total number of job devices; i,j N(i,j|m,n,d,t) represents the number of devices connected to the node with in-degree and out-degree of i, j in a delay transmission at time t, N t N(t) represents the number of all devices at time t;

[0048] At time t, the number of normal processing devices, delay state devices, delay transmission devices and recovery devices is represented as S t ,Et I t R t , and need to meet the following conditions:

[0049] S t E t I t R t =N t (2)

[0050]

[0051] Wherein, S t E t I t R t respectively represent the number of normal operation equipment, delay state equipment, delay transmission equipment and recovery equipment at t time; wherein, N t represent the number of all equipment at t time.

[0052] Beneficial effects:

[0053] The application proposes an automatic wharf operation delay propagation prediction method combining abnormal perception and transfer entropy analysis. In the traditional variable selection process, a time series set is constructed based on the abnormal characteristics of equipment operation behavior, and a selected condition variable set is constructed based on the recursive elimination algorithm, and a delay propagation map is constructed based on the transfer entropy. The application proposes a variable screening mechanism. The mechanism generates a second time series set by extracting equipment operation abnormal events (such as unplanned interruption, sudden increase of operation fluctuation rate, etc.), so that the system response characteristics of potential sudden events can be integrated in the delay propagation prediction process, thereby improving the adaptability of the prediction model to accidental influences in the actual operation scene. Secondly, unlike the mRMR variable screening method driven by mutual information in the prior art, the application introduces the RFE mechanism based on the importance score of the model, realizes a variable screening process with more dynamic adaptability, and uses the transfer entropy method to identify the potential causal relationship and structural characteristics between multi-level operation equipment. At the same time, according to the multi-level operation delay propagation model, the delay propagation trend under different initial delay scales and different control strategies can be simulated, which provides a theoretical basis and management reference for preventing and controlling large-area delay propagation in wharf internal operation.

[0054] The application evaluates the importance of candidate variables in the candidate variable set through a recursive feature elimination (RFE) algorithm, and step by step eliminates redundant variables with weak influence based on a regression model, so as to construct a variable subset with optimal prediction ability. This method can effectively reduce the calculation complexity and overfitting risk brought by high-dimensional variable space, and alleviate the problem of "dimension curse". At the same time, the transfer entropy calculation further reveals the dynamic causal relationship between the work devices, provides key support for constructing a more realistic and generalization ability of delay propagation graph, thereby improving the accuracy and robustness of system delay propagation modeling. At the same time, in the process of calculating the transfer entropy, by judging whether the source variable is included in the selected condition variable set, the problem that the calculation result of the transfer entropy loses meaning or degenerates to 0 when the source variable is included in the selected condition variable set is avoided. By constructing the delay propagation graph, the relationship between the work devices can be shown, which is intuitive; according to the transformation probability of the work devices in different states, a multi-level work delay transmission model is constructed, the work devices are divided into different states, and the influence of the devices on the port transportation process in different states can be more accurately reflected; meanwhile, the work devices that exit in the prediction process are also considered, so that the result of delay propagation prediction is closer to the actual situation; by constructing a multi-level work delay transmission model through the SEIR model, the quantitative analysis of the number change of work devices in different states and the influence of these changes on the port can be realized, so that the delay propagation prediction can be more accurately realized; when performing delay propagation prediction, the causal relationship between the work devices in the delay propagation graph can be used to more clearly understand how the delay propagates among the work devices, which helps to more accurately predict the occurrence and propagation range of the delay; meanwhile, the delay propagation trend under different initial delay scales and different control strategies can also be simulated, which provides a theoretical basis and management reference for preventing and controlling large-area delay propagation of the work in the terminal. BRIEF DESCRIPTION OF DRAWINGS

[0055] Figure 1 A flowchart of an automatic terminal work delay propagation prediction method combining abnormal perception and transfer entropy analysis.

[0056] Figure 2 A flowchart of an automatic terminal work delay propagation prediction method combining abnormal perception and transfer entropy analysis.

[0057] Figure 3 A schematic diagram of a delay propagation graph.

[0058] Figure 4 A flowchart of lag variable screening based on the RFE algorithm.

[0059] Figure 5 A schematic diagram of a multi-level work delay transmission model.

[0060] Figure 6Variable screening flowchart under abnormal auxiliary variable introduction.

[0061] Figure 7 Trend graph of each type of device node under different control measures over time.

[0062] Figure 8 Trend graph of each type of device node under different initial delay scale. DETAILED DESCRIPTION

[0063] The application will be further described below in conjunction with the accompanying drawings and examples.

[0064] As Figure 1 shown, the steps of an automatic port operation delay propagation prediction method combining abnormal perception and transfer entropy analysis include:

[0065] S1: Constructing a time series set: obtaining operation time data of internal operation equipment of the port and constructing a first time series set; obtaining abnormal index data of the operation equipment in the operation process; uniformly encoding the abnormal index data to obtain a second time series set; and fusing the first time series set and the second time series set into a time series set;

[0066] S2: Constructing a selected condition variable set: calculating the importance score of the variable in the time series set through a recursive feature elimination algorithm, and selecting the variable and constructing a selected condition variable set according to the importance score of the variable;

[0067] S3: Determining the causal relationship between the operation equipment and constructing a delay propagation graph: calculating the transfer entropy between the operation equipment according to the selected condition variable set; determining the causal relationship between the operation equipment according to the calculation result of the transfer entropy; and constructing a delay propagation graph through the causal relationship between the operation equipment;

[0068] S4: Constructing a multi-level operation delay propagation model: constructing a multi-level operation delay propagation model according to the transformation probability between operation equipment in different states and the SEIR model;

[0069] S5: Delay propagation prediction; the multi-level operation delay propagation model uses the causal relationship between the equipment in the delay propagation graph to perform delay propagation prediction.

[0070] Example one:

[0071] As Figure 2As shown, using transfer entropy and recursive feature elimination (RFE) variable screening algorithm, a delay propagation graph based on variable causal relationship can be constructed. Based on the actual operation data of the terminal equipment, the delay propagation graph is constructed by calculating the transfer entropy between the equipment, and the multi-level operation delay propagation model considering the equipment exit is constructed, and the delay propagation simulation analysis under different control strategies and delay scales is carried out by using Python.

[0072] An automatic terminal operation delay propagation prediction method combining abnormal perception and transfer entropy analysis is introduced in detail:

[0073] (1) Construction of time series:

[0074] According to the operation time data of the equipment inside the terminal, the time series of each operation equipment is constructed. The data is divided into 20 minutes as a time interval, and the average operation time of each equipment in each time interval is calculated as the time series data, so that the average operation time series of each equipment in the ship berthing and unloading period I=[1, 2, 3...i] is obtained. Where W i T is the average operation time of equipment e in the i-th time interval, T STA and T END are the operation start time and operation end time respectively, and F represents the number of container loading and unloading completed in the i-th time interval.

[0075]

[0076] According to formula (4), the average operation time of each equipment in the divided time interval is calculated, and the first time series set Q=[Q e1 ,Q e2 ,Q e3 ,...,Q eN ] of the operation equipment in the selected time period is constructed, wherein Q eN represents the average operation time series of the N-th operation equipment, and each variable in the first time series set corresponds to an operation equipment.

[0077] Further, the process of constructing the time series set also includes: the abnormal auxiliary variable sequence constructed based on the abnormal characteristics of the equipment behavior is taken as an optional input, and participates in the variable screening process together with the original operation time series, so as to enhance the perception ability of the variable selection to the sudden abnormality.

[0078] The construction of the abnormal auxiliary variable, i.e., the second time series set, includes encoding the abnormal indicators in the device operation process into structured time series variables. Specifically, abnormal behavior can include operation behavior abnormalities, devices not operating as planned, path deviations, unplanned starts, etc. Binary coding (whether abnormal), residual coding (difference from expected value), or frequency coding (number of abnormalities per unit time) are used to quantify these situations. The resulting second time series set is further structured with a multi-order lag structure consistent with the first time series set and is uniformly input as a candidate variable into the recursive feature selection module to achieve time series modeling and importance evaluation of abnormal information.

[0079] Specifically, the multi-order lag structure is constructed according to the internal operation experience of the terminal, with a lag order mlag=3 (i.e., only the causal relationship between operating devices within one hour is considered).

[0080] (2) Recursive Feature Elimination (RFE)

[0081] When calculating the transfer entropy between operating devices, both the influence of other devices and the influence of the historical time of the device need to be considered. With the expansion of the dimension of the conditional variable set, the "dimension curse" phenomenon will occur during the calculation process, which will significantly increase the required memory and calculation time. Therefore, dimensionality reduction is needed before calculating the transfer entropy. The dimensionality reduction process is equivalent to feature elimination, which can use the Recursive Feature Elimination (RFE) method.

[0082] Specifically, Recursive Feature Elimination (RFE) is a variable selection method based on model importance scoring. The principle is to train a regression model (such as ridge regression, SVM regression, etc.) to evaluate the predictive contribution of each variable to the target variable, and gradually eliminate the variables with the smallest contribution, ultimately obtaining a feature subset that best explains the prediction result. This method has good results in avoiding variable redundancy and improving model generalization ability, and is suitable for high-dimensional feature processing tasks in delay propagation prediction.

[0083] Specifically, in the recursive feature selection process, all variables participate in importance scoring and elimination judgment together without weight bias. This ensures the fairness of abnormal feature variables in multi-source information and is beneficial to screening key factors that have a practical impact on delay propagation.

[0084] The specific algorithm flow of variable selection by RFE algorithm is as follows:

[0085] Step one: select one variable in the first time series set as a target variable;

[0086] Step two: select variables in the time series set other than the target variable as candidate variables; build a candidate variable set corresponding to the target variable through the candidate variables;

[0087] Step three: score the importance of each variable in the candidate variable set through the recursive feature elimination algorithm and the target variable to obtain the importance score corresponding to the variable; delete the variable with the lowest importance score from the candidate variable set to form a new candidate variable set; determine whether the new candidate variable set meets the preset stopping condition;

[0088] Specifically, the method for scoring the importance of each variable in the candidate variable set through the recursive feature elimination algorithm and the target variable is:

[0089] a1: train a regression model through the candidate variable set and the target variable to obtain a trained regression model;

[0090] a2: score the importance of each variable in the candidate variable set through the trained regression model, and take the absolute value of the importance score to obtain the importance score corresponding to each variable; the importance score represents the contribution of the variable in the candidate variable set to the target variable.

[0091] Specifically, the importance score can be a weight value or a coefficient corresponding to each variable after training the regression model.

[0092] Step four: if the condition is not met, repeat the variable deletion process of step three using the new candidate variable set, and if the condition is met, stop iteration and take the new candidate variable set that meets the preset stopping condition as the selected condition variable set corresponding to the target variable;

[0093] Specifically, the preset stopping condition is whether the number of variables in the new candidate variable set meets the preset lower limit of the dimension, and whether the performance of the regression model decreases by more than the preset threshold.

[0094] Specifically, the performance of the regression model can be mean squared error or absolute error.

[0095] Step five: select the next variable in the first time series set as a target variable, and repeat the operations in steps two to four to screen out a corresponding selected condition variable set for each variable in the first time series set.

[0096] (3) Calculation method of transfer entropy

[0097] Since the operation processes of devices affect each other, the influence of other devices needs to be considered when calculating the transfer entropy between any two devices. For example, when the AGV transfers the container to the yard, it needs to consider not only the dynamic influence of other AGVs on its path, but also the scheduling and coordination with the yard bridge. Therefore, on the basis of two variables, the time series of the remaining operation devices are taken as the conditional variable set, and the transfer entropy method is proposed. The conditional variable set is Z={Z1,Z2,Z3...Z m}, and each time series variable in the conditional variable set is Z m =[z1,z2,z t-lag ...z t ]. The transfer entropy expression is as formula (5):

[0098]

[0099] Among them, X is the source variable, Y is the target variable; Y t =[y1,y2,y t-lag ...y t ] is the variable after time lag operation on Y; X t-lag =[x1,x2,...x t-lag ], Y t-lag =[y1,y2,...y t-lag ] represent the historical information of variables X, Y at t∈[1,t-lag] time; Z={Z1,Z2,Z3...Z m} represents the conditional variable set corresponding to Y in the variable combination; each time series variable in the conditional variable set is Z m =[z1,z2,z t-lag ...z t ], and Z t-lag represents the historical information of each variable in the conditional variable set at t∈[1,t-lag] time; H(Y t |Y t-lag ,Z t-lag ) represents the conditional entropy of Y t-lag under the condition of Y t-lag ,Z t ; H(Y t |X t-lag ,Y t-lag ,Z t-lag ) represents the conditional entropy of Y t-lag under the condition of X t-lag ,Y t-lag ,Z t ; P(y t ,x t-lag ,y t-lag ,zt-lag ) represents the joint probability distribution of x t ,x t-lag ,y t-lag ,z t-lag ; P(y t |x t-lag ,y t-lag ,z t-lag ) represents the conditional probability distribution of y t-lag ,y t-lag ,z t-lag given x t ; P(y t |y t-lag ,z t-lag ) represents the conditional probability distribution of y t-lag ,z t-lag given y t .

[0100] Specifically, in the calculation of the transfer entropy, due to the directionality of the transfer entropy, when calculating the transfer entropy of the variable combination (X, Y), it is necessary to distinguish and calculate the transfer entropy from X to Y and the transfer entropy from Y to X respectively.

[0101] Further, in the process of calculating the transfer entropy, if the source variable X is included in the selected conditional variable set, the calculation result of the transfer entropy will lose its meaning or degenerate to 0, so it is necessary to judge whether the source variable is included in the selected conditional variable set before calculating the transfer entropy; if included, the source variable in the selected conditional variable set is deleted to obtain the conditional variable set; if not included, the selected conditional variable set is directly used as the conditional variable set for calculation.

[0102] Further, suppose there are N job devices in total. To know whether there is a causal relationship between every two devices, the transfer entropy between them needs to be calculated. This is like pairing every two devices from the N devices. The variable of the first job device is taken as the target variable, and the variables of the other remaining N-1 devices are taken as the source variables and the target variable of the first job device for calculation; the variable of the second device is taken as the target variable, and the variables of the other remaining N-1 devices are taken as the source variables and the target variable of the second job device for calculation, and so on. Because the calculation result of the transfer entropy has directionality, the results obtained are different in direction due to the different positions of the variables, and therefore, N*(N-1) times of calculation are needed in total, so that the two-way directional causal relationship between all devices can be obtained.

[0103] (4) Delay propagation graph construction method

[0104] According to the calculation result of the transfer entropy, when TE>0, there is a causal relationship between the job equipment nodes i and j, and the source variable corresponds to the job equipment pointing to the target variable corresponding to the job equipment, otherwise there is no causal relationship between the job equipment nodes i and j. According to the obtained edge relationship between the job equipment nodes, a delay propagation graph is constructed;

[0105] Specifically, the delay propagation graph includes nodes, connections between nodes, and arrows between connections.

[0106] Specifically, the node represents a job equipment, the connection between the nodes indicates that there is a causal relationship between the two nodes, and the arrow between the connections indicates the direction of the influence between the nodes.

[0107] (5) Delay propagation model building

[0108] Based on the delay propagation graph constructed, the prediction analysis of delay propagation is carried out. It is known that the conduction of delay among multiple levels of job equipment is diffused from point to plane, which is similar to the propagation of disease in the network. At present, the SEIR model has been widely used in network information propagation, risk propagation and other aspects. Based on the SEIR model, the actual job characteristics of the terminal are combined, the exit of the equipment node is considered, and a dynamic multi-level job SEIR delay conduction model is constructed. The conversion relationship between various nodes is as shown in the following table: Figure 5

[0109] The multi-level job delay conduction model follows the following assumptions:

[0110] 1) When the normal job state node contacts the delay conduction node, it is converted into a delay node with a probability of α.

[0111] 2) Due to the existence of time windows between job links, they are not completely synchronized, and there is a mutual waiting situation between the shore crane, the yard crane and the AGV. Therefore, the equipment that appears delay will not immediately transmit the delay to the equipment that has a connection with it. The delay node will be converted into a delay conduction node with a probability of β, and the delayed equipment will automatically convert into a recovery state with a ratio of θ.

[0112] 3) Under the instant control of the terminal management party, the delay conduction node will be converted into a recovery node with a probability of γ; at the same time, part of the delay conduction nodes that cannot be restored to the normal state due to equipment failure will exit the job process with a probability of ρ.

[0113] 4) Unlike infectious disease immune nodes that will not be infected again, the immune state equipment will still be affected by the delay of other equipment in the job process and then delayed again. The immune node will be converted into a susceptible node with a probability of δ.

[0114] Based on the above assumptions, the basic SEIR model is improved to obtain the multi-level job delay conduction model, as shown in formula (1):​

[0115]

[0116] Among them, at time t, the normally operating equipment has an in-degree of m and an out-degree of n. The probability of being pointed to by a delayed transmission device is Define P((i,j)|(m,n)) as the connection probability, that is, the probability that a node with in-degree and out-degree of i and j connects to a node with in-degree and out-degree of m and n. This represents the number of devices with in-degree and out-degree of m and n that are operating normally at time t. This indicates the number of devices in a delayed state at time t, with in-degree and out-degree of m and n, respectively. This represents the number of devices whose in-degree and out-degree are delayed to propagate at time t; This represents the number of devices whose node connections with in-degree and out-degree of m and n have been restored at time t; m represents the in-degree of the selected node connections; n represents the out-degree of the selected node connections; N represents the total number of working devices; I i,j (t) represents the number of devices at time t that have delayed propagation to in-degree and out-degree i, j, where N is the number of such devices. t This represents the total number of all devices at time t.

[0117] P((i,j)|(m,n)) represents the connection probability between nodes i and j. Please provide the original text that needs translation. At time t, the number of normal processing devices, delayed state devices, delayed propagation devices, and recovery devices are respectively represented by S. t E t ,I t ,R t And it needs to meet the following conditions:

[0118] S t +E t +I t +R t =N t (2)

[0119]

[0120] Among them, S t E t ,I t ,R t These represent the number of normally operating devices, devices in a delayed state, devices transmitting delays, and devices recovering at time t, respectively; where N t This represents the total number of devices at time t.

[0121] Example 2:

[0122] Specifically, the second time series set is an abnormal auxiliary mechanism for enhancing the perception ability of the delay prediction model system, referred to as an equipment anomaly behavior modeling mechanism (EABM).

[0123] The core idea of the mechanism is to analyze the behavior characteristics of the equipment in the historical operation process, construct an abnormal feature variable set independent of the original operation time series, and use it as a supplementary input for the RFE conditional variable screening process. The abnormal features include, but are not limited to, the following three types:

[0124] (1) unplanned downtime or operation interruption frequency per unit time;

[0125] (2) standard deviation of operation duration sequence exceeding a certain time window;

[0126] (3) Z-score of the operation time of the adjacent equipment.

[0127] The above abnormal indicators are encoded as structured time series variables, i.e., the second time series set, which are uniformly screened with the first operation time series, and the screening results are used as the conditional variable set input for the transfer entropy calculation. The abnormal indicators are converted into structured time series variables, including binary encoding of abnormal occurrence state, numerical encoding of operation time residual, etc. The above abnormal auxiliary variable sequence and the original operation lag variable are uniformly constructed into a multi-order lag structure, and are jointly used as the input of the recursive feature elimination algorithm to evaluate the contribution of each type of variable to the prediction task. The overall structure of the enhanced screening process is shown in Figure 6 , which reflects the fusion process of abnormal auxiliary variables and traditional time series variables and their role path before transfer entropy calculation.

[0128] Specifically, the fusion process includes: uniformly constructing the original operation lag variable and the abnormal auxiliary variable into a multi-order lag structure, and aligning them in the same feature space before being used as the input of the variable screening module. All variables are uniformly preprocessed, scored, and screened to form a fusion candidate variable set, achieving cooperative modeling of equipment behavior time series dynamics and sudden abnormal behavior. This method enhances the perception ability of the conditional variable set to the system abnormal state, making the prediction model more robust and generalizable in dealing with sudden equipment failures or complex interactive disturbances.

[0129] Example Three:

[0130] 1. Delay propagation map construction

[0131] This invention constructs a delay propagation map based on the loading and unloading operation data of container ships berthing at a U-shaped automated terminal in China from 8:00 on May 12, 2023 to 16:00 on May 13, 2023. The data covers 54 pieces of equipment, including 6 quay cranes, 36 AGVs, and 12 yard cranes, and records the start and end times of operations for each piece of equipment. After outlier processing, the data format is shown in Table 1.

[0132] Table 1 Multi-level operation data of automated terminal

[0133]

[0134] Based on the operational time data of the equipment within the wharf, a time series is constructed for each piece of equipment, and the transfer entropy between any two pieces of equipment is calculated. A delay propagation map is then constructed based on the calculation results. The involved equipment is numbered, and the set of yard crane equipment is V. YC =[v1,v2,v3,...,v 12 The quay crane equipment assembly is V. QC =[v 13 ,v 14 ,v 15 ,...,v 18 ], AGV equipment collection is V AGV =[v 19 ,v 20 ,v 21 ,...,v 54 [The delayed transmission map is as follows] Figure 3 The characteristics of the delay propagation pattern are shown in Table 2. Overall, 128 direct causal relationships were formed among the 54 devices. The average degree was 4.74, meaning that each device interacted directly with an average of 4.74 other devices. The average path length between nodes was 2.656, indicating that causal relationships between devices could be established through an average of 2.656 other devices, showing that the mutual influence between devices was not limited to devices with direct operational task relationships. In addition, the average clustering coefficient was 0.042, and the low clustering indicates that the operations of each device were relatively independent, and no obvious clustering characteristics were formed between devices as operational needs and status changed.

[0135] Table 2. Structural Characteristics of Delayed Propagation Map

[0136]

[0137] 2. Multi-level operation delay propagation prediction

[0138] According to the constructed delay propagation graph, the delay propagation process in the terminal is simulated by Python, and the change trend of four types of nodes over time is analyzed. Combined with field research and literature analysis, the initial state is set as S0=51, E0=R0=0, I0=3, the parameters are α=0.5, β=0.3, γ=0.1, ρ=0.03, δ=0.2, and θ=0.1.

[0139] (1) Delay transmission analysis under different control strategies

[0140] In order to explore the transmission of delay among multi-level operation links under different control strategies, four control strategies are used for simulation analysis under the condition that other delay transmission parameters remain unchanged: no control ρ=0, γ=0; only dynamic control ρ=0, γ=0.1; only transmission device exit ρ=0.03, γ=0; and simultaneous control and device exit ρ=0.03, γ=0.1. Figure 7 The change of four types of state nodes over time under different control strategies is shown in Figure 7 (a), without any control, normal operation devices will be quickly affected by delay devices, leading to large-scale delay and persistent existence. As shown in Figure 7 (b)-(d), compared with Figure 7 (a), appropriate dynamic control of delay reduces the peak proportion of delay devices by 59.3%, and delay does not cause large-scale transmission, but only stable transmission in a small scale; only delay transmission device exit reduces the peak proportion of delay devices by 38.9%; under the double-strategy control, the peak proportion of delay devices is reduced by 66.7%, and compared with single strategy, delay dissipates faster. In summary, the measure of considering transmission device exit plays a positive role in inhibiting the spread of delay.

[0141] (2) Delay transmission analysis under different initial delay scales

[0142] In order to further explore the propagation of delay under different initial delay transmission device proportions, the initial delay transmission node numbers are set as 1, 3, 6, and 10 under the same device exit and control measures. The change of four types of state nodes under different initial delay scales is shown in Figure 8 .

[0143] As shown in Figure 8 , the more the initial delay transmission devices, the faster the delay propagation, and the more the affected normal operation devices. As shown in Figure 8 (a), when the initial delay scale is small, delay cannot be transmitted on a large scale among multi-level operation links, and will dissipate quickly; as shown in Figure 8(b)-(d) can be seen that with the increase of the initial delay ratio, the time to reach the peak of delay is faster. When the number of initial delay devices reaches 6 / 54 of the total number of devices, the change of delay propagation peak is no longer obvious with the increase of initial delay devices. In general, the number of normal operation devices gradually decreases over time, the number of delay devices and delay propagation devices increases rapidly in the initial stage, and then gradually decreases to 0 after reaching the highest value. It shows that the normal devices affected by the delay appear delay and transform into delay propagation state, and then gradually dissipate with the control strategy.

[0144] The embodiment is verified on the data set of a certain U-shaped automated wharf, and it is found that compared with the model without introducing the abnormal auxiliary mechanism, the key indicators such as delay peak identification accuracy and propagation range estimation are all improved by more than 10%. The method of the present application is not only suitable for the typical U-shaped automated wharf scene, but also can be applied to the task conflict identification and propagation modeling problem in the multi-source concurrent operation environment of the manufacturing system, the intelligent warehousing system and the like.

[0145] Finally, it should be noted that the above description is only used to illustrate the technical solutions of the present application and is not limited. Although the present application has been described in detail with reference to the preferred arrangement, those skilled in the art should understand that the technical solutions of the present application can be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present application.

Claims

1. An automated port operation delay propagation prediction method combining anomaly perception and transfer entropy analysis, characterized in that, S1: constructing a time series set: obtaining operation time data of internal operation equipment in the port and constructing a first time series set; obtaining abnormal index data of the operation equipment during operation; uniformly encoding the abnormal index data to obtain a second time series set; and fusing the first time series set and the second time series set into a time series set; S2: constructing a selected condition variable set: calculating the importance score of variables in the time series set by a recursive feature elimination algorithm, and selecting variables and constructing a selected condition variable set according to the importance score of the variables; S3: determining the causal relationship between operation equipment and constructing a delay propagation graph: calculating the transfer entropy between operation equipment according to the selected condition variable set; determining the causal relationship between operation equipment according to the calculation result of the transfer entropy; and constructing a delay propagation graph through the causal relationship between operation equipment; S4: constructing a multi-level operation delay propagation model: constructing a multi-level operation delay propagation model according to the transformation probability between operation equipment in different states and the SEIR model; S5: performing delay propagation prediction; the multi-level operation delay propagation model uses the causal relationship between equipment in the delay propagation graph to perform delay propagation prediction.

2. The automatic prediction method of port operation delay propagation based on the fusion of anomaly perception and transfer entropy analysis according to claim 1, characterized in that, The method for constructing the time series set in step S1 is: A1: dividing the operation time data at a fixed time interval to obtain divided operation time data; A2: calculating the average operation time of each operation equipment in each time interval in the divided operation time data; constructing an average operation time sequence of each operation equipment according to the average operation time; and constructing a first time series set through the average operation time sequence of each operation equipment; A3: obtaining abnormal index data of internal operation equipment; the abnormal index types include but are not limited to: operation behavior anomaly of operation equipment, operation equipment not operating according to the plan, path deviation of operation equipment, and non-planned start of operation equipment; A4: uniformly encoding the abnormal index data to obtain a second time series set consistent with the structure of the first time series variable set; the uniform encoding is binary encoding, numerical encoding, residual encoding or frequency encoding of the abnormal index data according to the characteristics of the abnormal index; A5: fusing the first time series set and the second time series set to obtain a time series set.

3. The method according to claim 1, wherein the method is characterized in that, The step of fusing the first time series set and the second time series set in step S1 is: B1: uniformly constructing the first time series set and the second time series set into a multi-order lag structure, and aligning them in the same feature space; B2: merging the aligned first time series set and second time series set to obtain a time series set.

4. The method according to claim 1, wherein the method is characterized in that, The method for constructing the selected condition variable set in step S2 is: C1: selecting any variable in the first time series set as a target variable; C2: selecting variables other than the target variable in the time series set as candidate variables; constructing a candidate variable set corresponding to the target variable through the candidate variables; C3: scoring the importance of each variable in the candidate variable set through a recursive feature elimination algorithm and the target variable to obtain an importance score corresponding to the variable; deleting the variable with the lowest importance score from the candidate variable set to form a new candidate variable set; determining whether the new candidate variable set meets a preset stopping condition; C4: if the new candidate variable set does not meet the preset stopping condition, repeating the variable deletion process of C3 using the new candidate variable set; if the new candidate variable set meets the preset stopping condition, stopping iteration and taking the new candidate variable set meeting the preset stopping condition as the selected condition variable set corresponding to the target variable; C5: selecting a next variable in the first time series set as the target variable and repeating steps C2-C4 to obtain the selected condition variable set corresponding to each variable in the first time series set.

5. The automated terminal operation delay propagation prediction method fusing anomaly perception and transfer entropy analysis according to claim 4, characterized in that, The method for scoring the importance of each variable in the candidate variable set through a recursive feature elimination algorithm and the target variable is: a1: training a regression model through the candidate variable set and the target variable to obtain a trained regression model; a2: scoring the importance of each variable in the candidate variable set through the trained regression model and taking the absolute value of the importance score to obtain an importance score corresponding to the variable; the importance score represents the contribution of the variable in the candidate variable set to the target variable.

6. The automatic terminal operation delay propagation prediction method based on the fusion of anomaly perception and transfer entropy analysis according to claim 4, characterized in that, The preset stopping condition includes whether the number of variables in the new candidate variable set meets a preset lower dimension limit and whether the performance of the regression model decreases by more than a preset threshold value.

7. The method according to claim 1, wherein the method is characterized in that, The method for determining the causal relationship between the job devices in step S3 is: D1: selecting a variable in the first time series set as the target variable; D2: selecting a variable other than the target variable from the first time series set as the source variable; determining whether the source variable exists in the selected condition variable set corresponding to the target variable, deleting the source variable in the selected condition variable set to obtain a condition variable set if the source variable exists, or taking the selected condition variable set as the condition variable set if the source variable does not exist; D3: taking the historical information of the target variable and the historical information of all variables in the condition variable set as conditions, calculating the conditional mutual information between the current information of the target variable and the historical information of the source variable to obtain the transition entropy between the job devices corresponding to the target variable and the source variable; D4: when the value of the transition entropy is greater than 0, the job devices corresponding to the target variable and the source variable have a causal relationship; otherwise, the job devices corresponding to the target variable and the source variable do not have a causal relationship; D5: reselecting the source variable from the first time series set; repeating steps D2-D4 to obtain the causal relationship between the job device corresponding to the target variable and all other job devices; D6: reselecting the target variable from the first time series set and repeating steps D2-D5 to obtain the causal relationship between all job devices.

8. The method according to claim 1, wherein the method is characterized in that, The delay propagation graph constructed in step S3 comprises nodes, connections between the nodes, and arrows between the connections; the nodes represent job equipment, the connections between the nodes indicate that there is a causal relationship between two nodes, and the arrows between the connections indicate the direction of the influence between the nodes.

9. The method according to claim 1, wherein the method is characterized in that, The conversion probability of the job equipment in different states in step S4 comprises: When the normal equipment is in contact with the delay conducting equipment, the normal job state node is converted into the delay node with a probability of α; The delay equipment is converted into the delay conducting equipment with a probability of β, and meanwhile, the delay equipment is converted into the recovery equipment with a probability of θ; The delay conducting equipment is converted into the recovery equipment with a probability of γ, and the delay conducting equipment exits the job process with a probability of ρ; The immune state equipment is converted into the susceptible state equipment with a probability of δ.

10. The method according to claim 1 or 9, wherein the method is characterized in that, The multi-stage job delay conducting model in step S4 is obtained based on the SEIR model, and the formula is: wherein at time t, the normal operating equipment with in-degree and out-degree m and n The probability of a delayed conducting equipment pointing to is Define P((i,j) \ (m,n)) as the connection probability, i.e. the probability of a node with in-degree and out-degree i,j connecting to a node with in-degree and out-degree m,n; denotes the number of nodes connecting to equipment with in-degree and out-degree m,n that are in normal operation at time t; denotes the number of nodes connecting to equipment with in-degree and out-degree m,n that are in a delayed state at time t: denotes the number of nodes connecting to equipment with in-degree and out-degree m,n that are in a delayed propagation at time t; denotes the number of nodes connecting to equipment with in-degree and out-degree m,n that are recovered at time t; m denotes the in-degree of the screened node; n denotes the out-degree of the screened node; N denotes the total number of operating equipment; I i,j (t) denotes the number of nodes connecting to equipment with in-degree and out-degree i,j that are in a delayed propagation at time t, N t denotes the number of all equipment at time t; At time t, the number of normal processing devices, delayed state devices, delayed propagation devices, and recovery devices are denoted as S t t t t and need to satisfy the following conditions:​​​ S t +E t +I t +R t =N t (2) where S t , E t , I t , and R t denote the number of normal operating equipment, delayed equipment, delayed conducting equipment, and recovery equipment at time t, respectively; and N t denotes the total number of equipment at time t.

Images