Intelligent wharf AGV operation time prediction method and system

Abstract

The invention relates to an intelligent wharf AGV operation time prediction method and system, and the method comprises the steps: obtaining the operation historical data and plan data of a wharf AGV, and the environment weather data of corresponding time; performing data cleaning on the acquired data, and performing classification, merging and coding of AGV operation logic place data, classification and identification of AGV performance, feature extraction of AGV load data, feature extraction of traffic jam conditions in an intelligent wharf area, matching and labeling of environment weather data and AGV operation data on the cleaned data to form a data set; constructing an AGV matrix based on the data set; and based on the AGV matrix, calculating AGV operation arrival time compensation value prediction under any working condition by using a Monte-Carlo algorithm, and outputting an AGV operation time prediction result based on the current working condition. Compared with the prior art, the AGV operation time prediction method has the advantages of quicker and more accurate AGV operation time prediction and the like.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent terminal operation, and in particular to an intelligent terminal AGV operation time prediction method and system. Background Art

[0002] An Automatic Guided Vehicle (AGV) is an unmanned container transport tool that works in automated container terminals. It is responsible for the transportation task of connecting the yard and the terminal front. The prediction of its operation time is crucial to the operation planning and scheduling of the entire terminal operation system.

[0003] In order to achieve accurate prediction of AGV operation time, several technical solutions are currently being studied. For example, Ji Mingjun et al., "Optimization of coordinated scheduling of container trucks and quay cranes at container terminals", Journal of Fudan University (Natural Science Edition), 2007 (4): 476-480+488, proposed a container truck route optimization system based on the shortest time. The system breaks down the container truck operation time into the delivery time, the loading time and the transportation time between the yard and the berth. The transportation time between the two points comes from the field measurement data of a container terminal and is stored in the time matrix for direct use each time the container truck operation is predicted. In the technical solution proposed in this document, the elements in the time matrix are static single values, which do not consider the uncertainty in the container truck transportation process, and do not distinguish between the influencing factors such as the type of the container truck itself, weather conditions, and traffic conditions. Therefore, it is impossible to make real-time and refined predictions of the operating machinery according to the actual working conditions of the terminal.

[0004] For example, HYK Lau et al., "Integrated Scheduling of Handling Equipment at Automated Container Terminals", International Journal of Production Economics, 2008, 112 (2): 665-682, proposed an AGV operation time prediction system that combines static and dynamic factors. The static factors of the AGV operation time are based on the fixed layout of the terminal and the static measurement values ​​of the AGV moving between different points, such as the travel time of the AGV from the yard interaction point to each quay crane, etc. These times are the shortest time required to go from one point to another under ideal conditions without interference; the dynamic factors refer to the uncertainty of the operation time, which is simulated by introducing random variables and generated from a uniform distribution, in which the mean plus or minus a certain percentage of deviation reflects the time changes that may occur in actual operation. The system described in this document takes into account the uncertainty of the AGV operation time to a certain extent, but does not take into account the congestion of the AGV on the guide path, and cannot achieve real-time prediction.

[0005] For example, P. Angeloudis et al., “An Uncertainty-aware AGV assignment algorithm for automated container terminals”, Transportation Research Part E: Logistics and Transportation Review, 2010, 46(3): 354-366, proposed an AGV operation time prediction system based on upper and lower bounds. The upper and lower bounds are obtained through simulation and are used to determine the distribution of AGV operation time. The operation time of the AGV falls within a range. After sufficient operation data is collected and sorted, the values ​​on the predetermined percentile are selected to set the upper and lower bounds. The uncertainty index is calculated based on the upper and lower bounds, and ILOG's CPLEX solver is directly used to generate the interval distribution of AGV operation time based on the upper and lower bounds. The technical solution described in this document does not effectively utilize the historical operation data of the terminal to achieve prediction, so the prediction result will deviate from the actual operation status of the terminal, resulting in unreliable prediction results.

[0006] For example, Zhang Si et al., "Integrated Scheduling of Quay Cranes and Container Trucks Considering Uncertainty", Logistics Technology, 2021, 40(4): 52-58+93, proposed an AGV operation time prediction system considering uncertainty factors. The uncertainty factors include equipment conditions and weather conditions, which cause the AGV operation time to fluctuate within a certain range. However, in the system described in this document, the unloading time and container truck driving speed under different equipment and different weather conditions are preset as empirical values, without the support of actual operating data. The prediction results are different from the actual operating conditions of the terminal, resulting in unreliable prediction results.

[0007] For example, Li Xingchun et al. proposed a data-driven AGV operation time prediction system in "Data-driven Automated Terminal Quay Crane and AGV Double-layer Optimization Scheduling Model", Frontiers of Engineering Management Science and Technology, 2024: 1-11. The system introduces the distributed robust optimization (DRO) method, establishes a Kantorovich fuzzy set to characterize the AGV travel time, and infers the true distribution information through historical data. The Kantorovich fuzzy set is defined by the reference distribution constructed by historical data. The technical solution described in this document does not clearly consider the impact of characteristics such as load weight, weather, number of turns, congestion, conflict, etc., which makes the prediction model lack accuracy.

[0008] In summary, although the existing technical work has achieved the prediction of the AGV operation time at the terminal to a certain extent, the existing technical solutions have the problems of being divorced from the actual working conditions of the terminal, being divorced from the actual operation data support, and not fully mining the characteristics of historical data, making it impossible for the existing technical solutions to accurately predict the AGV operation time in real time under various working conditions.

[0009] How to achieve rapid and accurate prediction of AGV operation time in intelligent terminals has become a technical problem that needs to be solved. Summary of the invention

[0010] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and provide an intelligent terminal AGV operation time prediction method and system.

[0011] The purpose of the present invention can be achieved by the following technical solutions:

[0012] According to one aspect of the present invention, a method for predicting the operation time of an AGV in an intelligent terminal is provided, the method comprising:

[0013] Obtain the terminal AGV operation history data and plan data, as well as the environmental weather data at the corresponding time;

[0014] The acquired data is cleaned, and the cleaned data is classified, merged and coded for the AGV operation logic location data, classified and identified the AGV performance, feature extraction of AGV load data, feature extraction of traffic congestion in the intelligent terminal area, matching and labeling of environmental weather data and AGV operation data to form a data set;

[0015] Construct an AGV matrix based on the data set;

[0016] Based on the AGV matrix, the Monte-Carlo algorithm is used to calculate the prediction of the AGV operation arrival time compensation value under any working conditions, and the prediction result of the AGV operation time is output based on the current working conditions.

[0017] Preferably, the data cleaning includes eliminating default values, outliers and abnormal data in terminal business in the acquired data. The strategy for eliminating outliers is: numerically counting the time consumed by all AGV transportation services, and sorting them from small to large based on the time consumed, to obtain the 95% quantile TP95. All data whose transportation service time consumed is greater than TP95 are considered invalid outliers and are eliminated.

[0018] Preferably, the classification and identification process of AGV performance includes the following steps:

[0019] Identify and cluster different operation data of the same AGV, and count the arrival time of each AGV under different services;

[0020] Perform a pairwise cross T test on the data of different AGVs to form a T test matrix;

[0021] Spectral clustering was used to classify based on the T-test matrix and two types of AGVs with large performance differences were separated.

[0022] Preferably, the feature extraction of AGV load data includes extracting the empty and loaded vehicle status and box weight data of the AGV vehicle in the AGV operation data, and encoding the corresponding label of the AGV operation data based on the empty and loaded vehicle status and box weight of the AGV vehicle to form an AGV load label;

[0023] The classification and merging coding of the AGV operation logical location data is specifically as follows: the AGV operation logical location data includes the container area number and the quay crane interaction point number. By identifying the leading character string used to identify the location object type, the starting point and arrival point types of the AGV operation, as well as the corresponding number values, are extracted, wherein the location object type includes the container area number and the quay crane interaction point number.

[0024] Preferably, the matching and labeling process of the environmental weather data and the AGV operation data includes: discretizing the precipitation data of the automated terminal area accurate to the hour and equivalent to the effective precipitation per minute, establishing a corresponding precipitation data set according to the date and time corresponding to the AGV operation data, calculating the matching precipitation in the operation time period based on the operation time of the AGV, and adding a rainfall level classification label to the AGV operation data, wherein the label assignment rule is:

[0025]

[0026] Where LABEL R is the weather feature label value, and R is the matching precipitation.

[0027] Preferably, the construction of the AGV matrix based on the data set is specifically as follows: the AGV operation arrival time is modeled as a multidimensional matrix, and each dimension of the multidimensional matrix is ​​a factor that affects the prediction result of the AGV operation arrival time, wherein the factors include the AGV operation departure point, the AGV operation destination, the AGV model ID label, the precipitation label, the AGV load label and the current operation traffic congestion label, and the corresponding value of the multidimensional matrix is ​​the actual arrival time distribution of all historical operations that meet the current input factors.

[0028] Preferably, the terminal AGV operation history data includes the AGV's starting point, operation destination, departure time, terminal operating system estimated arrival time, AGV actual arrival time, AGV mileage at departure time, AGV mileage at arrival time, AGV's own ID, AGV load box number and its corresponding box weight for each operation.

[0029] Preferably, the method further comprises: based on the AGV matrix, using a deep neural network prediction model to output the prediction result of the AGV operation time in real time, the deep neural network prediction model comprises an input layer, a hidden layer and an output layer, each layer comprises a number of neurons, and two adjacent layers of neurons are interconnected in a fully connected manner;

[0030] The input features of each input dimension of the AGV matrix under the working condition to be predicted are input into the input layer of the trained deep neural network, and the prediction result of the AGV operation time is output.

[0031] According to another aspect of the present invention, an intelligent terminal AGV operation time prediction system is provided, the system comprising a data processing function module and an algorithm prediction function module, wherein the data processing function module performs data cleaning and feature extraction on the acquired terminal AGV operation data and the environmental weather data of the corresponding time to form a data set; the algorithm prediction function module constructs an AGV matrix based on the data set formed by the data processing function module, executes a prediction algorithm based on the AGV matrix, and outputs an AGV operation time prediction result based on the current working conditions;

[0032] The prediction algorithms include two implementation schemes: Monte-Carlo algorithm and deep neural network prediction model.

[0033] Preferably, the data processing function module includes a data cleaning submodule and a feature engineering submodule;

[0034] The data cleaning submodule is used to remove default values, outliers and abnormal data in terminal business in the acquired data;

[0035] The feature engineering submodule is used for classification and merging of AGV operation logical locations, classification and identification of AGV performance, feature extraction of AGV load data, feature extraction of traffic congestion in the intelligent terminal area, and matching and labeling of environmental weather data and AGV operation data.

[0036] Compared with the prior art, the present invention has the following beneficial effects:

[0037] 1) The present invention incorporates multiple factors that affect the operating conditions of AGVs, including weather conditions in the working environment, traffic conditions in the terminal area, AGV load conditions, etc., into the input data, and after multi-feature extraction, various multi-scale and multi-modal factors are characterized by the same data set and stored by the same storage medium; an AGV matrix is ​​constructed based on the data set, and a prediction algorithm is executed to predict the AGV operation time. Compared with the existing technology, it can quickly realize the AGV operation time prediction with lower error, provide effective data support for the terminal loading plan, and has huge potential market value.

[0038] 2) The present invention applies the Monte-Carlo algorithm to the prediction of AGV operation time in automated container terminals for the first time. Based on the Monte-Carlo algorithm, the predicted value of AGV operation time and its corresponding probability can be accurately given.

[0039] 3) The present invention can also use deep neural networks to efficiently predict real-time AGV operation time, and use GPUs for parallel acceleration to perform efficient data-driven model training, thereby achieving real-time and accurate prediction driven by real-time data. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 It is a schematic diagram of the principle of the AGV operation time in the present invention;

[0041] Figure 2 It is a schematic diagram of the structure of the prediction system in the present invention;

[0042] Figure 3 It is a schematic diagram of the structure of the electronic device and the storage medium in the present invention;

[0043] Figure 4 It is a schematic diagram of the algorithm flow of the data processing function module in the present invention;

[0044] Figure 5 It is a schematic diagram of the prediction algorithm flow in the present invention;

[0045] Figure 6 It is the AGV matrix definition diagram in the present invention;

[0046] Figure 7 The figure is a comparison diagram of the error percentiles of the present invention and the existing method. DETAILED DESCRIPTION

[0047] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0048] The present invention comprehensively considers the influencing factors such as AGV model ID (characterizing the type of AGV), precipitation, AGV load, current operation traffic congestion, etc., and forms a multidimensional matrix for AGV operation time prediction based on the actual arrival time distribution of historical operations to cover the AGV operation time distribution of various working conditions; and innovatively introduces two methods, the Monte-Carlo algorithm and the neural network, which can give the AGV operation time prediction value under the influence of uncertainty factors according to historical data under various working conditions. For the implementation plan based on the Monte-Carlo algorithm, the present invention can accurately provide a series of prediction values ​​and the corresponding probabilities of the prediction values, and perform high-confidence predictions; for the prediction scheme based on the neural network, the present invention can achieve fast and accurate predictions driven by real-time data under the operation data.

[0049] This embodiment relates to a method for predicting the operation time of an intelligent terminal AGV. Figure 2 , the method comprising:

[0050] Step 1: Obtain the historical operation data of the terminal AGV and the environmental weather data of the corresponding time;

[0051] Step 2: Clean the data obtained in step 1, extract features from the cleaned data, and form a data set;

[0052] Step 3, construct an AGV matrix based on the data set;

[0053] Step 4: Execute the prediction algorithm based on the AGV matrix and output the AGV running time prediction result based on the current working conditions.

[0054] This embodiment uses the intelligent terminal AGV operation time prediction system for prediction. This embodiment collects all AGV command data, operation records, and precipitation data accurate to hours for the corresponding date of a domestic intelligent container terminal from January 2024 to May 2024. The data recorded in the example include the departure point, arrival point, departure time, terminal operating system (TOS) estimated arrival time, AGV actual arrival time, AGV mileage at departure time, AGV mileage at arrival time, AGV's own ID (AGV vehicle's mechanical number), AGV load box number and its corresponding box weight. Through the intelligent terminal AGV operation time prediction system, the terminal AGV operation time prediction is realized based on historical records. In this embodiment, the original TOS system estimates the arrival time (T TOS The value) can be used to roughly estimate the arrival time of the AGV operation (i.e. Figure 7 However, the error is large. Therefore, in this embodiment, the prediction object is T TOS The error compensation value between the value and the actual arrival time, thus TOSThe error of the value is compensated to improve the accuracy of the prediction.

[0055] In the data processing process in step 2, the departure time and arrival time data are subtracted to restore the time and distance of each AGV transportation service; the default values, outliers, and abnormal data that do not conform to the business logic are gradually identified and eliminated. The outlier elimination strategy is based on quantile elimination. The time consumption of all AGV transportation services is numerically counted, and sorted from small to large based on the time consumption to obtain the 95% quantile TP95. All operation data with a transportation service time greater than TP95 are considered invalid outliers and are eliminated.

[0056] Perform feature engineering on the processed data, and merge and encode the data of the logical locations (departure point and arrival point) of the AGV operation; for the AGV on the intelligent terminal, the departure point and arrival point location data include: container area number and quay bridge interaction point number. By identifying the keywords of the location data field in the data set (i.e., the leading string used to identify the type of location object (location object includes container area number and quay bridge interaction point number) in the corresponding data field of the terminal TOS system), the departure point and arrival point type of each AGV transportation business, as well as the corresponding number value, can be extracted.

[0057] The AGV load information is extracted and encoded, and the load container weight is discretized into a numerical label. The label assignment rule is shown in the following formula.

[0058]

[0059] Analyze the performance of different AGVs based on the data distribution characteristics based on the AGV's own ID, identify AGVs with large performance differences and add labels;

[0060] Match the environmental weather data with the data characterizing the AGV service and encode them; match the precipitation data for the corresponding hour in the environmental weather data set accurate to the hour based on the time period of the start and end time of each AGV transportation service. Score the precipitation values ​​according to the national standard GB / T 28592-2012 of the China Meteorological Administration. Since the possibility of extreme weather such as rainstorms and heavy rainstorms is relatively small, in order to avoid data sparsity, the scores of extreme weather are merged. The label assignment rule is shown in the following formula, where LABEL R is the weather feature label value, and R is the matching precipitation value in millimeters per hour.

[0061]

[0062] The terminal working conditions corresponding to each piece of AGV business data are evaluated, the traffic congestion is analyzed, and the corresponding label code is attached. The evaluation method is to analyze each AGV transportation business data, extract the start and end time of the current operation, match the number of operation records that exist simultaneously in the intelligent terminal during the operation time period in the historical operation data, and attach the corresponding numerical label. The label encoding rule is shown in the following formula, where LABEL T is the traffic congestion status label value, N AGV The number of job records synchronized.

[0063]

[0064] After the above process, a data set is formed from the original data and stored in a storage medium, such as Figure 3 .

[0065] In step 3, the AGV matrix is ​​defined based on the job departure location, job destination, precipitation label, load label, and traffic congestion label of the data set, such as Figure 6 shown.

[0066] In step 4, the prediction algorithm can be implemented in two ways: Monte-Carlo algorithm and neural network prediction model. For the Monte-Carlo algorithm, any working condition of the intelligent terminal that affects the AGV arrival time prediction can be represented by a point on the AGV matrix. For any point on the AGV matrix, the distribution characteristics of all historical AGV operation data that meet the corresponding working conditions are statistically analyzed; based on the distribution characteristics, the distribution characteristics are used as the prediction object of this embodiment, that is, T TOS Relative to the probability distribution of the actual arrival time error, 1024 independent random experiments are conducted; the results of the random experiments are statistically analyzed, the mathematical expectation and standard deviation are calculated, and the predicted result x of the AGV operation arrival time compensation value under any working condition is given based on the mathematical expectation; the predicted results are organized in the form of a matrix, and the predicted result matrix is ​​output to the storage medium. For the neural network method, the deep neural network with input layer, hidden layer and output layer is trained by error back propagation using the data set, and the neural network weight parameters are formed through iteration, thereby constructing a neural network prediction model.

[0067] In the actual terminal business prediction process, for any business under the terminal working conditions, the corresponding AGV operation arrival time compensation value prediction result x can be found in the output matrix, and the AGV operation arrival time prediction value can be calculated by the following formula. In the formula, T TOS is the estimated arrival time of the TOS system input, x is the prediction result compensation value obtained by the present invention, T p is the predicted AGV operation arrival time in this embodiment.

[0068] T p =(x×T TOS )+T TOS

[0069] In the above example, the predicted arrival time of the intelligent terminal AGV from the container area to the temporary parking point of the quay crane is as follows: Figure 7 The operation data from January 1, 2024 to May 20, 2024 are used to establish the prediction method, and the operation data from May 20, 2024 to May 31, 2024 are used to verify the prediction results. After testing, the average absolute error of the Monte-Carlo method proposed in the present invention is 0.15, and the average absolute error of the neural network model is 0.09, which is significantly lower than the prior art (0.32). Compared with the prior art, the prediction error of the present invention is reduced by 53% when the Monte-Carlo algorithm is used, and the prediction error is reduced by about 71% when the neural network algorithm is used.

[0070] Figure 7 The results show that the method proposed by the present invention is significantly better than the prior art in terms of error percentile. For cases with an error greater than 0.5, the error percentile is 7.61% when the Monte-Carlo algorithm is used for prediction, and the error percentile is 1.65% when the neural network algorithm is used for prediction, which is significantly lower than 18.41% of the prior art. It can be seen that the introduction of the present invention in this embodiment significantly reduces the probability of large error misjudgment events. The error calculation method in this embodiment is shown in the following formula.

[0071]

[0072] In the formula, E is the error, T p is the predicted value of this method, T R is the actual value of the test set.

[0073] It can be seen from the verification results of this embodiment that the intelligent terminal AGV operation time prediction system realizes the prediction of AGV operation arrival time under various working conditions based on Monte-Carlo algorithm and neural network by collecting actual terminal operation data and environmental weather data. The technical method proposed in the present invention can significantly reduce the prediction error and improve the prediction accuracy. It has a wide range of applications in intelligent container terminals and can provide data support for intelligent container terminal loading plans, with huge potential market value.

[0074] This embodiment also relates to an intelligent terminal AGV operation time prediction system, including a data processing function module and an algorithm prediction function module, such as Figure 1 and Figure 2 .

[0075] Data processing function module: The acquired terminal AGV operation data (including operation history data and plan data, where the operation history data is used to train the model of the prediction algorithm) and the environmental weather data of the corresponding time are removed through the data cleaning submodule to remove outliers in the original data, and the feature engineering submodule is used to extract features in the original data to form a data set. AGV operation data includes AGV operation starting point, destination point, load information, etc., which comes from the physical and logical data related to AGV business recorded by the terminal's own terminal operating system (TOS).

[0076] like Figure 4 The data processing function module includes two sub-modules: data cleaning and feature engineering. The data cleaning sub-module can effectively remove default values, outliers and abnormal data in the original data and improve the accuracy of the original data in representing the AGV business. The feature engineering sub-module can realize the classification and merging of AGV operation logical locations, the classification and identification of AGV performance, the feature extraction of AGV load data, the feature extraction of traffic congestion in the intelligent terminal area, and the matching and labeling of environmental weather data and AGV operation data.

[0077] The classification and identification process of AGV performance includes the following steps:

[0078] 1) Identify and cluster different operation data of the same AGV, and count the arrival time of each AGV under different working conditions;

[0079] 2) Based on the results of 1), perform a pairwise cross T test on the data of different AGVs to form a T test matrix;

[0080] 3) Spectral clustering is used to classify based on the T-test matrix to separate two types of AGVs with large performance differences.

[0081] The feature extraction method of AGV load data is used to extract the empty and loaded vehicle status and box weight data of the AGV vehicle in the AGV operation data, and encode the corresponding labels of the AGV operation data based on the above data.

[0082] The matching and labeling method of environmental weather data and AGV operation data discretizes the precipitation data in the area where the automated terminal is located with accuracy to the hour and converts it into effective precipitation per minute. The corresponding precipitation dataset is established according to the date and time corresponding to the AGV operation data, and the equivalent precipitation in the operation time period is calculated based on the operation time of the AGV. In addition, rainfall level classification labels are added to the AGV operation data according to the national standard GB / T 28592-2012 of the China Meteorological Administration.

[0083] The algorithm prediction function module builds an AGV matrix based on the data set formed by the data processing function module, and executes the prediction algorithm based on the AGV matrix. The prediction algorithm includes two implementation schemes: Monte-Carlo algorithm and neural network prediction model; the prediction result output process outputs the AGV running time prediction result based on the current working conditions.

[0084] like Figure 5 The algorithm prediction process includes two steps: building the AGV matrix and executing the prediction algorithm. Figure 6 ,AGV matrix models the AGV operation arrival time as a multi-dimensional matrix. Each dimension of the multi-dimensional matrix is ​​the factor that affects the prediction result of the operation arrival time, including the AGV operation departure place, AGV operation destination, AGV model ID label, precipitation label, AGV load label and current operation traffic congestion label. The corresponding value of the multi-dimensional matrix is ​​the actual arrival time distribution of all historical operations that meet the current input factors. The prediction algorithm includes two implementation methods: Monte-Carlo based prediction algorithm and neural network prediction model.

[0085] The theoretical basis of the Monte-Carlo algorithm mainly includes the law of large numbers and the central limit theorem. The basic theory of the algorithm believes that a large number of independent and identically distributed random experiments will present an approximate normal distribution, that is, for a sequence of independent and identically distributed random variables, it satisfies:

[0086]

[0087] The Monte-Carlo algorithm is based on the distribution characteristics of the actual arrival time of historical jobs in the AGV matrix dimension. Driven by historical data, it constructs a probability model based on data distribution, and constructs a random sampling experiment based on the probability model to generate a large number of random sampling test samples. Finally, an approximate solution to the problem is obtained by statistical analysis of the random sampling experimental samples, that is, the job arrival time prediction for any job situation on the AGV matrix is ​​obtained.

[0088] This deep neural network is a classic multi-layer fully connected deep neural network, which includes: input layer, hidden layer and output layer; the number of input layer layers is 1, the number of neurons is the same as the input parameter dimension, which is 6 in the test case of the present invention; the hidden layer contains 3 layers, and the number of neurons in the three hidden layers decreases layer by layer, the number of neurons is 64, 64, and 32, and the output layer is 1 layer, the number of neurons is 1. Any two layers of the neural network are connected to each other in a fully connected manner. For each neuron in each layer, there is an input-output relationship:

[0089] Output=f(W t Input

[0090] Among them, f is the activation function, Input is the input vector, W t is the weight vector.

[0091] During the neural network training process, the input parameters of the input layer are the working condition characteristics corresponding to the historical operation data under the AGV matrix dimension, and the output layer label is the AGV historical operation time data. The training process is based on error back propagation, and the weight vector is iteratively updated until convergence. During the prediction process, the input layer inputs the working condition characteristics of each dimension of the AGV matrix under the working condition to be predicted, and the output layer outputs the predicted value of the operation time. The predicted value of the AGV operation time under the input working condition can be obtained by the neural network through layer-by-layer forward propagation calculation.

[0092] The two prediction models can be used for predictions in different application scenarios: the prediction results given by the Monte-Carlo method include the predicted value and the probability corresponding to the predicted value, which is suitable for predictions that require credibility and can assist terminals in making more accurate assessments of operating conditions; the deep neural network method can use GPU hardware for acceleration, greatly shortening the time required to train the prediction model, and can be used in application scenarios with higher timeliness requirements, achieving real-time training and real-time prediction.

[0093] Compared with the prior art, the present invention incorporates other factors that affect the operating conditions of AGVs, including weather conditions in the working environment, traffic conditions in the terminal area, AGV load conditions, etc., into input data for the first time, and has lower prediction errors than the prior art; the present invention represents the various multi-scale and multi-modal factors through the same data set and stores them through the same storage medium; the present invention applies the Monte-Carlo algorithm to the prediction of AGV operation time in automated container terminals for the first time, and can accurately give prediction values ​​and their corresponding probabilities based on the Monte-Carlo algorithm; the neural network described in the present invention can be parallelized and accelerated through the GPU in the system described in the present invention, and efficient data-driven model training can be performed, thereby achieving efficient and accurate predictions driven by real-time data.

[0094] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of various equivalent modifications or replacements within the technical scope disclosed by the present invention, and these modifications or replacements should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be based on the protection scope of the claims.

Claims

1. A method for predicting the operation time of AGV in an intelligent terminal, characterized in that: The method comprises: Obtain the terminal AGV operation history data and plan data, as well as the environmental weather data at the corresponding time; The acquired data is cleaned, and the cleaned data is classified, merged and coded for the AGV operation logic location data, classified and identified the AGV performance, feature extraction of AGV load data, feature extraction of traffic congestion in the intelligent terminal area, matching and labeling of environmental weather data and AGV operation data to form a data set; Construct an AGV matrix based on the data set; Based on the AGV matrix, the Monte-Carlo algorithm is used to calculate the prediction of the AGV operation arrival time compensation value under any working condition, and the predicted value and corresponding probability of the AGV operation time are output based on the current working conditions.

2. The method for predicting the operation time of an intelligent terminal AGV according to claim 1 is characterized in that: The data cleaning includes eliminating default values, outliers and abnormal data in the acquired data. The strategy for eliminating outliers is to perform numerical statistics on the time consumed by all AGV transportation services, and sort them from small to large based on the time consumed to obtain the 95% quantile TP95. All data with a transportation service time greater than TP95 are considered invalid outliers and are eliminated.

3. The method for predicting the operation time of an intelligent terminal AGV according to claim 1 is characterized in that: The classification and identification process of AGV performance includes the following steps: Identify and cluster different operation data of the same AGV, and count the arrival time of each AGV under different working conditions; Perform a pairwise cross T test on the data of different AGVs to form a T test matrix; Spectral clustering was used to classify based on the T-test matrix and two types of AGVs with large performance differences were separated.

4. The method for predicting the operation time of an intelligent terminal AGV according to claim 1 is characterized in that: The feature extraction of the AGV load data includes extracting the AGV vehicle empty and loaded vehicle status and box weight data in the AGV operation data, and encoding the corresponding label of the AGV operation data based on the AGV vehicle empty and loaded vehicle status and box weight to form an AGV load label; The classification and merging coding of the AGV operation logical location data is specifically as follows: the AGV operation logical location data includes the container area number and the quay crane interaction point number. By identifying the leading character string used to identify the location object type, the starting point and arrival point types of the AGV operation, as well as the corresponding number values, are extracted, wherein the location object type includes the container area number and the quay crane interaction point number.

5. The method for predicting the operation time of an intelligent terminal AGV according to claim 1 is characterized in that: The matching and labeling process of the environmental weather data and the AGV operation data includes: discretizing the precipitation data of the automated terminal area with accuracy to the hour and converting it into effective precipitation per minute, establishing a corresponding precipitation data set according to the date and time corresponding to the AGV operation data, calculating the matching precipitation in the operation time period based on the operation time of the AGV, and adding a rainfall level classification label to the AGV operation data, wherein the label assignment rule is: Where LABEL R is the weather feature label value, and R is the matching precipitation.

6. The method for predicting the operation time of an intelligent terminal AGV according to claim 1 is characterized in that: The specific steps of constructing the AGV matrix based on the data set are as follows: the AGV operation arrival time is modeled as a multidimensional matrix, and each dimension of the multidimensional matrix is ​​a factor that affects the prediction result of the AGV operation arrival time, wherein the factors include the AGV operation departure point, the AGV operation destination, the AGV model ID label, the precipitation label, the AGV load label and the current operation traffic congestion label, and the corresponding value of the multidimensional matrix is ​​the actual arrival time distribution of all historical operations that meet the current input factors.

7. The method for predicting the operation time of an intelligent terminal AGV according to claim 1 is characterized in that: The terminal AGV operation history data includes each operation starting point, operation destination, departure time, terminal operating system estimated arrival time, AGV actual arrival time, AGV mileage at departure time, AGV mileage at arrival time, AGV own ID, AGV load box number and its corresponding box weight.

8. The method for predicting the operation time of an intelligent terminal AGV according to claim 1 is characterized in that: The method further includes: based on the AGV matrix, using a deep neural network prediction model to output the prediction result of the AGV operation time in real time, the deep neural network prediction model includes an input layer, a hidden layer and an output layer, each layer includes a number of neurons, and two adjacent layers of neurons are connected to each other in a fully connected manner; The input features of each input dimension of the AGV matrix under the working condition to be predicted are input into the input layer of the trained deep neural network, and the prediction result of the AGV operation time is output.

9. A prediction system for AGV operation time of an intelligent terminal, the system executing the method as claimed in any one of claims 1 to 8, characterized in that: The system includes a data processing function module and an algorithm prediction function module, wherein the data processing function module performs data cleaning and feature extraction on the acquired terminal AGV operation data and the environmental weather data of the corresponding time to form a data set; the algorithm prediction function module constructs an AGV matrix based on the data set formed by the data processing function module, executes a prediction algorithm based on the AGV matrix, and outputs an AGV operation time prediction result based on the current working conditions; The prediction algorithms include two implementation schemes: Monte-Carlo algorithm and deep neural network prediction model.

10. The intelligent terminal AGV operation time prediction system according to claim 9 is characterized in that: The data processing function module includes a data cleaning submodule and a feature engineering submodule; The data cleaning submodule is used to remove default values, outliers and abnormal data in terminal business in the acquired data; The feature engineering submodule is used for classification and merging of AGV operation logical locations, classification and identification of AGV performance, feature extraction of AGV load data, feature extraction of traffic congestion in the intelligent terminal area, and matching and labeling of environmental weather data and AGV operation data.

Images