A method for predicting key operation paths of container ships in automated terminals

Abstract

The present invention belongs to the field of automation technology, and specifically is a method for predicting the key operation route of a container ship at an automated terminal. The present invention predicts the key operation route of each container ship in real time based on the operation data of the on-site bridge crane at the automated terminal and the operation plan of each bay. The steps include: in the feature extraction stage, sampling is performed from the historical data of the automated terminal to extract the historical operation features and planned operation features of each container; in the training stage, the extracted features are input into a neural network model to predict and train the bay operation time; in the online prediction stage, the model obtained in the training stage is used to predict the operation time of each bay of the bridge crane, and the operation route with the longest time is obtained, which is marked as the key operation route. The accuracy of the method of the present invention in predicting key operation routes is significantly better than that of existing methods, and has better effects in actual environments.

Description

Technical Field

[0001] The invention belongs to the field of automation technology, and in particular relates to a method for predicting a key operation path of a container ship at an automated terminal. Background Art

[0002] The operation route of a container ship is a set of bays that can only be loaded and unloaded by one crane at any time. Bays that are not in the same operation route can use multiple cranes to operate simultaneously to improve loading and unloading efficiency. All bays in the same operation route can only be operated by one crane at any time. Therefore, the operation route with the longest operation time is the minimum operation time of the ship, that is, the operation time bottleneck of the ship. The prediction of the key operation route of a container ship is the prediction of the operation route with the longest operation time under a given operation plan. The prediction of the key operation route of a container ship is a key technology for automated terminals and has an important impact on decisions such as AGV scheduling in automated terminals.

[0003] The existing method is to take the operation road with the maximum sum of the number of containers in adjacent bays as the key operation road. However, the accuracy of this method is low. The main reason is that there are a lot of factors such as the number of different types of containers, different operation process efficiency, operation plan and past operation efficiency that affect the efficiency of loading and unloading containers on the operation road, which may lead to the wrong determination of the key operation road. Summary of the invention

[0004] In view of the above problems, the present invention proposes a method for predicting the key operation path of container ships in an automated terminal. The method of the present invention can more accurately predict the bay operation time and improve the prediction accuracy of the key operation path by extracting the historical operation characteristics of the bridge crane and the planned operation characteristics of each bay.

[0005] The key operation path prediction method for automated terminal container ships proposed in the present invention is divided into the following three stages:

[0006] (I) Feature extraction stage: The historical data of the automated terminal is sliced ​​according to time, and a large amount of historical operation data and planned operation data at different time points are sampled. After extracting features from the data, it is preprocessed into multivariate time series data. The specific steps are as follows:

[0007] (1) Sampling at fixed time intervals. Extract container information, bridge crane operation data, AGV operation data, rail crane operation data and other data within a fixed time window before the sampling time point. Extract the operation characteristics of each container. Convert information such as container length type and each mechanical operation process into numerical variables, and subtract the time point of each step in the container assembly line loading and unloading operation from the time point of the previous step to obtain the time of the step as one of the features. Arrange the container feature data of past operations according to the order of container operations as the bridge crane historical operation time series X h .

[0008] (2) Extract the container data of the planned operation of the bridge crane at the bay after the sampling time point from the historical data. Filter the information that can be obtained in real time during the field operation and convert it into numerical variables. Arrange the container characteristic data of the planned operation according to the order of the container planned operation as the bay planned operation time series data X p .

[0009] (3) Extract the remaining operation time at the bay after the sampling time point. Read the operation data of all containers at the bay from the historical data to obtain the bay completion time. Remove the cases where the time interval between two adjacent container operations exceeds a certain time in the data, and subtract the current time and the interval time from the bay completion time as the actual operation time. Obtain the remaining operation time of the operation path at the bay at the sampling time point. Regularize the remaining operation time.

[0010] (II) In the training phase, the multivariate time series data is input into a multi-layer neural network model for training, and the distance between the historical operation characteristics of the bridge crane and the planned operation characteristics of the bayonet and the prediction error loss function are used as constraints. The specific steps are:

[0011] (1) Construct a neural network model based on the historical operation feature extraction network E of the bridge crane h , Bayer planning task feature extraction network E p and a fully connected neural network.

[0012] (2) Bridge crane historical operation feature extraction network E h Hebei planning task feature extraction network E p These are two multi-layer neural network models with the same structure but different parameters. The calculation formula for each layer of the neural network is as follows:

[0013] X i+1 =f cnn (X i +f attn (X i )), (1)

[0014] Where X i represents the input data of the i-th layer, X 0 is the training data. cnn Represents a one-dimensional convolutional neural network, which consists of a convolutional layer, an activation function, and a pooling layer. attn is the self-attention mechanism function, and its calculation formula is as follows:

[0015]

[0016] where d k is the characteristic dimension of the container.

[0017] Finally, the l-dimensional extraction of historical operation features of the bridge crane is as follows:

[0018] f d =E h (X h ), (3)

[0019] Similarly, extract the l-dimensional Bayer planning job features:

[0020] f p =E p (X p ), (4)

[0021] (3) Define the distance between features:

[0022] L d =||f h -f p || 2 , (5)

[0023] By combining the historical operation features of the bridge crane, the planned operation features and the distance between the features, we can get the new feature X = [f h , f p , L d ];

[0024] The prediction time is obtained by using a fully connected neural network. The calculation formula of the fully connected neural network is as follows:

[0025] t p =WX+b, (6)

[0026] Among them, t p To predict the time required for the bridge crane to complete the planned operation at this bay, X is the new feature after splicing, and are the learnable parameters of the fully connected neural network.

[0027] (4) Define the prediction loss function as the square error between the predicted operation time and the actual operation time:

[0028] L p =(t p -t w ) 2 , (7)

[0029] The loss function is defined as:

[0030] L=L d +L p , (8)

[0031] The goal of training is to minimize the loss function L, that is:

[0032]

[0033] Among them, θ represents the training parameters of the model, including the convolution kernel of the one-dimensional convolutional neural network of the multi-layer neural network and the weight matrix and bias matrix of the fully connected neural network, and D is the size of the data set. Finally, the model parameters are updated using an algorithm based on stochastic gradient descent.

[0034] (III) Online prediction stage: Use the model obtained in the training stage to predict the time it takes for the bridge crane to complete each bay operation. Calculate the prediction results for the key operation path and update the parameters regularly. The specific steps are:

[0035] (1) For each bridge crane, read the operation data of a certain time window in the past from the real-time operation data in the automated terminal management system and extract the characteristics of each container unit to obtain the multivariate time series X of historical operations h Read the operation plan of each bay, extract the characteristics of each container unit to obtain the bay b k Multivariate time series X of scheduled jobs p .

[0036] (2) Use the model obtained in the training phase to predict the bridge crane s to complete the bay position b i The planned loading and unloading task operation time is:

[0037] t(b i )=W·[E h (X h ), E p (X p ),||E h (X h )-E p (X p )|| 2 ]+b, (10)

[0038] (3) Read the ship chart of the container ship in operation in the automated terminal management system, the bridge crane s and the bay position b s , the minimum spacing distance of the bridge crane is w, and the bay position is b k The relative position p k .

[0039] Traverse and obtain the operation path B including the operation position where the bridge crane s is located i :

[0040] B i = {b k , b k+1 , ..., b k+n}, p k+n ≤p k +w,b s ∈B i , (11)

[0041] Among them, B i A set of continuous bays including those that cannot operate when any bay is operating is called an operating path.

[0042] (4) Enumerate and obtain the set of all operation paths B = {B 1 , B 2 , .., B m}. Complete the continuous bay set B i Time required Select the operation path with the longest time i = argmax t(B i ), this value indicates the bay b containing the bridge crane s The longest time of all operation routes is called the longest time of the bridge crane, and the operation route is marked as the key operation route, and the results are output to the automated terminal management system.

[0043] (5) When the bridge crane completes the bay operation, the runtime data and results are incorporated into the data set. When the increase in the data set reaches a certain threshold, the model is retrained according to the training stage to update the parameters.

[0044] The method of the present invention is significantly better than the existing methods in predicting the accuracy of key operation roads and has better effects in actual environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 This is a schematic diagram of the container loading process at a certain position on a container ship at a certain moment.

[0046] Figure 2 This is a schematic diagram of the operation of a container ship's bridge crane. DETAILED DESCRIPTION

[0047] The present invention is further described below with reference to examples and drawings.

[0048] 1. Feature extraction stage, the specific steps are:

[0049] (1) The historical data of an automated terminal is sliced ​​into 30-minute slices. For the sampling time point t, Figure 1 It is a schematic diagram of the loading process of a certain position in the container at time point t. The gray part indicates that the container has been loaded on the ship. Extract the characteristics of each container that has been loaded on the ship by the bridge crane in the past three hours, such as the attribute characteristics of the container, including but not limited to the length, height, box type and other attributes, the characteristics of the machinery operating the container, including but not limited to the operation process of the main trolley, gantry trolley, rail crane and other machinery, and the operation time of each process.

[0050] For example, a container feature x h，i=(0, 2, 1, 1, 2, 128, ...) indicates that the container is being loaded, the main trolley operation process is a double crane, the container is a 40-foot high box, a refrigerated container, and the main trolley has grabbed the container for 128 seconds since the last time it was released.

[0051] Arrange each container feature according to the order of container loading into a historical operation time series X h =[x h，1 , x h，2 , ..., x h，s ], indicating that a total of s containers were loaded in the past three hours, of which Represents the n-dimensional historical operation characteristics of the i-th container that has completed loading.

[0052] (2) Figure 1 The white part indicates the container that has been scheduled to operate but has not yet completed the operation. When extracting the features of each container, the attribute features of the container are extracted and arranged according to the operation plan as the planned operation time series X p =[x p，1 , x p，2 , ..., x p，t ], indicating that there are t containers planned to be loaded at this location but not yet completed, among which Represents the m-dimensional planning operation characteristics of the i-th container planned to be shipped.

[0053] (3) Read the time t when the last container is loaded and unloaded f Assume that due to a fault during the operation of the bridge crane, the time interval between the two container gantry trolleys grabbing the containers is 1800 seconds, and the remaining operation time is t f -t-1800 seconds. Regularize the time, assuming the mean value of the bay completion time is μ and the standard deviation is σ, then the regularized result of the remaining time is

[0054] 2. Training phase: The specific steps are:

[0055] (1) Construct a neural network model based on the historical operation feature extraction network E of the bridge crane h , Bayer planning task feature extraction network E p and a fully connected neural network. Construct two 3-layer neural network models with the same structure but different parameters E h and E p . E h and E p are two 3-layer neural network models with the same structure but different parameters. The structure of each layer of the neural network is

[0056] X i+1 =f cnn (Xi +f attn (X i ))

[0057] Where X i represents the input data of the i-th layer, X i+1 The output data of this layer is used as the input data of the next layer. cnn Represents a one-dimensional convolutional neural network, which consists of a convolution layer, an activation function, and a pooling layer. Here, convolution and pooling operations are performed on the time dimension, and the activation function is the ReLU function: ReLU(x)=max(0,x).

[0058] f attn is the self-attention mechanism function, which is calculated as follows:

[0059]

[0060] where d k is the feature dimension. In the bridge crane historical operation feature extraction model E h The feature dimension in is the feature dimension d of the historical job time series k = n. In the Bay Plan Operation Feature Extraction Model E p The feature dimension in the middle is the feature dimension of the time series of the planned operation, that is, d k =m.

[0061] Randomly initialize the parameters of the convolutional neural network.

[0062] When extracting historical job features, X 0 As input data, we get the historical operation characteristics of the bridge crane f h =E h (X h ); Similarly, extract the Bay Plan operation feature f p =E p (X p ).

[0063] (2) Calculate the distance L between features d =||f h -f p || 2 , splicing bridge crane historical operation characteristics, Bay position planning operation characteristics and feature distance are unified feature vector X = [f h , f p , L d ], and use the fully connected neural network to predict the predicted time. The calculation formula is as follows:

[0064] t p =WX+b

[0065] where t pTo predict the time required for the regularized bridge crane to complete the planned operation at this bay, X is the new feature after splicing, and W and b are the learnable parameters of the fully connected neural network.

[0066] (3) Define the prediction loss function as the square error between the predicted operation time and the actual operation time

[0067] L p =(t p -t w ) 2

[0068] Define the loss function as L = L d +L p . Update the model parameters using an algorithm based on stochastic gradient descent.

[0069] 3. Online prediction stage, the steps are as follows:

[0070] (1) Read the real-time data of the automated terminal management system to obtain the crane operation data and planned operation data in the past three hours. Use the feature extraction method in the first stage to extract the historical operation features and planned operation features of each container to obtain the historical operation time series X h and the time series X of the planned jobs at position i p，i .

[0071] (2) Figure 2 The schematic diagram of the bridge crane operation of a container ship at an automated terminal. The minimum spacing distance of the bridge cranes at this terminal is 28 meters. The ship chart of the terminal management system is read to obtain the positions of some bays 201, 202, 203, 204, 205, 206, and 207 of the ship relative to bay 201, which are 0, 15, 22.5, 37.5, 52.5, 67.5, and 82.5 meters respectively. Bay 202 is the operating bay of bridge crane 1, and bay 206 is the operating bay of bridge crane 2. According to the minimum spacing distance of bridge crane operation and the ship chart, bay 201, bay 202, and bay 203 are one of the operating routes, that is, bay 203 and bay 201 are less than 28 meters away and cannot operate at the same time. Suppose the model predicts that it will take 14800, 7200, 13600 and 24000 seconds for crane 1 to complete the remaining container loading and unloading operations at bays 201, 202, 203 and 204, and it predicts that it will take 0, 9600 and 28000 seconds for crane 2 to complete the remaining container loading and unloading operations at bays 205, 206 and 207.

[0072] (3) Traverse all operation paths, operation path B 1 ={201, 202, 203} The predicted operation time is 14800+7200+13600=35600 seconds, operation path B 2={202, 203, 204} The predicted operation time is 7200+13600+24000=44800 seconds, operation path B 3 ={205, 206} The predicted operation time is 0+9600=9600 seconds, operation path B 4 ={206, 207} The predicted operation time is 9600+28000=37600 seconds. 2 ={202, 203, 204} The predicted operation time is the longest, so set the operation path B 2 It is the key operation road.

[0073] This specific example is applied to the Shanghai Yangshan Port Phase IV automated terminal, and the data set is the Shanghai Yangshan Port Phase IV automated terminal operation data. The prediction accuracy of key operation paths is shown in Table 1.

[0074] Table 1

[0075] Accuracy (%) Manual settings 83.3 Linear Regression 84.7 The present invention 87.7

[0076] As shown in Table 1, the manual setting method is to set the key operation route for the on-site staff. The linear regression method uses the number of different types of containers as features to predict the operation time. The prediction accuracy of the method of the present invention is better than that of the original method.

[0077] (3) Update the prediction results of the key operation path every 10 minutes and return the results to the automated terminal management system, and record the operation data and incorporate it into the new data set. When the bridge crane completes the bay, update the previously recorded bay completion time. Retrain the model and update the model parameters every month or a fixed time, or when the port structure changes.

Claims

1. A method for predicting key operation paths of container ships in automated terminals. It is characterized in that It is divided into three stages: (i) Feature extraction stage: The historical data of the automated terminal is sliced ​​according to time, and a large amount of historical operation data and planned operation data at different time points are sampled. Features are extracted from the data and then preprocessed into multivariate time series data; (ii) Training phase: inputting multivariate time series data into a multi-layer neural network model for training, and taking the distance between the historical operation characteristics of the bridge crane and the planned operation characteristics of the bayonet and the prediction error loss function as constraints; (3) Online prediction stage; Use the model obtained in the training phase to predict the time it takes for the bridge crane to complete each bay operation; calculate the prediction results for key operation routes and update the parameters regularly; (I) The specific steps of the feature extraction stage are: (1) Sampling is performed at fixed time intervals; the container information within the fixed time window before the sampling time point is extracted, including the bridge crane operation data, AGV operation data, and rail crane operation data; the operation characteristics of each container are extracted, including the container length type and each mechanical operation process information, and converted into numerical variables. The time point of each step in the container assembly line loading and unloading operation is subtracted from the time point of the previous step to obtain the time of the step as one of the features; the container feature data of the past operation is arranged according to the container operation sequence as the bridge crane historical operation time series X h ; (2) Extract the container data of the planned operation of the bridge crane at the bay after the sampling time point from the historical data; filter the information that can be obtained in real time during the on-site operation and convert this information into numerical variables; arrange the container characteristic data of the planned operation according to the order of the container planned operation into the bay planned operation time series data X p ; (3) Extract the remaining operation time at the bay after the sampling time point; read the operation data of all containers at the bay from the historical data to obtain the bay completion time; remove the cases where the time interval between two adjacent container operations exceeds a certain time in the data, and subtract the current time and the interval time from the bay completion time as the actual operation time; obtain the remaining operation time of the operation path at the bay at the sampling time point; and regularize the remaining operation time; (II) The specific steps of the training phase are: (1) Construct a neural network model based on the historical operation feature extraction network E of the bridge crane h , Bayer planning task feature extraction network E p and a fully connected neural network; (2) Bridge crane historical operation feature extraction network E h Hebei planning task feature extraction network E p There are two multi-layer neural network models with the same structure but different parameters; the calculation formula of each layer of the neural network is as follows: X i+1 =f cnn (X i +f attn (X i )), (1) Among them, X i represents the input data of the i-th layer, X 0 is the training data; f cnn represents a one-dimensional convolutional neural network, which consists of a convolutional layer, an activation function, and a pooling layer; f attn is the self-attention mechanism function, and its calculation formula is as follows: Among them, d k is the characteristic dimension of the container; Finally, the l-dimensional extraction of historical operation features of the bridge crane is as follows: f h =E h (X h ), (3) Similarly, extract the l-dimensional Bayer planning operation features: f p =E p (X p ), (4) (3) Define the distance between features: L d =||f h -f p || 2 , (5) By combining the historical operation features of the bridge crane, the planned operation features and the distance between the features, we can get the new feature X = [f h ,f p ,L d ]; The prediction time is obtained by using a fully connected neural network. The calculation formula of the fully connected neural network is as follows: t p =WX+b, (6) Among them, t p To predict the time required for the bridge crane to complete the planned operation at this bay, X is the new feature after splicing, and are the learnable parameters of the fully connected neural network; (4) Define the prediction loss function as the square error between the predicted operation time and the actual operation time: L p =(t p -t w ) 2 , (7) The loss function is defined as: L=L d +L p , (8) The goal of training is to minimize the loss function L, that is: Among them, θ represents the training parameters of the model, including the convolution kernel of the one-dimensional convolutional neural network of the multi-layer neural network and the weight matrix and bias matrix of the fully connected neural network, and D is the size of the data set; finally, the model parameters are updated using an algorithm based on stochastic gradient descent; (III) The specific steps of the online prediction stage are: (1) For each bridge crane, read the operation data of a certain time window in the past from the real-time operation data in the automated terminal management system and extract the characteristics of each container unit to obtain the multivariate time series X of historical operations h ; Read the operation plan of each bay, extract the characteristics of each container unit to obtain the bay b k Multivariate time series X of scheduled jobs p ; (2) Use the model obtained in the training phase to predict the bridge crane s to complete the bay position b i The planned loading and unloading task operation time is: t(b i )=W·[E h (X h ),E p (X p ),||E h (X h )-E p (X p )|| 2 ]+b, (10) (3) Read the ship chart of the container ship in operation in the automated terminal management system, the bridge crane s and the bay position b s , the minimum spacing distance of the bridge crane is w, and the bay position is b k The relative position p k ; Traverse and obtain the operation path B including the operation bay where the bridge crane s is located i : B i ={b k ,b k+1 ,…,b k+n },p k+n ≤p k +w,b s ∈B i , (11) Among them, B i A continuous set of bays including those where other bays cannot operate when any bay is operating is called an operating path. (4) Enumerate and obtain the set of all operation paths B = {B 1 ,B 2 ,..,B m }; Complete the continuous bay set B i Time required Select the operation path with the longest time i = argmax t(B i ), this value indicates the bay b containing the bridge crane s The longest time of all operation routes is called the longest time of the bridge crane, and the operation route is marked as the key operation route, and the result is output to the automated terminal management system; (5) When the bridge crane completes the bay operation, the runtime data and results are incorporated into the data set; when the increase in the data set reaches a certain threshold, the model is retrained according to the training stage to update the parameters.

Images