Marine container position distribution method based on attribute double-view stack learning network

Abstract

The invention provides a marine container position distribution method based on an attribute double-view stack learning network, and the method comprises the steps: 101, carrying out the coding of the attributes of containers through employing a hybrid coding method for comprehensively integrating the attributes of various containers, and obtaining initial embedded data; 102: performing sequential view stack state learning with a gated loop unit (GRU)-based aggregator using the initial embedded data to capture sequential patterns or sequential views between the containers; 103, using the initial embedded data, and adopting a double-view container aggregator to capture common characteristics of the container to obtain double views; 104, integrating the sequential view and the double views to obtain a final stack representation and performing scoring; and 105, performing robustness training based on the adversarial network to ensure the robustness and generalization ability of the model. The method can adapt to the inherent variability and complexity of real world operation, effectively improves the performance of container position distribution, and is of great practical significance.

Description

Technical Field

[0001] The present disclosure relates to the field of information technology, and in particular to a method for allocating shipping container locations based on an attribute dual-view stack learning network. Background Art

[0002] With the growing demand for global trade transportation, the shipping container market occupies a pivotal position in marine terminals. Container location allocation is one of the most critical issues in container transportation, which is directly related to transportation efficiency and production safety. According to the United Nations Conference on Trade and Development, the container throughput of terminals has increased significantly in the past few decades. In 2020, the global container terminal throughput was 815.6 million TEUs, an increase of 53.5% over 2010. The substantial increase in container transportation has obviously increased the workload of terminals, further highlighting the need for efficient and automated operations at terminals. Summary of the invention

[0003] The present disclosure provides a method for location allocation of shipping containers based on attribute dual-view stack learning network, called DLSA, which is used to capture and utilize implicit relationships between containers in a stack to confirm the location for arriving containers, greatly improving the performance and efficiency of container location allocation.

[0004] The embodiment of the present disclosure provides a method for allocating shipping container locations based on an attribute dual-view stack learning network, comprising: 101: encoding container attributes by using a hybrid encoding method that comprehensively integrates various container attributes to obtain initial embedding data; 102: using the initial embedding data obtained in step 101, using a gated recurrent unit (GRU)-based aggregator to perform sequential view stack state learning to capture sequential patterns or sequential views between containers; 103: using the initial embedding data obtained in step 101, using a dual-view container aggregator to capture common features of containers to obtain a dual view; 104: integrating the sequential view obtained in step 102 and the dual view obtained in step 103 to obtain a final stack representation and score it; 105: performing robustness training based on an adversarial network to ensure the robustness and generalization ability of the model.

[0005] In some embodiments, the hybrid encoding method of integrating various container attributes is used for encoding in step 101 to obtain initial embedded data, specifically as follows: general encoding techniques are used for different types of attributes, including single-hot encoding, ordinal encoding, integer encoding, and binary encoding, which encode most attributes of the container to obtain numerical data input; the encoding of the departure time is one of the key issues in terminal scheduling in port operations, and the time encoding based on the Mercer kernel is used, and the departure time embedding obtained is as follows:

[0006]

[0007]

[0008] Where Φ(t) represents the high-dimensional feature vector at time t, w k and represents the learnable parameters in the kernel function, and N is the total number of time points considered in the encoding process;

[0009] To integrate these different types of container attributes, these encoding vectors are concatenated and a positional feed-forward network (FFN) is applied to capture non-linear dependencies as shown below:

[0010] a i =[a1||...||a n ]W 0

[0011]

[0012] where a m is the encoding vector of the mth container attribute, is a learnable parameter, || is a connection; finally, we get It is a container v i The initial embedding of .

[0013] In some embodiments, the initial embedding data obtained in step 101 is used in step 102 to learn the sequential view stack state using a gated recurrent unit-based aggregator, specifically as follows: there is an implicit relationship between the containers in the stack, and the sequential view stack state is first learned using a gated recurrent unit-based aggregator to capture the sequential pattern between the containers, as shown below:

[0014]

[0015] Among them, e k is the input of step k, is the stack state at step k, z k is the update gate output of step k, r k is the reset gate output of the kth step; W z , W r , W h is the learnable weight matrix; σ(·) is the sigmoid function, || is the concatenation operation, and ⊙ is the Hadamard product; thus, the current stack S is obtained k ={e1, e2, ..., e k} is a sequential view representation of .

[0016] In some embodiments, the initial embedding data obtained in step 101 is used in step 103 to capture the common features of containers using a dual-view container aggregator, specifically as follows: the containers in the stack show implicit correlations, and identifying such commonalities is beneficial for fine-grained stack representation; the dual-view container aggregator is used to capture the common features of the containers; first, the stack S is calculated k ={e1, e2, ..., e k The cross feature e of each container-container pair in ij , as shown below:

[0017] z ij =e i ⊙e j

[0018] where e i It is a container v i The embedding of , ⊙ represents the element-wise multiplication to emphasize the feature dimensions with similar values ​​between the two container embeddings; then, the stack representation is calculated by performing pairwise aggregation of the containers As shown below:

[0019]

[0020] Where σ is the ReLU function, which filters out irrelevant common features; the calculation can be simplified to:

[0021]

[0022] In some embodiments, the sequential view and the dual view are integrated as described in step 104 to obtain a final stack representation and score it as follows: The container stacked later is more representative of the current state:

[0023]

[0024] The pooling is implemented using average pooling; h k It is constructed by fusing multiple container-container relationships, including sequential transformation information and pairwise commonality information; after obtaining the current stack S k The expression h k Then, calculate the candidate container v in the stack x Fraction of container v x The score is:

[0025] r((v x , a k )|S k )=f(v x , S k )=f(a x , g(a1, a2, ..., ak ))=e x ·h k

[0026] where f(·) is the container stack scoring function, which applies to all stacks and allows candidate containers to be assigned to stacks with higher scores.

[0027] In some embodiments, the robustness training based on the adversarial network described in step 105 is performed to ensure the robustness and generalization ability of the model, specifically as follows: using the capabilities of generative adversarial networks (GANs) to generate synthetic data similar to real-world scenes and train it on DSLA, which has a generator and a discriminator;

[0028] The goal of the discriminator is to distinguish these generated samples from real data; this is done by minimizing the negative log-likelihood:

[0029]

[0030] where f φ (,) represents the score function of the discriminator; the objective function The purpose is to give the real container (v k+1 ) assigns a higher score to the generated container Assigning lower scores, thereby enhancing the discriminator’s ability to distinguish between real containers and generated fake data;

[0031] The goal of the generator is to generate high-quality negative containers that confuse the discriminator; a key aspect is how to sample a negative container Assume f φ (v i , S k ) indicates that a given stack S k The score function of the generator; container v i The sampling probability of is given by the softmax function:

[0032]

[0033] Fake container based on sampling The generator deceives the discriminator by minimizing the recognition score, which is defined as follows:

[0034]

[0035] Due to the discrete process of the generator sampling negative containers, gradient descent cannot be directly applied to parameter estimation; to address this challenge, the policy gradient method is used to update the parameters of the generator; it is defined as follows:

[0036]

[0037] Finally, the generator and discriminator participate in an iterative training process similar to the minimax game. This continuous adversarial training can effectively enhance the robustness of DSLA in practical scenarios.

[0038] The present disclosure proposes a dual-view stack state learning network (DSLA) for attribute-based container location allocation, which can effectively extract sequential patterns and feature commonalities in the current stack from sequential views and paired views, eliminating the need for limited and inflexible manual pre-definition. In addition, the present disclosure designs an efficient stack scorer that fully integrates the dual-view stack representation to evaluate candidate containers. In order to enhance the robustness in practical scenarios, an adversarial network is further used for model training. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] The above and other features, advantages and aspects of the embodiments of the present disclosure will become more apparent with reference to the following detailed description in conjunction with the accompanying drawings. Throughout the accompanying drawings, the same or similar reference numerals represent the same or similar elements. It should be understood that the drawings are schematic and that components and elements are not necessarily drawn to scale.

[0040] Figure 1 It is a flowchart of a method for allocating shipping container locations based on an attribute dual-view stack learning network according to an embodiment of the present disclosure.

[0041] Figure 2 is a schematic diagram of a marine terminal yard block according to an embodiment of the present disclosure.

[0042] Figure 3 It is an overview diagram of a method for allocating shipping container locations based on an attribute dual-view stack learning network according to an embodiment of the present disclosure. DETAILED DESCRIPTION

[0043] Embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although certain embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be construed as being limited to the embodiments described herein, which are instead provided for a more thorough and complete understanding of the present disclosure. It should be understood that the drawings and embodiments of the present disclosure are only for exemplary purposes and are not intended to limit the scope of protection of the present disclosure.

[0044] It should be understood that the various steps described in the method embodiments of the present disclosure can be performed in sequence and / or in parallel. In addition, the method embodiments may include additional steps and / or omit the steps shown. The scope of the present disclosure is not limited in this respect.

[0045] The term "including" and its variations used herein are open inclusions, i.e., "including but not limited to". The term "based on" means "based at least in part on". The term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one additional embodiment"; the term "some embodiments" means "at least some embodiments". The relevant definitions of other terms will be given in the following description.

[0046] It should be noted that the concepts such as "first" and "second" mentioned in the present disclosure are only used to distinguish different devices, modules or units, and are not used to limit the order or interdependence of the functions performed by these devices, modules or units.

[0047] It should be noted that the modification of “one” mentioned in the present disclosure is illustrative rather than restrictive, and those skilled in the art should understand that it should be understood as “one or more” unless otherwise clearly indicated in the context.

[0048] The names of the messages or information exchanged between multiple devices in the embodiments of the present disclosure are only used for illustrative purposes and are not used to limit the scope of these messages or information.

[0049] Figure 1 It is a flowchart of a method for allocating shipping container locations based on an attribute dual-view stack learning network according to an embodiment of the present disclosure. Figure 2 is a schematic diagram of a marine terminal yard block according to an embodiment of the present disclosure. Figure 3 It is an overview diagram of a method for allocating shipping container locations based on an attribute dual-view stack learning network according to an embodiment of the present disclosure.

[0050] The present disclosure provides a method for allocating shipping container locations based on an attributed dual-view stack learning network, such as Figure 1 and Figure 3 As shown, the method comprises the following steps:

[0051] 101: A hybrid encoding method that integrates various container attributes is used to encode and obtain initial embedded data, as follows:

[0052] First, common encoding techniques are used for different types of attributes, including single-hot encoding, ordinal encoding, integer encoding, and binary encoding, which can complete the encoding of most container attributes to obtain numerical data input. The encoding of departure time is one of the key issues in terminal scheduling in port operations. Therefore, accurate encoding of departure time is a necessary condition for improving operational efficiency. This disclosure adopts time encoding based on Mercer kernel, and the departure time embedding obtained is as follows:

[0053]

[0054] Where Φ(t) represents the high-dimensional feature vector at time t, w k and represents the learnable parameters in the kernel function, and N is the total number of time points considered in the encoding process. In this way, the present disclosure can convert each departure time t into a high-dimensional feature vector Φ(t). k, d, i represent the serial number of the container, that is, the number of the container.

[0055] To integrate these different types of container attributes, the present disclosure concatenates these encoding vectors and applies a position feed-forward network (FFN) to capture nonlinear dependencies as shown below:

[0056] a i =[a1|...||α n ]W 0

[0057]

[0058] where a m is the encoding vector of the mth container attribute (m is an integer from 1 to n), is a learnable parameter, and || is a connection. Finally, the disclosure can obtain It is a container v i The initial embedding of .

[0059] 102: Using the initial embedding data obtained in step 101, a gated recurrent unit (GRU)-based aggregator is used to perform sequential view stack state learning, as follows:

[0060] Through the previous steps, this paper discovers that there are implicit relationships between containers in the stack. First, a GRU-based aggregator is used to learn the sequential view stack state to capture the sequential patterns between containers, as shown below:

[0061]

[0062] Among them, e k is the input of step k, is the stack state at step k, z k is the update gate output of step k, r k is the reset gate output of the kth step. z , W r , W h is the learnable weight matrix. σ(·) is the sigmoid function, || is the concatenation operation, and ⊙ is the Hadamard product. This module can obtain the current stack S k ={e1, e2, ..., e k} is a sequential view representation of .

[0063] 103: Using the initial embedding data obtained in step 101, a dual-view container aggregator is used to capture the common features of the container to obtain a dual view, as follows:

[0064] Containers in a stack exhibit implicit correlations, and identifying such commonalities is beneficial for fine-grained stack representation. Therefore, dual-view feature commonalities can complement sequential view patterns to more comprehensively represent the relationships between containers. This paper adopts a dual-view container aggregator to capture the common features of containers. First, this paper calculates the stack S k ={e1, e2, ..., e k The cross feature e of each container-container pair in ij , as shown below:

[0065] z ij =e i ⊙e j

[0066] where e i It is a container v i The embedding of , ⊙ represents the element-wise multiplication to emphasize the feature dimensions with similar values ​​between the two container embeddings. Then, the stack representation is calculated by performing pairwise container aggregation As shown below:

[0067]

[0068] Where σ is the ReLU function, which filters out irrelevant common features. The calculation can be simplified to:

[0069]

[0070] 104: Integrate the sequential view obtained in step 102 and the dual view obtained in step 103 to obtain the final stack representation and score:

[0071] This module aims to integrate the sequential view and the dual view to obtain the final stack representation. In addition, the contributions of containers in a stack are usually different, and containers stacked later are usually more representative of the current state.

[0072]

[0073] The pooling is implemented using average pooling. k It is constructed by fusing multiple container-container relationships, including sequential transformation information and pairwise commonality information. k The expression h k Then, calculate the candidate container v in the stack x In form, container v x The score is:

[0074] r((v x , ax ) |S k )=f(v x , S k )=f(a x , g(a1, ax, ..., a k ))=e x ·h k

[0075] where f(·) is the container stack scoring function, which applies to all stacks and allows candidate containers to be assigned to stacks with higher scores.

[0076] 105: Conduct robustness training based on adversarial networks to ensure the robustness and generalization ability of the model. The specific steps are as follows:

[0077] Ship terminals encounter rapidly changing conditions that require robust scheduling methods, which typically require large amounts of high-quality data. However, in actual terminal operations, obtaining such data is challenging and costly. Therefore, the present disclosure utilizes the ability of generative adversarial networks (GANs) to generate synthetic data that closely resembles real-world scenarios and trains it on a DSLA, which has a generator and a discriminator.

[0078] Discriminator: The goal is to distinguish these generated samples from real data. This is done by minimizing the negative log-likelihood:

[0079]

[0080] where f φ (,) represents the score function of the discriminator. Objective function The purpose is to give the real container (v k+1 ) assigns a higher score to the generated container Assigning lower scores enhances the discriminator’s ability to distinguish between real containers and generated fake data.

[0081] Generator: The goal is to generate high-quality negative containers that can confuse the discriminator. A key aspect is how to sample a negative container. Let f φ (v i , S k ) indicates that a given stack S k The score function of the generator of the container v i The sampling probability of is given by the softmax function:

[0082]

[0083] Fake container based on sampling The generator deceives the discriminator by minimizing the recognition score, which is defined as follows:

[0084]

[0085] Due to the discrete process of the generator sampling negative containers, gradient descent cannot be directly applied to parameter estimation. To address this challenge, this paper uses the policy gradient method to update the parameters of the generator. The definition is as follows:

[0086]

[0087] Finally, the generator and discriminator participate in an iterative training process similar to the minimax game. This continuous adversarial training can effectively enhance the robustness of DSLA in practical scenarios.

[0088] The present disclosure can adapt to the inherent variability and complexity of real-world operations and effectively improve the performance of container location allocation, which is of great practical significance.

[0089] The above description is only a preferred embodiment of the present disclosure and an explanation of the technical principles used. Those skilled in the art should understand that the scope of disclosure involved in the present disclosure is not limited to the technical solutions formed by a specific combination of the above technical features, but should also cover other technical solutions formed by any combination of the above technical features or their equivalent features without departing from the above disclosed concept. For example, the above features are replaced with the technical features with similar functions disclosed in the present disclosure (but not limited to) by each other to form a technical solution.

[0090] In addition, although each operation is described in a specific order, this should not be understood as requiring these operations to be performed in the specific order shown or in a sequential order. Under certain circumstances, multitasking and parallel processing may be advantageous. Similarly, although some specific implementation details are included in the above discussion, these should not be interpreted as limiting the scope of the present disclosure. Some features described in the context of a separate embodiment can also be implemented in a single embodiment in combination. On the contrary, the various features described in the context of a single embodiment can also be implemented in multiple embodiments individually or in any suitable sub-combination mode.

[0091] Although the subject matter has been described in language specific to structural features and / or methodological logical actions, it should be understood that the subject matter defined in the appended claims is not necessarily limited to the specific features or actions described above. On the contrary, the specific features and actions described above are merely example forms of implementing the claims.

Claims

1. A method for allocating shipping container locations based on attributed dual-view stack learning networks, characterized in that: include: 101: Encode the container attributes by using a hybrid encoding method that comprehensively integrates various container attributes to obtain initial embedded data; 102: Using the initial embedded data, using a gated recurrent unit (GRU)-based aggregator to perform sequential view stack state learning to capture sequential patterns or sequential views between containers; 103: using the initial embedded data, using a dual-view container aggregator to capture common features of the container to obtain a dual view; 104: Integrate the sequential view and the dual view to obtain a final stack representation and score it; 105: Conduct robust training based on adversarial networks to ensure the robustness and generalization ability of the model.

2. A method for allocating shipping container locations based on attributed dual-view stack learning networks according to claim 1, wherein: The hybrid encoding method described in step 101 is used to encode the container attributes, which is as follows: general encoding techniques are used for different types of attributes, including single-hot encoding, ordinal encoding, integer encoding, and binary encoding, which encode most attributes of the container to obtain numerical data input; the encoding of the departure time is one of the key issues in terminal scheduling in port operations, and the time encoding based on the Mercer kernel is used, and the departure time embedding obtained is as follows: Where φ(t) represents the high-dimensional feature vector at time t, w k and represents the learnable parameters in the kernel function, and N is the total number of time points considered in the encoding process; To integrate these different types of container attributes, these encoding vectors are concatenated and a positional feed-forward network (FFN) is applied to capture non-linear dependencies as shown below: a i =[a1||...||a n ]W 0 where a m is the encoding vector of the mth container attribute, W 0 , is a learnable parameter, || is a connection; finally, we get It is a container v i The initial embedding data.

3. A method for allocating shipping container locations based on attributed dual-view stack learning networks according to claim 2, wherein: The use of the initial embedded data in step 102 and the use of a gated recurrent unit-based aggregator to learn the sequential view stack state are as follows: There is an implicit relationship between the containers in the stack. The gated recurrent unit-based aggregator is first used to learn the sequential view stack state to capture the sequential pattern between the containers, as shown below: Among them, e k is the input of step k, is the stack state at step k, z k is the update gate output of step k, r k is the reset gate output of the kth step; W z , W r , W h is the learnable weight matrix; σ(·) is the sigmoid function, || is the concatenation operation, and ⊙ is the Hadamard product; thus, the current stack S is obtained k ={e1, e2, ..., e k } is a sequential view representation of .

4. A method for allocating shipping container locations based on attributed dual-view stack learning networks according to claim 3, wherein: Step 103 uses the initial embedded data and uses a dual-view container aggregator to capture the common features of the containers to obtain a dual view, as follows: the containers in the stack show implicit correlations, and identifying such commonalities is beneficial for fine-grained stack representation; a dual-view container aggregator is used to capture the common features of the containers; first, the stack S is calculated k ={e1, e2, ..., e k The cross feature e of each container-container pair in ii , as shown below: with ij =e i ⊙e j where e i It is a container v i ⊙ denotes element-wise multiplication to emphasize feature dimensions with similar values ​​between two container embeddings. Then, by performing pairwise aggregation of containers, the stack representation is computed As shown below: Where σ is the ReLU function, which filters out irrelevant common features; The calculation can be simplified to:

5. A method for allocating shipping container locations based on attributed dual-view stack learning networks according to claim 4, wherein: The sequential view and the dual view are integrated as described in step 104 to obtain the final stack representation and score it as follows: The container stacked later is more representative of the current state: The pooling is implemented using average pooling; h k It is constructed by fusing multiple container-container relationships, including sequential transformation information and pairwise commonality information; after obtaining the current stack S k The expression h k Then, calculate the candidate container v in the stack x Fraction of container v x The score is: r((v x ,a x )|S k )=f(v x ,S k )=f(a x ,g(a1,a2,...,a k ))=e x ·h k where f(·) is the container stack scoring function, which applies to all stacks and allows candidate containers to be assigned to stacks with higher scores.

6. A method for allocating shipping container locations based on attributed dual-view stack learning networks according to claim 5, wherein: Step 105 is to perform robustness training based on adversarial networks to ensure the robustness and generalization ability of the model, specifically as follows: Generative adversarial networks (GANs) are used to s ) to generate synthetic data similar to real-world scenes and train it on DSLA, which has a generator and a discriminator; The goal of the discriminator is to distinguish these generated samples from real data; this is done by minimizing the negative log-likelihood: where f φ (,) represents the score function of the discriminator; the objective function The purpose is to give the real container (v k+1 ) assigns a higher score to the generated container Assigning lower scores, thereby enhancing the discriminator’s ability to distinguish between real containers and generated fake data; The goal of the generator is to generate high-quality negative containers that confuse the discriminator; a key aspect is how to sample a negative container Assume f φ (v i , S k ) indicates that a given stack S k The score function of the generator; container v i The sampling probability of is given by the softmax function: Fake container based on sampling The generator deceives the discriminator by minimizing the recognition score, which is defined as follows: Due to the discrete process of the generator sampling negative containers, gradient descent cannot be directly applied to parameter estimation; to address this challenge, the policy gradient method is used to update the parameters of the generator; it is defined as follows: Finally, the generator and discriminator participate in an iterative training process similar to the minimax game.

Images