Intelligent decision-making method and system for ship ballast water emergency receiving scheme based on deep learning

Abstract

This invention discloses a deep learning-based intelligent decision-making method and system for emergency ballast water reception schemes, relating to the field of artificial intelligence technology. The method includes: acquiring emergency ballast water reception events and collecting corresponding basic information; constructing a first target emergency operation chain starting with the target ship type and a second target emergency operation chain starting with the target emergency plan; extracting representation vectors from each operation chain and inputting them into a target intelligent decision-making model, outputting a target emergency plan representation vector and a target ship type representation vector; and determining the target emergency reception scheme based on the aforementioned representation vectors. This method dynamically integrates multi-dimensional information through deep learning to achieve intelligent matching of emergency plans, improving decision-making efficiency and accuracy.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and more specifically, to an intelligent decision-making method and system for emergency ballast water reception schemes for ships based on deep learning. Background Technology

[0002] Emergency ballast water reception is a crucial step in ensuring safe navigation, preventing the invasion of harmful organisms, and protecting the marine ecosystem. Traditional emergency reception plans rely heavily on human experience and judgment, requiring comprehensive consideration of complex information from multiple sources, including ship type, ballast water condition, and emergency scenario. This approach suffers from issues such as delayed response, low accuracy in matching plans, and insufficient dynamic adaptability, making it difficult to quickly generate scientifically sound emergency strategies in urgent situations. Summary of the Invention

[0003] The purpose of this invention is to provide an intelligent decision-making method and system for emergency ballast water reception schemes for ships based on deep learning.

[0004] In a first aspect, embodiments of the present invention provide an intelligent decision-making method for emergency ballast water reception schemes for ships based on deep learning, including:

[0005] Acquire ship ballast water emergency reception events, and collect basic information corresponding to the ship ballast water emergency reception events. The basic information includes emergency event type, ship basic parameters, ballast water status parameters, and emergency scenario environment parameters.

[0006] Based on the aforementioned basic information, multiple first-target emergency operation chains and multiple second-target emergency operation chains are constructed. The first-target emergency operation chain is an operation chain that starts with the target vessel type in the ballast water emergency reception scenario, and the second-target emergency operation chain is an operation chain that starts with the target emergency plan in the ballast water emergency reception scenario.

[0007] Extract the operation chain representation vector of each first target emergency operation chain and the operation chain representation vector of each second target emergency operation chain;

[0008] The operation chain representation vectors of multiple first target emergency operation chains and multiple second target emergency operation chains are determined as the inputs of the target intelligent decision-making model to obtain the target emergency plan representation vector and the target ship type representation vector.

[0009] The target emergency reception scheme is determined based on the target emergency scheme characterization vector and the target ship type characterization vector.

[0010] In a second aspect, embodiments of the present invention provide a server system, wherein the readable storage medium includes a computer program, and the computer program, when running, controls the computer device where the readable storage medium is located to execute the method described in the first aspect.

[0011] Compared to existing technologies, the beneficial effects provided by this invention include: The intelligent decision-making method and system for emergency ballast water reception schemes based on deep learning, disclosed in this invention, relates to the field of artificial intelligence technology. The method includes: acquiring emergency ballast water reception events and collecting corresponding basic information; constructing a first target emergency operation chain starting with the target ship type and a second target emergency operation chain starting with the target emergency plan; extracting the representation vectors of each operation chain and inputting them into a target intelligent decision-making model, outputting a target emergency plan representation vector and a target ship type representation vector; and determining the target emergency reception scheme based on the aforementioned representation vectors. This method dynamically integrates multi-dimensional information through deep learning to achieve intelligent matching of emergency plans, improving decision-making efficiency and accuracy. Attached Figure Description

[0012] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of the present invention and should not be considered as limiting the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 A flowchart illustrating the steps of the intelligent decision-making method for emergency ballast water reception scheme based on deep learning provided in this embodiment of the invention;

[0014] Figure 2 A schematic block diagram of the structure of a computer device provided in an embodiment of the present invention. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0016] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0017] In order to solve the technical problems mentioned in the background art Figure 1This is a flowchart illustrating the intelligent decision-making method for emergency ballast water reception of ships based on deep learning, as provided in this embodiment. The following is a detailed description of this intelligent decision-making method for emergency ballast water reception of ships based on deep learning.

[0018] Step S201: Obtain the ship ballast water emergency reception event and collect the basic information corresponding to the ship ballast water emergency reception event. The basic information includes the emergency event type, ship basic parameters, ballast water status parameters and emergency scenario environment parameters.

[0019] Step S202: Based on the basic information, construct multiple first target emergency operation chains and multiple second target emergency operation chains. The first target emergency operation chain is an operation chain starting with the target ship type in the ballast water emergency reception scenario. The second target emergency operation chain is an operation chain starting with the target emergency plan in the ballast water emergency reception scenario.

[0020] Step S203: Extract the operation chain representation vector of each first target emergency operation chain and the operation chain representation vector of each second target emergency operation chain;

[0021] Step S204: Determine the operation chain representation vectors of the multiple first target emergency operation chains and the operation chain representation vectors of the multiple second target emergency operation chains as inputs to the target intelligent decision-making model to obtain the target emergency plan representation vector and the target ship type representation vector.

[0022] Step S205: Determine the target emergency reception scheme based on the target emergency scheme characterization vector and the target ship type characterization vector.

[0023] In this embodiment of the invention, for example, the server acts as the execution entity, monitoring the ship's dynamics in real time through the Automatic Identification System (AIS), the port monitoring platform, and the ship's Internet of Things (IoT) sensor network. When an abnormal signal is detected, an emergency event acquisition process is triggered. For example, while the 300,000-ton bulk carrier "Yuanyang XX" was waiting to berth at a port anchorage, its No. 3 ballast tank suffered damage and leakage due to structural fatigue. The ship's crew sent a distress signal to the port's emergency command center via satellite communication. Simultaneously, the port monitoring platform captured the abnormal change in the ship's draft and the sudden drop in ballast tank level sensor data. Upon receiving the emergency event, the server immediately activated the basic information collection module: First, it analyzed the distress signal text using natural language processing and, combined with a pre-set event type database, determined the event type to be "ballast water tank rupture and leakage (Level 1 Emergency)," characterized by a leakage rate of 800 m³ / h and ballast water containing 0.5 mg / L of petroleum pollutants (exceeding port acceptance standards). Subsequently, it retrieved the ship's basic parameters through the AIS system, including bulk carrier type, gross tonnage of 320,000 DWT, length of 320m, beam of 55m, design draft of 18.2m, and information on the ballast water system's 8 tanks (total capacity 50,000 m³) and No. 3 double bottom. Data such as the ballast water status parameters were obtained through IoT sensors, including the current liquid level of No.3 tank (2.5m, remaining water volume 2500m³), salinity 32‰, water temperature 22℃, pH 7.8, and the location of the leak point (bottom of port side, 0.5m diameter). Finally, the port geographic information system and meteorological and hydrological database were retrieved to obtain the anchorage water depth of 22m, tidal current velocity of 1.2m / s (high tide, direction 315°), wind force of 4 (northeast wind), visibility of 10km, and the port emergency resource allocation (2 sets of mobile receiving devices, 3 tugboats, 1 set of pollutant filtration system).

[0024] Based on the collected basic information, combined with a historical emergency case database and an expert rule base, the server constructs multiple sets of first-objective emergency operation chains (starting with the ship type) and second-objective emergency operation chains (starting with the emergency plan). For the first-objective emergency operation chain, starting with a "300,000-ton bulk carrier," three differentiated operation chains are generated based on the ship's structural characteristics (double bottom, large capacity), leakage characteristics (bottom breach, high leakage rate), and environmental parameters: The first chain includes the steps of "structural characteristic analysis → leakage rate prediction → temporary sealing → ballast water transfer → receiving requirement assessment," for example, using an underwater robot carrying a sealing pad to seal the breach (85% success rate), while simultaneously planning the ballast water transfer path to empty compartments No. 5 / 6 (rate 600m³). / h); the second chain revolves around "ship stability calculation → berthing feasibility analysis → wharf compatibility check → pipeline connection → pollutant treatment", such as assessing the feasibility of berthing at an emergency wharf when the ship is heeling at 3° (wharf load ≥ 300,000 tons, equipped with DN200 interface); the third chain focuses on "tugboat-assisted attitude adjustment → leak point detection → inert gas protection → pump group arrangement → barge transfer", such as controlling the ship's roll ≤ 2° with two tugboats, and placing submersible pumps (flow rate 400m³ / h) in the cabin to transfer to the port's standby barge. For the second objective, the emergency operation chain is based on three emergency reception plans at the port: Plan A (berthing reception + direct transport) covers "receiving equipment compatibility assessment → ship berthing position adjustment → pipeline connection → pollutant monitoring → temporary storage", for example, using a 1000m³ / h mobile receiving device adapted to the DN200 interface of a bulk carrier; Plan B (anchorage reception + barge transfer) includes "tugboat assisting barge berthing → flexible pipeline installation → barge transport → filtration treatment → discharge", for example, connecting the barge and the leaking vessel through a flexible pipeline (80m in length, tensile strength 500N / mm²); Plan C (mixed reception) involves "in-tank transfer → anchorage reception → parallel transport by dual pumps → chemical treatment → compliance testing", for example, simultaneously implementing in-tank transfer of ballast water (1000m³) and anchorage reception (800m³ / h device), adding activated carbon adsorbent (200g / m³) to treat pollutants.

[0025] The server uses a pre-trained operation link embedding model and a sequence coding network (Bi-LSTM) to transform each operation chain into a 512-dimensional representation vector. First, seven core features are extracted and quantified for each operation chain link: equipment type (one-hot encoding), number of personnel, time consumption, success rate, environmental sensitivity (degree of influence from water flow / wind), risk level (consequences of operation failure), and cost coefficient (direct cost). For example, the quantified features of the "temporary sealing operation" link are: underwater robot equipment vector [1,0,...,0], 3 personnel, time consumption 1.5h, success rate 0.85, environmental sensitivity 0.6 (affected by tidal currents), risk level 0.8 (sealing failure exacerbates leakage), and cost coefficient 1500 yuan / h, concatenated into a 128-dimensional vector. Then, the quantified feature vector is input into the embedding layer (256 dimensions) to obtain the link embedding vector. The link sequence is input into the Bi-LSTM (2 layers, 256-dimensional hidden layer), and the output sequence's hidden state at the last moment (512-dimensional concatenation of forward and backward layers) is used as the operation chain representation vector. Ultimately, the six operation chain vectors are: the first target operation chain C1→V1, C2→V2, C3→V3; and the second target operation chain C4→V4, C5→V5, C6→V6.

[0026] The server inputs six operation chain vectors into a pre-trained target intelligent decision-making model (loss function value < 0.05), and outputs a target representation vector through a scheme encoding network and a ship encoding network. The scheme encoding network contains a first fusion unit (attention mechanism) and a first GRU (512-dimensional): the second operation chain vectors V4 / V5 / V6 are input into the fusion unit, the weights (V4: 0.4, V5: 0.3, V6: 0.3) are calculated and weighted to obtain the decision representation vectors W4 / W5 / W6 (representation scheme coupling logic), which are then input into the first GRU and linearly superimposed through gating coefficients (α1=0.6, α2=0.25, α3=0.15) to output the target emergency scheme representation vector U scheme (512-dimensional). The ship coding network includes a second fusion unit (graph attention mechanism), a second GRU (512-dimensional), and a collaborative vector S (512-dimensional): The first operation chain vectors V1 / V2 / V3 are integrated with S. The fusion unit takes "bulk carrier" as the target node, calculates the influence coefficients of related links (such as "temporary sealing" 0.3, "leakage rate prediction" 0.25), and weights them to obtain decision representation vectors W1 / W2 / W3 (representing ship type coupling logic). The inputs are linearly superimposed through gating coefficients (C1:0.5, C2:0.3, C3:0.2) and integrated with S to output the target ship type representation vector U ship type (512-dimensional).

[0027] The server processes the U-scheme and U-ship type through the decision fusion module, calculates the collaborative matching degree, and determines the final scheme by combining multi-dimensional evaluation. First, the cosine similarity calculation yields cos(U-scheme, U-ship type) = 0.89 (high matching). The similarity between each second operation chain and the U-ship type is: V4 (0.92) > V6 (0.85) > V5 (0.78), indicating that scheme A (corresponding to V4) has the highest matching degree with the characteristics of bulk carriers. Subsequently, the scheme is evaluated from three dimensions: feasibility, safety, and economy: Scheme A (berthing and receiving) has a feasibility score of 9.5 (equipment capacity of 1000m³ / h covers leakage rate), a safety score of 9.0 (roll ≤1° after berthing), and an economy of 80,000 yuan (equipment leasing cost); Scheme B (anchorage barge transport) has a feasibility score of 7.5 (barge berthing is affected by tidal current), a safety score of 7.0 (collision risk), and an economy of 150,000 yuan; Scheme C (mixed receiving) has a feasibility score of 8.0 (parallel operation complexity), a safety score of 8.5 (stability fluctuation), and an economy of 120,000 yuan. Based on the weights of overall matching degree (0.5), feasibility (0.2), safety (0.2), and economy (0.1), Plan A scored 8.96 points (the highest). The server determined it as the target emergency receiving plan and instructed the "Ocean XX" to berth at the emergency dock and use the 1000m³ / h device to receive the water. The ballast water was temporarily stored in storage tank T1 for processing within 3 hours.

[0028] This invention integrates ship characteristics and emergency response logic through a deep learning model to achieve rapid and accurate emergency response decision-making, thereby ensuring ship safety and the port environment.

[0029] In this embodiment of the invention, the intelligent decision-making model is trained and obtained in the following manner, and can be implemented through the following examples.

[0030] Multiple first sample emergency operation chains and multiple second sample emergency operation chains are obtained. Each first sample emergency operation chain is used to characterize the dependency relationship between the sample ship type and the sample emergency plan when the sample ship type is the starting link. Each second sample emergency operation chain is used to characterize the dependency relationship between the sample emergency plan and the sample ship type when the sample emergency plan is the starting link.

[0031] Extract the operation chain representation vector of each first sample emergency operation chain and the operation chain representation vector of each second sample emergency operation chain;

[0032] The operation chain representation vectors of multiple second sample emergency operation chains are determined as the input of the scheme coding network in the original intelligent decision model to obtain the sample emergency scheme representation vector. One or more operation chain representation vectors of the first sample emergency operation chains are determined as the input of the ship coding network in the original intelligent decision model to obtain the sample ship type representation vector. The first decision coordination vector is used to represent the coordination parameters when the ship coding network dynamically adapts to the output of the scheme coding network.

[0033] The original intelligent decision-making model is trained based on the sample emergency plan representation vector and the sample ship type representation vector to obtain the target intelligent decision-making model.

[0034] In this embodiment of the invention, for example, when training the target intelligent decision-making model, the server first extracts 1000 sample operation chains from the historical ship ballast water emergency reception case database, including 500 first sample emergency operation chains and 500 second sample emergency operation chains. The first sample emergency operation chains start with the sample ship type, such as "200,000-ton oil tanker", "150,000-ton container ship", "300,000-ton bulk carrier", etc. Each chain contains the dependency relationship between the sample ship type and the sample emergency plan. For example, the links of a sample chain for a bulk carrier are "300,000-ton bulk carrier → double bottom structural characteristics → ballast water leakage rate prediction (600 m³ / h) → underwater robot plugging → berthing receiving device connection", thus characterizing the relationship between "berthing receiving" and the large capacity and high leakage scenario of a bulk carrier. The adaptation logic of the scheme; the second sample emergency operation chain takes the sample emergency scheme as the starting link, such as "berthing and receiving plan", "anchorage barge transport plan", "pollutant filtration and treatment plan", etc. The link of a certain berthing and receiving sample chain is "berthing and receiving plan → mobile receiving device (800m³ / h) → ship berthing attitude adjustment (roll ≤2°) → DN200 pipeline connection → real-time pollutant monitoring (accuracy ±0.1mg / L)", which characterizes the coupling relationship between the scheme and different ship types such as bulk carriers and container ships.

[0035] Next, the server extracts the representation vector for each sample operation chain. For the "Emergency Operation Chain of Ballast Water for Oil Tankers" in the first sample chain, the server quantifies the characteristics of each link: equipment type (oil tanker-specific tank washing pump: [0,1,0,...0]), number of personnel (5 people), time (2h), success rate (0.92), environmental sensitivity (0.7, affected by crude oil viscosity), risk level (0.85, leakage may cause explosion), and cost coefficient (3000 yuan / h) in the "Crude Oil Tank Washing System Compatibility Assessment" link. These characteristics are concatenated into a 128-dimensional vector and input into the embedding layer (256 dimensions). Then, the link sequence is processed by a Bi-LSTM network (2 layers, 256-dimensional hidden layer) to output a 512-dimensional representation vector of the first sample operation chain. The "Anchorage Barge Transportation Plan Operation Chain" in the second sample chain is also quantified by features (such as tensile strength of 500N / mm² and length of 60m in the "Flexible Pipeline Laying" link), embedded and encoded by Bi-LSTM to obtain a 512-dimensional representation vector. Finally, all sample chains are transformed into vector matrices of the same dimension.

[0036] Subsequently, the server inputs the representation vector of the second sample emergency operation chain into the scheme coding network of the original intelligent decision model. The scheme coding network contains a first fusion unit and a first gated recurrent unit (GRU). The server first loads 500 second sample chain vectors into the first fusion unit, and at the same time introduces a second decision coordination vector (512-dimensional, initially randomly generated, used to dynamically adapt to the ship coding network output). By element-wise addition and integration of vectors (e.g., after integrating a berthing reception sample chain vector with the coordination vector, highlighting the "equipment adaptability" feature weight), the decision representation vector (512-dimensional) of each second sample chain is obtained, representing the operational logic of scheme coupling (e.g., the coordination relationship of "equipment capability - ship interface - environmental stability" in berthing reception). Next, these decision representation vectors are input into the first GRU. The GRU calculates the gating coefficients of each sample chain through the update gate (e.g., 0.6 for the berthing and receiving sample chain, 0.3 for the anchorage barge transport sample chain, and 0.1 for the mixed processing sample chain). Then, the weighted vectors are linearly superimposed to output the sample emergency scheme representation vector (512 dimensions). This vector encodes the common logic and differentiated features of all sample schemes.

[0037] The ship coding network receives the representation vector and the first decision collaboration vector (512-dimensional, output by the dynamically adapted coding network) of the first sample emergency operation chain. Taking the "200,000-ton oil tanker" sample chain as an example, the server extracts the representation vector of "oil tanker ballast water tank material (corrosion-resistant alloy)" and the vectors of related links ("explosion-proof pump group layout" and "crude oil pollutant treatment") from the chain. The decision influence coefficient of the related samples is calculated by multiplying the vectors (explosion-proof pump group has an influence coefficient of 0.4 due to its high safety; pollutant treatment has an influence coefficient of 0.3 due to environmental protection requirements). After weighted superposition, the decision representation vector of the chain is obtained. Subsequently, the decision representation vectors of all first sample chains are integrated with the first decision collaboration vector and input into the second GRU. The weights are adjusted by the gating coefficient (0.4 for oil tanker sample chain, 0.3 for container ship sample chain, and 0.3 for bulk carrier sample chain). After linear superposition, the sample ship type representation vector (512-dimensional) is output.

[0038] Finally, the server calculates the cosine similarity deviation between the sample emergency plan representation vector and the sample ship type representation vector (loss function L=1-cos(U plan, U ship type)), with an initial loss value of 0.35. The network parameters are updated using the backpropagation algorithm (learning rate 0.001, decaying by 10% every 100 iterations). After 5000 iterations, the loss value drops to 0.04 (convergence threshold ≤ 0.05). At this point, the collaborative parameters (first and second decision collaborative vectors) between the plan encoding network and the ship encoding network are stable, the target intelligent decision model is trained, and it can be used for plan decision-making in actual emergency scenarios.

[0039] In this embodiment of the invention, the scheme coding network includes a first fusion unit and a first gated loop unit; the operation chain representation vectors of multiple second sample emergency operation chains are determined as the input of the scheme coding network in the original intelligent decision-making model to obtain the sample emergency scheme representation vectors, which can be implemented through the following example.

[0040] The operation chain representation vector of each second sample emergency operation chain is loaded into the first fusion unit to obtain the decision representation vector of each second sample emergency operation chain, wherein each vector is used to represent the operation logic of scheme coupling in the corresponding second sample emergency operation chain.

[0041] The decision representation vectors of multiple second sample emergency operation chains are determined as the input of the first gated loop unit to obtain the gating coefficient of each second sample emergency operation chain.

[0042] Based on the first gating loop unit, the decision representation vectors of multiple second sample emergency operation chains are linearly superimposed with the gating coefficients to obtain the sample emergency scheme representation vector.

[0043] In this embodiment of the invention, for example, the input processing flow of the scheme coding network when the server trains the intelligent decision-making model is as follows: First, the representation vectors (512 dimensions each) of 500 second sample emergency operation chains are loaded into the first fusion unit. These second sample chains start with different emergency schemes, such as "berthing and receiving plan chain", "anchorage barge transport plan chain", "pollutant filtration and treatment plan chain", etc. Taking the "berthing and receiving plan chain" as an example, when its representation vector is processed by the first fusion unit, the server will extract the "equipment-scenario-risk" coupling logic unique to the scheme: such as the processing capacity of the mobile receiving device (1000m³ / h) needs to match the ship's leakage rate (800m³ / h), the ship's heel needs to be controlled at ≤2° when berthing to avoid pipe detachment, and the real-time pollutant monitoring accuracy needs to reach ±0.1mg / L to meet environmental protection requirements. The first fusion unit strengthens these key logical features through an attention mechanism, outputting a 512-dimensional decision representation vector. Compared to the original representation vector, this vector significantly increases the weight of core dimensions such as "device compatibility" and "environmental stability".

[0044] Subsequently, the server inputs the decision representation vectors of all second sample chains into the first gated recurrent unit (GRU) in batches. For example, when processing sample chains related to the "emergency scenario of a 300,000-ton bulk carrier with a high leakage rate (800 m³ / h) and pollutant concentration exceeding the standard," the GRU dynamically calculates the gating coefficient of each scheme chain through update and reset gates: the "berthing and receiving plan chain" is assigned a gating coefficient of 0.5 due to its high processing efficiency (1000 m³ / h > 800 m³ / h); the "pollutant filtration and treatment plan chain" is assigned a coefficient of 0.3 because it needs to prioritize controlling pollution diffusion; and the "anchorage barge transport plan chain" is assigned a coefficient of 0.2 because it is greatly affected by tidal currents. The gating coefficient directly reflects the contribution of each scheme in the current scenario; the higher the value, the stronger the match between its operational logic and the scenario requirements.

[0045] Finally, the first GRU linearly superimposes the decision representation vectors and gating coefficients. For example, the decision representation vector of the "berthing and receiving plan chain" (denoted as Vberthing) is multiplied by a gating coefficient of 0.5, the decision representation vector of the "pollutant filtration and treatment plan chain" (VFiltering) is multiplied by 0.3, and the decision representation vector of the "anchorage transport plan chain" (VTransport) is multiplied by 0.2, resulting in the sample emergency plan representation vector Uplan = 0.5 × Vberthing + 0.3 × VFiltering + 0.2 × VTransport. This vector integrates the core operational logic of different emergency plans and their priorities in specific scenarios, laying the foundation for subsequent matching training with ship type representation vectors. In actual training, the server processes the 500 second sample chains in batches according to scenario categories (such as high leakage rate, high pollutant concentration, complex weather conditions, etc.) to ensure that the generated sample emergency plan representation vectors can cover diverse emergency scenario features.

[0046] In this embodiment of the invention, before loading the operation chain representation vector of each second sample emergency operation chain into the first fusion unit to obtain the decision representation vector of each second sample emergency operation chain, this embodiment of the invention also provides the following implementation method.

[0047] Obtain a second decision collaboration vector, which is used to characterize the collaboration parameters when the scheme coding network dynamically adapts to the output of the ship coding network;

[0048] The operation chain representation vectors of each second sample emergency operation chain are loaded into the first fusion unit to obtain the decision representation vectors of each second sample emergency operation chain, including:

[0049] The second decision collaboration vector is integrated with the operation chain representation vector of each second sample emergency operation chain to obtain the operation chain integration vector of each second sample emergency operation chain.

[0050] The operation chain integration vector of each second sample emergency operation chain is loaded into the first fusion unit to obtain the decision representation vector of each second sample emergency operation chain.

[0051] In this embodiment of the invention, for example, before loading the operation chain representation vector of the second sample emergency operation chain into the first fusion unit, the server first obtains the second decision collaboration vector. This vector is the core collaboration parameter for the scheme coding network to dynamically adapt to the output of the ship coding network. It has 512 dimensions, and its initial value is randomly generated through a normal distribution (mean 0, standard deviation 0.1), and is dynamically updated with sample iterations during model training. For example, when processing the sample training for a "high leakage scenario (800 m³ / h) of a 300,000-ton bulk carrier", the features such as "large capacity", "double bottom structure", and "high stability requirements" have higher weights in the sample ship type representation vector output by the ship coding network. The second decision collaboration vector will synchronously adjust the parameters of these dimensions to enhance the adaptability of the scheme coding network to the ship structural characteristics.

[0052] Subsequently, the server integrates the second decision-making collaboration vector with the operation chain representation vectors of each second sample emergency operation chain. Taking the "berthing and receiving plan operation chain" as an example, in its original representation vector (512 dimensions), the inherent features of the scheme, such as "mobile receiving device capacity (1000m³ / h)" and "pipeline connection pressure resistance level (1.0MPa)," account for 60% of the weight, while features related to the ship type, such as "roll adaptability (≤2°)" and "cabin capacity matching degree," account for only 20%. The server integrates the second decision-making collaboration vector with this representation vector through element-level addition: in the integrated vector, the weight of the "roll adaptability" dimension increases from 0.2 to 0.4, and the weight of the "cabin capacity matching degree" dimension increases from 0.15 to 0.3, while retaining core scheme features such as "equipment processing capacity" (weight maintained at 0.5), finally obtaining a 512-dimensional operation chain integrated vector. This process ensures that the input vector of the scheme coding network not only includes the scheme's own logic but also incorporates the ship type adaptation requirements, avoiding scheme decision-making bias due to ignoring ship characteristics.

[0053] After integration, the server loads the operation chain integration vectors of each second sample chain into the first fusion unit. For example, in the integration vector of the "pollutant filtration and treatment plan operation chain," features such as "filtration accuracy (5μm)" and "treatment efficiency (500m³ / h)" are synergistically enhanced with features such as "pollutant concentration (0.5mg / L)" and "leakage rate (600m³ / h)" that the ship coding network focuses on. The first fusion unit further focuses on key coupling logic through an attention mechanism: when the sample ship is an oil tanker, due to the high viscosity of crude oil pollutants, the fusion unit will increase the weight of the "filtration membrane anti-clogging capability" dimension (from 0.3 to 0.5); when the sample ship is a bulk carrier, due to the large tank capacity and high leakage rate, it will strengthen the weight of the "matching degree between treatment efficiency and leakage rate" dimension (from 0.4 to 0.6). Finally, the first fusion unit outputs a 512-dimensional decision representation vector, which accurately encodes the dynamic adaptation relationship between the scheme operation logic and the characteristics of the ship type, providing reliable input for the subsequent gating loop unit to calculate the gating coefficient.

[0054] In the initial stage of model training, the adjustment range of the second decision collaboration vector is relatively large (e.g., an update step size of 0.01 for each iteration in the first 1000 iterations). As the training converges (the loss value drops below 0.05), the adjustment step size is gradually reduced to 0.001 to ensure that the collaboration parameters stably adapt to various ship-solution coupling scenarios. For example, for the sample pair of "200,000-ton oil tanker + pollutant filtration plan", the collaboration vector finally converges to: "corrosion-resistant material adaptation" dimension weight 0.45, "explosion-proof equipment compatibility" dimension weight 0.35, and "high-viscosity pollutant treatment" dimension weight 0.2, making the decision representation vector output by the scheme coding network highly matched with the oil tanker characteristic vector of the ship coding network (cosine similarity increased to 0.92).

[0055] In this embodiment of the invention, the decision representation vectors of multiple second sample emergency operation chains are linearly superimposed with the gating coefficients based on the first gating loop unit to obtain the sample emergency scheme representation vector, which can be implemented through the following example.

[0056] Based on the first gated loop unit, the decision representation vectors of multiple second sample emergency operation chains are linearly superimposed with the gate coefficients to obtain the target operation chain integration vector;

[0057] The target operation chain integration vector is integrated with the second decision collaboration vector to obtain the sample emergency solution characterization vector.

[0058] In an embodiment of the invention, for example, the server linearly superimposes the decision representation vectors and gating coefficients of multiple second-sample emergency operation chains through a first gating loop unit (GRU). For instance, in the training of the "high leakage scenario (800 m³ / h) for a 300,000-ton bulk carrier" sample, the GRU calculates gating coefficients for three core second-sample chains: the decision representation vector of the "berthing and receiving plan chain" is assigned a gating coefficient of 0.5 because its processing efficiency (1000 m³ / h) matches the leakage rate well; the "pollutant filtration and treatment plan chain" has a coefficient of 0.3 because it needs to prioritize controlling pollution diffusion; and the "anchorage barge transport plan chain" has a coefficient of 0.2 because it is greatly affected by tidal currents. GRU linearly superimposes each decision representation vector (512 dimensions) with its corresponding coefficient: Target Operation Chain Integration Vector = 0.5 × Berthing and Receiving Decision Vector + 0.3 × Filtering and Processing Decision Vector + 0.2 × Anchorage Barge Transportation Decision Vector. This vector integrates the core logic of each scheme, with the "Equipment Processing Capacity" dimension having a weight of 0.45, the "Environmental Stability" dimension having a weight of 0.3, and the "Pollutant Control" dimension having a weight of 0.25.

[0059] Subsequently, the server integrates the target operation chain vector with the second decision collaboration vector (element-wise addition). The second decision collaboration vector, now adapted to the "large bulk carrier capacity (6000m³), double-bottom structure" parameters output by the ship coding network, is integrated. The weight of the "capacity matching degree" dimension in the integrated vector increases from 0.2 to 0.4, and the weight of the "heel adaptability (≤2°)" dimension increases from 0.25 to 0.35, while retaining inherent characteristics of the scheme such as "processing efficiency" (weight 0.4). This ultimately generates a 512-dimensional sample emergency scheme representation vector. This vector encodes the scheme's operational logic and enhances its adaptability to the structural characteristics of the bulk carrier, providing accurate scheme feature representations for subsequent model training.

[0060] In this embodiment of the invention, the ship coding network includes a second fusion unit and a second gated loop unit; the operation chain representation vector of one or more of the first sample emergency operation chains is determined as the input of the ship coding network in the original intelligent decision model to obtain the sample ship type representation vector. This can be implemented through the following example.

[0061] The first decision collaboration vector is integrated with the operation chain representation vector of each first sample emergency operation chain to obtain the first operation chain integration vector of each first sample emergency operation chain.

[0062] The first operation chain integration vector of each first sample emergency operation chain is loaded into the second fusion unit to obtain the decision representation vector of each first sample emergency operation chain. The decision representation vector of each first sample emergency operation chain is used to represent the ship-shaped coupling operation logic in the corresponding first sample emergency operation chain.

[0063] The decision representation vector of each first sample emergency operation chain is determined as the input of the second gated loop unit to obtain the gate coefficient of each first sample emergency operation chain.

[0064] The decision representation vectors of multiple emergency operation chains of the first samples are linearly superimposed with the gating coefficients based on the second gating loop unit to obtain the first gating representation vector;

[0065] The first gating representation vector is integrated with the first decision collaboration vector to obtain the sample ship type representation vector.

[0066] In this embodiment of the invention, for example, the input processing flow of the ship coding network when the server trains the original intelligent decision-making model is as follows: First, the first decision collaboration vector is obtained. This vector is the core parameter output by the ship coding network's dynamic adaptation scheme coding network. It has a dimension of 512, and its initial value is randomly generated through a normal distribution and updated iteratively with samples during training. For example, when the "berthing reception" feature (processing efficiency 1000m³ / h, heel adaptability ≤2°) has a high weight in the sample emergency scheme representation vector output by the scheme coding network, the first decision collaboration vector will simultaneously strengthen the parameters of ship feature dimensions such as "large cabin capacity adaptation" and "structural stability" to ensure that the ship coding network focuses on the adaptation requirements of the scheme to the ship type.

[0067] Subsequently, the server integrates the first decision-making collaboration vector with the operation chain representation vectors of each first sample emergency operation chain. Taking the "300,000-ton bulk carrier emergency operation chain" as an example, in its original representation vector (512 dimensions), the inherent ship type features such as "double bottom structure," "total tank capacity of 50,000 m³," and "ballast water system design pressure of 0.2 MPa" account for 50% of the weight, while features related to the emergency plan such as "berthing attitude control" and "tank capacity and receiving device matching degree" account for only 25%. After the server integrates the first decision-making collaboration vector through element-level addition, the weight of the dimension "berthing attitude control (heel ≤ 2°)" in the vector increases from 0.2 to 0.4, and the weight of the dimension "tank capacity matching degree (fitting of 6,000 m³ leak tank and receiving device)" increases from 0.15 to 0.3, while retaining core ship type features such as "double bottom structure strength" (weight maintained at 0.45), finally obtaining the 512-dimensional first operation chain integrated vector. This process ensures that the input to the ship coding network not only includes the ship type's own characteristics but also incorporates the output requirements of the scheme coding network, thus avoiding deviations in ship type representation due to neglecting scheme logic.

[0068] Next, the server loads the integration vector of the first operation chain of each first sample chain into the second fusion unit. For example, in the integration vector of the "200,000-ton oil tanker emergency operation chain", the tanker-specific operation logic such as "crude oil tank washing system compatibility" and "explosion-proof ballast pump configuration" has been synergistically strengthened with the scheme features such as "pollutant treatment efficiency (500m³ / h)" and "leakage rate control (600m³ / h)" that the scheme coding network focuses on. The second fusion unit focuses on the key logic of ship type coupling through an attention mechanism: when the sample scheme is "pollutant filtration and treatment plan", because the oil tanker ballast water contains high-viscosity crude oil pollutants, the fusion unit will increase the weight of the "anti-clogging filter membrane adaptation" dimension (from 0.25 to 0.45); when the sample scheme is "berthing and receiving plan", because the oil tanker has a draft (design draft 19m), the weight of the "terminal water depth adaptability (≥20m)" dimension will be strengthened (from 0.3 to 0.5). For the "emergency operation chain of a 150,000-ton container ship", the features such as "the impact of deck container layout on the arrangement of receiving devices" and "rapid ballast water transfer system (800m³ / h)" in its integrated vector are coupled with the "anchorage barge transport plan" feature (flexible pipeline construction efficiency) output by the scheme coding network after being processed by the second fusion unit. The weight of the "spatial adaptability" dimension in the output decision representation vector is increased from 0.2 to 0.35, which accurately represents the adaptation logic of the anchorage scheme in the scenario of limited deck space of container ships.

[0069] After extracting the decision representation vectors, the server inputs the decision representation vectors of all first sample chains into the second gated recurrent unit (GRU) in batches. For example, when processing the sample training for "high leakage rate (800 m³ / h) + berthing reception plan", the GRU calculates the gating coefficients for three typical first sample chains: the "300,000-ton bulk carrier operation chain" is assigned a gating coefficient of 0.4 due to its large cargo hold (6,000 m³ leakage tank) and high berthing stability requirements (heel ≤ 1°); the "200,000-ton oil tanker operation chain" is assigned a coefficient of 0.3 due to its high pollutant risk (oil spill explosion risk); and the "150,000-ton container ship operation chain" is assigned a coefficient of 0.3 due to the deck space constraints requiring special piping arrangements. The gating coefficient directly reflects the contribution of different ship types in the current scenario; a higher value indicates a stronger match between its operational logic and the requirements of the plan.

[0070] Subsequently, the second GRU linearly superimposes the decision representation vector and the gating coefficient: the first gating representation vector = 0.4 × bulk carrier decision representation vector + 0.3 × tanker decision representation vector + 0.3 × container ship decision representation vector. In this vector, the "large hold capacity adaptation" dimension has a weight of 0.35 (from the bulk carrier chain), the "pollutant risk control" dimension has a weight of 0.25 (from the tanker chain), and the "spatial adaptability" dimension has a weight of 0.2 (from the container ship chain), while retaining common features such as "berthing stability" (weight 0.2).

[0071] Finally, the server integrates the first gating representation vector with the first decision-coordination vector (element-wise addition). At this point, the parameters of the first decision-coordination vector have been adjusted according to the "berthing and receiving plan" features output by the scheme coding network. After integration, the weight of the "berthing attitude control" dimension in the vector increases from 0.3 to 0.4, and the weight of the "receiving device interface adaptation (DN200)" dimension increases from 0.2 to 0.3. At the same time, the inherent ship type features such as "structural strength (EH36 steel)" are strengthened (weight remains at 0.3). Finally, a 512-dimensional sample ship type representation vector is generated. This vector encodes the core structural characteristics (capacity, draft, spatial layout) of bulk carriers, tankers, and container ships, and also incorporates the coupling logic with the berthing and receiving scheme (stability, interface adaptation, pollutant control), providing accurate ship type feature representation for matching the sample emergency scheme representation vector with the sample ship type representation vector in subsequent model training. During the iterative process of model training, the first decision collaboration vector will be continuously optimized. For example, when the sample emergency scheme representation vector output by the scheme coding network shifts to the "anchorage barge transport plan", the collaboration vector will increase the weight of dimensions such as "tidal adaptability" and "barge berthing space" to ensure that the ship coding network dynamically adapts to changes in scheme logic.

[0072] In this embodiment of the invention, the first operation chain integration vector of each first sample emergency operation chain is loaded into the second fusion unit to obtain the decision representation vector of each first sample emergency operation chain. This can be implemented through the following example.

[0073] From the first operation chain integration vector of the first emergency operation branch, extract the representation vector of the target sample ship type and the representation vector of each associated sample that has a connection relationship with the target sample ship type. The first emergency operation branch is any one of the plurality of first sample emergency operation chains, and the target sample ship type is any sample ship type of the first emergency operation branch.

[0074] The representation vector of the target sample ship type is multiplied with the representation vector of each associated sample to obtain the sample decision influence coefficient of each associated sample in the first emergency operation branch.

[0075] Based on the sample decision influence coefficients of each of the associated samples, the representation vectors of the corresponding associated samples are linearly superimposed to obtain the decision representation vector of the first emergency operation branch.

[0076] In an embodiment of the invention, for example, when the server loads the first operation chain integration vector of the first sample emergency operation chain into the second fusion unit, it uses the "300,000-ton bulk carrier first sample emergency operation chain" as the first emergency operation branch. The operation steps of this branch include "300,000-ton bulk carrier → double bottom structure characteristics → ballast water leakage rate prediction → underwater robot plugging operation → berthing receiving device connection". First, the server extracts the representation vector of the target sample ship type and the representation vector of each associated sample from the first operation chain integration vector (512 dimensions) of this branch: the target sample ship type is "300,000-ton bulk carrier", and its representation vector focuses on the core features of the ship type, including total cargo capacity of 50,000 m³ (weight 0.3), double bottom structure (material EH36 steel, weight 0.25), design draft of 18.2 m (weight 0.2), heel control requirement ≤2° (weight 0.15), and ballast water system design pressure of 0. 2MPa (weight 0.1); the associated samples are the subsequent links in the operation chain, including "double bottom tank structural characteristics" (characterization vector includes tank thickness of 20mm, fatigue strength index, etc.), "ballast water leakage rate prediction" (characterization vector includes breach area of ​​0.2m², leakage rate of 600m³ / h, etc.), "underwater robot plugging operation" (characterization vector includes success rate of 0.85, time of 1.5h, etc.), and "berthing receiving device connection" (characterization vector includes DN200 interface compatibility, connection time of 30 minutes, etc.).

[0077] Next, the server performs vector multiplication on the representation vector of the target sample's ship type and the representation vectors of each associated sample to obtain the sample decision influence coefficient. For example, the "structural strength" dimension in the target vector has a weight of 0.25, which, when multiplied by the corresponding dimension (hull thickness 20mm, weight 0.4) in the "double bottom tank structural characteristics" vector, yields an influence coefficient of 0.25 × 0.4 = 0.1 for this associated sample; the "leakage control" dimension in the target vector has a weight of 0.2, which, when multiplied by the "rate accuracy" dimension (error ±50m³ / h, weight 0.5) in the "ballast water leakage rate prediction" vector, yields an influence coefficient of 0.2 × 0.5 = 0. 0.1; The weight of the "Operational Feasibility" dimension in the target vector is 0.15, which is multiplied by the "Success Rate" dimension (0.85, weight 0.6) in the "Underwater Robot Blocking Operation" vector to obtain an influence coefficient of 0.15×0.6=0.09; The weight of the "Equipment Adaptability" dimension in the target vector is 0.1, which is multiplied by the "Interface Matching Degree" dimension (DN200 adaptation, weight 0.7) in the "Boating Receiver Connection" vector to obtain an influence coefficient of 0.1×0.7=0.07.

[0078] Finally, the server performs linear superposition of the corresponding representation vectors based on the sample decision influence coefficients of each associated sample. The representation vectors for "double bottom structural characteristics" are multiplied by 0.1, "leakage rate prediction" by 0.1, "underwater robot plugging operation" by 0.09, and "berthing receiving device connection" by 0.07, and then summed element-wise to obtain the decision representation vector (512 dimensions) for this first emergency operation branch. In this vector, the coupling logic weight of "structural characteristics-leakage control-operational feasibility" accounts for 0.36, significantly higher than the independent feature weights of each individual link. This accurately represents the deep coupling logic between the hull characteristics and emergency operation links of a 300,000-ton bulk carrier in a high-leakage scenario, providing a focused input vector for the subsequent second gating loop unit to calculate the gating coefficients based on the hull coupling relationship.

[0079] In this embodiment of the invention, the following implementation methods are also provided.

[0080] Calculate the first Huber error between the sample ship type representation vector and the second decision collaboration vector;

[0081] The first decision collaboration vector and the second decision collaboration vector are updated based on the first Huber error.

[0082] In this embodiment of the invention, for example, during model training, the server calculates the first Huber error between the sample ship type representation vector and the second decision collaboration vector for a sample of "high leakage scenario of a 300,000-ton bulk carrier". The sample ship type representation vector (512 dimensions) includes features such as a cargo capacity of 50,000 m³ (weight 0.3) and a heel of ≤2° (0.25). The second decision collaboration vector (512 dimensions) is adapted to the "berthing and receiving" scheme, focusing on equipment adaptability (0.3) and environmental stability (0.25). δ=1.0 is set, the sum of squares of the absolute values ​​of the element-level errors of the two vectors ≤1.0 is 0.3, the sum of linear errors >1.0 is 0.1, and the total Huber error is 0.4. The server updates the first decision collaboration vector (increasing the weight of the "cargo capacity matching" dimension by 0.02) and the second decision collaboration vector (strengthening the weight of the "heel control" dimension by 0.03) with a step size of 0.001 through backpropagation, thereby improving the collaboration between the ship and the scheme coding network.

[0083] In this embodiment of the invention, the method further includes:

[0084] Calculate the second Huber error between the sample emergency response scheme representation vector and the first decision collaboration vector;

[0085] The second decision collaboration vector and the first decision collaboration vector are updated based on the second Huber error.

[0086] In this embodiment of the invention, for example, the server calculates the second Huber error based on the emergency response vector (512 dimensions, including features such as processing efficiency of 1000 m³ / h, heel adaptability ≤2°, and DN200 interface compatibility, with weights of 0.3, 0.25, and 0.2 respectively) of the "berthing and receiving contingency plan" sample emergency response vector and the first decision-making coordination vector (focusing on the large cargo hold capacity and structural stability requirements of a 300,000-ton bulk carrier, with corresponding dimension weights of 0.28 and 0.22). Setting δ=1.0, the sum of squares of the absolute values ​​of the element-level errors of the two vectors ≤1.0 is 0.25, the sum of linear errors >1.0 is 0.08, and the total error is 0.33. The second decision-making coordination vector (increasing the weight of the "cargo hold capacity matching" dimension by 0.02) and the first decision-making coordination vector (strengthening the weight of the "heel control" dimension by 0.03) are updated with a step size of 0.001 through backpropagation, enhancing the coordination and adaptability between the plan and the ship type.

[0087] In this embodiment of the invention, the ship coding network includes multiple strategy units, strategy scheduling units, and decision fusion units; the operation chain representation vector of one or more of the first sample emergency operation chains is determined as the input of the ship coding network in the original intelligent decision model to obtain the sample ship type representation vector. This can be implemented through the following example.

[0088] The operation chain representation vector and multiple policy guidance vectors of each first sample emergency operation chain are loaded into the multiple policy units to obtain multiple policy output vectors of each first sample emergency operation chain. Each policy unit outputs one policy output vector, and each policy guidance vector is used to represent the preference of the corresponding policy unit for the sample emergency plan.

[0089] The multiple policy output vectors, the multiple policy guidance vectors, and the sample emergency scheme characterization vector of each first sample emergency operation chain are loaded into the policy scheduling unit to obtain the second operation chain integration vector of each first sample emergency operation chain.

[0090] The second operation chain integration vector of multiple first sample emergency operation chains is loaded into the decision fusion unit to obtain the sample ship type characterization vector.

[0091] In this embodiment of the invention, for example, when the server trains the original intelligent decision-making model, the ship coding network includes three strategy units (efficiency priority, safety priority, and cost priority), one strategy scheduling unit, and one decision fusion unit. Taking the sample emergency plan representation vector of the "berthing and receiving plan" (512 dimensions, including features such as processing efficiency of 1000 m³ / h, heel adaptability ≤2°, and DN200 interface adaptation, with weights of 0.3 / 0.25 / 0.2) and three first sample emergency operation chains (300,000-ton bulk carrier, 200,000-ton oil tanker, and 150,000-ton container ship) as an example, the specific process is as follows:

[0092] The server first initializes three policy-oriented vectors (512 dimensions), which respectively represent the preferences of each policy unit for the sample emergency plan: efficiency-first policy-oriented vector (focusing on the dimensions of "processing rate" and "operation time", with a weight of 0.4 / 0.3), security-first policy-oriented vector (focusing on the dimensions of "structural stability" and "risk level", with a weight of 0.35 / 0.3), and cost-first policy-oriented vector (focusing on the dimensions of "equipment rental cost" and "consumable consumption", with a weight of 0.3 / 0.25).

[0093] Subsequently, the operational chain representation vectors and three strategy guidance vectors of each first sample chain are loaded into the corresponding strategy units. Taking the "first sample chain of a 300,000-ton bulk carrier" as an example, its operational chain representation vector includes features such as "double bottom tank capacity 6000m³ (weight 0.25), leakage rate 800m³ / h (0.2), berthing heel ≤1° (0.18), equipment adaptation cost 20,000 yuan / h (0.15)".

[0094] Efficiency-first strategy unit: Multiply the operation chain representation vector and the efficiency-oriented vector element-wise (e.g., the weight of the dimension "leakage rate 800m³ / h" is 0.2 × the weight of "processing rate" in the efficiency-oriented vector is 0.4 = 0.08), and output the strategy output vector (512 dimensions), which strengthens efficiency features such as "processing rate matching degree (1000m³ / h > 800m³ / h, weight 0.35)" and "berthing time (0.5h, weight 0.25)".

[0095] Safety Priority Strategy Unit: Integrates the operation chain vector with the safety guidance vector, and the output vector strengthens safety characteristics such as "structural stability (fatigue resistance of double bottom tank, weight 0.3)", "roll control accuracy (≤1°, weight 0.28)" and "leakage risk level (0.7, weight 0.22)".

[0096] Cost-first strategy unit: integrates operation chain vector and cost-oriented vector, and the output vector strengthens cost characteristics such as "equipment rental cost (20,000 yuan / h, weight 0.3)" and "pipeline consumable cost (5,000 yuan, weight 0.25)".

[0097] Similarly, after processing by three strategy units, the 200,000-ton oil tanker sample chain (including features such as "crude oil pollutant concentration of 0.6 mg / L" and "explosion-proof equipment requirements") has the following output vectors: efficiency strategy output vector emphasizes "filtration rate of 500 m³ / h", safety strategy output vector emphasizes "explosion-proof rating of ExdⅡBT4", and cost strategy output vector emphasizes "activated carbon addition cost of 200 yuan / m³". The 150,000-ton container ship sample chain (including features such as "deck space limitation" and "rapid connection requirements") has corresponding strategy output vectors that emphasize dimensions such as "connection time of 20 minutes", "space adaptability", and "lightweight equipment cost".

[0098] The server loads the three policy output vectors, three policy guidance vectors, and sample contingency plan representation vectors from each first sample chain into the policy scheduling unit. Taking a 300,000-ton bulk carrier sample chain as an example:

[0099] The strategy scheduling unit first calculates the cosine similarity between the output vector of each strategy and the representation vector of the sample emergency plan: the similarity between the output vector of the efficiency strategy and the plan vector is 0.85 (because the plan emphasizes "1000m³ / h processing efficiency"), the similarity between the safety strategy and the plan is 0.78 (the plan requires "tilt ≤2°"), and the similarity between the cost strategy and the plan is 0.65 (the plan has low sensitivity to cost).

[0100] The similarity-based allocation decision impact coefficients are: efficiency strategy 0.5, security strategy 0.35, and cost strategy 0.15.

[0101] The three strategy output vectors are linearly superimposed: Second operation chain integrated vector = 0.5 × Efficiency strategy output vector + 0.35 × Safety strategy output vector + 0.15 × Cost strategy output vector. In the integrated vector, the weight of "processing rate matching degree" is 0.3 (0.5 × 0.35 + 0.35 × 0.1 + 0.15 × 0.05), and the weight of "roll control accuracy" is 0.25 (0.5 × 0.1 + 0.35 × 0.28 + 0.15 × 0.08), which retains the core requirements of the scheme while incorporating ship type strategy preferences.

[0102] After processing by the scheduling unit, the tanker sample chain, due to the need to prioritize the control of crude oil pollution, has its safety strategy output vector (including "explosion-proof level" and "pollutant filtration accuracy") weight increased to 0.45, efficiency strategy to 0.35, and cost strategy to 0.2, with the integrated vector highlighting the "safety-efficiency" synergy. The container ship sample chain, on the other hand, due to the need for rapid response, has its efficiency strategy weighted at 0.55, safety strategy at 0.3, and cost strategy at 0.15, with the integrated vector strengthening the "time-space adaptation" feature.

[0103] The server loads the second operation chain integration vectors (bulk carrier, tanker, container ship) of the three first sample chains into the decision fusion unit. The fusion unit calculates the scenario contribution of each chain through an attention mechanism.

[0104] The bulk carrier chain, due to its largest cargo capacity (6000m³) and highest leakage rate (800m³ / h), best matches the "efficient handling" requirement of the solution and is assigned a weight of 0.45.

[0105] Due to the high risk of pollutants (oil spill and explosion risk), the tanker chain is coupled with the "safety control" requirement of the plan and is assigned a weight of 0.35.

[0106] Due to deck space constraints, the container ship chain is related to the "rapid connection" requirement of the solution and is assigned a weight of 0.2.

[0107] Finally, the decision fusion unit weighted and superimposed the three integrated vectors: Sample ship type representation vector = 0.45 × bulk carrier integrated vector + 0.35 × tanker integrated vector + 0.2 × container ship integrated vector. This vector (512 dimensions) integrates the core features of different ship types: "large cargo capacity adaptation" (0.3, from bulk carriers), "pollutant risk control" (0.25, from tankers), and "spatial adaptability" (0.2, from container ships), while retaining the scheme-oriented "efficiency-safety" synergistic feature (0.25), providing accurate ship type feature representation for model training.

[0108] In this embodiment of the invention, the operation chain representation vectors of multiple first sample emergency operation chains and multiple policy guidance vectors are loaded into the multiple policy units to obtain multiple policy output vectors for each first sample emergency operation chain. This can be implemented through the following example.

[0109] From the operation chain representation vector of the second emergency operation branch, extract the representation vector of the target sample ship type and the representation vector of each associated sample that has a connection relationship with the target sample emergency plan. The second emergency operation branch is any one of the plurality of first sample emergency operation chains, and the target sample ship type is any sample emergency plan of the second emergency operation branch.

[0110] The representation vector of the target sample ship type is integrated with multiple strategy-oriented vectors to obtain multiple first integrated vectors;

[0111] The first integrated vectors are multiplied by the representation vectors of each of the associated samples to obtain the sample decision influence coefficients of each of the associated samples in the second emergency operation branch.

[0112] Based on the sample decision influence coefficient of each of the associated samples, the representation vectors of the corresponding associated samples are linearly superimposed to obtain multiple strategy output vectors of the second emergency operation branch.

[0113] In this embodiment of the invention, for example, the server uses the "emergency operation chain of the first sample of a 300,000-ton bulk carrier" as the second emergency operation branch. The operation steps of this branch are "300,000-ton bulk carrier → double bottom structural characteristics → ballast water leakage rate prediction (800 m³ / h) → underwater robot plugging operation → berthing receiving device connection". The sample emergency plan is "berthing receiving plan" (core requirements: processing efficiency ≥ 800 m³ / h, heel control ≤ 2°, pollutant treatment meets standards). The ship coding network contains 3 strategy units (efficiency priority, safety priority, cost priority) and corresponding 3 strategy guidance vectors (512 dimensions). The specific processing flow is as follows: The server extracts the representation vector of the target sample ship type "300,000-ton bulk carrier" and the representation vectors of each associated sample from the operation chain representation vector (512 dimensions) of the second emergency operation branch. The representation vector of the target sample ship type focuses on the core characteristics of the ship type: total cargo capacity of 50,000 m³ (weight 0.25), double bottom structure (material EH36 steel, weight 0.2), design draft of 18.2 m (weight 0.15), ballast water system treatment capacity of 1,000 m³ / h (weight 0.15), berthing heel control requirement of ≤1° (weight 0.15), and equipment adaptation cost coefficient of 20,000 yuan / h (weight 0.1). The associated samples represent subsequent links in the operational chain, and their characterization vectors are as follows: Double-bottom tank structural characteristics: including tank thickness 20mm (weight 0.3), fatigue strength index (weight 0.25), welding quality grade (weight 0.2), and damage tolerance area 0.3m² (weight 0.25); Ballast water leakage rate prediction: including breach area 0.2m² (weight 0.3), current leakage rate 800m³ / h (weight 0.35), predicted rate after 2 hours 500m³ / h (weight 0.2), and error range ±50m³ / h. (Weight 0.15); Underwater robot plugging operation: including plugging success rate 0.85 (weight 0.3), operation time 1.5h (weight 0.25), equipment adaptability (dedicated robot for bulk carriers, weight 0.2), environmental sensitivity (affected by tidal currents 0.6, weight 0.25); Mooring receiving device connection: including DN200 interface adaptability (weight 0.35), connection time 30 minutes (weight 0.25), pipeline pressure resistance 1.0MPa (weight 0.2), and manual operation complexity (weight 0.2).

[0114] The server initializes three policy orientation vectors, each representing the preference of a policy unit for the "berthing and receiving plan": Efficiency-first policy orientation vector: focusing on "processing rate matching degree" (weight 0.4), "operation time" (weight 0.3), and "equipment response speed" (weight 0.3); Safety-first policy orientation vector: focusing on "structural stability" (weight 0.35), "risk level control" (weight 0.3), and "pollutant diffusion prevention" (weight 0.35); Cost-first policy orientation vector: focusing on "equipment rental cost" (weight 0.3), "consumable consumption" (weight 0.25), "labor cost" (weight 0.2), and "post-maintenance cost" (weight 0.25). The server integrates the representation vector of the target sample ship type with the three strategy-oriented vectors element-wise to obtain three first integrated vectors (512 dimensions): Efficiency-first integrated vector: The target vector's "upper limit of processing capacity 1000m³ / h" (weight 0.15) is added to the efficiency-oriented vector's "processing rate matching degree" (weight 0.4), increasing the weight of this dimension to 0.55; "berthing connection time 30 minutes" (target vector weight 0.1) is added to the efficiency-oriented vector's "operation time" (0.3), increasing the weight to 0.4; basic features such as "cabin capacity 50000m³" (0.25) in the target vector are retained, and the final vector highlights the dimensions of "processing efficiency" and "time" (total weight percentage 0.7). Safety-First Integrated Vector: The target vector "Tilting Control ≤1°" (0.15) is added to the safety-oriented vector "Structural Stability" (0.35), increasing the weight to 0.5; "Double-Deck Bottom Structure" (0.2) is added to "Risk Level Control" (0.3), increasing the weight to 0.5; features such as "Fatigue Resistance" (0.2) are retained, highlighting the dimensions of "Stability" and "Risk Control" (total weight 0.65). Cost-First Integrated Vector: The target vector "Equipment Adaptation Cost 20,000 RMB / h" (0.1) is added to the cost-oriented vector "Equipment Rental Cost" (0.3), increasing the weight to 0.4; "Pipeline Consumable Costs" (implied in the target vector 0.05) is added to "Consumable Consumption" (0.25), increasing the weight to 0.3; features such as "Manual Operation Complexity" (0.2) are retained, highlighting the dimension of "Cost Control" (total weight 0.6).

[0115] The server performs vector multiplication on the three first integrated vectors and the representation vectors of each associated sample to obtain the sample decision influence coefficients for each associated sample under different strategies. Taking the efficiency-first strategy as an example: For the double-bottom structural characteristics associated sample: the "damage tolerance area" dimension in the efficiency-first first integrated vector has a weight of 0.2 (target vector 0.25 + efficiency-oriented vector 0), which is multiplied by the corresponding dimension (0.25) in the associated sample vector, resulting in an influence coefficient of 0.2 × 0.25 = 0.05; For the leakage rate prediction associated sample: the "processing rate matching degree" dimension in the efficiency-first first integrated vector has a weight of 0.55, which is multiplied by the "current leakage rate 800m³ / h" dimension in the associated sample vector with a weight of 0.35, resulting in an influence coefficient of 0. 0.55 × 0.35 = 0.1925; Underwater robot blocking operation associated samples: The weight of the "operation time" dimension in the first integration vector with efficiency priority is 0.4, which is multiplied by the weight of the "time taken 1.5h" dimension in the associated sample vector with 0.25, resulting in an influence coefficient of 0.4 × 0.25 = 0.1; Mooring receiving device connection associated samples: The weight of the "equipment response speed" dimension in the first integration vector with efficiency priority is 0.3, which is multiplied by the weight of the "connection time taken 30 minutes" dimension in the associated sample vector with 0.25, resulting in an influence coefficient of 0.3 × 0.25 = 0.075. Similarly, under the safety-first strategy, the multiplication of "tilt control ≤1°" (integrated vector weight 0.5) with "interface compatibility" (weight 0.35) in the "berthing connection" associated sample results in an impact coefficient of 0.5 × 0.35 = 0.175; under the cost-first strategy, the multiplication of "equipment rental cost" (integrated vector weight 0.4) with "equipment compatibility" (weight 0.2) in the "robot blocking operation" associated sample results in an impact coefficient of 0.4 × 0.2 = 0.08.

[0116] The server linearly superimposes the representation vectors of each associated sample's decision influence coefficient to obtain multiple strategy output vectors for the second emergency operation branch. Taking the efficiency-first strategy output vector as an example: double-bottom structural characteristic vector × 0.05 + leakage rate prediction vector × 0.1925 + robot plugging operation vector × 0.1 + berthing connection vector × 0.075. In the superimposed vector, the weight of the dimension "leakage rate 800m³ / h" increases from 0.35 to 0.35 × 0.1925 ≈ 0.067, and the weight of the dimension "connection time 30 minutes" increases from 0.25 to 0.25 × 0.075 ≈ 0.019. At the same time, core strategy features such as "processing rate matching degree" (0.55) are retained. The final efficiency-first strategy output vector (512 dimensions) highlights efficiency-oriented logic such as "leakage rate and equipment capacity matching" and "operation time control". The safety-first strategy output vector reinforces safety features such as "heel control accuracy" and "damage tolerance" by overlaying related sample vectors of "structural stability" and "risk level" (influence coefficients of 0.175 and 0.15, respectively). The cost-first strategy output vector, on the other hand, highlights cost features such as "leasing cost control" and "low-consumable solutions" by overlaying related sample vectors of "equipment cost" and "consumable consumption" (influence coefficients of 0.08 and 0.06, respectively). Through this process, the server generates three strategy output vectors for the first sample chain of a 300,000-ton bulk carrier, respectively adapting to the preferences of efficiency, safety, and cost priority, providing precise strategy-oriented feature expressions for subsequent strategy scheduling unit integration.

[0117] In this embodiment of the invention, multiple policy output vectors, multiple policy guidance vectors, and sample emergency scheme characterization vectors of each first sample emergency operation chain are loaded into the policy scheduling unit to obtain the second operation chain integration vector of each first sample emergency operation chain. This can be implemented through the following example.

[0118] For each of the first sample emergency operation chains, the multiple policy guidance vectors are multiplied by the sample emergency plan representation vector to obtain the decision influence coefficient of each policy output vector in each of the first sample emergency operation chains.

[0119] For each of the first sample emergency operation chains, the corresponding multiple policy output vectors are linearly superimposed based on the decision influence coefficients of the multiple policy output vectors to obtain the second operation chain integration vector of each of the first sample emergency operation chains.

[0120] In this embodiment of the invention, for example, the server takes the "first sample emergency operation chain of a 300,000-ton bulk carrier" as an example. After processing by the strategy unit, the chain generates three strategy output vectors (efficiency priority, safety priority, and cost priority, all 512-dimensional). The sample emergency plan representation vector is the "berthing and receiving plan" vector (including features such as processing efficiency of 1000 m³ / h, heel control ≤2°, DN200 interface adaptation, and equipment cost of 20,000 yuan / h, with weights of 0.3, 0.25, 0.2, and 0.15, respectively). The strategy guidance vectors include efficiency priority (focusing on processing efficiency and time consumption, with weights of 0.4 / 0.3), safety priority (focusing on heel control and risk level, with weights of 0.35 / 0.3), and cost priority (focusing on equipment cost and consumables, with weights of 0.3 / 0.25). The specific process is as follows: For this first sample chain, the server performs vector multiplication operations on the three strategy guidance vectors and the sample emergency plan representation vectors to obtain the decision influence coefficients. Efficiency-first strategy-oriented vector: Multiplied by the "processing efficiency 1000 m³ / h" dimension (weight 0.3) in the sample emergency plan representation vector (0.4 × 0.3 = 0.12), multiplied by the "operation time 30 minutes" dimension (weight 0.2) (0.3 × 0.2 = 0.06), and multiplied by other dimensions (such as cost), then summed. The total decision influence coefficient is 0.5 (the efficiency strategy has the highest matching degree with the plan requirements); Safety-first strategy-oriented vector: Multiplied by the "tilt control ≤ 2°" dimension (weight 0.25) in the sample emergency plan representation vector (0.35 × 0.2). The total decision impact coefficient is 0.3 (second best in terms of safety strategy matching degree). The cost-priority strategy orientation vector is multiplied by the dimension of "equipment cost of 20,000 yuan / h" (weight 0.15) in the sample emergency plan representation vector (0.3 × 0.15 = 0.045) and the dimension of "consumable consumption of 5,000 yuan" (weight 0.1) (0.25 × 0.1 = 0.025), resulting in a low cost strategy matching degree.

[0121] Based on the aforementioned decision influence coefficients, the server linearly superimposes the output vectors of the three strategies. The efficiency-first strategy output vector (strengthening "processing rate matching 800m³ / h" and "time taken 30 minutes", weights 0.35 / 0.25) is multiplied by a coefficient of 0.5 to obtain the efficiency contribution vector; the safety-first strategy output vector (strengthening "tilt control ≤1°" and "risk level 0.7", weights 0.3 / 0.25) is multiplied by a coefficient of 0.3 to obtain the safety contribution vector; and the cost-first strategy output vector (strengthening "equipment cost 20,000 yuan / h" and "consumables 5,000 yuan", weights 0.3 / 0.2) is multiplied by a coefficient of 0.2 to obtain the cost contribution vector. The three contribution vectors are summed element-wise to obtain the second operation chain integration vector (512 dimensions): the weight of the "processing rate matching degree" dimension is increased from 0.35 to 0.35×0.5+0.1×0.3+0.05×0.2=0.2, the weight of the "tilt control" dimension is increased from 0.3 to 0.3×0.3+0.25×0.5+0.1×0.2=0.225, while retaining features such as "equipment cost" (weight 0.15×0.2+0.1×0.5+0.3×0.3=0.17). The final vector integrates the advantages of each strategy and accurately adapts to the core requirements of the "berthing and receiving plan".

[0122] In this embodiment of the invention, the operation chain representation vector of each second sample emergency operation chain is loaded into the first fusion unit to obtain the decision representation vector of each second sample emergency operation chain. This can be implemented through the following example.

[0123] From the operation chain representation vector of the third emergency operation branch, extract the representation vector of the target sample emergency plan and the representation vector of each associated sample that has a connection relationship with the target sample emergency plan. The third emergency operation branch is any one of the plurality of second sample emergency operation chains, and the target sample emergency plan is any one of the sample emergency plans in the third emergency operation branch.

[0124] The representation vector of the emergency plan of the target sample is multiplied by the representation vector of each of the associated samples to obtain the sample decision influence coefficient of each of the associated samples in the third emergency operation branch.

[0125] Based on the sample decision influence coefficients of each of the associated samples, the representation vectors of the corresponding associated samples are linearly superimposed to obtain the decision representation vector of the third emergency operation branch.

[0126] In this embodiment of the invention, for example, the server uses the "berthing and receiving plan second sample emergency operation chain" as the third emergency operation branch. The operation steps of this branch are "berthing and receiving plan → mobile receiving device evaluation → ship berthing attitude adjustment → ballast water delivery pipeline connection → real-time pollutant monitoring". The target sample emergency plan is the "berthing and receiving plan" (core requirements: processing efficiency ≥800m³ / h, interface adaptation DN200, roll control ≤2°, pollutant monitoring accuracy ±0.1mg / L). The server extracts the representation vector of the target sample emergency plan and the representation vector of related samples from the operation chain representation vector (512 dimensions) of this branch, and generates the decision representation vector through vector multiplication and linear superposition. The specific process is as follows: The server extracts the representation vector of the target sample emergency plan "berthing and receiving plan" and the representation vector of each related sample from the operation chain representation vector of the third emergency operation branch. The characterization vector of the target sample emergency plan focuses on the core logical features of the plan: mobile receiving device processing efficiency of 1000 m³ / h (weight 0.3), interface compatibility (DN200 standard interface, weight 0.25), roll adaptability (≤2°, weight 0.2), pollutant monitoring accuracy ±0.1 mg / L (weight 0.15), and maximum operation time of 2 hours (weight 0.1). Related samples are subsequent links in the operation chain, and their characterization vectors are as follows: Mobile receiving device evaluation: including equipment model (MRT-1000 type, weight 0.3), processing capacity of 1000 m³ / h (weight 0.35), power consumption of 50 kW / h (weight 0.15), and equipment failure rate of 0.05 (weight 0.2); Ship berthing attitude adjustment: including roll control target ≤2° (weight 0.3), adjustment time of 40 minutes (weight 0.25), tidal current adaptability (current velocity ≤1.5 m / s, weight 0.2), and tugboat assistance requirements (2 vessels, 400 m³ / h). 0HP tugboat, weight 0.25); Ballast water transport pipeline connection: including interface type (DN200 flange, weight 0.35), connection time 30 minutes (weight 0.25), pipeline pressure rating 1.0MPa (weight 0.2), sealing performance (leakage rate ≤0.1m³ / h, weight 0.2); Real-time pollutant monitoring: including monitoring indicators (petroleum, suspended solids, weight 0.3), detection frequency 1 time / minute (weight 0.25), accuracy ±0.1mg / L (weight 0.3), data transmission delay ≤10s (weight 0.15).

[0127] The server performs vector multiplication on the representation vector of the target sample's emergency plan "berthing and receiving plan" and the representation vectors of each associated sample to obtain the sample decision influence coefficient. The coefficient value reflects the contribution of the associated sample to the coupled operation logic of the plan: Mobile receiving device evaluation: The weight of the "processing efficiency 1000m³ / h" dimension in the target plan vector is 0.3, which is multiplied by the weight of the "processing capacity 1000m³ / h" dimension in the associated sample vector (0.3×0.35=0.105). The "equipment failure rate" dimension (target plan weight 0.1) is multiplied by the associated sample vector. The sample vector corresponding dimension (0.2) is multiplied (0.1×0.2=0.02), and other dimensions (such as power consumption) are multiplied and then summed. The total sample decision influence coefficient is 0.3 (this step directly determines whether the scheme can meet the leakage rate requirement); Ship berthing attitude adjustment: The "roll adaptability ≤2°" dimension in the target scheme vector has a weight of 0.2, which is multiplied by the "roll control target ≤2°" dimension in the associated sample vector with a weight of 0.3 (0.2×0.3=0.06). The "tidal current adaptability" dimension (target scheme weight 0.15) is multiplied by the corresponding dimension (0.2) in the associated sample vector. .2) Multiply by (0.15 × 0.2 = 0.03), the total influence coefficient is 0.25 (attitude adjustment directly affects the stability of pipeline connection); Ballast water transportation pipeline connection: The weight of the "interface compatibility DN200" dimension in the target scheme vector is 0.25, multiplied by the weight of the "interface type DN200 flange" dimension in the associated sample vector (0.35) (0.25 × 0.35 = 0.0875), and the "connection time 30 minutes" dimension (target scheme weight 0.1) is multiplied by the corresponding dimension (0.25) of the associated sample vector (0.1 × 0.25 = 0.02). 5) The total impact coefficient is 0.25 (pipeline connection is a key link in the implementation of the plan); Real-time pollutant monitoring: The weight of the dimension "monitoring accuracy ±0.1mg / L" in the target plan vector is 0.15, which is multiplied by the weight of the dimension "accuracy ±0.1mg / L" in the associated sample vector (0.3) (0.15×0.3=0.045). The "detection frequency" dimension (target plan weight 0.1) is multiplied by the corresponding dimension (0.25) of the associated sample vector (0.1×0.25=0.025). The total impact coefficient is 0.2 (pollutant control is the environmental constraint of the plan).

[0128] The server performs linear superposition of the representation vectors of each associated sample's decision influence coefficient to obtain the decision representation vector (512 dimensions) for the third emergency operation branch. The specific superposition process is as follows: Mobile receiving device evaluation vector: weighted with an influence coefficient of 0.3, emphasizing the "processing capacity 1000 m³ / h" dimension (original weight 0.35 × 0.3 = 0.105) and the "equipment failure rate 0.05" dimension (0.2 × 0.3 = 0.06), ensuring the solution's processing efficiency covers the leakage rate (800 m³ / h); Ship berthing attitude adjustment vector: weighted with an influence coefficient of 0.25, strengthening the "roll control ≤ 2°" dimension (0.3 × 0.25 = 0.075) and the "tidal current adaptability" dimension (0.2 × 0.25 = 0.075). .05), ensuring berthing stability; Ballast water transport pipeline connection vector: weighted with an influence coefficient of 0.25, improving the "DN200 interface compatibility" dimension (0.35×0.25=0.0875) and the "sealing performance" dimension (0.2×0.25=0.05) to avoid secondary leakage caused by connection failure; Real-time pollutant monitoring vector: weighted with an influence coefficient of 0.2, retaining the "monitoring accuracy ±0.1mg / L" dimension (0.3×0.2=0.06) and the "detection frequency" dimension (0.25×0.2=0.05) to meet environmental monitoring requirements. Through linear superposition, each associated sample vector contributes to the final decision representation vector according to its influence coefficient. Compared with the original operation chain representation vector, this vector (512 dimensions) significantly increases the weight of the coupled operation logic of schemes such as "processing efficiency-leakage rate matching", "interface adaptation-ship interface compatibility" and "environmental stability-operational feasibility" (the total proportion increases from 40% to 65%). It accurately represents the collaborative operation logic of "equipment-ship-environment" in the berthing and receiving plan, and provides an input vector that focuses on the core coupling relationship for the subsequent calculation of the gating coefficient in the first gating loop unit.

[0129] In this embodiment of the invention, the original intelligent decision-making model is trained based on the sample emergency plan representation vector and the sample ship type representation vector to obtain the target intelligent decision-making model, which can be implemented through the following example.

[0130] Calculate the deviation between the sample emergency response scheme characterization vector and the sample ship type characterization vector to obtain the target error value;

[0131] The network parameters of the original intelligent decision-making model are updated based on the target error value to obtain the target intelligent decision-making model.

[0132] In an embodiment of the present invention, for example, the server takes the training sample of "high leakage scenario of a 300,000-ton bulk carrier (leakage rate of 800 m³ / h)" as an example. The sample emergency plan representation vector (U scheme) and the sample ship type representation vector (U ship type) have been generated by the coding network. The U-type scheme focuses on the core features of the "berthing and receiving plan": processing efficiency of 1000 m³ / h (weight 0.3), heel adaptability ≤2° (weight 0.25), DN200 interface compatibility (weight 0.2), pollutant monitoring accuracy ±0.1 mg / L (weight 0.15), and operation time of 1.5 h (weight 0.1). The U-type vessel focuses on the core features of a "300,000-ton bulk carrier": total cargo capacity of 50,000 m³ (weight 0.25), double-bottom structure (weight 0.2), heel control requirement ≤1° (weight 0.18), ballast water system processing capacity upper limit of 1000 m³ / h (weight 0.15), equipment adaptation cost coefficient of 20,000 yuan / h (weight 0.12), and berthing stability requirement (weight 0.1). The server completes model training by calculating the two-vector deviation and backpropagating to update network parameters. The specific process is as follows:

[0133] The server calculates the target error value by combining "overall matching degree + local feature deviation" to comprehensively evaluate the adaptability of the two vectors. The server evaluates the overall matching degree of the two vectors by cosine similarity: the vector dot product of the U scheme and the U-shaped vessel is the sum of the products of the weights of each dimension. For example, the product of "processing efficiency 1000m³ / h (U scheme 0.3)" and "processing limit 1000m³ / h (U-shaped vessel 0.15)" is 0.045, the product of "heel adaptability ≤2° (U scheme 0.25)" and "heel requirement ≤1° (U-shaped vessel 0.18)" is 0.045, the product of "DN200 interface adaptation (U scheme 0.2)" and "equipment adaptability (U-shaped vessel 0.12)" is 0.024, and the total dot product of other dimensions (such as pollutant monitoring and operation time) is 0.32. The two vectors have magnitudes of 0.7 and 0.65 respectively, and their cosine similarity is 0.32 ÷ (0.7 × 0.65) = 0.70. The overall matching deviation is 1 - 0.70 = 0.30, indicating that the overall adaptability of the solution to the ship type needs to be improved. To capture the detailed deviations in key dimensions, the server uses Huber error (with an error threshold of 1.0): for low-error dimensions (absolute error ≤ 1.0), such as the "processing efficiency matching" dimension, the error is |0.3 - 0.15| = 0.15, and the "interface adaptation" error is |0.2 - 0.12| = 0.08, with a cumulative sum of squares of 0.25; for no high-error dimensions (absolute error > 1.0), the linear error sum is 0, and the total Huber error is 0.25. The server merges the overall matching deviation (0.30) and the local Huber error (0.25) with a weight of 0.5 to obtain the target error value 0.30×0.5+0.25×0.5=0.275 (initial error, which needs to be reduced to ≤0.05 threshold).

[0134] The server uses the Adam optimizer (learning rate 0.001) to backpropagate the target error and update the network parameters of the original model (weights, gating coefficients, and collaborative vectors of the scheme coding network and the ship coding network): Scheme coding network parameter adjustments: First fusion unit attention weight: Error feedback shows that the "roll adaptability" dimension is not well matched (scheme requirement ≤ 2°, ship type requirement ≤ 1°), so the fusion unit increases the attention weight of this dimension from 0.2 to 0.3 to strengthen the scheme's adaptability to ship type roll control; First GRU gating coefficient: The original "berthing and receiving plan" gating coefficient of 0.6 was too high, causing interference from other scheme features, so it was adjusted to 0.55 to make the gating weight more in line with the actual needs of the ship type; Second decision collaborative vector: The original weight of the "cabin capacity matching degree" dimension parameter was 0.2, and the error gradient showed that it needed to be increased, so it was updated to 0.20003 (fine-tuned by 0.00003) to enhance the adaptability to the large capacity of bulk carriers. Ship coding network parameter adjustments: Second fusion unit influence coefficient: The influence coefficient of the "leakage rate prediction" associated sample was originally 0.19, which was too low, and was adjusted to 0.22 to strengthen the coupling between ship type leakage characteristics and solution processing efficiency; Second GRU reset gate weight: The weight of the tanker sample chain was originally 0.35, which was too high and interfered with the characteristics of bulk carriers, and was reduced to 0.3 to highlight the core characteristics of bulk carriers such as "double bottom structure" and "large capacity"; First decision collaboration vector: The weight of the "equipment adaptation cost" dimension parameter was originally 0.15, and the error gradient display needs to be reduced, so it was updated to 0.14998 (fine-tuned by 0.00002) to reduce the excessive influence of cost factors on ship type characterization.

[0135] The server iteratively processes 1000 samples from multiple ship types (bulk carriers, tankers, and container ships), calculating the target error every 100 iterations: initial 0.275 → 0.15 after 200 iterations → 0.08 after 500 iterations → 0.04 after 800 iterations (≤0.05 threshold), at which point the model converges. At this point, the cosine similarity between the sample emergency response plan and the ship type representation vector increases to 0.92, the Huber error decreases to 0.03, network parameters (such as GRU gating coefficients and collaborative vector dimension weights) stabilize, and the original intelligent decision-making model is trained, resulting in a target intelligent decision-making model that can accurately adapt to emergency scenarios involving multiple ship types.

[0136] In this embodiment of the invention, the target error value includes a comparison error value.

[0137] In this embodiment of the invention, for example, the server incorporates the comparison error value when calculating the target error value. Taking the "300,000-ton bulk carrier + berthing and receiving plan" sample as an example, a preset standard comparison vector (ideal matching state) is used: processing efficiency matching degree 0.3 (1000m³ / h adapts to 800m³ / h), heel control adaptability 0.25 (≤1°), and interface adaptability 0.2 (DN200 fully matches). The server compares the fusion vector (actual matching state) of the sample emergency plan representation vector and the sample ship type representation vector with the standard comparison vector. The sum of squared deviations in each dimension is 0.04 (processing efficiency deviation 0.05, heel deviation 0.03, interface deviation 0.02), resulting in a comparison error value of 0.04, which is used as a component of the target error value (accounting for 20%), and is used in conjunction with the cosine similarity deviation and Huber error to optimize the model parameters.

[0138] In this embodiment of the invention, obtaining multiple first sample emergency operation chains and multiple second sample emergency operation chains can be implemented through the following example.

[0139] An emergency operation knowledge graph is obtained, which includes multiple first sample emergency operation chains and multiple second sample emergency operation chains. The emergency operation knowledge graph is used to characterize the interdependencies between the sample ship types and the sample emergency plans.

[0140] In this embodiment of the invention, for example, the server integrates a historical case database in the field of emergency ballast water reception (containing records of over 1000 emergency events globally from 2018 to 2023), IMO Ballast Water Management Convention technical guidelines, port emergency resource allocation manuals, and an expert rule base (compiled by 15 marine engineering and environmental emergency experts) to construct an emergency operation knowledge graph. This graph uses "sample ship type," "sample emergency plan," and "operational steps" as core nodes, and "adaptation," "dependency," and "constraint" as edge relationships to dynamically represent the interdependencies among the three. It includes 500 first-sample emergency operation chains (starting with ship type) and 500 second-sample emergency operation chains (starting with emergency plan), with the specific structure as follows: In the graph, the sample ship type node includes multi-dimensional attributes: 300,000-ton bulk carrier (total cargo capacity 50,000 m³, double bottom structure, heel control ≤1°), 200,000-ton oil tanker (crude oil washing system, explosion-proof equipment requirements, pollution...). The sample emergency response plan nodes include operational logic attributes such as: berthing and receiving plan (processing efficiency 1000 m³ / h, DN200 interface), pollutant filtration plan (filtration accuracy 5 μm, processing capacity 500 m³ / h), and anchorage barge transport plan (flexible pipeline tensile strength 500 N / mm²). Operational link nodes are intermediate related nodes, such as "leakage rate prediction," "underwater robot plugging," and "roll attitude adjustment," containing attributes such as success rate, time consumption, and environmental sensitivity. In the edge relationships, "fit" represents the matching degree between the ship type and the plan (e.g., bulk carrier → berthing and receiving plan, fit degree 0.9), "dependency" represents the support of the operational link for the ship / plan (e.g., the berthing and receiving plan depends on the "DN200 interface connection" link), and "constraint" represents the limitations of the environment or resources on the operation (e.g., the anchorage barge transport plan is constrained by tidal current velocity > 1.5 m / s).

[0141] The graph stores the first sample emergency operation chain through the path "ship type → operation link → emergency plan". For example, the node for a 300,000-ton bulk carrier is connected to the "double bottom structure analysis" operation link (success rate 0.92) via the "adaptation" edge. This link is then connected to the "leakage rate prediction (800m³ / h)" link via the "dependency" edge, and finally connected to the "berthing and reception plan" node via the "constraint" edge, forming a complete first sample chain. The edge weights in the chain (e.g., "adaptation" edge 0.9, "dependency" edge 0.85) quantify the dependence strength. The second sample emergency operation chain is stored through the path "emergency plan → operation link → ship type". For example, the pollutant filtration plan node is connected to the "activated carbon addition treatment" link (treatment efficiency 500m³ / h) via the "dependency" edge. This link is connected to the 200,000-ton oil tanker node (crude oil pollutant concentration 0.6mg / L is adapted to activated carbon adsorption) via the "adaptation" edge, forming a second sample chain. The edge weights (such as "dependency" edge 0.8, "adaptation" edge 0.88) reflect the coupling requirements of the plan for the ship type.

[0142] The graph dynamically updates dependencies based on temporal attributes: when the leakage rate of a 300,000-ton bulk carrier increases from 600 m³ / h to 800 m³ / h, the "fitting" edge weight between the "berthing and receiving plan" node and the "bulk carrier" node decreases from 0.9 to 0.85, while simultaneously activating the "dual-pump parallel delivery" operation node (added to the first sample chain); when the pollutant concentration on an oil tanker increases from 0.5 mg / L to 0.8 mg / L, the "dependency" edge weight between the "pollutant filtration plan" node and the "oil tanker" node increases from 0.8 to 0.9, reinforcing the requirement for filtration accuracy. The server directly extracts these weighted sample chains through the graph query interface, providing structured dependency data for subsequent model training.

[0143] This invention provides a computer device 100, which includes a processor and a non-volatile memory storing computer instructions. When the computer instructions are executed by the processor, the computer device 100 executes the aforementioned intelligent decision-making method for emergency ballast water reception schemes based on deep learning. Figure 2 As shown, Figure 2 This is a structural block diagram of a computer device 100 provided in an embodiment of the present invention. The computer device 100 includes a memory 111, a processor 112, and a communication unit 113. To enable data transmission or interaction, the memory 111, processor 112, and communication unit 113 are electrically connected to each other directly or indirectly. For example, these components can be electrically connected to each other through one or more communication buses or signal lines.

[0144] For illustrative purposes, the foregoing description has been made with reference to specific embodiments. However, the foregoing illustrative discussions are not intended to be exhaustive or to limit the present disclosure to the precise forms disclosed. Numerous modifications and variations are possible in accordance with the foregoing teachings. These embodiments were chosen and described in order to best illustrate the principles of the present disclosure and its practical application, thereby enabling those skilled in the art to best utilize the disclosure and to employ various embodiments with different modifications to suit a particular intended application.

Claims

1. A deep learning-based intelligent decision-making method for emergency ballast water reception schemes on ships, characterized in that, include: Acquire ship ballast water emergency reception events, and collect basic information corresponding to the ship ballast water emergency reception events. The basic information includes emergency event type, ship basic parameters, ballast water status parameters, and emergency scenario environment parameters. Based on the aforementioned basic information, and combined with the historical emergency case database and expert rule database, multiple first-target emergency operation chains and multiple second-target emergency operation chains are constructed. The first-target emergency operation chain is an operation chain that starts with the target vessel type in the ballast water emergency reception scenario, and the second-target emergency operation chain is an operation chain that starts with the target emergency plan in the ballast water emergency reception scenario. Extract the operation chain representation vector of each first target emergency operation chain and the operation chain representation vector of each second target emergency operation chain; The operation chain representation vectors of multiple first target emergency operation chains and multiple second target emergency operation chains are determined as the inputs of the target intelligent decision-making model to obtain the target emergency plan representation vector and the target ship type representation vector. The matching degree is calculated based on the target emergency response plan representation vector and the target ship type representation vector. Based on the preset weights corresponding to the matching degree, feasibility, safety, and economy, the comprehensive score of the contingency plans corresponding to multiple second target emergency operation chains is calculated, and the contingency plan with the highest comprehensive score is determined as the target emergency reception plan. The intelligent decision-making model is trained and acquired through the following methods: Multiple first sample emergency operation chains and multiple second sample emergency operation chains are obtained. Each first sample emergency operation chain is used to characterize the dependency relationship between the sample ship type and the sample emergency plan when the sample ship type is the starting link. Each second sample emergency operation chain is used to characterize the dependency relationship between the sample emergency plan and the sample ship type when the sample emergency plan is the starting link. Extract the operation chain representation vector of each first sample emergency operation chain and the operation chain representation vector of each second sample emergency operation chain; The operation chain representation vectors of multiple second sample emergency operation chains are determined as the input of the scheme coding network in the original intelligent decision model to obtain the sample emergency scheme representation vector. One or more operation chain representation vectors of the first sample emergency operation chains are determined as the input of the ship coding network in the original intelligent decision model to obtain the sample ship type representation vector. The first decision coordination vector is used to represent the coordination parameters when the ship coding network dynamically adapts to the output of the scheme coding network. The original intelligent decision-making model is trained based on the sample emergency plan representation vector and the sample ship type representation vector to obtain the target intelligent decision-making model.

2. The method according to claim 1, characterized in that, The scheme coding network includes a first fusion unit and a first gated loop unit; the operation chain representation vectors of multiple second sample emergency operation chains are determined as the input of the scheme coding network in the original intelligent decision-making model to obtain sample emergency scheme representation vectors, including: The operation chain representation vector of each second sample emergency operation chain is loaded into the first fusion unit to obtain the decision representation vector of each second sample emergency operation chain, wherein each vector is used to represent the operation logic of scheme coupling in the corresponding second sample emergency operation chain. The decision representation vectors of multiple second sample emergency operation chains are determined as the input of the first gated loop unit to obtain the gating coefficient of each second sample emergency operation chain. Based on the first gating loop unit, the decision representation vectors of multiple second sample emergency operation chains are linearly superimposed with the gating coefficients to obtain the sample emergency scheme representation vector.

3. The method according to claim 2, characterized in that, Before loading the operation chain representation vectors of each second sample emergency operation chain into the first fusion unit to obtain the decision representation vectors of each second sample emergency operation chain, the method further includes: Obtain a second decision collaboration vector, which is used to characterize the collaboration parameters when the scheme coding network dynamically adapts to the output of the ship coding network; The operation chain representation vectors of each second sample emergency operation chain are loaded into the first fusion unit to obtain the decision representation vectors of each second sample emergency operation chain, including: The second decision collaboration vector is integrated with the operation chain representation vector of each second sample emergency operation chain to obtain the operation chain integration vector of each second sample emergency operation chain. The operation chain integration vector of each second sample emergency operation chain is loaded into the first fusion unit to obtain the decision representation vector of each second sample emergency operation chain.

4. The method according to claim 3, characterized in that, Based on the first gated loop unit, the decision representation vectors of multiple second sample emergency operation chains are linearly superimposed with the gate coefficients to obtain the sample emergency scheme representation vector, including: Based on the first gated loop unit, the decision representation vectors of multiple second sample emergency operation chains are linearly superimposed with the gate coefficients to obtain the target operation chain integration vector; The target operation chain integration vector is integrated with the second decision collaboration vector to obtain the sample emergency solution characterization vector.

5. The method according to claim 3, characterized in that, The ship coding network includes a second fusion unit and a second gated loop unit; it determines one or more operation chain representation vectors of the first sample emergency operation chain from the sample emergency scheme representation vector or the first decision coordination vector as the input to the ship coding network in the original intelligent decision model, to obtain the sample ship type representation vector, including: The first decision collaboration vector is integrated with the operation chain representation vector of each first sample emergency operation chain to obtain the first operation chain integration vector of each first sample emergency operation chain. From the first operation chain integration vector of the first emergency operation branch, extract the representation vector of the target sample ship type and the representation vector of each associated sample that has a connection relationship with the target sample ship type. The first emergency operation branch is any one of the plurality of first sample emergency operation chains, and the target sample ship type is any sample ship type of the first emergency operation branch. The representation vector of the target sample ship type is multiplied with the representation vector of each associated sample to obtain the sample decision influence coefficient of each associated sample in the first emergency operation branch. Based on the sample decision influence coefficient of each of the associated samples, the representation vector of the corresponding associated sample is linearly superimposed to obtain the decision representation vector of the first emergency operation branch. The decision representation vector of each first sample emergency operation chain is used to represent the ship-shaped coupling operation logic in the corresponding first sample emergency operation chain. The decision representation vector of each first sample emergency operation chain is determined as the input of the second gated loop unit to obtain the gate coefficient of each first sample emergency operation chain. The decision representation vectors of multiple emergency operation chains of the first samples are linearly superimposed with the gating coefficients based on the second gating loop unit to obtain the first gating representation vector; The first gating representation vector is integrated with the first decision collaboration vector to obtain the sample ship type representation vector.

6. The method according to claim 1, characterized in that, The ship coding network includes multiple strategy units, strategy scheduling units, and decision fusion units; one or more operation chain representation vectors of the first sample emergency operation chain are determined as inputs to the ship coding network in the original intelligent decision model to obtain the sample ship type representation vector, including: From the operation chain representation vector of the second emergency operation branch, extract the representation vector of the target sample ship type and the representation vector of each associated sample that has a connection relationship with the target sample emergency plan. The second emergency operation branch is any one of the plurality of first sample emergency operation chains, and the target sample ship type is any sample emergency plan of the second emergency operation branch. The representation vector of the target sample ship type is integrated with multiple strategy-oriented vectors to obtain multiple first integrated vectors; The first integrated vectors are multiplied by the representation vectors of each of the associated samples to obtain the sample decision influence coefficients of each of the associated samples in the second emergency operation branch. Based on the sample decision influence coefficient of each of the associated samples, the representation vectors of the corresponding associated samples are linearly superimposed to obtain multiple strategy output vectors of the second emergency operation branch. Each strategy unit outputs a strategy output vector, and each strategy guidance vector is used to represent the preference of the corresponding strategy unit for the sample emergency plan. For each of the first sample emergency operation chains, the multiple policy guidance vectors are multiplied by the sample emergency plan representation vector to obtain the decision influence coefficient of each policy output vector in each of the first sample emergency operation chains. For each of the first sample emergency operation chains, the corresponding multiple policy output vectors are linearly superimposed based on the decision influence coefficients of the multiple policy output vectors to obtain the second operation chain integration vector of each of the first sample emergency operation chains. The second operation chain integration vector of multiple first sample emergency operation chains is loaded into the decision fusion unit to obtain the sample ship type characterization vector.

7. The method according to claim 2, characterized in that, The operation chain representation vectors of each second sample emergency operation chain are loaded into the first fusion unit to obtain the decision representation vectors of each second sample emergency operation chain, including: From the operation chain representation vector of the third emergency operation branch, extract the representation vector of the target sample emergency plan and the representation vector of each associated sample that has a connection relationship with the target sample emergency plan. The third emergency operation branch is any one of the plurality of second sample emergency operation chains, and the target sample emergency plan is any one of the sample emergency plans in the third emergency operation branch. The representation vector of the emergency plan of the target sample is multiplied by the representation vector of each of the associated samples to obtain the sample decision influence coefficient of each of the associated samples in the third emergency operation branch. Based on the sample decision influence coefficients of each of the associated samples, the representation vectors of the corresponding associated samples are linearly superimposed to obtain the decision representation vector of the third emergency operation branch.

8. The method according to claim 1, characterized in that, The original intelligent decision-making model is trained based on the sample emergency response scheme representation vector and the sample ship type representation vector to obtain the target intelligent decision-making model, including: Calculate the deviation between the sample emergency response scheme characterization vector and the sample ship type characterization vector to obtain the target error value; The network parameters of the original intelligent decision-making model are updated based on the target error value to obtain the target intelligent decision-making model.

9. A server system, characterized in that, The server system includes a readable storage medium, which includes a computer program. When the computer program is executed, it controls the computer device where the readable storage medium is located to perform the method described in any one of claims 1-8.

Images