Integrated monitoring and automatic handling system for unattended engine room of bulk carrier

Abstract

The present application relates to the technical field of ship automation and intelligent monitoring, in particular to an integrated monitoring and automatic handling system for unmanned engine room of bulk carrier, comprising a data acquisition layer, an edge computing and decision-making layer, a self-healing execution layer and a network outage self-sustaining control module. The data acquisition layer acquires engine room equipment operation data; the edge computing and decision-making layer comprises a dynamic baseline generation module and a multi-agent game decision-making module, the dynamic baseline generation module dynamically generates an abnormality judgment benchmark according to the working condition by using a generative adversarial network, the multi-agent game decision-making module encapsulates core equipment as a decision-making agent, solves a global optimal handling scheme in a multi-fault scenario through collaborative game and generates a control instruction sequence; the self-healing execution layer is automatically executed after the control instruction is simulated and verified by a hardware-in-the-loop verification gateway; the present application realizes autonomous closed-loop control, multi-fault resource conflict coordination and strategy self-optimization in a network outage environment, and significantly improves the safety and reliability of the unmanned engine room.

Description

Technical Field

[0001] This invention relates to the field of ship automation and intelligent monitoring technology, specifically to an integrated monitoring and automatic handling system for the unmanned engine room of a bulk carrier. Background Technology

[0002] With the rapid development of the shipping industry towards intelligent and less-staffed operations, unmanned engine rooms have become an important trend for improving operational efficiency and reducing labor costs. As the "heart" of a ship, the engine room houses core equipment such as the main engine, generators, steering gear, and pump sets, and its operational status directly affects the ship's navigation safety. To achieve unmanned engine room operations, various monitoring and alarm systems have been developed.

[0003] In existing technologies, some solutions collect multi-source information from engine room equipment using sensors, transmit it to a shore-based facility via satellite communication, and reconstruct the engine room scene at the shore-based monitoring station, providing operators with a remote monitoring experience. Other solutions employ a redundant network structure to upload ship data to a cloud platform, achieving integrated ship-shore data sharing. Still other solutions introduce handheld wireless terminals, allowing engine room personnel to access monitoring data at any time.

[0004] However, the applicant found that the aforementioned existing technologies still have the following problems in practical applications. First, existing solutions heavily rely on satellite communication or shore-based support. When ships are sailing in high-latitude regions or encounter severe weather leading to communication interruptions, or when the network is attacked, the system will degenerate into a "blind box" state, unable to make autonomous decisions, making it difficult to achieve true unmanned operation. Second, existing alarm systems are mostly single-point threshold triggers. When multiple devices fail simultaneously, the system cannot automatically balance resource allocation and can only pile up chaotic alarm information on the crew, leading to delays in handling. Third, some systems that introduce artificial intelligence are still at the "strategy recommendation" level, only able to provide suggestions, and ultimately still require human confirmation. In unmanned scenarios, if personnel are not on-site or cannot react in time, the suggestions are meaningless. Fourth, traditional alarm methods based on fixed thresholds cannot adapt to dynamic baseline drift caused by factors such as equipment aging and changes in sea conditions, easily generating a large number of false alarms or missed alarms, reducing the reliability of the system.

[0005] More importantly, none of the aforementioned existing technologies have solved the problem of coordinating resource conflicts when multiple faults occur concurrently, nor have they achieved autonomous closed-loop control under communication interruption conditions, making it difficult to meet the comprehensive requirements of unmanned engine rooms of bulk carriers for system autonomy, safety, and reliability. Summary of the Invention

[0006] The purpose of this invention is to provide an integrated monitoring and automatic handling system for the unmanned engine room of a bulk carrier, so as to solve the shortcomings mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: an integrated monitoring and automatic handling system for an unmanned engine room of a bulk carrier, comprising:

[0008] The data acquisition layer, deployed inside the cabin, is used to collect operational data from cabin equipment.

[0009] An edge computing and decision-making layer, deployed at the ship's bow and connected to the data acquisition layer, includes:

[0010] The dynamic baseline generation module is used to generate dynamically changing anomaly judgment benchmarks based on equipment operating conditions.

[0011] The multi-agent game decision-making module encapsulates core equipment as decision-making agents, solves the global optimal handling scheme under multiple fault scenarios through collaborative game among agents, and generates control command sequences.

[0012] A self-healing execution layer, deployed at the ship's end, is connected to the edge computing and decision layer and is used to receive and automatically execute the control command sequence.

[0013] The self-sustaining control module, connected to the edge computing and decision-making layer, is used to switch the system to an offline autonomous operation mode when a communication interruption is detected.

[0014] Preferably, the dynamic baseline generation module is constructed using a generative adversarial network. The generator is trained using historical operating data under healthy conditions, so that the generator outputs corresponding theoretical health state parameters based on the current operating condition parameters. A dynamically changing standard parameter range is constructed with the theoretical health state parameters as the center. The real-time collected equipment status parameters are compared with the standard parameter range, and the equipment is determined to be abnormal when there is a continuous deviation.

[0015] Preferably, the multi-agent game decision-making module includes:

[0016] The intelligent agent management unit is used to encapsulate the core equipment in the cabin into independent decision-making intelligent agents. Each intelligent agent is configured with a state space and an action space. The state space includes the operating parameters and health status of the equipment, and the action space includes the self-healing operations that the equipment can perform.

[0017] A local policy storage unit, connected to the agent management unit, pre-configures a local self-healing policy library for each agent, storing the mapping relationship between fault types and handling operations; when any agent detects a single point of failure, it directly matches the corresponding handling operation from the local policy storage unit and executes it.

[0018] The global resource monitoring unit connects to each intelligent agent and is used to record the global resource occupancy status in real time, and to mark resource conflict events when multiple intelligent agents request to use the same restricted resource.

[0019] The collaborative game solving unit is connected to the global resource monitoring unit and each intelligent agent involved in resource conflicts. It is used to solve the globally optimal solution through collaborative game among intelligent agents when resource conflicts occur, with the overall security of the system as the goal.

[0020] The instruction generation and priority sorting unit, connected to the collaborative game solving unit, is used to prioritize the handling schemes according to the importance of equipment safety, generate a sequence of control instructions and send them to the self-healing execution layer.

[0021] The feedback processing unit, connected to the self-healing execution layer and the local policy storage unit, is used to receive instruction execution results and update policy confidence and resource status.

[0022] Preferably, the resource conflicts marked by the global resource monitoring unit include at least one of power capacity conflicts, standby equipment preemption conflicts, and cooling medium flow allocation conflicts; the instruction generation and priority sorting unit prioritizes the equipment according to the importance of its navigation safety, identifies the timing constraints in the operation sequence, and generates a control instruction sequence that satisfies the timing constraints.

[0023] Preferably, the self-sustaining control module for network outages further includes a strategy sandbox simulation unit, which includes:

[0024] The virtual environment construction subunit is used to build a digital twin virtual environment that is synchronized with the real cabin environment;

[0025] The fault scenario generation subunit is used to generate diverse fault scenarios;

[0026] The simulation execution engine subunit is used to simulate the multi-agent game decision-making process in a virtual environment and generate candidate disposal strategies;

[0027] The strategy evaluation and selection subunit is used to evaluate candidate disposal strategies and select the optimal strategy.

[0028] The policy solidification interface subunit is used to write the optimal policy into the local policy storage unit.

[0029] Preferably, the network outage self-sustaining control module further includes a decision log compression and storage unit, the decision log compression and storage unit comprising:

[0030] The log collection interface subunit is used to capture key decision-making process data in offline autonomous operation mode;

[0031] The compressed storage subunit is used to compress and block-store log data;

[0032] The communication status monitoring subunit is used to monitor the recovery status of the ship-to-shore communication link;

[0033] The backhaul control subunit is used to send the stored log data back to the shore-based center after communication is restored.

[0034] Preferably, the self-healing execution layer includes a hardware-in-the-loop verification gateway, the hardware-in-the-loop verification gateway comprising:

[0035] The real-time mirror synchronization subunit is connected to the data acquisition layer and is used to build a real-time digital twin model that keeps synchronized with the physical device state based on real-time operating data. The digital twin model includes the physical characteristics, dynamic response characteristics and topological connection relationships between the devices.

[0036] The instruction receiving interface subunit is connected to the instruction generation and priority sorting unit and is used to receive control instruction sequences;

[0037] The virtual execution engine subunit is connected to the instruction receiving interface subunit and the real-time mirror synchronization subunit, and is used to sequentially extract control instructions from the control instruction sequence, simulate execution on the real-time digital twin model, and predict the system response after instruction execution;

[0038] The security boundary storage sub-unit is used to store security boundary conditions;

[0039] The security determination subunit, connected to the virtual execution engine subunit and the security boundary storage subunit, is used to compare the predicted system response with the security boundary conditions to determine whether there is a risk exceeding the security boundary.

[0040] The instruction interception and feedback subunit is connected to the security determination subunit and is used to intercept instructions when a risk is determined to exist, and send a re-decision trigger signal to the multi-agent game decision module.

[0041] The instruction release subunit is connected to the security determination subunit and the physical execution mechanism, and is used to issue instructions to the physical execution mechanism for execution when it is determined that there is no risk.

[0042] Preferably, the data acquisition layer includes:

[0043] Multiple monitoring subunits include several sensors, including at least one of vibration sensors, temperature sensors, pressure sensors, speed sensors, current sensors, vision acquisition devices, and oil sensors, used to collect vibration parameters, thermal parameters, speed and torque parameters, electrical parameters, visual image parameters, and oil status parameters of the engine room equipment;

[0044] The multi-source data fusion subunit is connected to each monitoring subunit and is used to perform spatiotemporal alignment and correlation fusion of heterogeneous data collected by different monitoring subunits to generate a multi-dimensional state vector in a unified format, which serves as the input to the dynamic baseline generation module and the multi-agent game decision-making module.

[0045] An adaptive sampling control subunit, connected to each monitoring subunit, is used to dynamically adjust the sampling frequency of the monitoring subunits according to the equipment operating status.

[0046] The redundancy and self-diagnosis subunit is connected to each monitoring subunit and is used to perform health self-diagnosis of the sensors. When a sensor failure is detected, redundant sensors are activated or soft measurement reconstruction is performed based on relevant parameters to ensure the reliability of data acquisition.

[0047] Preferably, the local policy storage unit includes:

[0048] The strategy structured storage subunit is used to maintain an independent local self-healing strategy library for each agent, storing the mapping relationship between fault modes and handling operations;

[0049] A mapping relationship construction subunit is connected to the strategy structured storage subunit and is used to construct a mapping relationship between fault modes and handling operations.

[0050] The policy confidence management subunit is connected to the policy structured storage subunit and the feedback processing unit, and is used to maintain dynamically updated confidence parameters for each mapping relationship;

[0051] The applicable condition evaluation subunit is connected to the strategy structured storage subunit and the agent management unit, and is used to filter candidate disposal strategies that meet the current system state when a fault occurs.

[0052] The multi-strategy optimization subunit, connected to the applicable condition evaluation subunit and the strategy structured storage subunit, is used to select the optimal strategy from multiple candidate disposal strategies for the agent to execute.

[0053] Preferably, the optimal selection strategy includes:

[0054] The candidate disposal strategies are comprehensively scored based on preset evaluation indicators, which include at least one of the following: system security after strategy execution, resource consumption, fault recovery time, and degree of protection for critical equipment.

[0055] The comprehensive score is calculated using a weighted summation method, where:

[0056] The system security is calculated based on the security boundary conditions in the hardware-in-the-loop verification gateway;

[0057] The resource consumption is calculated based on the global resource occupancy recorded by the global resource monitoring unit;

[0058] The fault recovery time is based on statistics of the instruction execution results returned by the feedback processing unit;

[0059] The degree of protection for critical equipment is calculated based on the importance of the equipment's navigation safety in the instruction generation and priority sorting unit;

[0060] The optimal strategy is selected based on the overall score.

[0061] Compared with the prior art, the beneficial effects of the present invention are:

[0062] Compared with existing technologies, this invention, by setting up a self-sustaining control module in the event of a network outage, automatically switches to an offline autonomous operation mode when a communication interruption is detected. This allows edge computing and the decision-making layer to independently complete the entire decision-making process based entirely on the ship's computing power and local data, while the self-healing execution layer continues to execute closed-loop control. This achieves absolutely autonomous operation with "uninterrupted control even when the network is down," fundamentally solving the decision-making vacuum problem caused by the reliance on communication links in existing technologies. Furthermore, by using a multi-agent game theory decision-making module, the core equipment in the engine room is encapsulated as an independent decision-making agent, and a global resource monitoring unit records the global resource occupancy in real time. When multiple failures occur and resource conflicts exist, the collaborative game theory solution unit, with the overall system security as the goal, solves the global optimal solution through collaborative game theory among agents, thus resolving the decision implosion problem when multiple failures occur concurrently. Finally, by constructing a real-time digital twin model synchronized with the physical equipment through a hardware-in-the-loop verification gateway, the control command is simulated and compared with safety boundary conditions before being issued. The command is only released when no risk is detected; if a risk is detected, it is intercepted and a re-decision is triggered, eliminating the security risks caused by model misoperation.

[0063] The dynamic baseline generation module uses generative adversarial networks to construct standard parameter ranges that adapt to changing operating conditions, replacing traditional fixed thresholds. This effectively solves the problem of dynamic baseline drift caused by equipment aging and sea state changes, significantly reducing false alarm and false negative rates. The strategy sandbox simulation unit uses a digital twin virtual environment to simulate and simulate optimal handling strategies for various fault scenarios when communication is normal and shipboard computing power is idle. These strategies are then stored in the local strategy storage unit. Simultaneously, the strategy confidence management subunit dynamically adjusts the strategy confidence based on actual execution results, achieving continuous optimization and self-evolution of the strategy library. The decision log compression storage unit compresses and stores key decision process data in blocks in offline autonomous operation mode. After communication is restored, the data is transmitted back to the shore-based center through a breakpoint resume mechanism. This solves the bottleneck of limited satellite communication bandwidth and ensures the integrity and traceability of decision data. Attached Figure Description

[0064] Figure 1 This is a diagram of the overall system structure of the present invention. Detailed Implementation

[0065] 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 embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0066] Please see Figure 1 This invention provides a technical solution: an integrated monitoring and automatic response system for an unmanned engine room of a bulk carrier. Deployed at the end of the bulk carrier, the system enables real-time monitoring, autonomous decision-making, and automatic response of engine room equipment. The system mainly includes a data acquisition layer, an edge computing and decision-making layer, a self-healing execution layer, and a self-sustaining control module for network outages.

[0067] Furthermore, in order to achieve accurate perception and reliable collection of the operating status of cabin equipment, this embodiment sets up a data acquisition layer.

[0068] The data acquisition layer is deployed within the engine room to collect operational data from the engine room equipment. Specifically, the data acquisition layer includes multiple monitoring sub-units, each containing several sensors. These sensors include at least one of vibration sensors, temperature sensors, pressure sensors, speed sensors, current sensors, vision sensors, and oil level sensors, used to collect vibration parameters, thermal parameters, speed and torque parameters, electrical parameters, visual image parameters, and oil level status parameters of the engine room equipment.

[0069] In this embodiment, vibration sensors are deployed at the bearings, bases, and housings of core equipment such as the main unit, generator, and pump set to collect the vibration amplitude, frequency, and phase characteristics of the equipment. Temperature sensors, pressure sensors, flow sensors, and level sensors are deployed at key measuring points of each core equipment to collect temperature, pressure, flow, and level parameters. Speed ​​sensors and torque sensors are deployed on the rotating shafts of the main unit, generator, and pump set to collect speed, direction, and torque parameters. Current transformers, voltage transformers, and power transmitters are deployed in the power supply circuits of the main distribution board, generator output terminals, and key loads to collect current, voltage, power factor, and active power parameters. High-definition vision acquisition units and infrared thermal imaging acquisition units are deployed in key areas of the engine room. The high-definition vision acquisition unit is used to acquire visible light images of the equipment appearance, instrument readings, and pipeline status, while the infrared thermal imaging acquisition unit is used to acquire thermal imaging images of the surface temperature distribution of the equipment to identify local overheating anomalies. Oil sensors are deployed on the circulation pipelines of the main unit's lubricating oil system, steering gear hydraulic system, and auxiliary engine oil system to monitor the viscosity, density, moisture content, and abrasive particle condition of the oil in real time.

[0070] The raw signals collected by each monitoring subunit first enter the intelligent acquisition preprocessing subunit. This subunit performs preprocessing on the raw signals, including signal amplification, filtering and noise reduction, analog-to-digital conversion, feature extraction, timestamp marking, and data encapsulation, forming a standardized data format. Specifically, feature extraction is performed as follows: extracting time-domain features such as root mean square value, peak value, and kurtosis from vibration signals, or extracting spectral features through Fourier transform; extracting features such as fundamental amplitude and harmonic content from current signals; and extracting features such as rate of change and fluctuation range from temperature and pressure signals. Subsequently, the multi-source data fusion subunit connects to each monitoring subunit, performing spatiotemporal alignment and correlation fusion of heterogeneous data from different monitoring subunits to generate a unified format multidimensional state vector. This multidimensional state vector contains the vibration characteristics, thermal parameters, electrical parameters, visual features, environmental parameters, and oil state of each device at the same timestamp, serving as input to the subsequent dynamic baseline generation module and multi-agent game decision-making module.

[0071] To optimize data acquisition efficiency and computing resources, the adaptive sampling control subunit is connected to each monitoring subunit, dynamically adjusting the sampling frequency of the monitoring subunits based on the current operating status and fault risk level of the equipment. When the equipment is in normal operating condition, the basic sampling frequency is used to reduce the amount of data; when the equipment shows abnormal signs or enters the fault handling stage, the sampling frequency and resolution of the relevant monitoring subunits are automatically increased to obtain more detailed status data, providing support for accurate decision-making.

[0072] Meanwhile, the redundancy and self-diagnosis subunit performs health self-diagnosis on each sensor, identifying sensor drift, failure, or communication faults. When a sensor failure is detected, redundant sensors are automatically activated or soft measurement reconstruction is performed based on relevant parameters to ensure the reliability of data acquisition. The specific implementation of soft measurement reconstruction is as follows: when the temperature sensor fails, the temperature value is estimated using a pre-trained neural network model based on other related parameters of the same device, such as current, speed, and pressure. This neural network model takes current, speed, and pressure as inputs and temperature as output, and is trained using historical normal operating condition data. When the pressure sensor fails, a pressure estimation model is established based on the pump's speed and current.

[0073] Furthermore, to address the problem of false alarms and missed alarms caused by the inability of fixed threshold alarms to adapt to changes in operating conditions in existing technologies, this embodiment sets up a dynamic baseline generation module.

[0074] The edge computing and decision-making layer is deployed at the ship's end and connected to the data acquisition layer, including a dynamic baseline generation module and a multi-agent game decision-making module.

[0075] The dynamic baseline generation module generates dynamically changing anomaly detection benchmarks based on equipment operating conditions. This module is constructed using a generative adversarial network, and its specific implementation is as follows:

[0076] First, collect historical operating data of the equipment in a healthy state. This historical operating data includes operating parameters and their corresponding equipment status parameters. Operating parameters include main engine speed, load, seawater temperature, etc.; equipment status parameters include vibration amplitude, temperature value, pressure value, etc.

[0077] Secondly, a generative adversarial network (GAN) is constructed, consisting of a generator and a discriminator. The generator takes the current operating condition parameters as input and outputs the simulated health state parameters corresponding to those parameters. The discriminator takes the simulated health state parameters output by the generator and the actual historical health state parameters as input and outputs the discrimination result. The historical health data is divided into training and validation sets, and the generator and discriminator are optimized alternately: the generator minimizes the discriminator's discrimination probability of its output, while the discriminator maximizes its ability to distinguish between real and generated data. Iterative training continues until the discriminator's accuracy stabilizes at around 50%, at which point the generator has learned the mapping relationship between the operating condition parameters and the health state parameters.

[0078] After the trained generator is deployed on the ship's deck, during actual operation, it takes real-time collected operating parameters as input and outputs corresponding theoretical health status parameters. Centered on this theoretical health status parameter μ, the standard deviation σ is calculated based on the residual distribution of historical health data, constructing a dynamically changing standard parameter interval [μ-2σ, μ+2σ], which serves as the benchmark for judging anomalies under the current operating conditions. This interval covers 95% of the normal fluctuation range. The real-time collected equipment status parameters are compared with this standard parameter interval. When the real-time status parameters continuously deviate from this interval and the deviation exceeds a preset threshold, the equipment is determined to be abnormal.

[0079] Furthermore, to address the problem in existing technologies that the inability to coordinate resource conflicts and the limitation of single-point alarms when multiple faults occur concurrently, this embodiment sets up a multi-agent game decision-making module.

[0080] The multi-agent game theory decision-making module encapsulates core devices as decision-making agents. Through collaborative game theory among these agents, it solves the globally optimal handling scheme for multiple fault scenarios and generates a sequence of control commands. This module specifically includes an agent management unit, a local policy storage unit, a global resource monitoring unit, a collaborative game theory solving unit, a command generation and priority sorting unit, and a feedback processing unit.

[0081] The agent management unit is used to encapsulate the core equipment in the cabin into independent decision agents. Each agent corresponds to a physical device or functional unit, and each agent is configured with a unique identifier, state space, action space and communication interface. The state space includes the device's real-time operating parameters, health status and remaining lifespan. The action space includes the set of self-healing operations that the device can perform, such as device switching, parameter adjustment, load transfer, start and stop control, etc.

[0082] The local policy storage unit is connected to the agent management unit and pre-configures a local self-healing policy library for each agent. This policy library uses a multi-dimensional index structure to store the mapping relationship between fault modes and handling operations. The index dimensions include fault type, device status, operating parameters, and execution cost. When any agent detects a single point of failure through the dynamic baseline generation module, it directly matches the corresponding handling operation from the local policy storage unit and executes it without consulting with other agents, thus achieving a rapid response.

[0083] The global resource monitoring unit connects with each agent, maintains a shared global resource status table, and records the occupancy of global resources such as power capacity, standby equipment status, and cooling medium flow rate in real time. When multiple agents simultaneously trigger a fault and request the use of the same restricted resource, the global resource monitoring unit identifies and marks the resource conflict event, triggering a collaborative game process. The types of resource conflicts include power capacity conflicts, standby equipment preemption conflicts, and cooling medium flow rate allocation conflicts. Power capacity conflicts refer to multiple agents simultaneously requesting to increase power load, and the total amount requested exceeds a preset safety threshold of the currently available power capacity. Standby equipment preemption conflicts refer to multiple faulty agents simultaneously requesting to use the same standby equipment, and the standby equipment cannot satisfy all requests at the same time. Cooling medium flow rate allocation conflicts refer to multiple agents simultaneously requesting to increase cooling medium flow rate, and the total amount requested exceeds a preset safety threshold of the maximum supply capacity of the cooling system.

[0084] The collaborative game solving unit is connected to the global resource monitoring unit and the agents involved in resource conflicts. Upon receiving a resource conflict trigger signal, the collaborative game solving unit receives the current state, fault severity, required resource amount, and expected payoff function broadcast by each conflicting agent. With maximizing the overall security of the system as the objective function and the actions of each conflicting agent as the decision variables, a non-cooperative game model is constructed.

[0085] In this embodiment, an iterative optimization algorithm is used to solve the Nash equilibrium, that is, to find a set of policy combinations such that no agent can obtain better system security benefits by unilaterally changing its own policy. Based on the Nash equilibrium policy combination obtained, a globally optimal disposal scheme including multi-device collaborative operation is generated. This scheme includes the operation sequence, execution timing and resource allocation scheme to be executed by each relevant agent.

[0086] In one specific implementation of this embodiment, after the global resource monitoring unit marks a resource conflict event, the cooperative game solving unit initiates the Nash equilibrium solving process, as follows:

[0087] Let the set of agents involved in the conflict be A={a1, a2, …, an}. Each agent ai has a finite set of actions Si, where action si∈Si represents the disposal options that agent ai can choose, such as switching to backup equipment, adjusting operating parameters, transferring load, or starting or stopping equipment.

[0088] First, the collaborative game-solving unit receives the current state information broadcast by each conflicting agent, including: the current fault severity score Fi of agent AI, where Fi ranges from 0 to 1, with a larger value indicating a more severe fault; the required resource quantity Ri, which includes the required power capacity ΔPi, the required cooling medium flow rate ΔQi, or the required backup equipment identifier; and the expected reward function Ui(si, si), which represents the reward gained by agent AI when it chooses action si and other agents choose action combination si. The reward function Ui(si, si) is defined as Ui = Fi - Ci, where Fi is the reduction in fault severity, and Ci is the cost of executing action si, including resource consumption and the impact on other devices.

[0089] Secondly, the collaborative game solving unit constructs a non-cooperative game model G={A, {Si}, {Ui}}, with the goal of maximizing the overall system security. The overall system security function is defined as the weighted sum of the payoffs of each agent: J(s1, s2, …,sn) = Σ wi · Ui(si, si), where wi is the navigation safety weight of agent ai, and wi is the product of the static base weight and the dynamic adjustment factor. The base weight is pre-set according to the degree of impact of equipment failure on the ship's navigation safety. The first-level equipment such as the main engine and steering gear has the highest weight, followed by the second-level equipment such as auxiliary engines, and the third-level equipment such as auxiliary equipment has the lowest weight. The dynamic adjustment factor is calculated in real time according to the ship's navigation status. For example, when the ship is in a narrow channel, the dynamic adjustment factor is 1.2, when it is in open water, it is 0.8, and when it is in bad weather, it is 1.5.

[0090] Then, the collaborative game solving unit uses an iterative optimization algorithm to solve for the Nash equilibrium, specifically including the following steps:

[0091] The first step is to initialize the policy combination. For each agent ai, an initial action si0 ∈ Si is randomly selected to obtain the initial policy combination σ0 = (s10, s20, …, sn0), and the initial overall system security score J0 is calculated.

[0092] The second step is iterative optimization. For the k-th iteration, each agent ai is iterated sequentially: The current action combination si(k) of other agents is fixed, and the optimal response strategy for agent ai is solved, i.e., the action si(k+1) = argmax_{si∈Si} Ui(si, si(k)) is found to maximize the agent's gain under the current strategies of other agents. If the solved si(k+1) differs from the current action si(k), the action of agent ai is updated, and the overall system safety score J(k+1) is recalculated.

[0093] The third step is convergence determination. After traversing all agents in one iteration, the change in the policy combination before and after this iteration is calculated as Δ = ||σ(k+1) - σ(k)||, where ||·|| represents the Euclidean distance of the policy combination. When Δ is less than the preset convergence threshold ε, the policy combination σ(k+1) is determined to be in Nash equilibrium; if it has not converged, the next iteration continues until the maximum number of iterations is reached.

[0094] The fourth step is to output the Nash equilibrium policy combination. The converged policy combination σ = (s1, s2, …, sn) is taken as the globally optimal solution.

[0095] Based on the Nash equilibrium strategy combination obtained from the solution, the collaborative game solving unit generates a globally optimal disposal plan that includes multi-device collaborative operations. This plan specifically includes: the sequence of operations to be executed by each relevant agent, such as: the host agent switching to the backup oil pump, the auxiliary agent reducing load, and the cooling water pump agent increasing speed; the execution sequence of each operation, clarifying the order and synchronization of operations, for example: first switching the backup oil pump, then after a 3-second delay, reducing the load on the auxiliary agent while simultaneously starting the cooling water pump to increase speed; and a resource allocation plan, clarifying the allocation of limited resources such as power capacity, backup equipment, and cooling medium, for example: allocating the backup generator to the host agent, allocating battery power to the auxiliary agent, and allocating the total cooling medium flow rate at a ratio of 60% for the host and 40% for the auxiliary agent.

[0096] Through the above iterative optimization process, the collaborative game solving unit can converge to Nash equilibrium within a finite number of steps, generating a stable strategy combination in which each agent has no incentive to deviate unilaterally, thereby achieving global optimal resource coordination and disposal decision-making in multi-fault concurrent scenarios.

[0097] The instruction generation and priority ranking unit is connected to the collaborative game solving unit, prioritizing the operation sequences in the global optimal solution based on the importance of each device for navigation safety. The importance of navigation safety is categorized based on the impact of device failure on ship navigation safety: Level 1 devices include the main engine, steering gear, and navigation equipment; their failure will directly lead to the ship losing its navigation capability or causing a major safety accident. Level 2 devices include the main generator, main switchboard, and cooling water pumps; their failure will severely affect the ship's power or safety redundancy capabilities. Level 3 devices include auxiliary generators, ballast water systems, and sewage treatment facilities; their failure will affect the ship's economy or comfort but will not affect navigation safety in the short term. The safety criticality level is dynamically weighted and adjusted according to the ship's current navigation status, including water depth, channel width, meteorological and sea conditions, distance from shore, and traffic density. Under complex navigation conditions, the real-time weights of Level 2 and Level 3 devices are increased. Based on the dynamically adjusted safety criticality weights, the operation sequence is sorted in multiple levels: the first priority is the operation involving first-level equipment and the operation is necessary to ensure the basic function of the equipment; the second priority is the non-essential operation involving first-level equipment and the essential operation involving second-level equipment; the third priority is the non-essential operation involving second-level equipment and the essential operation involving third-level equipment; and the fourth priority is the non-essential operation involving third-level equipment.

[0098] The instruction generation and priority sorting unit also identifies temporal dependencies in the operation sequence, including operation order dependencies, operation synchronization dependencies, and operation time window constraints. Based on these temporal dependencies, the priority sorting is adjusted a second time to generate a final operation sequence that satisfies the temporal constraints. Each operation is then converted into a standard format control instruction according to the final operation sequence. The control instruction includes the target device identifier, operation type, operation parameters, execution timestamp, and timeout threshold. The generated instruction sequence is then sequentially distributed to the self-healing execution layer.

[0099] The feedback processing unit is connected to the self-healing execution layer and the local policy storage unit. It receives the instruction execution results returned by the self-healing execution layer, updates the confidence level of the corresponding policy in the local policy storage unit, and synchronously updates the global resource status table in the global resource monitoring unit to form a closed-loop optimization.

[0100] Furthermore, to achieve secure verification of decision-making instructions and prevent security risks caused by AI misoperation, this embodiment sets up a hardware-in-the-loop verification gateway.

[0101] The self-healing execution layer is deployed at the ship's end and connects to the edge computing and decision layer. It is used to receive and automatically execute control command sequences. The self-healing execution layer includes a hardware-in-the-loop verification gateway.

[0102] The hardware-in-the-loop verification gateway is a key component for ensuring the secure execution of control commands. It includes a real-time mirror synchronization subunit, a command receiving interface subunit, a virtual execution engine subunit, a security boundary storage subunit, a security decision subunit, a command interception and feedback subunit, and a command release subunit.

[0103] The real-time mirror synchronization subunit connects to the data acquisition layer, receiving real-time operational data from the cabin equipment and constructing a real-time digital twin model that remains synchronized with the physical equipment's state based on this data. This digital twin model includes the equipment's geometry, physical characteristics, dynamic response characteristics, and their interconnections, enabling it to map the current state of the physical equipment in real time and predict its dynamic response under external influences. The real-time digital twin model is updated at a fixed time step Δt. Within each step, it receives the latest multidimensional state vector from the data acquisition layer, updates the operating parameters of each virtual device in the model, and extrapolates the state at the next moment based on the physical model equations.

[0104] The instruction receiving interface subunit is connected to the instruction generation and priority sorting unit. It receives the sequence of control instructions issued by the multi-agent game decision module, performs format parsing and integrity verification on the instructions, and temporarily stores the parsed instructions in the instruction queue to be verified.

[0105] The virtual execution engine subunit is connected to the instruction receiving interface subunit and the real-time mirror synchronization subunit. It sequentially retrieves control instructions from the instruction queue to be verified and performs simulated execution on the real-time digital twin model. The simulation execution process includes: applying the control instructions to the corresponding virtual devices in the digital twin model, deduce the change process of the device state according to the preset simulation step size, and predict the complete response trajectory of the system after the instruction is executed.

[0106] The safety boundary storage subunit pre-stores multi-dimensional safety boundary conditions, including physical limits of equipment parameters, parameter ranges for stable system operation, constraints for multi-device collaborative operation, and dynamic safety margins derived from ship navigation safety. The dynamic safety margin is calculated in real time based on the ship's current navigation status. For example, in narrow channels or severe weather conditions, it narrows the allowable fluctuation range of each parameter to 80% of the normal value.

[0107] The safety determination subunit is connected to the virtual execution engine subunit and the safety boundary storage subunit. It compares the predicted system response after simulation with the safety boundary conditions item by item to determine whether there is a risk exceeding the safety boundary. The determination includes whether the predicted parameters exceed physical limits, whether the system enters an unstable state, whether conflicts occur in multi-device collaboration, and whether the safety margin is lower than a preset threshold.

[0108] The instruction interception and feedback subunit is connected to the safety determination subunit. When the safety determination subunit determines that the simulation result exceeds the safety boundary, it immediately intercepts the corresponding control instruction, prohibits it from being sent to the physical execution mechanism, and sends a re-decision trigger signal to the multi-agent game decision module. At the same time, it feeds back the intercepted instruction and the determination result to the instruction generation and priority sorting unit so that it can adjust the subsequent instruction generation strategy.

[0109] The instruction release subunit is connected to the safety determination subunit and the physical actuator. When the safety determination subunit determines that all simulation results are within the safety boundary, it converts the corresponding control instructions into a format that the physical actuator can recognize, and sends them to the physical actuator for execution according to the execution timestamp and timing requirements in the instructions, thereby realizing closed-loop automatic control of the cabin equipment.

[0110] Furthermore, to enable the system to operate autonomously and optimize its strategies when communication is interrupted, this embodiment includes a network outage self-sustaining control module.

[0111] The self-sustaining control module is connected to the edge computing and decision-making layer. It is used to switch the system to an offline autonomous operation mode when a communication interruption is detected. This module includes a strategy sandbox simulation unit and a decision log compressed storage unit.

[0112] To enable the system to continuously optimize the strategy library without relying on shore-based support, this embodiment includes a strategy sandbox simulation unit. When communication is normal and ship-side computing power is idle, the strategy sandbox simulation unit utilizes available computing power for strategy optimization.

[0113] The virtual environment construction subunit constructs a digital twin virtual environment synchronized with the real cabin environment based on the physical parameters, topology, and operating mechanisms of the cabin equipment. This digital twin virtual environment contains virtual device models that correspond one-to-one with the physical devices, and these virtual device models can simulate the dynamic response characteristics and fault propagation behavior of the real equipment.

[0114] The fault scenario generation subunit automatically generates a diverse set of fault scenarios based on historical fault data, equipment health status, and current operating conditions. These scenarios include single-point faults, multi-point concurrent faults, and resource conflict faults. Each fault scenario includes fault type, fault occurrence time, and fault intensity parameters.

[0115] The simulation scheduling subunit monitors the status of ship-to-shore communication links and ship-side computing power load. When it detects that communication is normal and the ship-side computing power utilization rate is lower than the preset idle threshold, it triggers the strategy simulation process.

[0116] The simulation execution engine subunit loads the fault scenario generated by the fault scenario generation subunit into the digital twin virtual environment, and calls a copy of the multi-agent game decision-making module to perform simulated decision-making in the virtual environment. The simulated decision-making process includes: agents perceiving anomalies in the virtual environment, executing local self-healing strategies, initiating cooperative game when resource conflicts occur, generating disposal plans, and simulating their execution.

[0117] The strategy evaluation and selection subunit evaluates multiple candidate handling strategies generated from multiple simulations under the same fault scenario. The evaluation indicators include system security, resource consumption, fault recovery time, and degree of protection for critical equipment after strategy execution.

[0118] To achieve quantitative comparison and selection of multiple candidate strategies, this embodiment adopts a comprehensive scoring method. Specifically, the strategy evaluation and selection sub-unit performs a comprehensive score on each candidate disposal strategy based on preset evaluation indicators, and the comprehensive score is calculated using a weighted summation method.

[0119] The calculation formula is: Score = α·S + β·(1 / R) + γ·(1 / T) + δ·P, where S is the system security score after policy execution, which is calculated based on the security boundary conditions in the hardware-in-the-loop verification gateway. This score is normalized based on the deviation between the predicted system parameters after policy execution and the security boundary; the smaller the deviation, the higher the score. R is the resource consumption, calculated based on the global resource occupancy recorded by the global resource monitoring unit, including the power capacity required for policy execution, cooling medium flow rate, and the number of backup devices. T is the fault recovery time, calculated based on the historical instruction execution results returned by the feedback processing unit, representing the time span from the occurrence of a fault to the system's return to normal operation. P is the threshold value for the system's security performance. The protection level of key equipment is calculated based on the importance of equipment for navigation safety in the instruction generation and priority sorting unit. That is, after the strategy is executed, the health status protection effect of first-level equipment, second-level equipment, and third-level equipment is weighted and scored. α, β, γ, and δ are the weight coefficients of each evaluation index. Each weight coefficient is dynamically adjusted according to the current navigation status of the ship. When the ship is in a narrow channel, in bad weather, or in a complex navigation condition with high traffic density, the weight of system safety α and the weight of protection level of key equipment δ are increased. When the ship is in open water and communication is normal, the weight of resource consumption β and the weight of fault recovery time γ can be appropriately increased.

[0120] The strategy evaluation and optimization subunit calculates a comprehensive score for multiple candidate handling strategies generated from multiple simulations under the same fault scenario. The strategy with the highest comprehensive score is output as the optimal strategy to the strategy solidification interface subunit. Through the above comprehensive scoring and weighted summation mechanism, a scientific and quantitative evaluation of candidate handling strategies is achieved, ensuring that the selected strategy achieves the optimal balance between system security, resource efficiency, recovery speed, and protection of critical equipment.

[0121] The strategy solidification interface subunit converts the candidate solidification strategies into a format supported by the local strategy storage unit and writes them into the local self-healing strategy library of the corresponding agent. At the same time, it records the source of the strategy, the applicable conditions, and the initial confidence value.

[0122] The incremental learning management subunit tracks the execution effect of the solidified strategy in practical applications. Based on the execution results returned by the feedback processing unit, it dynamically adjusts the confidence level of the solidified strategy. When the confidence level falls below a preset threshold, it triggers a re-analysis or strategy elimination. The confidence level update uses an exponentially weighted moving average: Ct+1 = λ·Ct + (1-λ)·Rt, where Rt is the execution effect score, and λ is the forgetting factor, ranging from 0 to 1, with a typical value of 0.9.

[0123] Furthermore, to address the issue of reliable storage and transmission of decision data in offline mode, this embodiment includes a decision log compression storage unit. This unit compresses and stores key decision process data in offline autonomous operation mode.

[0124] The log collection interface subunit captures key decision-making process data in real time in offline mode, including the state change sequence of each agent, the communication content between agents, the game process data of the collaborative game solving unit, the intermediate results of Nash equilibrium solving, the generated global optimal disposal scheme, the control instruction sequence output by the instruction generation and priority sorting unit, the instruction verification results of the hardware-in-the-loop verification gateway, the instruction execution feedback of the self-healing execution layer, and the intercepted instruction records.

[0125] The data tiered filtering subunit filters the collected raw log data based on data type, event type, equipment safety criticality, and anomaly severity, removing redundant and low-value data and retaining key information for subsequent model optimization and incident analysis. The specific filtering rules are as follows: all event logs for Level 1 equipment are retained; only fault event logs for Level 2 equipment are retained; and only severe fault logs for Level 3 equipment are retained; data is forcibly retained when the anomaly severity exceeds a preset threshold.

[0126] The streaming compression coding subunit performs real-time compression processing on the filtered log data using streaming compression algorithms. It adaptively selects the compression method according to the data type, performs differential coding and entropy coding on time-series data, dictionary compression on text data, and structured coding on game process data.

[0127] The block storage management subunit stores the compressed log data in blocks according to time windows. Each data block contains the complete decision log within a fixed time window, and a unique identifier, timestamp range, and data integrity check code are generated for each data block. The block storage management subunit adopts a circular buffer mechanism, automatically overwriting the oldest data block according to the first-in, first-out principle when the storage space reaches a preset threshold. The preset threshold can be set to 90% of the total storage capacity, or dynamically set according to the expected communication recovery time.

[0128] The communication status monitoring subunit monitors the recovery status of the ship-to-shore communication link in real time. When it detects that the communication link has been restored to normal and the bandwidth meets the transmission requirements, it triggers the log back transmission process.

[0129] After communication is restored, the breakpoint resume control subunit transmits the stored log data blocks back to the shore-based center one by one according to the time sequence of the data blocks. If the communication link is interrupted again, it records the identifier of the data block that has been successfully transmitted and continues to transmit from the point of interruption after communication is restored, thus avoiding duplicate transmission.

[0130] The data encryption and integrity verification subunit encrypts the log data before it is transmitted back and attaches a digital signature to each data block. After receiving the data, the shore-based center verifies the integrity and authenticity of the data through the digital signature to ensure that the transmitted data has not been tampered with.

[0131] The data reception confirmation and cleanup subunit receives the data reception confirmation signal returned by the shore-based center, marks or cleans up the data blocks that have been successfully transmitted and confirmed, and releases storage space.

[0132] To more clearly illustrate the overall working process of this invention, a specific application scenario will be used as an example below:

[0133] Assuming a bulk carrier is sailing in high-latitude waters and satellite communication is interrupted, the self-sustaining control module detects the communication interruption and automatically switches the system to offline autonomous operation mode.

[0134] The data acquisition layer continuously collects operational data from the cabin equipment, while the dynamic baseline generation module uses a generative adversarial network to determine in real time whether the equipment is malfunctioning. When the main engine lubricating oil pressure is abnormal, the corresponding main engine agent detects a single point of failure and directly matches a handling operation from the local policy storage unit, such as switching to the backup lubricating oil pump, without needing to negotiate with other agents.

[0135] If the auxiliary engine fails at the same time and the backup generator needs to be switched, but only one backup generator is available, the global resource monitoring unit marks the resource conflict event and triggers the collaborative game solving unit. The main engine agent and the auxiliary engine agent play a game with the overall system safety as the goal and solve the Nash equilibrium. Considering that the failure of the main engine will directly cause the ship to lose power, while the failure of the auxiliary engine can be temporarily powered by the battery, the game result determines that the backup generator should be allocated to the main engine first, and the auxiliary engine should implement a load reduction operation strategy.

[0136] The instruction generation and priority sorting unit generates a sequence of control instructions based on the importance of navigation safety. The hardware-in-the-loop verification gateway simulates the execution on the digital twin model. After confirming that there are no safety risks, the instructions are sent to the physical actuators for execution, thus achieving autonomous closed-loop control.

[0137] Meanwhile, the decision log compression storage unit compresses and stores the data of the entire decision-making process, which will be automatically transmitted back to the shore-based center after communication is restored. After communication is restored, the strategy sandbox simulation unit uses idle computing power to simulate and optimize strategies in the virtual environment based on the experience of handling this fault, and solidifies the better handling solution into the local strategy storage unit to achieve continuous optimization of the strategy library.

[0138] Through the above technical solution, this invention constructs a complete intelligent closed loop of "perception-decision-verification-execution-self-sustainability", realizing a qualitative change from "assisting human monitoring" to "machine autonomous decision-making and execution", and providing reliable technical support for unmanned engine rooms of bulk carriers.

[0139] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An integrated monitoring and automatic handling system for an unmanned engine room of a bulk carrier, characterized in that, include: The data acquisition layer, deployed inside the cabin, is used to collect operational data from cabin equipment. An edge computing and decision-making layer, deployed at the ship's bow and connected to the data acquisition layer, includes: The dynamic baseline generation module is used to generate dynamically changing anomaly judgment benchmarks based on equipment operating conditions. The multi-agent game decision-making module encapsulates core equipment as decision-making agents, solves the global optimal handling scheme under multiple fault scenarios through collaborative game among agents, and generates control command sequences. A self-healing execution layer, deployed at the ship's end, is connected to the edge computing and decision layer and is used to receive and automatically execute the control command sequence. The self-sustaining control module, connected to the edge computing and decision-making layer, is used to switch the system to an offline autonomous operation mode when a communication interruption is detected.

2. The integrated monitoring and automatic handling system for an unmanned engine room of a bulk carrier according to claim 1, characterized in that, The dynamic baseline generation module is constructed using a generative adversarial network. The generator is trained using historical operating data under healthy conditions. The generator outputs corresponding theoretical health state parameters based on the current operating conditions. A dynamically changing standard parameter range is constructed with the theoretical health state parameters as the center. The real-time collected equipment status parameters are compared with the standard parameter range. When there is a continuous deviation, the equipment is determined to be abnormal.

3. The integrated monitoring and automatic handling system for an unmanned engine room of a bulk carrier according to claim 2, characterized in that, The multi-agent game decision-making module includes: The intelligent agent management unit is used to encapsulate the core equipment in the cabin into independent decision-making intelligent agents. Each intelligent agent is configured with a state space and an action space. The state space includes the operating parameters and health status of the equipment, and the action space includes the self-healing operations that the equipment can perform. A local policy storage unit, connected to the agent management unit, pre-configures a local self-healing policy library for each agent, storing the mapping relationship between fault types and handling operations; when any agent detects a single point of failure, it directly matches the corresponding handling operation from the local policy storage unit and executes it. The global resource monitoring unit connects to each intelligent agent and is used to record the global resource occupancy status in real time, and to mark resource conflict events when multiple intelligent agents request to use the same restricted resource. The collaborative game solving unit is connected to the global resource monitoring unit and each intelligent agent involved in resource conflicts. It is used to solve the globally optimal solution through collaborative game among intelligent agents when resource conflicts occur, with the overall security of the system as the goal. The instruction generation and priority sorting unit, connected to the collaborative game solving unit, is used to prioritize the handling schemes according to the importance of equipment safety, generate a sequence of control instructions and send them to the self-healing execution layer. The feedback processing unit, connected to the self-healing execution layer and the local policy storage unit, is used to receive instruction execution results and update policy confidence and resource status.

4. The integrated monitoring and automatic handling system for an unmanned engine room of a bulk carrier according to claim 3, characterized in that, The resource conflicts marked by the global resource monitoring unit include at least one of power capacity conflicts, standby equipment preemption conflicts, and cooling medium flow allocation conflicts; the instruction generation and priority sorting unit prioritizes the equipment according to the importance of its navigation safety, identifies the timing constraints in the operation sequence, and generates a control instruction sequence that satisfies the timing constraints.

5. The integrated monitoring and automatic handling system for an unmanned engine room of a bulk carrier according to claim 4, characterized in that, The network outage self-sustaining control module also includes a strategy sandbox simulation unit, which includes: The virtual environment construction subunit is used to build a digital twin virtual environment that is synchronized with the real cabin environment; The fault scenario generation subunit is used to generate diverse fault scenarios; The simulation execution engine subunit is used to simulate the multi-agent game decision-making process in a virtual environment and generate candidate disposal strategies; The strategy evaluation and selection subunit is used to evaluate candidate disposal strategies and select the optimal strategy. The policy solidification interface subunit is used to write the optimal policy into the local policy storage unit.

6. The integrated monitoring and automatic handling system for an unmanned engine room of a bulk carrier according to claim 5, characterized in that, The network outage self-sustaining control module further includes a decision log compression and storage unit, which comprises: The log collection interface subunit is used to capture key decision-making process data in offline autonomous operation mode; The compressed storage subunit is used to compress and block-store log data; The communication status monitoring subunit is used to monitor the recovery status of the ship-to-shore communication link; The backhaul control subunit is used to backhaul the stored log data to the shore-based center after communication is restored.

7. The integrated monitoring and automatic handling system for an unmanned engine room of a bulk carrier according to claim 6, characterized in that, The self-healing execution layer includes a hardware-in-the-loop verification gateway, which includes: The real-time mirror synchronization subunit is connected to the data acquisition layer and is used to build a real-time digital twin model that keeps synchronized with the physical device state based on real-time operating data. The digital twin model includes the physical characteristics, dynamic response characteristics and topological connection relationships between the devices. The instruction receiving interface subunit is connected to the instruction generation and priority sorting unit and is used to receive control instruction sequences; The virtual execution engine subunit is connected to the instruction receiving interface subunit and the real-time mirror synchronization subunit, and is used to sequentially extract control instructions from the control instruction sequence, simulate execution on the real-time digital twin model, and predict the system response after instruction execution; The security boundary storage sub-unit is used to store security boundary conditions; The security determination subunit, connected to the virtual execution engine subunit and the security boundary storage subunit, is used to compare the predicted system response with the security boundary conditions to determine whether there is a risk exceeding the security boundary. The instruction interception and feedback subunit is connected to the security determination subunit and is used to intercept instructions when a risk is determined to exist, and send a re-decision trigger signal to the multi-agent game decision module. The instruction release subunit is connected to the security determination subunit and the physical execution mechanism, and is used to issue instructions to the physical execution mechanism for execution when it is determined that there is no risk.

8. The integrated monitoring and automatic handling system for an unmanned engine room of a bulk carrier according to claim 7, characterized in that, The data acquisition layer includes: Multiple monitoring subunits, including several sensors, are used to collect vibration parameters, thermal parameters, speed and torque parameters, electrical parameters, visual image parameters, and oil condition parameters of the engine room equipment; The multi-source data fusion subunit is connected to each monitoring subunit and is used to perform spatiotemporal alignment and correlation fusion of heterogeneous data collected by different monitoring subunits to generate a multi-dimensional state vector in a unified format, which serves as the input to the dynamic baseline generation module and the multi-agent game decision module. An adaptive sampling control subunit, connected to each monitoring subunit, is used to dynamically adjust the sampling frequency of the monitoring subunits according to the equipment operating status. The redundancy and self-diagnosis subunit is connected to each monitoring subunit and is used to perform health self-diagnosis of the sensors, and to enable redundant sensors or perform soft measurement reconstruction based on relevant parameters when a sensor failure is detected.

9. The integrated monitoring and automatic handling system for an unmanned engine room of a bulk carrier according to claim 8, characterized in that, The local policy storage unit includes: The strategy structured storage subunit is used to maintain an independent local self-healing strategy library for each agent, storing the mapping relationship between fault modes and handling operations; A mapping relationship construction subunit is connected to the strategy structured storage subunit and is used to construct a mapping relationship between fault modes and handling operations. The policy confidence management subunit is connected to the policy structured storage subunit and the feedback processing unit, and is used to maintain dynamically updated confidence parameters for each mapping relationship; The applicable condition evaluation subunit is connected to the strategy structured storage subunit and the agent management unit, and is used to filter candidate disposal strategies that meet the current system state when a fault occurs. The multi-strategy optimization subunit, connected to the applicable condition evaluation subunit and the strategy structured storage subunit, is used to select the optimal strategy from multiple candidate disposal strategies for the agent to execute.

10. An integrated monitoring and automatic handling system for an unmanned engine room of a bulk carrier according to claim 9, characterized in that, The optimal selection strategy includes: The candidate disposal strategies are comprehensively scored based on preset evaluation indicators, which include at least one of the following: system security after strategy execution, resource consumption, fault recovery time, and degree of protection for critical equipment. The comprehensive score is calculated using a weighted summation method, where: The system security is calculated based on the security boundary conditions in the hardware-in-the-loop verification gateway; The resource consumption is calculated based on the global resource occupancy recorded by the global resource monitoring unit; The fault recovery time is based on statistics of the instruction execution results returned by the feedback processing unit; The degree of protection for critical equipment is calculated based on the importance of the equipment's navigation safety in the instruction generation and priority sorting unit; The optimal strategy is selected based on the overall score.

Images