A ship engine room equipment fault diagnosis method based on a multi-agent system

Abstract

The application discloses a kind of ship engine room equipment fault diagnosis methods based on multi-agent system, it is related to multi-agent system technical field, comprising: acquisition operating state data, generate engine room multivariate state sequence;Construct equipment agent, system agent and global diagnosis agent, generate engine room equipment initial association graph;Improved MTGNN model is constructed, and equipment node quaternion spectrum domain hidden feature and energy modulation association graph are generated;Event-triggered graph state evolution processing is executed, and engine room equipment dynamic time sequence coupling graph is constructed;Distribution offset value is calculated based on JS divergence, and equipment distribution offset score is generated;Fault propagation consistency score is generated;Ship engine room equipment fault diagnosis result is generated.The application realizes ship engine room equipment primary fault positioning, abnormal propagation relationship identification and fault diagnosis result generation by introducing improved MTGNN model and JS divergence.

Description

Technical Field

[0001] This invention relates to the field of multi-agent system technology, and in particular to a method for diagnosing faults in ship engine room equipment based on multi-agent systems. Background Technology

[0002] The ship's engine room is the core area for power generation, energy conversion, and auxiliary equipment operation. It contains various key equipment such as main engine equipment, cooling equipment, lubrication equipment, fuel processing equipment, and electrical equipment. These devices form complex operational relationships through mechanical transmission, media transport, and power supply. With the continuous improvement of ship automation, numerous sensors and monitoring terminals have been deployed in the engine room to collect operating parameters such as speed, vibration, temperature, pressure, and current in real time. The monitoring system analyzes and manages the equipment's operating status. To prevent equipment failures from escalating and causing downtime or navigational safety risks, it is necessary to utilize monitoring data to promptly detect abnormal states and diagnose faults. Therefore, ship engine room equipment fault diagnosis technology based on operational data has become an important research direction in the field of intelligent ship operation and maintenance.

[0003] Existing fault diagnosis methods for ship engine room equipment typically rely on threshold alarm analysis, statistical feature analysis, machine learning classification models, or graph neural network models to identify equipment status. These methods generally begin by collecting equipment operating parameters and constructing status features, then use classification or correlation analysis models to determine if abnormal conditions exist. Some solutions further integrate the connections between equipment to establish an equipment correlation network, thereby improving fault identification accuracy. However, in actual engine room environments, there are continuous energy transfer and status influence relationships between different devices. Equipment anomalies often propagate gradually along the equipment coupling chain, leading to simultaneous anomalies in multiple devices. Existing methods mostly focus on single-device status identification or local correlation analysis, failing to adequately utilize the propagation process of abnormal states within the equipment coupling chain. This makes it difficult to accurately distinguish between the primary faulty equipment and associated abnormal equipment, thus affecting the accuracy of fault diagnosis results in complex engine room environments.

[0004] Therefore, how to provide a fault diagnosis method for ship engine room equipment based on a multi-agent system is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a fault diagnosis method for ship engine room equipment based on a multi-agent system. This invention utilizes multi-agent collaborative analysis technology, graph neural network modeling technology, temporal correlation analysis technology, and fault propagation diagnosis technology to perform unified modeling and collaborative diagnosis of the operating status of the power system, cooling system, lubrication system, fuel system, and electrical system within the ship's engine room. It details the implementation process of equipment operating status data acquisition, equipment correlation graph construction, equipment node status representation, equipment coupling relationship analysis, anomaly propagation chain identification, and fault diagnosis result generation. This method enables the differentiation and diagnosis of primary faulty equipment and associated anomaly equipment in the complex equipment network of the ship's engine room, and achieves fault propagation path tracing and fault level assessment. Compared to traditional fault diagnosis methods that rely on single-equipment status analysis or simple threshold judgment, this invention can fully utilize the coupling relationships between equipment and anomaly propagation characteristics, improving the accuracy of fault identification under complex operating conditions, enhancing the ability to locate primary faults, improving fault propagation analysis capabilities, and increasing the reliability of fault diagnosis results. This provides a more accurate decision-making basis for the operation, maintenance, and fault handling of ship engine room equipment.

[0006] A method for diagnosing ship engine room equipment based on a multi-agent system according to an embodiment of the present invention includes:

[0007] Collect operational status data of key equipment in the ship's engine room, perform preprocessing on the operational status data, and generate a multivariable state sequence of the engine room.

[0008] Construct equipment intelligent agents, system intelligent agents, and global diagnostic intelligent agents; establish equipment nodes and connection edges based on the physical coupling relationship of equipment; and generate an initial association graph of cabin equipment.

[0009] An improved MTGNN model is constructed based on the multivariate state sequence of the cabin and the initial association graph of the cabin equipment. Quaternion joint encoding, spectral domain graph convolution and edge weight modulation based on energy transfer deviation are performed on the state of the equipment nodes to generate the quaternion spectral domain latent features and energy modulation association graph of the equipment nodes.

[0010] Based on the quaternion spectral domain latent features of equipment nodes within adjacent time windows, event-triggered graph state evolution processing is performed, and a dynamic temporal coupling graph of cabin equipment is constructed by combining the energy modulation correlation graph.

[0011] Each device agent converts the quaternion spectral domain latent features of the device nodes within the current time window and the historical normal operation latent features into probability distributions, and calculates the distribution offset values ​​between the corresponding probability distributions based on JS divergence to generate a device distribution offset score;

[0012] Based on the dynamic temporal coupling diagram of cabin equipment and equipment distribution offset score, each system agent analyzes the propagation relationship of abnormal offset in the equipment coupling chain and generates a fault propagation consistency score.

[0013] The global diagnostic agent generates fault diagnosis results for ship engine room equipment based on equipment distribution offset scores, fault propagation consistency scores, dynamic temporal coupling diagrams of engine room equipment, and alarm logs.

[0014] Optionally, the operating status data specifically includes rotational speed data, load rate data, vibration amplitude data, vibration phase data, vibration frequency data, temperature data, pressure data, current data, alarm log data, and acquisition timestamp data.

[0015] Optionally, generating the cabin multivariate state sequence includes:

[0016] The data including rotational speed, load rate, vibration amplitude, vibration phase, vibration frequency, temperature, pressure, current, and alarm logs are time-aligned according to the timestamp data collected, and then arranged in chronological order according to the timestamp data collected.

[0017] Missing data is filled by interpolation, abnormal data is removed, and the data of each state is standardized. Then, the data is recombined according to the collection timestamp data to generate a multivariate state sequence of the cabin.

[0018] Optionally, generating the initial association diagram of the cabin equipment includes:

[0019] Read the equipment identifier, equipment type and equipment location of the key equipment in the ship's engine room, set each key piece of equipment in the ship's engine room as an equipment node, and bind the corresponding operating status data to the corresponding equipment node;

[0020] Equipment intelligent agents are constructed according to equipment nodes, system intelligent agents are constructed according to power system, cooling system, lubrication system, fuel system and electric system, and a global diagnostic intelligent agent is constructed based on the set of equipment intelligent agents corresponding to each system intelligent agent;

[0021] Read the physical coupling relationships of the devices, which include cooling medium transmission relationships, lubrication medium transmission relationships, fuel supply relationships, mechanical transmission relationships, power supply relationships, and spatial proximity relationships, and establish connection edges between the corresponding device nodes;

[0022] Establish node mapping relationships among equipment nodes, connection edges, equipment agents, system agents, and global diagnostic agents to generate an initial association graph of cabin equipment.

[0023] Optionally, the generation of the device node quaternion spectral domain latent features and energy modulation correlation graph includes:

[0024] An improved MTGNN model is constructed, which includes a quaternion joint encoding module, a hyperbolic spinor spectral domain propagation module, and a thermal-work-electrical-energy-flow constraint module.

[0025] The quaternion joint encoding module, based on the operating status data corresponding to each equipment node in the multivariable state sequence of the engine room, writes the vibration amplitude data into the real part of the quaternion and writes the vibration phase data, vibration frequency data and temperature and pressure fusion data into the imaginary part of the quaternion to generate the quaternion state features of the equipment node.

[0026] The hyperbolic screw spectral domain propagation module is based on the quaternion state features of equipment nodes and the initial association graph of cabin equipment. It performs Riemann hyperbolic-screw waveguide convolution processing to decompose the quaternion state features of equipment nodes into screw state features and map them to hyperbolic space. Waveguide convolution processing is performed along the geodesic path corresponding to the connecting edge to generate hyperbolic screw spectral domain features of equipment nodes.

[0027] A symmetric antisymmetric dual propagation structure is constructed. Symmetric antisymmetric decomposition and dual graph convolution processing are performed on the connection edge relationship to generate symmetric propagation features and directional propagation features. These features are then fused with the hyperbolic spiral spectral domain features of the device node to generate the quaternion spectral domain latent features of the device node.

[0028] The thermal-electric energy flow constraint module reads load rate data, temperature data, pressure data, current data, and physical coupling relationship of the equipment, calculates heat flow value, shaft work value, electric power value, enthalpy difference value, and entropy difference value, constructs thermal-electric joint energy flow constraints based on heat flow value, shaft work value, and electric power value, and performs enthalpy-entropy double conservation regularization processing based on enthalpy difference value and entropy difference value to generate connection edge energy flow constraint value;

[0029] Based on the energy flow constraint values ​​of the connection edges, the weights of the connection edges in the initial correlation diagram of the cabin equipment are modulated to generate an energy modulation correlation diagram.

[0030] The improved MTGNN model was trained by using the combination of device node quaternion spectral domain latent feature prediction error and energy flow constraint error as the optimization objective. The parameters of the quaternion joint encoding module, hyperbolic spinor spectral domain propagation module and thermal-electrical energy flow constraint module were continuously optimized. Training was stopped when the comprehensive loss difference corresponding to 5 consecutive training cycles was less than the convergence threshold.

[0031] Optionally, the construction of the dynamic timing coupling diagram of cabin equipment by combining the energy modulation correlation diagram includes:

[0032] Calculate the difference of device node quaternion spectral domain latent features between adjacent time windows for the same device node, and summarize the difference of device node quaternion spectral domain latent features for each device node to generate the change in device node quaternion spectral domain latent features.

[0033] Event-triggered graph state evolution processing is performed based on the changes in the latent features of the quaternion spectrum domain of the device node. When the changes in the latent features of the quaternion spectrum domain of the device node reach the state trigger threshold, the corresponding latent state of the device node is updated. When the changes in the latent features of the quaternion spectrum domain of the device node do not reach the state trigger threshold, the latent state of the device node in the previous time window is maintained, and an event-triggered latent state sequence is generated.

[0034] Read the event-triggered hidden state sequence and energy modulation association graph, write the event-triggered hidden state sequence into the corresponding device node, write the modulated connection edge weights in the energy modulation association graph into the corresponding connection edges, and establish time window identifiers according to the collected timestamp data to generate a dynamic timing coupling graph of the cabin equipment.

[0035] Optionally, the generation of device distribution offset score includes:

[0036] Each device agent reads the quaternion spectral domain latent features and historical normal operation latent features of the device node within the current time window according to the device node identifier and the collected timestamp data. It then performs interval division, frequency statistics and normalization on the two to generate the current basic probability distribution and the normal basic probability distribution.

[0037] Map the current basic probability distribution and the normal basic probability distribution to hyperbolic space, calculate the corresponding Riemann potential energy values ​​respectively, and perform potential energy weighting on the current basic probability distribution and the normal basic probability distribution to generate the current potential energy enhanced probability distribution and the normal potential energy enhanced probability distribution.

[0038] According to the device coupling chain, read the device node quaternion spectral domain latent features of adjacent device nodes, calculate the decrease between the coupling chain mutual information value and the historical normal coupling chain mutual information value in the current time window, and perform link association weighting processing on the current potential energy enhancement probability distribution and the normal potential energy enhancement probability distribution to generate the current link association enhancement probability distribution and the normal link association enhancement probability distribution.

[0039] Feature masking and reconstruction processing is performed on the quaternion spectral domain latent features of the device nodes. The error difference between the current reconstruction error and the historical normal reconstruction error is calculated, and reconstruction confidence correction processing is performed to generate the current enhancement probability distribution and the normal enhancement probability distribution.

[0040] Vibration mode sub-features, temperature mode sub-features, and pressure mode sub-features are extracted from the quaternion spectral domain latent features of equipment nodes. The Wasserstein distance of each mode sub-feature relative to the historical normal latent features is calculated. The maximum value is taken to generate the multimodal Wasserstein wavefront value. Dynamic weight modulation and averaging fusion processing are performed on the current enhancement probability distribution and the normal enhancement probability distribution to generate the enhancement average probability distribution.

[0041] Based on the JS divergence, the first KL divergence value of the current enhancement probability distribution relative to the enhancement average probability distribution and the second KL divergence value of the normal enhancement probability distribution relative to the enhancement average probability distribution are calculated respectively. The first KL divergence value and the second KL divergence value are averaged to generate the JS divergence value, and the JS divergence value is used as the device distribution offset score.

[0042] Optionally, generating a fault propagation consistency score includes:

[0043] Each system agent reads the dynamic temporal coupling diagram of the cabin equipment and the equipment distribution offset score, and extracts the corresponding equipment nodes, connection edges, modulated connection edge weights and collection timestamp data according to the power system, cooling system, lubrication system, fuel system and electric system to generate a system-level abnormal node set;

[0044] Based on the physical coupling relationship of the devices and the collected timestamp data, the device nodes in the system-level abnormal node set are sorted in the propagation order, and the propagation direction between adjacent abnormal device nodes is determined according to the weight of the modulated connection edge, thus generating a system-level abnormal propagation chain.

[0045] Read the device distribution offset scores of adjacent device nodes in the system-level anomaly propagation chain, calculate the offset change between upstream and downstream device nodes, and combine the modulated connection edge weights to calculate the propagation strength value, generating an anomaly propagation strength sequence.

[0046] Each system agent verifies the progressive relationship of the device distribution offset score in the device coupling chain based on the system-level anomaly propagation chain, propagation direction, propagation order, and anomaly propagation intensity sequence, and generates a fault propagation consistency score.

[0047] Optionally, generating fault diagnosis results for ship engine room equipment includes:

[0048] The global diagnostic agent reads the equipment distribution offset score, fault propagation consistency score, cabin equipment dynamic timing coupling diagram and alarm log, aligns it with the equipment node identifier and the collected timestamp data, and extracts abnormal equipment nodes, abnormal time, modulated connection edge weights, system-level abnormal propagation chains, propagation directions and alarm event records.

[0049] Based on the device distribution offset score, fault propagation consistency score, modulated connection edge weight and alarm event record corresponding to the abnormal device node, the primary fault score is calculated. The abnormal device node with the highest primary fault score is identified as the primary fault device. The abnormal device node located downstream of the system-level abnormal propagation chain and whose device distribution offset score reaches the abnormal judgment threshold is identified as the associated abnormal device.

[0050] Based on the primary faulty equipment, associated abnormal equipment, primary fault score, abnormal time, system-level abnormal propagation chain, and alarm event records, the fault propagation path, fault level, and maintenance priority are determined, and the fault diagnosis results of the ship's engine room equipment are generated.

[0051] The beneficial effects of this invention are:

[0052] Compared with existing technologies, the present invention proposes a fault diagnosis method for ship engine room equipment based on a multi-agent system. First, it constructs equipment agents, system agents, and global diagnostic agents based on the physical coupling relationship of the equipment. Then, it establishes a unified diagnostic framework by combining the correlation relationship of engine room equipment. This enables the operation status information of the power system, cooling system, lubrication system, fuel system, and electrical system to be analyzed under a unified correlation structure. This avoids the problem of one-sided diagnostic results caused by traditional methods that rely only on the status judgment of a single device, and improves the completeness of equipment status analysis in complex engine room environments.

[0053] This invention constructs an improved MTGNN model to perform quaternion joint encoding, spectral domain correlation propagation, and energy flow constraint modeling on equipment operating status data. This enables a unified representation of the correlation features between vibration amplitude data, vibration phase data, vibration frequency data, temperature data, pressure data, and current data. Furthermore, by combining the physical coupling relationships between equipment, a quaternion spectral domain latent feature and energy modulation correlation graph of equipment nodes are formed, thereby improving the expressive power of equipment status features and the ability to characterize equipment correlation relationships.

[0054] This invention further combines the quaternion spectral domain latent features of equipment nodes, the dynamic temporal coupling diagram of engine room equipment, and equipment distribution offset scoring to analyze the propagation process of abnormal states in the equipment coupling chain. It also identifies abnormal propagation relationships through fault propagation consistency scoring, enabling the differentiation and diagnosis of primary faulty equipment and associated abnormal equipment. Simultaneously, it combines system-level abnormal propagation chains and alarm event records to determine fault propagation paths, fault levels, and maintenance priorities, thereby improving the accuracy of primary fault location, enhancing fault propagation analysis capabilities, and increasing the reliability of fault diagnosis results. This provides an accurate basis for the handling and operation and maintenance of ship engine room equipment faults. Attached Figure Description

[0055] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0056] Figure 1 This is a flowchart of a fault diagnosis method for ship engine room equipment based on a multi-agent system proposed in this invention;

[0057] Figure 2This is a schematic diagram of the structure of an improved MTGNN model for a fault diagnosis method of ship engine room equipment based on a multi-agent system proposed in this invention.

[0058] Figure 3 This is a functional flowchart of the JS divergence of a fault diagnosis method for ship engine room equipment based on a multi-agent system proposed in this invention. Detailed Implementation

[0059] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0060] refer to Figure 1 , Figure 2 and Figure 3 A method for fault diagnosis of ship engine room equipment based on a multi-agent system, comprising:

[0061] Collect operational status data of key equipment in the ship's engine room, perform preprocessing on the operational status data, and generate a multivariable state sequence of the engine room.

[0062] Construct equipment intelligent agents, system intelligent agents, and global diagnostic intelligent agents; establish equipment nodes and connection edges based on the physical coupling relationship of equipment; and generate an initial association graph of cabin equipment.

[0063] An improved MTGNN model is constructed based on the multivariate state sequence of the cabin and the initial association graph of the cabin equipment. Quaternion joint encoding, spectral domain graph convolution and edge weight modulation based on energy transfer deviation are performed on the state of the equipment nodes to generate the quaternion spectral domain latent features and energy modulation association graph of the equipment nodes.

[0064] Based on the quaternion spectral domain latent features of equipment nodes within adjacent time windows, event-triggered graph state evolution processing is performed, and a dynamic temporal coupling graph of cabin equipment is constructed by combining the energy modulation correlation graph.

[0065] Each device agent converts the quaternion spectral domain latent features of the device nodes within the current time window and the historical normal operation latent features into probability distributions, and calculates the distribution offset values ​​between the corresponding probability distributions based on JS divergence to generate a device distribution offset score;

[0066] Based on the dynamic temporal coupling diagram of cabin equipment and equipment distribution offset score, each system agent analyzes the propagation relationship of abnormal offset in the equipment coupling chain and generates a fault propagation consistency score.

[0067] The global diagnostic agent generates fault diagnosis results for ship engine room equipment based on equipment distribution offset scores, fault propagation consistency scores, dynamic temporal coupling diagrams of engine room equipment, and alarm logs.

[0068] In this embodiment, the operating status data specifically includes rotational speed data, load rate data, vibration amplitude data, vibration phase data, vibration frequency data, temperature data, pressure data, current data, alarm log data, and acquisition timestamp data.

[0069] In this embodiment, generating the cabin multivariate state sequence includes:

[0070] The data, including rotational speed data, load rate data, vibration amplitude data, vibration phase data, vibration frequency data, temperature data, pressure data, current data, and alarm log data, are time-aligned according to the timestamp data collected and arranged in chronological order.

[0071] Missing data is filled by interpolation, abnormal data is removed, and the data of each state is standardized. Then, the data is recombined according to the collection timestamp data to generate a multivariate state sequence of the cabin.

[0072] In this embodiment, generating the initial association diagram of the cabin equipment includes:

[0073] Read the equipment identifier, equipment type and equipment location of key equipment in the ship's engine room, set each key piece of equipment in the ship's engine room as an equipment node, and bind the corresponding operating status data to the corresponding equipment node;

[0074] Equipment intelligent agents are constructed according to equipment nodes, system intelligent agents are constructed according to power system, cooling system, lubrication system, fuel system, and electric system, and a global diagnostic intelligent agent is constructed based on the set of equipment intelligent agents corresponding to each system intelligent agent, wherein:

[0075] The device intelligent agent is constructed according to the device node, specifically as follows:

[0076] Read the device identifier, device type and running status data corresponding to the device node, create a device agent for each device node, write the device node identifier into the device agent index field, write the running status data into the device agent data field, and establish a one-to-one correspondence between device nodes and device agents.

[0077] The system intelligent agent is constructed according to the power system, cooling system, lubrication system, fuel system, and electrical system, specifically as follows:

[0078] Read the device type corresponding to each device agent, classify the device agents belonging to the same system into the corresponding system agent, and write the classification result into the device agent set of the system agent.

[0079] A global diagnostic agent is constructed based on the set of device agents corresponding to each system agent, specifically as follows:

[0080] Read the sets of device agents corresponding to the power system agent, cooling system agent, lubrication system agent, fuel system agent and electric system agent, write each set of device agents into the global diagnostic agent, and establish a hierarchical mapping relationship between the global diagnostic agent and each system agent.

[0081] Read the physical coupling relationships of the devices, including cooling medium transmission relationships, lubrication medium transmission relationships, fuel supply relationships, mechanical transmission relationships, power supply relationships, and spatial proximity relationships, and establish connection edges between corresponding device nodes, wherein:

[0082] Cooling medium transmission relationship refers to the connection relationship between cooling equipment according to the direction of cooling medium flow; lubrication medium transmission relationship refers to the connection relationship between lubrication equipment according to the direction of lubrication medium flow; fuel supply relationship refers to the connection relationship between fuel equipment according to the direction of fuel delivery; mechanical transmission relationship refers to the connection relationship between power equipment according to the direction of power transmission; power supply relationship refers to the connection relationship between electrical equipment according to the direction of electrical energy transmission; spatial proximity relationship refers to the positional relationship between equipment locations according to spatial distance.

[0083] Establish node mapping relationships among device nodes, connection edges, device agents, system agents, and global diagnostic agents to generate an initial association graph of cabin equipment, where:

[0084] Generate the initial association diagram of the cabin equipment, specifically as follows:

[0085] Using equipment nodes as graph nodes and equipment physical coupling relationships as connecting edges, the cooling medium transmission relationship, lubrication medium transmission relationship, fuel supply relationship, mechanical transmission relationship, power supply relationship or spatial proximity relationship corresponding to each connecting edge are read. The corresponding relationship is written into the connecting edge attribute field, and the initial weight of each connecting edge is uniformly assigned to 1. The mapping relationship between equipment nodes and corresponding equipment agents, the affiliation relationship between equipment agents and system agents, and the hierarchical mapping relationship between system agents and global diagnostic agents are written into the graph structure to obtain the initial association graph of cabin equipment.

[0086] In this embodiment, the generation of the device node quaternion spectral domain latent feature and energy modulation correlation graph includes:

[0087] An improved MTGNN model is constructed, which includes a quaternion joint encoding module, a hyperbolic spinor spectral domain propagation module, and a thermal-work-electrical-flow constraint module, wherein:

[0088] The improved MTGNN model is constructed as follows:

[0089] A quaternion joint encoding module is added between the graph construction layer and the graph convolutional layer of the traditional MTGNN model. The real node feature encoding method of the traditional MTGNN model is improved to the quaternion joint encoding method, resulting in the quaternion joint encoding module. The graph convolutional layer of the traditional MTGNN model is improved to the hyperbolic screw spectrum domain propagation module. The Euclidean space graph propagation method is improved to the hyperbolic space screw spectrum domain propagation method, resulting in the hyperbolic screw spectrum domain propagation module. A thermal, mechanical, and electrical energy flow constraint module is added between the hyperbolic screw spectrum domain propagation module and the time convolutional layer. The device energy flow constraint information is written into the connection edge weights, resulting in the thermal, mechanical, and electrical energy flow constraint module, thus obtaining the improved MTGNN model.

[0090] The quaternion joint encoding module, based on the operating state data corresponding to each equipment node in the multivariate state sequence of the engine room, writes the vibration amplitude data into the real part of the quaternion and the vibration phase data, vibration frequency data, and temperature and pressure fusion data into the imaginary part of the quaternion, generating quaternion state features of the equipment nodes, where:

[0091] The quaternion joint encoding module includes:

[0092] Temperature and pressure fusion converter: Reads temperature and pressure data corresponding to the same device node and splices them according to the same acquisition timestamp to generate temperature and pressure fused data;

[0093] Quaternion distributor: writes vibration amplitude data into the real part of a quaternion, and writes vibration phase data, vibration frequency data, and temperature and pressure fusion data into the imaginary parts of three quaternions in sequence to generate quaternion state characteristics;

[0094] Node mapper: Writes quaternion state features to the corresponding device node to generate device node quaternion state features;

[0095] Generate the quaternion state features of the device nodes, specifically:

[0096] Read the vibration amplitude data, vibration phase data, vibration frequency data, temperature data and pressure data corresponding to the same device node, splice the temperature data and pressure data according to the same acquisition timestamp to obtain temperature and pressure fusion data, write the vibration amplitude data into the real part of the quaternion, and write the vibration phase data, vibration frequency data and temperature and pressure fusion data into the imaginary parts of the three quaternions in sequence to form the quaternion state characteristics of the device node.

[0097] The hyperbolic screw spectral domain propagation module, based on the quaternion state features of equipment nodes and the initial association graph of cabin equipment, performs Riemann hyperbolic-screw waveguide convolution processing. This process decomposes the quaternion state features of equipment nodes into screw state features and maps them to hyperbolic space. Waveguide convolution processing is then performed along the geodesic paths corresponding to the connecting edges to generate the hyperbolic screw spectral domain features of the equipment nodes, where:

[0098] The hyperbolic spinor spectral domain propagation module includes:

[0099] Screw splitter: Reads the real part and three imaginary parts from the quaternion state characteristics of the device node, and combines the real part with each imaginary part to generate the first screw vector, the second screw vector and the third screw vector;

[0100] Hyperbolic mapper: Performs normalization mapping on the first, second, and third screw vectors to generate hyperbolic screw features;

[0101] Geodetic waveguide: Establish a feature propagation path based on the connection edge relationship in the initial association diagram of the cabin equipment, and propagate hyperbolic spiral features along the connection edge;

[0102] Neighborhood aggregator: Performs aggregation processing on hyperbolic spinor features received from the same terminating device node to generate hyperbolic spinor spectral domain features of the device node;

[0103] Symmetric propagator: Performs bidirectional propagation processing based on the symmetric propagation structure to generate symmetric propagation features;

[0104] Antisymmetric propagator: Performs directed propagation processing based on the antisymmetric propagation structure to generate directional propagation features;

[0105] Feature Fuder: Performs fusion processing on hyperbolic spinor spectral domain features, symmetric propagation features, and directional propagation features of device nodes to generate quaternion spectral domain latent features of device nodes;

[0106] The Riemann hyperbolic-spinor waveguide convolution process is performed as follows:

[0107] Read the quaternion state features of the equipment nodes, and combine the real part of the quaternion with the first imaginary part, the second imaginary part, and the third imaginary part to form the first screw vector, the second screw vector, and the third screw vector, respectively, to generate screw state features; perform normalization mapping on the screw state features to generate hyperbolic screw features; read the connection edge relationship in the initial association graph of the cabin equipment, and take the feature propagation direction between the starting and ending equipment nodes of the connection edge as the geodesic path corresponding to the connection edge; perform waveguide convolution processing along the geodesic path corresponding to the connection edge.

[0108] Hyperbolic space refers to the negative curvature feature space used to represent the spinor state characteristics of device nodes;

[0109] The quaternion state features of the device nodes are decomposed into screw state features and mapped to hyperbolic space, specifically as follows:

[0110] Read the real part, first imaginary part, second imaginary part, and third imaginary part from the quaternion state feature. Combine the data at the corresponding positions of the real part and the first imaginary part to form the first screw vector, combine the data at the corresponding positions of the real part and the second imaginary part to form the second screw vector, and combine the data at the corresponding positions of the real part and the third imaginary part to form the third screw vector. Combine the first screw vector, the second screw vector, and the third screw vector to form the screw state feature. Read the element values ​​in each screw vector and divide each element value by the sum of the absolute values ​​of all element values ​​in the screw vector to obtain the normalized first screw vector, the normalized second screw vector, and the normalized third screw vector. Use the normalized first screw vector, the normalized second screw vector, and the normalized third screw vector as the hyperbolic screw feature of the corresponding device node.

[0111] The geodesic path corresponding to the connecting edge refers to the feature propagation path between the starting device node and the ending device node of the connecting edge in the initial association graph of the cabin equipment.

[0112] Waveguide convolution is performed along the geodesic path corresponding to the connecting edge, specifically as follows:

[0113] Read the hyperbolic spiral features corresponding to the starting device node and the ending device node of the connection edge. Transmit the hyperbolic spiral features of the starting device node to the ending device node along the propagation direction of the connection edge. Add the transmitted hyperbolic spiral features to the original hyperbolic spiral features of the ending device node according to the same element positions and divide by 2 to obtain the connection edge aggregation features. Average all the connection edge aggregation features received by the same ending device node according to the same element positions to generate the hyperbolic spiral spectral domain features of the device node.

[0114] Construct a symmetric-antisymmetric dual-propagation structure, specifically as follows:

[0115] Read the connection edges in the initial association diagram of the cabin equipment. For each connection edge, write the connection relationship from the starting device node to the ending device node and the connection relationship from the ending device node to the starting device node in the symmetric propagation structure. Write the connection relationship from the starting device node to the ending device node in the antisymmetric propagation structure to obtain the symmetric propagation structure and the antisymmetric propagation structure.

[0116] The connection edge relationships are subjected to symmetric antisymmetric decomposition and bigraph convolution processing, specifically as follows:

[0117] Read the bidirectional connection relationship in the symmetric propagation structure, transfer the hyperbolic spiral spectral domain features of the device nodes at both ends of the connection relationship to each other, and perform element-wise averaging on the hyperbolic spiral spectral domain features received by each device node to generate symmetric propagation features. Read the directed connection relationship in the antisymmetric propagation structure, transfer the hyperbolic spiral spectral domain features of the starting device node to the terminating device node, and perform element-wise addition on the hyperbolic spiral spectral domain features of the terminating device node itself and the received hyperbolic spiral spectral domain features to generate directional propagation features.

[0118] The generation of latent features in the spectral domain of quaternions for device nodes is as follows:

[0119] Read the hyperbolic spinor spectral domain features, symmetric propagation features, and directional propagation features of the same device node, splice the hyperbolic spinor spectral domain features, symmetric propagation features, and directional propagation features of the device node according to the same element position, and perform element-wise averaging on the spliced ​​features to generate the quaternion spectral domain latent features of the device node.

[0120] The thermal-electrical energy flow constraint module reads load rate data, temperature data, pressure data, current data, and the physical coupling relationship of the equipment. It calculates heat flux, shaft work, electrical power, enthalpy difference, and entropy difference. Based on the heat flux, shaft work, and electrical power values, it constructs a combined thermal-electrical energy flow constraint. Then, based on the enthalpy difference and entropy difference, it performs enthalpy-entropy double conservation regularization to generate connection edge energy flow constraint values, where:

[0121] The thermal-electrical energy flow constraint module includes:

[0122] Energy Flow Calculator: Reads load rate data, temperature data, pressure data, current data, and physical coupling relationships of equipment to calculate heat flow, shaft work, electrical power, enthalpy difference, and entropy difference.

[0123] Deviation estimator: Calculates the energy flow deviation value of the connection side based on the heat flow value, shaft work value and electric power value, and calculates the conservation deviation value based on the enthalpy difference value and entropy difference value;

[0124] Constraint Fusioner: Performs fusion processing on the energy flow deviation value and conservation deviation value of the connection edge to generate the energy flow constraint value of the connection edge;

[0125] Weight modulator: Modulates the initial weights of the connecting edges according to the energy flow constraint values ​​of the connecting edges, and generates modulated connecting edge weights;

[0126] Calculate the heat flux, shaft work, electrical power, enthalpy difference, and entropy difference values, specifically as follows:

[0127] Read the temperature and pressure data corresponding to the starting and ending device nodes of the connection edge. Subtract the temperature data of the starting device node from the temperature data of the ending device node to obtain the temperature difference value. Subtract the pressure data of the starting device node from the pressure data of the ending device node to obtain the pressure difference value. Multiply the temperature difference value and the pressure difference value to obtain the heat flux value. Read the load rate data corresponding to the device nodes at both ends of the connection edge. Average the load rate data of the device nodes at both ends of the connection edge to obtain the shaft power value. Read the current data corresponding to the device nodes at both ends of the connection edge. Average the current data of the device nodes at both ends of the connection edge to obtain the electric power value. Use the temperature difference value as the enthalpy difference value and the pressure difference value as the entropy difference value.

[0128] A combined heat-work-electricity energy flow constraint is constructed based on heat flux, shaft work, and electrical power values, specifically as follows:

[0129] Read the heat flow value, shaft work value, and electric power value corresponding to the same connection edge. Add the heat flow value, shaft work value, and electric power value together and divide by 3 to obtain the reference energy flow value of the connection edge. Calculate the absolute value of the difference between the heat flow value and the reference energy flow value of the connection edge, the absolute value of the difference between the shaft work value and the reference energy flow value of the connection edge, and the absolute value of the difference between the electric power value and the reference energy flow value of the connection edge. Add the three absolute values ​​of the difference to obtain the energy flow deviation value of the connection edge. Use the energy flow deviation value of the connection edge as the combined heat, work, and electric energy flow constraint.

[0130] The combined heat, work, and electricity energy flow constraint refers to the energy flow deviation value of the connection edge calculated based on the heat flow value, shaft work value, and electric power value of the connection edge. The energy flow deviation value of the connection edge is used to constrain the adjustment range of the corresponding connection edge weight. The larger the energy flow deviation value of the connection edge, the larger the adjustment range of the corresponding connection edge weight; the smaller the energy flow deviation value of the connection edge, the smaller the adjustment range of the corresponding connection edge weight.

[0131] The enthalpy-entropy double conservation canonical refers to the conservation deviation value calculated based on the enthalpy difference and entropy difference between the device nodes at both ends of the connection edge. The enthalpy difference is used to represent the degree of temperature change between the device nodes at both ends of the connection edge, and the entropy difference is used to represent the degree of pressure change between the device nodes at both ends of the connection edge.

[0132] The enthalpy-entropy double conservation regularization is performed based on the enthalpy difference and entropy difference, specifically as follows:

[0133] Read the enthalpy difference and entropy difference values ​​of the connecting edge, calculate the absolute values ​​of the enthalpy difference and entropy difference respectively, and add the absolute values ​​of the enthalpy difference and entropy difference to obtain the conservation deviation value. Read the energy flow deviation value of the connecting edge, add the conservation deviation value to the energy flow deviation value of the connecting edge to obtain the comprehensive energy flow deviation value of the connecting edge, and use the comprehensive energy flow deviation value of the connecting edge as the energy flow constraint value of the connecting edge.

[0134] Based on the energy flow constraint values ​​of the connection edges, the weights of the connection edges in the initial correlation graph of the naval equipment are modulated, specifically as follows:

[0135] Read the initial weights and corresponding energy flow constraints of the connecting edges in the initial association diagram of the cabin equipment. Divide the initial weights of the connecting edges by the sum of the energy flow constraints and 1 to obtain the modulated weights of the connecting edges. Write the modulated weights of the connecting edges into the corresponding connecting edges.

[0136] Generate an energy modulation correlation graph, specifically as follows:

[0137] The improved MTGNN model was trained using a combination of device node quaternion spectral domain latent feature prediction error and energy flow constraint error as the optimization objective. The parameters of the quaternion joint encoding module, hyperbolic spinor spectral domain propagation module, and thermal-electrical energy flow constraint module were continuously optimized. Training was stopped when the combined loss difference for five consecutive training cycles was less than the convergence threshold.

[0138] The improved MTGNN model is trained as follows:

[0139] The system reads the multivariate state sequence of the cabin and the initial association graph of the cabin equipment, inputs them into the improved MTGNN model, and outputs the quaternion spectral domain latent features of the equipment nodes and the energy flow constraint values ​​of the connection edges. It reads the quaternion spectral domain latent features of the equipment nodes corresponding to the current time window and the next time window, calculates the average squared difference between the corresponding elements, and obtains the prediction error of the quaternion spectral domain latent features of the equipment nodes. It reads the energy flow constraint value and the energy flow deviation value of the connection edges, calculates the average of the absolute values ​​of the corresponding differences, and obtains the energy flow constraint error. It then weights and sums the prediction error of the quaternion spectral domain latent features of the equipment nodes and the energy flow constraint error to obtain the comprehensive loss value. Based on the comprehensive loss value, it calculates the parameter gradient and updates the parameters in the quaternion joint encoding module, the hyperbolic spinor spectral domain propagation module, and the thermal-electrical energy flow constraint module. Training stops when the comprehensive loss difference corresponding to five consecutive training cycles is less than 0.001.

[0140] In this embodiment, the step of constructing a dynamic temporal coupling diagram of cabin equipment by combining the energy modulation correlation diagram includes:

[0141] Calculate the difference in latent features of the quaternion spectral domain for the same device node between adjacent time windows, and summarize the differences in latent features of the quaternion spectral domain for each device node to generate the change in latent features of the quaternion spectral domain for each device node, where:

[0142] The generation of latent feature variations in the quaternion spectral domain of device nodes is as follows:

[0143] Read the quaternion spectral domain latent features of the same device node in the current time window and the previous time window, calculate the absolute value of the difference between the corresponding elements, add the absolute values ​​of the differences between all elements to obtain the change amount of the device node quaternion spectral domain latent features corresponding to the device node, summarize the change amount of the device node quaternion spectral domain latent features corresponding to all device nodes according to the device node identifier, and generate a set of device node quaternion spectral domain latent feature changes.

[0144] Event-triggered graph state evolution processing is performed based on the changes in the quaternion spectral domain latent features of device nodes. When the changes in the quaternion spectral domain latent features of a device node reach the state trigger threshold, the corresponding device node latent state is updated. When the changes in the quaternion spectral domain latent features of a device node do not reach the state trigger threshold, the device node latent state of the previous time window is maintained, and an event-triggered latent state sequence is generated, where:

[0145] Event-triggered graph state evolution processing is performed based on the changes in latent features of quaternions in the spectral domain of device nodes. Specifically:

[0146] Read the set of hidden feature changes in the quaternion spectrum domain of the device node and the hidden state of the device node corresponding to the previous time window. Compare the hidden feature changes in the quaternion spectrum domain of the device node corresponding to each device node with the state trigger threshold. When the hidden feature changes in the quaternion spectrum domain of the device node are greater than or equal to the state trigger threshold, perform hidden state update processing on the corresponding device node. When the hidden feature changes in the quaternion spectrum domain of the device node are less than the state trigger threshold, retain the hidden state of the device node corresponding to the previous time window.

[0147] The state trigger threshold is specifically 1.2 times the average value of the latent feature change in the quaternion spectral domain of all corresponding device nodes.

[0148] Update the hidden state of the corresponding device node, specifically:

[0149] Read the quaternion spectral domain latent feature of the device node corresponding to the current time window and the device node latent state corresponding to the previous time window. Add the device node quaternion spectral domain latent feature and the device node latent state according to the position of the same element, divide by 2, and obtain the updated device node latent state. Write it to the corresponding device node.

[0150] The event triggers a hidden state sequence, specifically:

[0151] Read all updated hidden states of device nodes and the retained hidden states of device nodes, sort them according to the collected timestamp data and device node identifiers, and form an event-triggered hidden state sequence by sorting the results.

[0152] Read the event-triggered hidden state sequence and energy modulation association graph, write the event-triggered hidden state sequence into the corresponding device node, write the modulated connection edge weights in the energy modulation association graph into the corresponding connection edges, and establish time window identifiers according to the collected timestamp data to generate a dynamic timing coupling graph of the cabin equipment, wherein:

[0153] Generate a dynamic timing coupling diagram of the cabin equipment, specifically as follows:

[0154] Read the event-triggered hidden state sequence and energy modulation association graph, write the event-triggered hidden state sequence into the corresponding device node, write the modulated connection edge weight into the corresponding connection edge, establish a time window identifier for each time window according to the collected timestamp data, and associate the time window identifier with the corresponding device node and connection edge. Connect the device nodes, modulated connection edge weights and time window identifiers corresponding to all time windows in chronological order to generate a dynamic time-series coupling graph of cabin equipment.

[0155] In this embodiment, the generation of device distribution offset score includes:

[0156] Each device agent, based on the device node identifier and collected timestamp data, reads the quaternion spectral domain latent features and historical normal operation latent features of the device nodes within the current time window, performs interval partitioning, frequency statistics, and normalization on both, and generates the current basic probability distribution and the normal basic probability distribution, wherein:

[0157] Historical normal operation latent features refer to the set of device node quaternion spectral domain latent features corresponding to the device node during the historical fault-free operation period.

[0158] The two processes are divided into intervals, frequency statistics, and normalization, specifically as follows:

[0159] Read the quaternion spectral domain latent features of the device nodes within the current time window, determine the maximum and minimum values, divide the numerical range between the maximum and minimum values ​​into 10 intervals, count the number of features falling into each interval, obtain the current frequency statistics, divide the frequency of each interval by the sum of the frequencies of all intervals, and generate the current basic probability distribution.

[0160] Read the hidden features of normal operation in history, determine the maximum and minimum values, divide the range between the maximum and minimum values ​​into 10 intervals, count the number of features falling into each interval, obtain the normal frequency statistics, divide the frequency of each interval by the sum of the frequencies of all intervals, and generate the normal basic probability distribution.

[0161] Map the current fundamental probability distribution and the normal fundamental probability distribution to hyperbolic space, calculate the corresponding Riemann potential energy values ​​respectively, and perform potential energy weighting on the current fundamental probability distribution and the normal fundamental probability distribution to generate the current potential energy enhancement probability distribution and the normal potential energy enhancement probability distribution, where:

[0162] Mapping the current fundamental probability distribution and the normal fundamental probability distribution to hyperbolic space is as follows:

[0163] Read the probability values ​​of each interval in the current basic probability distribution, divide each interval probability value by the sum of all interval probability values ​​in the current basic probability distribution to obtain the current normalized probability value, write the current normalized probability value into the corresponding coordinate position in hyperbolic space, and generate the current hyperbolic space probability distribution.

[0164] Read the probability values ​​of each interval in the normal basic probability distribution, divide each interval probability value by the sum of the probability values ​​of all intervals in the normal basic probability distribution to obtain the normalized probability value, write the normalized probability value into the corresponding coordinate position in hyperbolic space to generate the normal hyperbolic space probability distribution.

[0165] Calculate the corresponding Riemann potential values ​​separately, as follows:

[0166] Read all coordinate values ​​in the current hyperbolic space probability distribution, calculate the average of all coordinate values, calculate the absolute value of the difference between each coordinate value and the average value, add up all the absolute values ​​of the difference, and generate the current Riemann potential energy value.

[0167] Read all coordinate values ​​in the normal hyperbolic space probability distribution, calculate the average of all coordinate values, calculate the absolute value of the difference between each coordinate value and the average value, add up all the absolute values ​​of the difference to generate the normal Riemann potential value;

[0168] The current fundamental probability distribution and the normal fundamental probability distribution are subjected to potential energy weighting, specifically as follows:

[0169] Read the current basic probability distribution and the current Riemann potential energy value, multiply the probability values ​​of each interval of the current basic probability distribution by the current Riemann potential energy value, add all the product results to get the current weighted total value, divide the product results of each interval by the current weighted total value to generate the current potential energy enhancement probability distribution;

[0170] Read the normal basic probability distribution and the normal Riemann potential value, multiply the probability value of each interval of the normal basic probability distribution by the normal Riemann potential value, add all the product results to get the normal weighted total value, divide the product results of each interval by the normal weighted total value to generate the normal potential enhanced probability distribution.

[0171] According to the device coupling chain, the device node quaternion spectral domain latent features of adjacent device nodes are read. The decrease between the coupling chain mutual information value and the historical normal coupling chain mutual information value within the current time window is calculated. Then, link association weighting processing is performed on the current potential energy enhancement probability distribution and the normal potential energy enhancement probability distribution to generate the current link association enhancement probability distribution and the normal link association enhancement probability distribution, where:

[0172] Equipment coupling chain refers to the propagation path of equipment nodes formed by sequentially connecting edges in the dynamic temporal coupling diagram of engine room equipment;

[0173] Calculate the decrease in the mutual information value of coupled chains within the current time window compared to the historical normal mutual information value of coupled chains, specifically as follows:

[0174] Read the quaternion spectral domain latent features of adjacent device nodes in the device coupling chain, perform interval statistics on the latent features of adjacent device nodes, count the joint frequency when the latent features of two device nodes fall into the same interval, and divide the joint frequency by the sum of all statistical frequencies to obtain the joint probability value. Count the independent frequency when the latent features of two device nodes fall into each interval, and divide the independent frequency by the sum of all statistical frequencies to obtain the first independent probability value and the second independent probability value. Divide the joint probability value by the product of the first independent probability value and the second independent probability value to obtain the probability ratio. Calculate the logarithm of the probability ratio, and multiply the logarithm of each interval by the joint probability value and sum them to generate the mutual information value of the coupling chain in the current time window.

[0175] Read the quaternion spectral domain latent features of the device nodes of the corresponding device coupling chain under the historical normal operation state, calculate the joint probability value, the first independent probability value and the second independent probability value in the same way, divide the joint probability value by the product of the first independent probability value and the second independent probability value to obtain the historical probability ratio, calculate the logarithm of the historical probability ratio, and multiply the logarithm of each interval with the joint probability value and sum them to generate the mutual information value of the historical normal coupling chain.

[0176] Subtract the current time window's mutual information value from the historical normal coupled chain mutual information value to generate the mutual information decrease amount;

[0177] The link correlation weighting process is performed on the current potential energy enhancement probability distribution and the normal potential energy enhancement probability distribution, specifically as follows:

[0178] Read the current potential energy enhancement probability distribution and mutual information decrease, multiply the probability value of each interval of the current potential energy enhancement probability distribution by the sum of mutual information decrease and 1, add all the product results to get the current weighted total value, divide the product results of each interval by the current weighted total value to generate the current link association enhancement probability distribution;

[0179] Read the normal potential energy enhancement probability distribution and mutual information decrease, multiply the probability value of each interval of the normal potential energy enhancement probability distribution by the sum of mutual information decrease and 1, add all the product results to get the normal weighted total value, divide the product results of each interval by the normal weighted total value to generate the normal link association enhancement probability distribution.

[0180] Feature masking and reconstruction are performed on the latent features of the quaternion spectral domain of the device nodes. The error difference between the current reconstruction error and the historical normal reconstruction error is calculated, and reconstruction confidence correction is performed to generate the current enhancement probability distribution and the normal enhancement probability distribution, where:

[0181] Feature masking and reconstruction are performed on the latent features of the quaternion spectral domain of the device nodes, specifically as follows:

[0182] Read the quaternion spectral domain latent features of the device node, select one element every 5 elements according to the feature arrangement order and replace it with 0 to generate occlusion latent features, read the values ​​of the adjacent elements before and after each occlusion position, write the average value of the adjacent elements before and after into the corresponding occlusion position to generate reconstructed latent features.

[0183] Calculate the error difference between the current reconstruction error and the historical normal reconstruction error, specifically as follows:

[0184] Read the quaternion spectral domain latent features and reconstruction latent features of the device node corresponding to the current time window, calculate the average value of the absolute values ​​of the differences between corresponding elements, and generate the current reconstruction error.

[0185] Read the historical normal operation latent features and the corresponding reconstructed latent features, calculate the average of the absolute values ​​of the differences between corresponding elements, and generate the historical normal reconstruction error;

[0186] Subtract the historical normal reconstruction error from the current reconstruction error to generate the error difference;

[0187] Perform the reconstruction confidence correction process, specifically as follows:

[0188] Read the current link association enhancement probability distribution and error difference, divide the probability value of each interval of the current link association enhancement probability distribution by the sum of the error difference and 1, and re-normalize all the interval probability values ​​after correction to generate the current enhancement probability distribution;

[0189] Read the normal link association enhancement probability distribution and error difference, divide the probability value of each interval of the normal link association enhancement probability distribution by the sum of the error difference and 1, and re-normalize all the interval probability values ​​after correction to generate the normal enhancement probability distribution.

[0190] Vibration mode sub-features, temperature mode sub-features, and pressure mode sub-features are extracted from the quaternion spectral domain latent features of equipment nodes. The Wasserstein distance of each mode sub-feature relative to the historical normal latent features is calculated, and the maximum value is used to generate a multimodal Wasserstein wavefront value. Dynamic weight modulation and averaging fusion processing are then performed on the current enhancement probability distribution and the normal enhancement probability distribution to generate an enhanced average probability distribution, where:

[0191] The vibration mode sub-features, temperature mode sub-features, and pressure mode sub-features are extracted from the quaternion spectral domain latent features of the equipment nodes, specifically as follows:

[0192] Read the feature data formed by mapping vibration amplitude data, vibration phase data and vibration frequency data in the quaternion spectrum domain latent features of the equipment node, generate vibration mode sub-features; read the feature data formed by mapping temperature data in the quaternion spectrum domain latent features of the equipment node, generate temperature mode sub-features; read the feature data formed by mapping pressure data in the quaternion spectrum domain latent features of the equipment node, generate pressure mode sub-features.

[0193] Historical normal latent features refer to the set of quaternion spectral domain latent features of equipment nodes corresponding to equipment nodes in historical fault-free operation states.

[0194] Calculate the Wasserstein distance of each modality sub-feature relative to the historical normal latent features, specifically as follows:

[0195] Read all feature values ​​from the vibration mode sub-features, sort them in ascending order of value. Read all feature values ​​from the historical normal latent features corresponding to the vibration mode sub-features, sort them in ascending order of value. Calculate the absolute value of the difference between the feature values ​​at corresponding positions after sorting, and add all the absolute values ​​of the differences to generate the vibration mode Wasserstein distance. Read all feature values ​​from the temperature mode sub-features, sort them in ascending order of value. Read all feature values ​​from the historical normal latent features corresponding to the temperature mode sub-features, sort them in ascending order of value, and calculate the absolute value of the difference between the feature values ​​at corresponding positions after sorting. The absolute values ​​of the differences are calculated and summed to generate the temperature mode Wasserstein distance. All feature values ​​in the pressure mode sub-features are read and sorted in ascending order of value. All feature values ​​in the historical normal latent features corresponding to the pressure mode sub-features are read and sorted in ascending order of value. The absolute values ​​of the differences between the feature values ​​at the corresponding positions after sorting are calculated and summed to generate the pressure mode Wasserstein distance. The vibration mode Wasserstein distance, temperature mode Wasserstein distance and pressure mode Wasserstein distance are compared and the maximum value is taken to generate the multimodal Wasserstein wavefront value.

[0196] Dynamic weight modulation and averaging fusion are performed on the current enhancement probability distribution and the normal enhancement probability distribution, specifically as follows:

[0197] Read the current enhancement probability distribution and the multimodal Wasserstein wavefront value. Multiply the probability value of each interval of the current enhancement probability distribution by the multimodal Wasserstein wavefront value to obtain the current modulation probability value. Add all the current modulation probability values ​​to obtain the current total modulation value. Divide each current modulation probability value by the current total modulation value to generate the current weighted modulation probability distribution. Read the normal enhancement probability distribution and the multimodal Wasserstein wavefront value. Multiply the probability value of each interval of the normal enhancement probability distribution by the multimodal Wasserstein wavefront value to obtain the normal modulation probability value. Add all the normal modulation probability values ​​to obtain the normal modulation total value. Divide each normal modulation probability value by the normal modulation total value to generate the normal weighted modulation probability distribution. Read the current weighted modulation probability distribution and the normal weighted modulation probability distribution. Add the corresponding interval probability values ​​and divide by 2 to obtain the fusion probability value. Combine all the fusion probability values ​​to form the enhancement average probability distribution.

[0198] Based on the JS divergence, the first KL divergence value of the current enhancement probability distribution relative to the enhancement average probability distribution and the second KL divergence value of the normal enhancement probability distribution relative to the enhancement average probability distribution are calculated respectively. The first KL divergence value and the second KL divergence value are averaged to generate the JS divergence value, which is then used as the device distribution offset score.

[0199] Based on JS divergence, the first KL divergence value of the current enhancement probability distribution relative to the enhancement average probability distribution, and the second KL divergence value of the normal enhancement probability distribution relative to the enhancement average probability distribution are calculated respectively, as follows:

[0200] Read the current augmentation probability distribution and the augmentation average probability distribution, calculate the ratio between the probability values ​​of each interval of the current augmentation probability distribution and the corresponding interval probability values ​​of the augmentation average probability distribution, calculate the logarithm of each ratio, multiply the corresponding interval probability values ​​of the current augmentation probability distribution by the corresponding logarithm and sum them to generate the first KL divergence value; read the normal augmentation probability distribution and the augmentation average probability distribution, calculate the ratio between the probability values ​​of each interval of the normal augmentation probability distribution and the corresponding interval probability values ​​of the augmentation average probability distribution, calculate the logarithm of each ratio, multiply the corresponding interval probability values ​​of the normal augmentation probability distribution by the corresponding logarithm and sum them to generate the second KL divergence value;

[0201] The first KL divergence value and the second KL divergence value are averaged as follows:

[0202] Read the first KL divergence value and the second KL divergence value, add the first KL divergence value and the second KL divergence value and divide by 2 to generate the JS divergence value, and use the JS divergence value as the device distribution offset score of the corresponding device node.

[0203] In this embodiment, generating a fault propagation consistency score includes:

[0204] Each system agent reads the dynamic temporal coupling diagram of the cabin equipment and the equipment distribution offset score. Based on the power system, cooling system, lubrication system, fuel system, and electrical system, it extracts the corresponding equipment nodes, connection edges, modulated connection edge weights, and data collection timestamps to generate a system-level set of abnormal nodes, including:

[0205] Generate a system-level set of exception nodes, specifically:

[0206] Each system agent reads the device nodes, connection edges, modulated connection edge weights, collection timestamp data, and device distribution offset scores belonging to its own system. Device nodes with device distribution offset scores greater than 0.6 are marked as abnormal device nodes. All abnormal device nodes are summarized according to device node identifiers and collection timestamp data to generate a system-level abnormal node set.

[0207] Based on the physical coupling relationships of the devices and the collected timestamp data, the device nodes in the system-level abnormal node set are sorted according to their propagation order. The propagation direction between adjacent abnormal device nodes is determined based on the modulated connection edge weights, generating a system-level abnormal propagation chain, where:

[0208] Generate a system-level exception propagation chain, specifically as follows:

[0209] Read the abnormal device nodes in the system-level abnormal node set, sort them according to the collected timestamp data from smallest to largest to obtain the abnormal device node sorting sequence, read the device physical coupling relationship and the weight of the modulated connection edge between the abnormal device nodes, starting from the first abnormal device node in the abnormal device node sorting sequence, read two adjacent abnormal device nodes in sequence, when there is a device physical coupling relationship between two adjacent abnormal device nodes and the corresponding modulated connection edge weight is greater than 0.3, establish a propagation edge between the two abnormal device nodes, connect all propagation edges according to the abnormal device node sorting sequence to generate a system-level abnormal propagation chain;

[0210] Read the device distribution offset scores of adjacent device nodes in the system-level anomaly propagation chain, calculate the offset change between upstream and downstream device nodes, and combine this with the modulated connection edge weights to calculate the propagation strength value, generating an anomaly propagation strength sequence, where:

[0211] Calculate the offset change between upstream and downstream device nodes, specifically:

[0212] Read the device distribution offset scores of the two abnormal device nodes corresponding to each propagation edge in the system-level anomaly propagation chain, and subtract the device distribution offset score of the starting point of the propagation edge from the device distribution offset score of the end point of the propagation edge to generate the offset change.

[0213] The propagation strength value is calculated by combining the modulated connection edge weights, specifically as follows:

[0214] Read the offset change and the modulated connection weight corresponding to the propagation edge, multiply the absolute value of the offset change by the modulated connection weight to generate the propagation strength value, arrange all the propagation strength values ​​according to the propagation edge order in the system-level anomaly propagation chain to generate an anomaly propagation strength sequence.

[0215] Each system agent verifies the progressive relationship of the device distribution offset score in the device coupling chain based on the system-level anomaly propagation chain, propagation direction, propagation order, and anomaly propagation intensity sequence, and generates a fault propagation consistency score, where:

[0216] The progressive relationship of the verification device distribution offset score in the device coupling chain is as follows:

[0217] Read the device distribution offset scores of the two abnormal device nodes corresponding to each propagation edge in the system-level anomaly propagation chain. When the device distribution offset score corresponding to the end point of the propagation edge is greater than or equal to the device distribution offset score corresponding to the start point of the propagation edge, mark the corresponding propagation edge as a progressively consistent propagation edge. When the device distribution offset score corresponding to the end point of the propagation edge is less than the device distribution offset score corresponding to the start point of the propagation edge, mark the corresponding propagation edge as a progressively inconsistent propagation edge.

[0218] Generate a fault propagation consistency score, specifically as follows:

[0219] Read the progressive consistency propagation edge count, the total number of propagation edges, and the abnormal propagation intensity sequence. Divide the progressive consistency propagation edge count by the total number of propagation edges to generate the progressive consistency ratio. Average all propagation intensity values ​​in the abnormal propagation intensity sequence to generate the average propagation intensity. Multiply the progressive consistency ratio by the average propagation intensity to generate the fault propagation consistency score.

[0220] In this embodiment, generating fault diagnosis results for ship engine room equipment includes:

[0221] The global diagnostic agent reads the device distribution offset score, fault propagation consistency score, dynamic timing coupling diagram of cabin equipment, and alarm logs. Aligning the data by device node identifier and acquisition timestamp, it extracts abnormal device nodes, abnormal times, modulated connection edge weights, system-level abnormal propagation chains, propagation directions, and alarm event records. Among these:

[0222] Alarm logs refer to alarm event data recorded by the ship's engine room monitoring system during equipment operation, specifically including alarm device identification, alarm occurrence time, alarm type, alarm level, and alarm duration;

[0223] Based on the device distribution offset score, fault propagation consistency score, modulated connection edge weights, and alarm event records corresponding to the abnormal device nodes, the primary fault score is calculated. The abnormal device node with the highest primary fault score is identified as the primary fault device. Abnormal device nodes located downstream in the propagation direction of the system-level abnormality propagation chain and whose device distribution offset scores reach the abnormality judgment threshold are identified as associated abnormal devices.

[0224] The primary fault score is calculated as follows:

[0225] Read the device distribution offset score, fault propagation consistency score, modulated connection edge weight, and alarm event record corresponding to the abnormal device node. Convert the alarm level in the alarm event record into an alarm score, where a level 1 alarm is recorded as 1, a level 2 alarm as 0.7, and a level 3 alarm as 0.4. Add the device distribution offset score, fault propagation consistency score, modulated connection edge weight, and alarm score together and divide by 4 to obtain the original fault score.

[0226] The anomaly detection threshold is specifically 0.6;

[0227] Abnormal device nodes located downstream in the propagation direction of the system-level anomaly propagation chain and whose device distribution offset scores reach the anomaly judgment threshold are identified as associated abnormal devices. Specifically:

[0228] Read the propagation edge order in the system-level anomaly propagation chain, take the abnormal device node after the propagation edge where the original fault device is located as the candidate associated abnormal device, read the device distribution offset score corresponding to the candidate associated abnormal device, and determine the candidate associated abnormal device with a device distribution offset score greater than or equal to 0.6 as the associated abnormal device.

[0229] Based on the primary faulty equipment, associated abnormal equipment, primary fault score, abnormal time, system-level abnormal propagation chain, and alarm event records, the fault propagation path, fault level, and maintenance priority are determined, generating fault diagnosis results for the ship's engine room equipment.

[0230] Determine the fault propagation path, fault level, and maintenance priority, specifically as follows:

[0231] Read the system-level anomaly propagation chain, connect the propagation edges between the primary faulty device and related faulty devices in sequence to obtain the fault propagation path, read the primary fault score, and determine the fault as a Level 1 fault if the primary fault score is greater than or equal to 0.8, a Level 2 fault if the primary fault score is greater than or equal to 0.6 and less than 0.8, and a Level 3 fault if the primary fault score is less than 0.6; determine the maintenance priority according to the order of Level 1, Level 2, and Level 3 faults.

[0232] The following are the results of the fault diagnosis of the ship's engine room equipment:

[0233] Write the original faulty equipment, associated abnormal equipment, abnormal time, fault propagation path, fault level, maintenance priority, and alarm event records into the diagnostic result field to generate the ship's engine room equipment fault diagnosis result.

[0234] Example 1: To verify the feasibility of this invention in practice, it was applied to a ship engine room equipment fault diagnosis scenario. The engine room is equipped with a main engine, shafting, generator, cooling pump, lubricating oil pump, fuel pump, heat exchanger, and air compressor. Each device continuously collects speed data, load rate data, vibration amplitude data, vibration phase data, vibration frequency data, temperature data, pressure data, current data, alarm log data, and timestamp data. The acquisition cycle is 1 second, with each 60-second window forming a time window. 120 time windows are continuously collected, with the first 80 windows used as historical normal operation samples and the last 40 windows used as samples to be diagnosed.

[0235] During normal historical operation, the cooling pump outlet pressure remained stable between 0.42 MPa and 0.46 MPa, the current remained stable between 28.4 A and 31.2 A, the vibration amplitude remained stable between 1.7 mm / s and 2.3 mm / s, the heat exchanger outlet temperature remained stable between 57.1°C and 60.8°C, and the main unit cooling channel temperature remained stable between 81.2°C and 84.5°C.

[0236] Within the 84th time window, the cooling pump outlet pressure dropped to 0.34 MPa, the current increased to 35.8 Amperes, and the vibration amplitude increased to 3.6 mm / s. Within the 87th time window, the heat exchanger outlet temperature rose to 70.4 degrees Celsius. Within the 90th time window, the main unit cooling channel temperature rose to 95.1 degrees Celsius, and the main unit vibration amplitude increased to 4.4 mm / s. At this point, the traditional threshold alarm system simultaneously issued abnormal alarms for the cooling pump, heat exchanger, and main unit, making it impossible to determine the faulty device.

[0237] After inputting the above data into the method of this invention, the system first establishes an initial association diagram of the nacelle equipment based on the physical coupling relationships of the equipment. This diagram establishes a cooling medium transfer connection between the cooling pump and the heat exchanger, a cooling medium transfer connection between the heat exchanger and the main engine, a mechanical transmission connection between the main engine and the shaft system, and a power supply connection between the generator and the cooling pump. Subsequently, it constructs equipment intelligent agents, system intelligent agents, and a global diagnostic intelligent agent, and writes the operating status data to the corresponding equipment nodes.

[0238] After improving the MTGNN model to characterize the equipment status, the change in the quaternion spectral domain latent feature of the equipment node corresponding to the cooling pump reached 0.71 in the 84th time window, 0.58 for the heat exchanger in the 87th time window, and 0.49 for the main unit in the 90th time window, while the changes in the lubricating oil pump, fuel pump, and generator were all below 0.15. The results indicate that the cooling chain-related equipment has shown significant abnormalities, while other equipment remains in normal condition.

[0239] The system further constructs connection edge energy flow constraints based on heat flux, shaft work, and electrical power values. The connection edge energy flow constraint value between the cooling pump and the heat exchanger reaches 39.80, corresponding to a modulated connection edge weight of 0.0245; the connection edge energy flow constraint value between the heat exchanger and the main unit reaches 32.60, corresponding to a modulated connection edge weight of 0.0298. Compared with the connection edge weights above 0.08 under historical normal conditions, the relevant connection edge weights have decreased significantly, indicating that the consistency of energy transfer in the cooling chain has been disrupted.

[0240] The device agent continues to calculate the device distribution offset score. In the 84th time window, the cooling pump device distribution offset score is 0.62. As the anomaly continues to develop, in the 90th time window, the cooling pump device distribution offset score increases to 0.91, the heat exchanger device distribution offset score increases to 0.73, and the main unit device distribution offset score increases to 0.64, while the scores for the lubricating oil pump, fuel pump, generator, air compressor, and valve assembly are all below 0.25. Based on the anomaly determination threshold of 0.60, the system identifies the cooling pump, heat exchanger, and main unit as anomalous device nodes.

[0241] After sorting the collected timestamp data, the cooling system agent found that the cooling pump was the first to malfunction, followed by the heat exchanger, and finally the main unit. Based on the physical coupling relationships between the cooling pump and the heat exchanger, and between the heat exchanger and the main unit, the system established an anomaly propagation chain. The propagation intensity value for the propagation edge from the cooling pump to the heat exchanger was 0.00441, and the propagation intensity value for the propagation edge from the heat exchanger to the main unit was 0.00268. The direction of anomaly propagation is consistent with the actual operation of the equipment.

[0242] The global diagnostic agent reads device distribution offset scores, fault propagation consistency scores, and alarm log data. The alarm logs show that the cooling pump first generated pressure and current anomaly alarms; the heat exchanger subsequently generated a temperature anomaly alarm, and the main unit last generated a temperature anomaly alarm. The system calculated the primary fault score for the cooling pump to be 0.4095, the primary fault score for the heat exchanger to be 0.2908, and the primary fault score for the main unit to be 0.2683. Since the cooling pump has the highest primary fault score, it is identified as the primary faulty device, and the heat exchanger and main unit are identified as associated faulty devices. The fault propagation path is output as cooling pump → heat exchanger → main unit.

[0243] To verify the diagnostic effectiveness of this invention, 7200 training samples and 1080 test samples were constructed, including scenarios such as cooling system failures, lubrication system failures, fuel system failures, electrical system failures, and mechanical transmission failures. Test results show that the traditional single-device threshold alarm method achieved a primary fault location accuracy of 72.31%, a related abnormal device identification accuracy of 61.48%, and a fault propagation path identification accuracy of 48.06%, respectively. Ordinary machine learning classification methods achieved 78.89%, 66.57%, and 55.74%, respectively; the unmodified MTGNN model achieved 84.72%, 73.33%, and 68.52%, respectively; and the method of this invention achieved 93.61%, 88.24%, and 86.39%, respectively.

[0244] The results show that the present invention can identify the primary faulty device, associated faulty devices, and fault propagation paths when multiple devices malfunction simultaneously, effectively solving the problem that traditional methods have difficulty distinguishing between the fault source device and the propagating faulty device, and improving the accuracy of fault diagnosis in the complex coupled environment of the ship's engine room.

[0245] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for fault diagnosis of ship engine room equipment based on a multi-agent system, characterized in that, include: Collect operational status data of key equipment in the ship's engine room, perform preprocessing on the operational status data, and generate a multivariable state sequence of the engine room. Construct equipment intelligent agents, system intelligent agents, and global diagnostic intelligent agents; establish equipment nodes and connection edges based on the physical coupling relationship of equipment; and generate an initial association graph of cabin equipment. An improved MTGNN model is constructed based on the multivariate state sequence of the cabin and the initial association graph of the cabin equipment. Quaternion joint encoding, spectral domain graph convolution and edge weight modulation based on energy transfer deviation are performed on the state of the equipment nodes to generate the quaternion spectral domain latent features and energy modulation association graph of the equipment nodes. Based on the quaternion spectral domain latent features of equipment nodes within adjacent time windows, event-triggered graph state evolution processing is performed, and a dynamic temporal coupling graph of cabin equipment is constructed by combining the energy modulation correlation graph. Each device agent converts the quaternion spectral domain latent features of the device nodes within the current time window and the historical normal operation latent features into probability distributions, and calculates the distribution offset values ​​between the corresponding probability distributions based on JS divergence to generate a device distribution offset score; Based on the dynamic temporal coupling diagram of cabin equipment and equipment distribution offset score, each system agent analyzes the propagation relationship of abnormal offset in the equipment coupling chain and generates a fault propagation consistency score. The global diagnostic agent generates fault diagnosis results for ship engine room equipment based on equipment distribution offset scores, fault propagation consistency scores, dynamic temporal coupling diagrams of engine room equipment, and alarm logs.

2. The method for fault diagnosis of ship engine room equipment based on a multi-agent system according to claim 1, characterized in that, The operating status data specifically includes speed data, load rate data, vibration amplitude data, vibration phase data, vibration frequency data, temperature data, pressure data, current data, alarm log data, and acquisition timestamp data.

3. The method for fault diagnosis of ship engine room equipment based on a multi-agent system according to claim 1, characterized in that, The generation of the cabin multivariate state sequence includes: The data, including rotational speed data, load rate data, vibration amplitude data, vibration phase data, vibration frequency data, temperature data, pressure data, current data, and alarm log data, are time-aligned according to the timestamp data collected and arranged in chronological order. Missing data is filled by interpolation, abnormal data is removed, and the data of each state is standardized. Then, the data is recombined according to the collection timestamp data to generate a multivariate state sequence of the cabin.

4. The method for fault diagnosis of ship engine room equipment based on a multi-agent system according to claim 1, characterized in that, The generation of the initial association diagram of cabin equipment includes: Read the equipment identifier, equipment type and equipment location of key equipment in the ship's engine room, set each key piece of equipment in the ship's engine room as an equipment node, and bind the corresponding operating status data to the corresponding equipment node; Equipment intelligent agents are constructed according to equipment nodes, system intelligent agents are constructed according to power system, cooling system, lubrication system, fuel system and electric system, and a global diagnostic intelligent agent is constructed based on the set of equipment intelligent agents corresponding to each system intelligent agent; Read the physical coupling relationships of the devices, which include cooling medium transmission relationships, lubrication medium transmission relationships, fuel supply relationships, mechanical transmission relationships, power supply relationships, and spatial proximity relationships, and establish connection edges between the corresponding device nodes; Establish node mapping relationships among equipment nodes, connection edges, equipment agents, system agents, and global diagnostic agents to generate an initial association graph of cabin equipment.

5. The method for fault diagnosis of ship engine room equipment based on a multi-agent system according to claim 1, characterized in that, The generated device node quaternion spectral domain latent features and energy modulation correlation graph include: An improved MTGNN model is constructed, which includes a quaternion joint encoding module, a hyperbolic spinor spectral domain propagation module, and a thermal-work-electrical-energy-flow constraint module. The quaternion joint encoding module, based on the operating status data corresponding to each equipment node in the multivariable state sequence of the engine room, writes the vibration amplitude data into the real part of the quaternion and writes the vibration phase data, vibration frequency data and temperature and pressure fusion data into the imaginary part of the quaternion to generate the quaternion state features of the equipment node. The hyperbolic screw spectral domain propagation module is based on the quaternion state features of equipment nodes and the initial association graph of cabin equipment. It performs Riemann hyperbolic-screw waveguide convolution processing to decompose the quaternion state features of equipment nodes into screw state features and map them to hyperbolic space. Waveguide convolution processing is performed along the geodesic path corresponding to the connecting edge to generate hyperbolic screw spectral domain features of equipment nodes. A symmetric antisymmetric dual propagation structure is constructed. Symmetric antisymmetric decomposition and dual graph convolution processing are performed on the connection edge relationship to generate symmetric propagation features and directional propagation features. These features are then fused with the hyperbolic spiral spectral domain features of the device node to generate the quaternion spectral domain latent features of the device node. The thermal-electric energy flow constraint module reads load rate data, temperature data, pressure data, current data, and physical coupling relationship of the equipment, calculates heat flow value, shaft work value, electric power value, enthalpy difference value, and entropy difference value, constructs thermal-electric joint energy flow constraints based on heat flow value, shaft work value, and electric power value, and performs enthalpy-entropy double conservation regularization processing based on enthalpy difference value and entropy difference value to generate connection edge energy flow constraint value; Based on the energy flow constraint values ​​of the connection edges, the weights of the connection edges in the initial correlation diagram of the cabin equipment are modulated to generate an energy modulation correlation diagram. The improved MTGNN model was trained by using the combination of device node quaternion spectral domain latent feature prediction error and energy flow constraint error as the optimization objective. The parameters of the quaternion joint encoding module, hyperbolic spinor spectral domain propagation module and thermal-electrical energy flow constraint module were continuously optimized. Training was stopped when the comprehensive loss difference corresponding to 5 consecutive training cycles was less than the convergence threshold.

6. The method for fault diagnosis of ship engine room equipment based on a multi-agent system according to claim 1, characterized in that, The construction of the dynamic timing coupling diagram of cabin equipment by combining the energy modulation correlation diagram includes: Calculate the difference of device node quaternion spectral domain latent features between adjacent time windows for the same device node, and summarize the difference of device node quaternion spectral domain latent features for each device node to generate the change in device node quaternion spectral domain latent features. Event-triggered graph state evolution processing is performed based on the changes in the latent features of the quaternion spectrum domain of the device node. When the changes in the latent features of the quaternion spectrum domain of the device node reach the state trigger threshold, the corresponding latent state of the device node is updated. When the changes in the latent features of the quaternion spectrum domain of the device node do not reach the state trigger threshold, the latent state of the device node in the previous time window is maintained, and an event-triggered latent state sequence is generated. Read the event-triggered hidden state sequence and energy modulation association graph, write the event-triggered hidden state sequence into the corresponding device node, write the modulated connection edge weights in the energy modulation association graph into the corresponding connection edges, and establish time window identifiers according to the collected timestamp data to generate a dynamic timing coupling graph of the cabin equipment.

7. The method for fault diagnosis of ship engine room equipment based on a multi-agent system according to claim 1, characterized in that, The generated device distribution offset score includes: Each device agent reads the quaternion spectral domain latent features and historical normal operation latent features of the device node within the current time window according to the device node identifier and the collected timestamp data. It then performs interval division, frequency statistics and normalization on the two to generate the current basic probability distribution and the normal basic probability distribution. Map the current basic probability distribution and the normal basic probability distribution to hyperbolic space, calculate the corresponding Riemann potential energy values ​​respectively, and perform potential energy weighting on the current basic probability distribution and the normal basic probability distribution to generate the current potential energy enhanced probability distribution and the normal potential energy enhanced probability distribution. According to the device coupling chain, read the device node quaternion spectral domain latent features of adjacent device nodes, calculate the decrease between the coupling chain mutual information value and the historical normal coupling chain mutual information value in the current time window, and perform link association weighting processing on the current potential energy enhancement probability distribution and the normal potential energy enhancement probability distribution to generate the current link association enhancement probability distribution and the normal link association enhancement probability distribution. Feature masking and reconstruction processing is performed on the quaternion spectral domain latent features of the device nodes. The error difference between the current reconstruction error and the historical normal reconstruction error is calculated, and reconstruction confidence correction processing is performed to generate the current enhancement probability distribution and the normal enhancement probability distribution. Vibration mode sub-features, temperature mode sub-features, and pressure mode sub-features are extracted from the quaternion spectral domain latent features of equipment nodes. The Wasserstein distance of each mode sub-feature relative to the historical normal latent features is calculated. The maximum value is taken to generate the multimodal Wasserstein wavefront value. Dynamic weight modulation and averaging fusion processing are performed on the current enhancement probability distribution and the normal enhancement probability distribution to generate the enhancement average probability distribution. Based on the JS divergence, the first KL divergence value of the current enhancement probability distribution relative to the enhancement average probability distribution and the second KL divergence value of the normal enhancement probability distribution relative to the enhancement average probability distribution are calculated respectively. The first KL divergence value and the second KL divergence value are averaged to generate the JS divergence value, and the JS divergence value is used as the device distribution offset score.

8. The method for fault diagnosis of ship engine room equipment based on a multi-agent system according to claim 1, characterized in that, The generation of the fault propagation consistency score includes: Each system agent reads the dynamic temporal coupling diagram of the cabin equipment and the equipment distribution offset score, and extracts the corresponding equipment nodes, connection edges, modulated connection edge weights and collection timestamp data according to the power system, cooling system, lubrication system, fuel system and electric system to generate a system-level abnormal node set; Based on the physical coupling relationship of the devices and the collected timestamp data, the device nodes in the system-level abnormal node set are sorted in the propagation order, and the propagation direction between adjacent abnormal device nodes is determined according to the weight of the modulated connection edge, thus generating a system-level abnormal propagation chain. Read the device distribution offset scores of adjacent device nodes in the system-level anomaly propagation chain, calculate the offset change between upstream and downstream device nodes, and combine the modulated connection edge weights to calculate the propagation strength value, generating an anomaly propagation strength sequence. Each system agent verifies the progressive relationship of the device distribution offset score in the device coupling chain based on the system-level anomaly propagation chain, propagation direction, propagation order, and anomaly propagation intensity sequence, and generates a fault propagation consistency score.

9. A method for fault diagnosis of ship engine room equipment based on a multi-agent system according to claim 1, characterized in that, The generated ship engine room equipment fault diagnosis results include: The global diagnostic agent reads the equipment distribution offset score, fault propagation consistency score, cabin equipment dynamic timing coupling diagram and alarm log, aligns it with the equipment node identifier and the collected timestamp data, and extracts abnormal equipment nodes, abnormal time, modulated connection edge weights, system-level abnormal propagation chains, propagation directions and alarm event records. Based on the device distribution offset score, fault propagation consistency score, modulated connection edge weight and alarm event record corresponding to the abnormal device node, the original fault score is calculated. The abnormal device node with the highest original fault score is identified as the original fault device. The abnormal device node located downstream of the system-level abnormal propagation chain and whose device distribution offset score reaches the abnormal judgment threshold is identified as the associated abnormal device. Based on the primary faulty equipment, associated abnormal equipment, primary fault score, abnormal time, system-level abnormal propagation chain, and alarm event records, the fault propagation path, fault level, and maintenance priority are determined, and the fault diagnosis results of the ship's engine room equipment are generated.

Images