Intelligent fault diagnosis and active control system for ship multi-source heterogeneous data fusion

Abstract

The invention discloses an intelligent fault diagnosis and active control system for ship multi-source heterogeneous data fusion, and relates to the technical field of ship control, and the system comprises a feedback optimization module which is used for calculating the deviation between an expected value and heterogeneous data of a current period, and judging whether the deviation exceeds a tolerance interval; if yes, the control parameters are updated again based on the deviation value and fed back to the output simulation calculation module to calculate the expected value till the deviation value is within the tolerance interval, the control parameters updated at the last time are input into the expert rule base, and real-time dynamic autonomous updating of the logic tree is achieved. Through an iterative correction mechanism of the feedback optimization module, full life cycle evolution of diagnosis-decision-optimization is realized, the intelligent level is remarkably improved, high reliability, high adaptability and intelligent autonomous management of a ship system are finally realized, the updating period is short, the efficiency is high, the probability of missed diagnosis and misdiagnosis is effectively reduced, and the system is suitable for popularization and application. And the operation safety of the ship is effectively improved.

Description

Technical Field

[0001] The present invention relates to the field of ship control technology, and in particular to an intelligent fault diagnosis and active control system for ship multi-source heterogeneous data fusion. Background Art

[0002] As a maritime transport vehicle integrating complex systems such as power, navigation, and monitoring, the operational safety of ships directly impacts human life, property, and the marine environment. During ship operation, the power, navigation, and monitoring systems generate large amounts of multi-source, heterogeneous data in real time. Analyzing this heterogeneous data enables fault diagnosis and, based on the diagnostic results, adjustments to ship control parameters to ensure safe operation.

[0003] Currently, existing ship monitoring solutions mostly adopt the simple logic of "sensor acquisition-threshold alarm-manual disposal", or diagnostic methods based on a single model (such as expert system or machine learning). Since the ship's operating environment (such as sea conditions, load, equipment status) is dynamically changing, control strategies based on fixed rules (such as preset engine speed-fuel consumption curve) cannot adapt to operating condition fluctuations in real time. It is necessary to regularly update system control parameters based on historical operating data to adapt to operating condition fluctuations. Traditional methods rely on manually set threshold rules, and trial-and-error adjustment of control parameters relies on manual experience. Manual rule updates lag behind equipment aging or environmental changes. The system cannot automatically adjust diagnostic rules and autonomous purification, which may lead to missed diagnosis or false alarms, and the update efficiency is low. Summary of the Invention

[0004] In view of the shortcomings of the existing technology, the present invention provides an intelligent fault diagnosis and active control system for ships based on multi-source heterogeneous data fusion.

[0005] In order to achieve the above object, the technical solution of the present invention is as follows:

[0006] Intelligent fault diagnosis and active control system based on multi-source heterogeneous data fusion for ships, including:

[0007] A data acquisition module is used to obtain heterogeneous data collected by the power system, navigation system, and monitoring system in the current cycle and the previous cycle, wherein the heterogeneous data includes engine data, position data, and vibration data;

[0008] The rule retrieval module is used to retrieve the corresponding expert rules from the expert rule library based on the heterogeneous data of the previous cycle and form a logic tree;

[0009] The fault diagnosis module is used to calculate the fault probability cloud using the spatiotemporal graph convolutional network and logic tree for the heterogeneous data of the previous cycle, and to parse out the probability of failure of each system;

[0010] A control decision module is used to generate control parameters for each system in the current cycle based on the probability of failure of each system;

[0011] Simulation calculation module: used to input control parameters into the pre-built twin digital simulation model to simulate and calculate the expected values ​​of the heterogeneous data of various ship systems in the current cycle;

[0012] Feedback optimization module:

[0013] Used to calculate the deviation between the expected value and the heterogeneous data of the current period and determine whether the deviation exceeds the tolerance range;

[0014] If yes, the control parameters are updated based on the deviation value and fed back to the output simulation calculation module to calculate the expected value until the deviation value is within the tolerance range. The last updated control parameters are input into the expert rule library to realize real-time dynamic autonomous update of the logic tree.

[0015] Preferably, the logic tree is a differentiable function that encodes the expert rule base:

[0016] ;

[0017] in, : sigmoid function;

[0018] : Confidence of proposition A / B;

[0019] : rule weight;

[0020] : Bias term.

[0021] Preferably, the optimization process of the feedback optimization module is:

[0022] Encode the control parameters of the previous cycle and the control parameters of the current cycle into quantum states and construct quantum transformations that depend on rule weights;

[0023] The quantum gradient value is calculated based on the quantum state of the control parameters of the previous cycle and the control parameters of the current cycle, and the rule weight is adjusted using quantum gradient descent.

[0024] Preferably, the spatiotemporal graph convolutional network includes designing an edge weight model of the host node and the bearing node based on the coupling characteristics of physical distance and force transmission efficiency, and introducing a time decay factor. When the edge weight exceeds a preset threshold, the hidden fault path is predicted, a new virtual node is added to represent the potential fault point, and the edge weight is dynamically updated based on the fusion of the time decay effect and the fault probability of the hidden fault path.

[0025] Preferably, the pre-built twin digital simulation model is a multi-physics field coupling model based on fluid-structure coupling and heat-force coupling, and is solved in real time through a ship-specific reduced-order model to obtain the expected values ​​of the heterogeneous data of each ship system in the current period.

[0026] Preferably, it also includes a brain-computer collaboration module:

[0027] Calculating the risk entropy of the control parameters of the current cycle based on the pre-built twin digital simulation model and in combination with the fault probability cloud, and determining whether the entropy value exceeds a threshold;

[0028] If yes, the control parameters of the previous cycle are parsed through the security verification mechanism and overwriting instructions are generated.

[0029] Preferably, the risk entropy is calculated as follows:

[0030] Extract typical failure modes throughout the ship's life cycle from the heterogeneous database of the previous cycle and construct a failure mode probability distribution model:

[0031] The risk entropy is calculated based on the failure mode probability distribution model.

[0032] Preferably, the risk entropy threshold setting method is:

[0033] The risk entropy value is 0: the probability of only one failure mode is 1;

[0034] The risk entropy value increases: the probability distribution becomes more dispersed;

[0035] The maximum value of risk entropy is ;

[0036] Set risk entropy thresholds in combination with safety standards.

[0037] Preferably, the brain-computer collaboration module includes a fuse mechanism:

[0038] Used to monitor in real time whether the expert fatigue index is greater than the fuse threshold after the cover instruction is generated and before it is executed. If so, it will automatically switch to the AI ​​autonomous decision-making mode;

[0039] The fuse threshold is dynamically adjusted according to the sea condition level.

[0040] Preferably, the data acquisition module further includes a quantum enhanced fusion unit:

[0041] A quadratic unconstrained binary optimization model for constructing ship equipment topology constraints is used to represent the activation state of the collected heterogeneous data as binary variables;

[0042] The optimal space-time alignment path is quickly solved through quantum annealing parallel search capabilities.

[0043] Compared with the prior art, the present invention has the following beneficial effects:

[0044] The present invention uses the iterative correction mechanism of the feedback optimization module to enable the system to continuously self-adjust control parameters and dynamically update expert rules, so that the system no longer relies on static rules or requires manual updates. By directly feeding the output of control decisions and the results of feedback optimization (updated control parameters) back into the rule generation process, the system realizes closed-loop learning and autonomous evolution of the knowledge base, adapts to ship aging, environmental changes, new fault modes, etc., realizes the full life cycle evolution of "diagnosis-decision-optimization", and significantly improves the intelligence level, ultimately achieving high reliability, high adaptability and intelligent autonomous management of ship systems, with short update cycles and high efficiency, effectively reducing the probability of missed diagnosis and misdiagnosis, and effectively improving the safety of ship operation. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] The disclosure of the present invention is described with reference to the accompanying drawings. It should be understood that the drawings are for illustrative purposes only and are not intended to limit the scope of protection of the present invention. In the drawings, the same reference numerals are used to refer to the same components. Among them:

[0046] Figure 1 This is the system flow of the present invention;

[0047] Figure 2 This is a system flow chart of the brain-computer collaboration module of the present invention. DETAILED DESCRIPTION

[0048] It is easy to understand that according to the technical solution of the present invention, without changing the essential spirit of the present invention, a person skilled in the art can propose a variety of interchangeable structural modes and implementation modes. Therefore, the following specific embodiments and drawings are only exemplary descriptions of the technical solution of the present invention and should not be regarded as the entire invention or as a limitation or restriction of the technical solution of the present invention.

[0049] like Figure 1-2 As shown in the figure, the intelligent fault diagnosis and active control system for ship multi-source heterogeneous data fusion includes:

[0050] The data acquisition module is used to obtain heterogeneous data collected by the power system, navigation system, and monitoring system in the current cycle and the previous cycle. The heterogeneous data includes engine data, position data, and vibration data;

[0051] Specifically, the data collection module is responsible for collecting raw data on the operation of ship systems and is the "input source" of the entire system. The data acquisition module collects operational data of key ship systems, including power systems (such as engines), navigation systems (such as positioning equipment), and monitoring systems (such as vibration sensors). The collected data types are heterogeneous data from different sources and formats, including:

[0052] Engine data (such as engine speed, fuel flow, boost pressure, exhaust temperature and other core parameters);

[0053] Position data (such as GPS position, heading angle, wave angle, speed and other environmental and navigation parameters);

[0054] Vibration data (such as vibration amplitude / frequency of main engine bearings and superchargers, equipment status data such as temperature sensors and pressure sensors);

[0055] The collection cycle not only obtains the real-time data of the operation of each system in the current cycle of the ship, but also needs to retrieve the historical data of the operation of each system in the previous cycle of the ship to provide a data basis for subsequent comparative analysis and fault diagnosis.

[0056] The rule retrieval module is used to retrieve the corresponding expert rules from the expert rule library based on the heterogeneous data of the previous cycle and form a logic tree;

[0057] Specifically, the rule retrieval module generates a logic tree for ship fault diagnosis. It is a bridge connecting the historical data collected from the ship's operation to the current cycle and the diagnostic logic. The rule retrieval module retrieves the corresponding rules from the expert rule library based on the heterogeneous data of the previous cycle and forms a logic tree, where:

[0058] The expert rule base stores fault judgment rules learned from historical data obtained from the ship's operation to the current period, including domain expert experience, equipment physical models, and historical fault cases.

[0059] The logic tree visualizes the extracted expert rules in a tree structure, where each node represents a judgment condition and branches represent different fault possibilities, providing a structured reasoning path for fault diagnosis.

[0060] The fault diagnosis module is used to calculate the fault probability cloud using the spatiotemporal graph convolutional network and logic tree for the heterogeneous data of the previous cycle, and to parse out the probability of failure of each system;

[0061] Specifically, the fault diagnosis module is used to analyze whether the system has faults and the probability of faults. It is the diagnostic core of the system and combines two methods to make a comprehensive judgment:

[0062] The spatiotemporal graph convolutional network is a deep learning model that excels at processing spatiotemporal correlation data and can capture complex patterns hidden in the data.

[0063] The logic tree is a structured reasoning based on expert rules, which can quickly locate known fault types;

[0064] The fault diagnosis module uses the spatiotemporal graph convolutional network and logic tree to make a comprehensive judgment on the heterogeneous data of the previous cycle and output:

[0065] Fault probability cloud: represents the possible fault range of the system in the form of probability distribution;

[0066] Further analyze the probability of failure of each system to provide a quantitative basis for subsequent decision-making.

[0067] A control decision module is used to generate control parameters for each system in the current cycle based on the probability of failure of each system;

[0068] Specifically, the control decision module generates the control strategy for the current cycle based on the fault diagnosis results. It is the decision center of the system and outputs the probability of failure of each system according to the output of the fault diagnosis module:

[0069] Generate corresponding control parameters for the power system, navigation system, and monitoring system of the current cycle. These parameters are specific instructions, such as:

[0070] Power system: target speed, throttle opening limit;

[0071] Navigation system: heading fine-tuning instructions, speed setting;

[0072] Monitoring system: sensor sampling frequency adjustment, key monitoring instructions for specific areas;

[0073] When the probability of failure of a system is high, its parameters need to be adjusted first.

[0074] Simulation calculation module: used to input control parameters into the pre-built twin digital simulation model to simulate and calculate the expected values ​​of the heterogeneous data of various ship systems in the current cycle;

[0075] Specifically, the simulation calculation module simulates the execution effect of control parameters through digital models, and is a virtual verification link for decision-making. Through the pre-built twin digital simulation model, which is a digital mirror of the ship's physical system, it can accurately simulate the operating status of the system under different control parameters, and simulate and calculate the expected value of the heterogeneous data of various ship systems in the current cycle, that is, the data that should be generated in theory if the current control parameters are executed.

[0076] Feedback optimization module:

[0077] Used to calculate the deviation between the expected value and the heterogeneous data of the current period and determine whether the deviation exceeds the tolerance range;

[0078] If yes, the control parameters are updated based on the deviation value and fed back to the output simulation calculation module to calculate the expected value until the deviation value is within the tolerance range. The last updated control parameters are input into the expert rule library to realize real-time dynamic autonomous update of the logic tree.

[0079] Specifically, the feedback optimization module acts as the system's self-optimizer, correcting control parameters through closed-loop iteration to ensure that decisions match actual operating conditions and dynamically updating expert rules. The specific process includes the following:

[0080] Deviation calculation: Compare the expected value of the simulation calculation module with the actual heterogeneous data of the current cycle to obtain the deviation between the two;

[0081] Deviation judgment: Check whether the deviation is within the preset tolerance range;

[0082] Iterative optimization:

[0083] If the deviation exceeds the tolerance interval, it means that the current control parameters are unreasonable and need to be readjusted based on the deviation value. The updated control parameters are output to the simulation calculation module to calculate the expected value and the deviation value again. This process is repeated, and the control parameters are updated and the simulation calculation deviation is repeated until the deviation value is within the tolerance interval. The final optimized control parameters are input into the expert rule library and the corresponding rules are updated, thereby realizing real-time dynamic autonomous update of the logic tree, so that the expert rules can adapt to new situations such as equipment aging and environmental changes, and the system can learn from actual deviations, optimize its own knowledge base and decision-making ability, and improve the accuracy of subsequent diagnosis.

[0084] If the deviation does not exceed the tolerance interval: the current control parameters are valid and no adjustment is required. The system operates as expected, the control parameters are adopted and executed, and the process enters the next cycle.

[0085] Furthermore, through the full process design of data collection → rule reasoning → intelligent diagnosis → decision generation → simulation verification → closed-loop optimization, a dynamic and adaptive intelligent management closed loop is formed:

[0086] It not only uses deep learning to mine implicit patterns in data, but also combines expert rules to ensure the interpretability of diagnosis; it verifies decision-making effects in advance through digital twin simulation, reducing the trial-and-error costs of actual operations; most importantly, the iterative correction mechanism of the feedback optimization module enables the system to continuously self-adjust control parameters and dynamically update expert rules, so that the system no longer relies on static rules or requires manual updates. By directly feeding the output of control decisions and the results of feedback optimization (updated control parameters) back into the rule generation process, the system realizes closed-loop learning and autonomous evolution of the knowledge base, adapts to ship aging, environmental changes, new fault modes, etc., realizes the full life cycle evolution of "diagnosis-decision-optimization", and significantly improves the level of intelligence, ultimately achieving high reliability, high adaptability and intelligent autonomous management of ship systems, with short update cycles and high efficiency, effectively reducing the probability of missed diagnosis and misdiagnosis, and effectively improving the safety of ship operation.

[0087] The logic tree is a differentiable function that encodes the expert rule base:

[0088] ;

[0089] in, : sigmoid function;

[0090] : The confidence level of proposition A / B;

[0091] : rule weight;

[0092] : Bias term.

[0093] Specifically, the logic tree is mathematically defined as a differentiable function, which essentially converts the logical relationships in the expert rule base into computable and optimizable mathematical expressions, so that the expert rules originally based on discrete logic can be jointly trained and optimized with deep learning models (such as spatiotemporal graph convolutional networks) while retaining the interpretability of the rules.

[0094] Further, : sigmoid function, which is a commonly used activation function, the mathematical expression is , whose output value range is between (0,1), and the result of the linear combination is mapped to a "probability value" or "confidence level", which meets the probabilistic output requirements of the "fault probability cloud";

[0095] : The confidence level of proposition A / B. A proposition refers to the judgment condition in the expert rule. The confidence level indicates the probability that the proposition is true (expressed as a value between 0 and 1). It is usually calculated based on heterogeneous data from the previous cycle.

[0096] Rule weight: This represents the importance of propositions A and B in logical reasoning. The larger the weight, the stronger the influence of the proposition on the final result. As a learnable parameter, it can be dynamically adjusted according to the actual deviation through the feedback optimization module, which is different from the traditional expert rules that are fixed;

[0097] : Bias term, used to adjust the "baseline" of the logic function, similar to the constant term in a linear equation, used to compensate for systematic deviations in rule modeling (for example, when the confidence levels of A and B are both 0, the bias term can set a base probability to avoid the output being 0).

[0098] Further, the "logic tree" is transformed from the traditional "tree branch judgment" to a "continuous differentiable mathematical function", preserving the logical relationship of the expert rules while the sigmoid function and linear combination are both continuous differentiable functions, so that the logic tree can be combined with the space-time graph convolution network (deep learning model) to optimize the parameters through gradient descent and other methods, solving the problem that traditional discrete rules cannot be trained with neural networks, realizing "differentiability" and deep learning compatibility, and the rule weight and bias can be adjusted in real time through the feedback optimization module (based on the deviation between expected value and actual data), so that the logic tree can adapt to new working conditions (such as adjusting the weight of "vibration amplitude" to improve the diagnostic sensitivity after equipment aging), realizing "dynamic and autonomous update of expert rule base". The design of defining the logic tree as a "linear combination function with sigmoid activation" is a typical way of combining symbolism (expert rules) and connectionism (deep learning): it not only preserves the interpretability of expert rules (reflecting the importance of each condition through weights), but also gives it the property of continuous differentiability, enabling it to be optimized with neural networks, and ultimately achieving a balance between accuracy and adaptability in fault diagnosis.

[0099] Example One

[0100] Suppose an expert rule is: "If engine temperature exceeds standard (A) and vibration is abnormal (B), then judge as mechanical failure", the corresponding logic tree function is:

[0101] Mechanical failure probability ;

[0102] When the sensor data shows (Temperature exceeds standard with 90% confidence), (Vibration is abnormal with 80% confidence):

[0103] The linear combination result is: 0.7x0.9+0.5x0.8+0.1=0.63+0.4+0.1=1.13;

[0104] The sigmoid function output is , i.e. the probability of mechanical failure is about 75%;

[0105] If the feedback optimization module finds that the actual failure probability is higher than the predicted value, it may increase the weight or (such as adjusting from 0.7 to 0.8), making the logic tree more sensitive.

[0106] The optimization process of the feedback optimization module is:

[0107] Encode the control parameters of the last period and the current period into quantum states, and construct a quantum transformation that depends on the rule weight;

[0108] Quantum state calculates quantum gradient value based on control parameters of last period and control parameters of current period, and adjusts rule weight using quantum gradient descent.

[0109] Specifically, traditional feedback optimization usually adjusts parameters based on classical algorithms (such as gradient descent), while here the control parameters are encoded as quantum states, and the superposition and entanglement of quantum mechanics are used for calculation. The core is to optimize the rule weight (i.e. , and other parameters) in the logic tree through quantum gradient descent, improve optimization efficiency, empower feedback optimization with quantum computing, and realize the optimization of rule weight through quantum state encoding and quantum gradient descent, making the feedback adjustment of the system more efficient and more accurate.

[0110] Further, the optimization process is as follows:

[0111] Encoding control parameters as quantum states:

[0112] Control parameters: refer to control parameters of last period (historical decisions) and control parameters of current period (decisions to be optimized, such as engine speed, navigation correction value, etc.);

[0113] Quantum state encoding: convert classical control parameters (numerical data) into "quantum states" in quantum mechanics (described by wave functions, with superposition). For example:

[0114] A control parameter x can be encoded as a quantum state ;

[0115] where and are complex numbers (probability amplitudes) that satisfy , representing the probability distribution of the parameter in quantum space;

[0116] A single quantum state can represent multiple parameters in "superposition state" at the same time, laying the foundation for parallel computing.

[0117] Construct quantum transformation dependent on rule weight:

[0118] Quantum transformation: a mathematical transformation based on quantum gate operation, used to describe the evolution of quantum states. The quantum transformation here is designed as a function of rule weight ( , , etc.), that is:

[0119] ;

[0120] where is a quantum transformation operator determined by the rule weight, which transforms the quantum state of the control parameters of the last period Quantum state mapping to current period control parameters By simulating "how the rule weight affects the iteration of the control parameter" through quantum transformation, the adjustment logic of the expert rule is embedded in the quantum computing framework.

[0121] Calculate quantum gradient value:

[0122] The gradient is an indicator of "the degree to which parameter changes affect the objective function", and in feedback optimization, the objective function is "the deviation between the expected value and the actual data", and the quantum gradient represents "how a small change in the rule weight affects the change in the deviation".

[0123] Quantum gradient calculation:

[0124] Quantum state based on last period and current period control parameters And Using quantum algorithms (such as quantum phase estimation, variational quantum algorithms) to calculate the gradient value, compared with classical gradient calculation, quantum gradient can evaluate the gradient of multiple parameter combinations simultaneously using quantum superposition, theoretically more efficient (especially in high-dimensional parameter space).

[0125] Adjust the rule weight using quantum gradient descent:

[0126] Quantum gradient descent: is the quantum version of the classical gradient descent algorithm, the core logic is "adjust the parameters in the opposite direction of the gradient to minimize the objective function (deviation)". Specifically:

[0127] If the quantum gradient shows that "increasing will reduce the deviation", then increase in the direction of the gradient;

[0128] If "decreasing can reduce the deviation", then decrease in the direction of the gradient;

[0129] Repeat the iteration until the deviation falls within the tolerance interval, quantum parallelism can accelerate the convergence speed of gradient descent, especially suitable for scenarios with high-dimensional rule weights and complex optimization.

[0130] Furthermore, the superposition of quantum states allows the simultaneous processing of multiple combinations of control parameters. Quantum gradient descent can find the optimal solution more quickly in high-dimensional parameter space, reducing the number of iterations of feedback optimization. There may be strong coupling between the rule weights of the ship system (such as the mutual influence of the parameters of the power system and the navigation system). The characteristics of quantum entanglement can naturally describe this relationship and improve the accuracy of optimization. When the expert rule base is huge (the number of rule weights is large), classical computing may face a "dimensionality disaster", while quantum computing has the potential to handle high-dimensional problems in theory. The feedback optimization module integrates the characteristics of quantum computing into the adjustment process of rule weights through the process of "quantum state encoding → quantum transformation construction → quantum gradient calculation → quantum gradient descent", breaking through the efficiency bottleneck of classical optimization algorithms. It is especially suitable for complex scenarios such as ship systems that are high-dimensional, strongly coupled, and require real-time response, and ultimately achieves faster and more accurate dynamic updates of rule weights (logic trees), thereby improving the adaptability of the entire system.

[0131] The spatiotemporal graph convolutional network includes designing an edge weight model of the host node and the bearing node based on the coupling characteristics of physical distance and force transmission efficiency, and introducing a time decay factor. When the edge weight exceeds the preset threshold, the hidden fault path is predicted, and new virtual nodes are added to represent potential fault points. The edge weight is dynamically updated based on the fusion of the time decay effect and the fault probability of the hidden fault path.

[0132] Specifically, in complex mechanical systems such as ships, there are strong physical connections between devices (for example, the main engine and bearings are connected through a drive shaft, and vibrations and forces are transmitted to each other), and faults have spatiotemporal propagation characteristics (for example, bearing wear may spread to the main engine over time, and early fault signals are weak and hidden). Traditional graph convolutional networks only focus on static topological relationships and have difficulty capturing this dynamic coupling and hidden faults. The core optimization strategy of the spatiotemporal graph convolutional network designed for ship systems focuses on improving the diagnostic capabilities of complex system faults by modeling the physical coupling relationship between devices, time attenuation effects, and hidden fault paths.

[0133] Furthermore, the edge weight model based on physical coupling characteristics:

[0134] ;

[0135] in, For physical distance, is the force transfer angle;

[0136] The basic structure of graph networks: In spatiotemporal graph convolutional networks, key devices in the system (such as hosts and bearings) are abstracted as "nodes", the physical connections or mutual influence relationships between devices are abstracted as "edges", and the "edge weight" represents the strength of this association.

[0137] Among them, the distance term ( ): is the physical straight-line distance between the host and the bearing node. In mechanical systems, the intensity of force and vibration transmission tends to decay with distance. The inverse of the distance directly quantifies the characteristic of "strong correlation at close distance and weak correlation at long distance", i.e., the closer the spatial distance between devices (e.g., the host and the adjacent bearing), the higher the transmission efficiency of vibration, heat, etc., and the larger the edge weight; the farther the distance, the smaller the edge weight (e.g., the host and the remote bearing).

[0138] Force transmission angle term ( ): is the angle between the actual transmission direction of the host output force and the direction of the host-bearing node connection line. The effective transmission of force depends on the component along the connection line direction:

[0139] When the force is transmitted along the connection line direction ( ), , the transmission efficiency is the highest;

[0140] When the force is perpendicular to the connection line ( ), , there is almost no effective transmission;

[0141] The square term further amplifies the influence of the direction, highlighting the role of "direction consistency" on the correlation strength. Based on the mechanical structure characteristics (such as the rigidity of the transmission shaft, the meshing efficiency of the gear), the transmission loss of force / vibration between devices is quantified, and the higher the transmission efficiency, the larger the edge weight (e.g., the edge weight of rigidly connected devices is larger than that of flexibly connected devices), so that the topology of the network is consistent with the actual coupling characteristics of the physical system, ensuring that the network can capture the real fault propagation path (e.g., the rule of vibration transmission from the bearing to the host).

[0142] Further, the intensity of the fault influence between devices will change over time (e.g., the vibration signal of early-stage faults may decay or strengthen over time). For example, the slight wear of the bearing has a weak impact on the host at the initial stage (the signal decays quickly), but as time goes by (wear intensifies), the impact gradually increases (decay slows down), and a time-dependent decay function is introduced into the edge weight to dynamically adjust the edge weight at different times:

[0143] The edge weight of recent data decays small (large impact), and the edge weight of historical data decays large (small impact);

[0144] Different decay coefficients can be set according to the type of equipment (e.g., high-speed rotating parts decay quickly, and static parts decay slowly);

[0145] This allows the network to distinguish the dynamic changes of fault signals in the time dimension, avoid interference from outdated weak signals, and improve the accuracy of time-series correlation analysis.

[0146] Further, implicit fault path prediction and virtual node addition:

[0147] Because early faults (such as microcracks and slight looseness) have weak signals and no clear fault path, traditional graph networks are difficult to identify (because the edge weight does not exceed the threshold and cannot trigger fault judgment). The edge weight model can predict hidden fault paths:

[0148] When the edge weight (after time decay) exceeds the preset threshold (e.g., the edge weight of "host-bearing" is greater than 0.6), but has not yet reached the explicit fault standard, the network determines that there is a "hidden fault path" (i.e., the fault is latently propagating);

[0149] A new "virtual node" is added to the hidden fault path to represent the potential fault point (such as "micro-cracks on the transmission shaft between the host and the bearing"). This node does not correspond to a physical device, but carries the characteristic information of the hidden fault.

[0150] The virtual node establishes a temporary edge connection with the relevant physical nodes (host, bearing). The initial value of the edge weight is set based on the probability of latent faults, which transforms latent faults from "invisible" to "modelable", captures the budding state of faults in advance, and avoids missed diagnosis.

[0151] Furthermore, after adding a virtual node, the network adjusts the edge weights in real time based on the following factors:

[0152] Propagation trend of hidden fault paths: If the fault characteristics of a virtual node (such as changes in vibration frequency) increase, it indicates that the fault is spreading, and the weight of related edges (such as the edge weights between host and virtual node, and virtual node and bearing) will be increased.

[0153] Time decay effect: Combined with the time decay factor, the edge weights at different times are updated (for example, the longer the fault duration, the slower the edge weight decays, reflecting the cumulative effect of the fault);

[0154] Fault probability feedback: By integrating the fault probability output by the fault diagnosis module (such as an increase in the fault probability corresponding to a virtual node), the edge weights are dynamically modified (the higher the probability, the larger the edge weight). This allows the topology of the graph network to evolve dynamically as the fault develops, upgrading from "static modeling" to "dynamic tracking" to achieve full cycle coverage of faults from latent to explicit.

[0155] Furthermore, through a four-layer design consisting of physical coupling modeling (edge ​​weights) – dynamic time adjustment (attenuation factor) – latent fault capture (virtual nodes) – and real-time weight updates, accurate modeling of complex ship systems is achieved:

[0156] It not only complies with physical laws (based on distance and force transmission), but also captures dynamic changes (time decay);

[0157] It can identify explicit faults (clear paths) and predict implicit risks (virtual nodes);

[0158] Ultimately, the network can more accurately mine the spatiotemporal correlation features in the data, provide more reliable "data-driven" analysis results for the fault diagnosis module, and support early fault warning and accurate diagnosis of ship systems.

[0159] The pre-built twin digital simulation model is a multi-physics coupling model based on fluid-structure coupling and thermal-mechanical coupling. It uses a ship-specific reduced-order model to perform real-time solution to obtain the expected values ​​of the heterogeneous data of various ship systems in the current cycle.

[0160] Specifically, the essence of a twin digital simulation model is a digital mirror image of a physical system. A "twin digital simulation model" (digital twin) is a full-element digital reproduction of a ship's physical system (such as the hull, power plant, navigation equipment, etc.). It can map the operating status of the physical system in real time and predict the expected values ​​of the system under different control parameters (such as engine temperature, hull vibration, speed, etc.) through simulation calculations, providing a virtual verification basis for actual decision-making. Ship operation is a typical multi-physical field interaction process (multiple physical phenomena occur simultaneously and influence each other). The model must accurately depict these coupling relationships to ensure the authenticity of the simulation results. Specifically, there are two types of key couplings:

[0161] Fluid-Structure Coupling (Fluid-Structure Interaction):

[0162] Physical scenario: When a ship sails in the water, there is interaction between the water flow (fluid) and structural components such as the hull and propeller.

[0163] Coupling mechanism:

[0164] The effect of fluid on the structure: water flow impacts the hull to generate resistance and pressure loads. When the propeller rotates and pushes the water flow, it is subjected to reaction force (thrust). Wave loads cause the hull to vibrate or deform.

[0165] The impact of structure on fluid: the shape of the hull changes the direction of water flow (such as wave-making effect), and structural vibration (such as hull rolling) in turn interferes with the flow field distribution (such as vortex changes).

[0166] By simulating flow field characteristics (flow velocity, pressure, wave shape) through computational fluid dynamics and combining it with structural mechanics to calculate the stress, deformation, and vibration response of the hull and propeller, a dynamic coupling solution is achieved between the two, accurately predicting the ship's power performance (such as propulsion efficiency), structural safety (such as hull fatigue damage), and navigation resistance at different speeds and sea conditions.

[0167] Thermal and mechanical coupling (thermo-mechanical coupling):

[0168] Physical scenario: The interaction between thermal and mechanical forces when the ship's power system (such as the engine and drive shaft) is in operation.

[0169] Coupling mechanism:

[0170] Thermal effects on mechanical forces: Mechanical components generate heat through friction (e.g., bearing rotation, gear meshing), and structural stress changes (e.g., stress concentration due to thermal expansion and contraction) can alter heat conduction paths.

[0171] Mechanical effects on thermal forces: Temperature increases can cause material properties to change (e.g., reduced strength, decreased stiffness), and thermal stresses (internal stresses caused by temperature differences) can exacerbate structural deformation or wear, even leading to failures (e.g., engine overheating causing piston seizure).

[0172] By simulating temperature field distribution (e.g., engine block temperature, bearing temperature rise) through heat conduction equations, and analyzing thermal stress and thermal deformation effects on mechanical component stress states through structural mechanics, a two-way coupling of thermal and mechanical forces is achieved, accurately predicting the thermal stability (e.g., heat dissipation efficiency) of the power system, mechanical losses (e.g., changes in friction coefficient with temperature), and failure risks (e.g., performance degradation due to overheating).

[0173] Furthermore, multi-physics coupling models contain a large number of physical parameters (e.g., flow field grid, structural elements, temperature nodes), and direct solving can face problems of excessive computational load and long time consumption (difficult to meet the "current cycle real-time simulation" requirement). Therefore, "reduced order model" (ROM) technology needs to be introduced:

[0174] Dimension reduction: Under the premise of ensuring simulation accuracy, reduce the degrees of freedom of the model (e.g., ignore minor physical details, combine similar calculation units) through mathematical methods (e.g., modal decomposition, machine learning approximation), and reduce computational complexity.

[0175] Ship-specific optimization: Tailor the reduction strategy to the characteristics of the ship system (e.g., periodicity of navigation working conditions, structural symmetry):

[0176] For ship flow field: Preserve key speed and draft characteristic modes of the flow field, and ignore details of atypical working conditions;

[0177] For power system: Focus on thermal coupling of core components (e.g., engine block, main shaft bearing), and simplify modeling of secondary accessories;

[0178] Introduce prior knowledge of ship operation (e.g., typical sea state database, equipment aging curve) to improve the generalization ability of the reduced order model.

[0179] The reduced order model shortens the high-precision simulation time from hours to milliseconds or seconds by simplifying the calculation dimension, meeting the "current cycle" (real-time or quasi-real-time) simulation requirement.

[0180] After inputting the "control parameters" generated by the control decision module (e.g., engine speed, rudder angle), the reduced order model can quickly output the corresponding "heterogeneous data expected values", such as:

[0181] Power system: engine temperature, fuel consumption, output power;

[0182] Navigation system: speed, course deviation, hull swing angle;

[0183] Monitoring system: vibration frequency of key components, temperature gradient, etc.

[0184] Furthermore, the model accurately restores the complex physical process of ship operation through multi-physical field coupling (fluid-solid coupling characterizes navigation characteristics, and thermal-solid coupling captures the state of the power system), and then realizes real-time solution through ship-specific order reduction technology, providing a "virtual testing ground" for control decision-making, predicting the execution effect (expected value) of control parameters in advance, avoiding direct trial and error on the physical system (such as equipment damage caused by parameter adjustment), reducing testing costs, and providing a comparison benchmark (expected value vs. actual data) for the feedback optimization module, supporting the dynamic correction of control parameters. Ultimately, the model becomes a key bridge connecting "decision-making" and "actual operation", effectively improving the intelligent control level of ship systems.

[0185] Also includes brain-computer collaboration module:

[0186] Based on the pre-built twin digital simulation model and combined with the failure probability cloud computing, the risk entropy of the control parameters of the current cycle is calculated, and whether the entropy value exceeds the threshold is determined;

[0187] If yes, the control parameters of the previous cycle are parsed through the security verification mechanism and overwriting instructions are generated.

[0188] Specifically, risk assessment and human (or intelligent) intervention mechanisms are introduced on the basis of the system's autonomous decision-making. Through the logic of "risk entropy judgment + safety verification + override instructions," the safety and reliability of ship control decisions are improved. In particular, the dual protection of "machine autonomy + human / intelligent collaboration" is achieved in high-risk scenarios. "Brain-computer collaboration" is not narrowly defined as "direct interaction between the human brain and machine," but rather a broader mechanism of "autonomous decision-making by intelligent systems and collaborative intervention from external parties (such as expert experience, safety rules, and emergency logic)." As the system's "safety redundancy and decision-making checkpoint," this module intervenes when autonomous decisions may pose high risks to ensure the safety of control parameters.

[0189] Further, entropy is a physical quantity that measures "uncertainty", and "risk entropy" is used here to quantify the potential risk level of the current cycle control parameters - the higher the entropy value, the greater the uncertainty of the system operating state corresponding to the control parameters, and the higher the risk (such as fault probability dispersion, simulation result fluctuation), using the expected value and its fluctuation range obtained by simulation (such as parameter deviation distribution under different working conditions), and combining the fault probability distribution output by the fault diagnosis module (such as the dispersion degree of the probability cloud), the risk entropy of the current control parameters is calculated through comprehensive analysis of the two (for example: if the simulation expected value fluctuates greatly and the fault probability cloud is dispersed, the risk entropy increases).

[0190] Further, it is judged whether the risk entropy exceeds the threshold value, and the threshold value is a preset "risk warning line" set by domain experts or safety specifications, representing the maximum risk uncertainty that the system can accept (such as triggering intervention when the entropy value > 0.8), the judgment logic is:

[0191] If the risk entropy does not exceed the threshold value: it means that the risk of the current control parameters is within a controllable range, and no intervention is needed, and the system executes according to the self-determined parameters;

[0192] If the risk entropy exceeds the threshold value: it means that the current control parameters may have high uncertainty (such as extreme working conditions not covered by the simulation model, and insufficient decision basis due to fuzzy fault probability cloud), and the collaborative intervention mechanism needs to be started.

[0193] Further, the safety verification mechanism is used to analyze the control parameters of the last cycle, and when high risk occurs, the safety verification mechanism is used to backtrack and verify the historical decisions to ensure the safety and rationality of the intervention instructions, and the analysis content is:

[0194] Retrieve and analyze the "control parameters of the last cycle" (i.e. the decision results of the last cycle that have been proven effective), including:

[0195] The execution effect of the parameters of the last cycle (such as whether it makes the system run stably, and whether the deviation is within the tolerance interval);

[0196] The safety boundary corresponding to the parameters (such as the safety range of engine speed, and the obstacle avoidance logic of navigation route);

[0197] Associated expert rules or emergency plans (such as successful intervention cases in historical high-risk scenarios);

[0198] Through analysis, the safety and applicability of the parameters of the last cycle are confirmed to provide a basis for subsequent intervention.

[0199] Further, when the current cycle control parameter risk is too high, the "safety instruction" used to replace or modify the original parameter ensures that the system enters a more stable operating state. By analyzing the results of the last cycle control parameter, a cover instruction is generated, which is usually a "conservative and safe parameter adjustment scheme", for example:

[0200] If the current parameter leads to excessive engine load due to pursuit of efficiency (risk entropy exceeds the standard), the cover instruction may reduce the engine speed to the historical safe value;

[0201] If the navigation parameter leads to high failure probability and large uncertainty of the route, the cover instruction may switch to the preset safe route or slow down for inspection;

[0202] The cover instruction directly replaces the autonomous decision-making parameter of the current cycle, allowing the system to switch from a high-risk state to a low-risk state, avoiding potential failures or accidents.

[0203] Further, the brain-machine collaboration module adds a "safety line" to the system through the process of "risk entropy evaluation → threshold judgment → safety verification → cover instruction". When the risk uncertainty of autonomous decision-making is too high, it no longer relies on the current parameter, but instead uses verified historical safe parameters. Through the collaborative mode of "machine autonomous evaluation + safety rules / experience intervention", the safety and stability of ship control decisions in complex working conditions are ensured, which is an important supplement to the "intelligence" and "reliability" of the entire system. When the system's autonomous decision-making may lead to high risk due to model limitations, data noise, or extreme working conditions, the intervention is triggered through risk entropy judgment to avoid dangerous "machine misjudgment" and effectively ensure decision safety. As a "safety backup" for autonomous decision-making, especially in high-risk scenarios such as ships, single autonomous decision-making may have blind spots. Brain-machine collaboration provides reliable emergency solutions through the use of historical safe parameters, achieving redundancy protection. The "risk entropy calculation" reflects the quantitative analysis ability of the machine, and the "safety verification and cover instruction" combines historical experience (parameters from the last cycle) and safety rules to achieve a collaborative decision-making process that combines "data-driven" and "experience-based" safeguards.

[0204] The calculation method of risk entropy is as follows:

[0205] From the heterogeneous database of the last cycle, extract typical failure modes of the ship throughout its life cycle, and build a failure mode probability distribution model:

[0206] Based on the calculation of risk entropy.

[0207] Specifically, extract typical failure modes and build a probability distribution model

[0208] Step 1: Extract typical failure modes:

[0209] From the "previous cycle heterogeneous database" (containing historical data from the entire ship life cycle, such as fault records, maintenance logs, sensor data, etc.), screen and summarize "typical failure modes" - that is, the types of failures or failure scenarios that repeatedly occur in the long-term operation of various ship systems (power, navigation, monitoring, etc.). For example:

[0210] Power system: engine overheating failure, fuel pump blockage, bearing wear and fracture;

[0211] Navigation system: GPS signal loss, heading sensor drift, and route deviation warning;

[0212] Monitoring system: vibration sensor failure, data transmission delay, false alarm failure.

[0213] Step 2: Construct failure mode probability distribution model:

[0214] For the extracted typical failure modes, their probability of occurrence in the entire life cycle is calculated (e.g., "engine overheating failure occurs 20 times in 1,000 voyages, with a probability of 2%)" and a probability distribution model (e.g., discrete probability distribution) is constructed. The model can be expressed as: ; where n is the total number of typical failure modes, is the probability of occurrence of the i-th failure mode, and satisfies:

[0215] (The sum of the probabilities of all failure modes is 1).

[0216] Furthermore, the risk entropy is calculated based on the probability distribution model:

[0217] The calculation of risk entropy adopts the Shannon entropy formula in information theory, and its mathematical expression is: ;

[0218] in:

[0219] is the risk entropy, the larger the value, the higher the uncertainty of the risk;

[0220] is the probability of occurrence of the i-th typical failure mode;

[0221] Usually the natural logarithm or the logarithm with base 2 is taken (the unit is nat or bit respectively), which does not essentially affect the relative size of the entropy value.

[0222] Calculation logic example:

[0223] If the system has only one typical failure mode (n=1, ),but:

[0224] (The risk is completely certain and there is no uncertainty);

[0225] If there are two failure modes, the probabilities are 、 ,but:

[0226] (Risk uncertainty is medium);

[0227] If there are four failure modes, each with a probability of 0.25, then:

[0228] (Risk uncertainty is higher).

[0229] The calculation of risk entropy transforms the risk of the ship system from "qualitative description" to "quantitative value" through the process of "extracting typical failure modes from historical data → statistical probability distribution → calculation of Shannon entropy formula", providing an objective risk uncertainty indicator for the brain-computer collaboration module. When the risk entropy exceeds the threshold, the safety intervention mechanism is triggered, ultimately improving the safety and robustness of system decision-making.

[0230] The risk entropy threshold setting method is:

[0231] The risk entropy value is 0: the probability of only one failure mode is 1;

[0232] The risk entropy value increases: the probability distribution becomes more dispersed;

[0233] The maximum value of risk entropy is ;

[0234] Set risk entropy thresholds in combination with safety standards.

[0235] Specifically, the value range and change pattern of risk entropy are determined by its definition (Shannon entropy formula). These characteristics are the premise for setting thresholds:

[0236] The risk entropy value is 0: the risk of a single failure is completely certain:

[0237] When the system has only one possible failure mode and its probability of occurrence is 1 ( , when the probability of all other failure modes is 0), according to the Shannon entropy formula:

[0238] ;

[0239] At this point, the risk entropy is 0, meaning the risk state is completely certain (there is only one known failure mode) and the system faces minimal uncertainty. For example, if a ship's engine is bound to wear out due to insufficient lubrication (a single failure mode with a 100% probability), the risk entropy is 0.

[0240] Risk entropy increases: the failure probability distribution is more dispersed, when the probability distribution of multiple failure modes changes from "concentration" to "dispersion", the risk entropy increases accordingly. For example:

[0241] If the probability of 2 failure modes is (0.9, 0.1), the entropy value H ≈ 0.47;

[0242] If the probability becomes (0.5, 0.5), the entropy value H ≈ 1 (increases);

[0243] The higher the entropy value, the more uniform the distribution of failure risk among multiple modes, and the higher the uncertainty faced by the system (it is difficult to predict which failure will occur).

[0244] The maximum risk entropy is : complete uniform probability distribution, when there are N typical failure modes in the system, and the occurrence probability of each mode is completely equal ( ) when the risk entropy reaches the maximum value:

[0245] ;

[0246] For example: if there are 4 failure modes, the maximum entropy value is ; if there are 8 failure modes, the maximum entropy value is , which represents the "extreme uncertainty" of risk - all failure modes have equal probability of occurrence, and the system is most difficult to predict and control.

[0247] Further, the risk entropy threshold is the critical value for judging "whether the system risk is acceptable", and its setting needs to consider both mathematical characteristics and actual safety needs:

[0248] The role of the threshold is clear: the threshold is the "trigger switch" of the brain-machine cooperation module - when the actual calculated risk entropy exceeds the threshold, it means that the risk uncertainty corresponding to the current control parameter is too high, and the override instruction needs to be started (such as switching to the historical safety parameter); if it does not exceed, the system can operate autonomously.

[0249] Quantitative setting based on safety standards:

[0250] The threshold is not set arbitrarily, but is determined in combination with safety standards such as safety specifications for the ship industry, historical accident data, and expert experience:

[0251] Reference industry specifications: such as the International Maritime Organization (IMO) safety operation standards for ship power systems and navigation systems, which clearly define the risk dispersion level that different systems can tolerate;

[0252] Historical accident analysis: statistics of the accident rate of similar ships in high-risk entropy scenarios in the past, set the threshold below the "critical point where the accident rate significantly increases";

[0253] Expert assessment: Ship engineers and safety experts adjust thresholds based on system importance (e.g., power system thresholds are stricter than auxiliary system thresholds) to ensure higher safety (lower thresholds) for core systems.

[0254] For example, assuming that a ship power system has four typical failure modes (N=4), the maximum risk entropy is .

[0255] If the safety standard requires that “failure risk must be centrally controlled,” experts may set the threshold to 1.2—when the entropy value is greater than 1.2 (the probability distribution is more dispersed), intervention is triggered;

[0256] If the system redundancy is high (such as dual engines), the threshold can be appropriately relaxed to 1.5, but still lower than the maximum value of 2 to reserve a safety margin.

[0257] The setting of the risk entropy threshold is a combination of mathematical properties (probability dispersion law) and actual security requirements (industry standards, expert experience):

[0258] An entropy value of 0 represents a “completely certain single risk”, and an entropy value of stands for "extreme uncertainty";

[0259] The threshold needs to fall between 0 and the maximum value, neither too strict (frequently triggering unnecessary intervention) nor too loose (ignoring high risks);

[0260] The ultimate goal is to ensure that the system operates within an "acceptable risk uncertainty range" through thresholds, initiate collaborative intervention when the range is exceeded, ensure the safety of the ship, and transform the "risk entropy judgment" from an abstract "high or low risk" to an executable "quantitative standard", thereby improving the objectivity and reliability of the brain-computer collaborative module.

[0261] The brain-computer collaboration module includes a fuse mechanism:

[0262] Used to monitor in real time whether the expert fatigue index is greater than the fuse threshold after the cover instruction is generated and before it is executed. If so, it will automatically switch to the AI ​​autonomous decision-making mode;

[0263] The fuse threshold is dynamically adjusted according to the sea condition level.

[0264] Specifically, the fuse mechanism is the "emergency switch" in the brain-computer collaboration module. It mainly targets the key node "after the coverage instruction is generated and before it is executed" - when it is found that the expert status is not suitable for decision-making, the manual intervention process is automatically terminated and switched to the AI ​​autonomous decision-making mode to prevent risks caused by human errors. Its workflow is as follows:

[0265] Real-time monitoring of expert fatigue index:

[0266] Expert Fatigue Index: Quantifies the expert's fatigue level through biosensors (such as eye tracking, brain wave monitoring, heart rate variability analysis, etc.) or behavioral data (such as operation response time, instruction modification frequency). Higher values ​​indicate more severe fatigue (e.g., 0-100 points, with 80 points or above indicating severe fatigue).

[0267] Monitoring timing: Start monitoring after the coverage instruction is generated and before it is executed. At this time, the expert may need to make final confirmation or adjustments to the instruction. If the expert is fatigued, he may misjudge the risk or make an erroneous operation.

[0268] Compare with the fuse threshold:

[0269] Circuit breaker threshold: A preset fatigue threshold (such as 70 points). When the expert fatigue index is greater than this threshold, the expert is judged to be unsuitable for decision-making and a circuit breaker is triggered.

[0270] Core logic: Fatigue can lead to inattention and decreased judgment accuracy, especially in high-risk ship scenarios (such as severe sea conditions). Expert fatigue may make it impossible to guarantee the safety of overriding instructions. Therefore, it is necessary to terminate manual intervention in a timely manner through threshold judgment.

[0271] Automatically switch to AI autonomous decision-making mode after triggering:

[0272] If the fuse condition is met (fatigue index > fuse threshold), the system immediately terminates the execution process of the original overriding instruction and automatically switches to "AI autonomous decision-making mode" - that is, the system regenerates control parameters based on data such as failure probability and simulation results to ensure that decisions are not interrupted and rely on verified algorithm logic.

[0273] The purpose of the switch is to replace human decision-making with the stable output of AI when it is unreliable, to avoid the expansion of risks caused by waiting or erroneous human instructions.

[0274] The fuse threshold is dynamically adjusted according to the sea condition level:

[0275] Sea state level: An indicator that measures the severity of the marine environment (e.g. calm sea is level 1, strong winds and huge waves is level 9). The higher the level, the greater the environmental risk faced by the ship and the higher the requirement for decision-making accuracy.

[0276] Dynamic adjustment logic:

[0277] Low sea state level (e.g., 1-3): The environment is stable, the decision-making margin for error is large, and the fuse threshold can be appropriately relaxed (e.g., from 70 to 80 points) to allow experts to participate in decision-making even when they are slightly fatigued.

[0278] High sea condition level (such as level 7-9): The environment is dangerous, and decisions must be absolutely reliable. The circuit breaker threshold is strictly tightened (for example, from 70 points to 60 points) - even if the expert is slightly fatigued, the circuit breaker is triggered, and the safety of decision-making is prioritized.

[0279] The essence of the adjustment is to dynamically balance "human experience" and "AI reliability" according to environmental risks, relying more on the stability of AI in high-risk scenarios and leaving more room for human intervention in low-risk scenarios.

[0280] Furthermore, the fuse mechanism is the "safety net design" of the brain-computer collaboration module. Through the logic of "monitoring expert fatigue → comparing dynamic thresholds → triggering AI switching", it solves the problem of "human intervention may fail due to poor status". By monitoring the expert's fatigue status, it prevents the execution of erroneous instructions due to fatigue, makes up for the uncertainty shortcomings of "people" in "brain-computer collaboration", and avoids the risk of human error. When manual intervention is not feasible, it automatically switches to AI mode to ensure the uninterrupted generation and execution of control parameters, especially in emergency scenarios to avoid decision-making vacuums and ensure decision continuity. The fuse threshold is adjusted with the sea condition level to make the mechanism more in line with the actual operation scenario, taking into account flexibility and safety. While retaining the value of expert experience, it uses intelligent fuse logic to ensure that the decision-making system can output safe and reliable control parameters under any circumstances, ultimately improving the operation safety of ships in complex environments.

[0281] The data acquisition module also includes a quantum enhanced fusion unit:

[0282] A quadratic unconstrained binary optimization model for constructing ship equipment topology constraints is used to represent the activation state of the collected heterogeneous data as binary variables;

[0283] The optimal space-time alignment path is quickly solved through quantum annealing parallel search capabilities.

[0284] Specifically, heterogeneous data from various ship systems (such as engine data, position data, and vibration data) comes from different sensors and exhibits spatiotemporal asynchrony (e.g., different sampling times and installation locations). Direct fusion can lead to errors. Quantum computing technology can quickly find the optimal spatiotemporal alignment path, ensuring consistency in both temporal and spatial dimensions for heterogeneous data, providing a high-quality data foundation for subsequent diagnosis and decision-making. The quantum-enhanced fusion unit, introduced at its core, leverages the characteristics and optimization capabilities of quantum computing to address the spatiotemporal alignment issues for heterogeneous ship data, improving the efficiency and accuracy of data fusion.

[0285] Furthermore, a quadratic unconstrained binary optimization model (QUBO model) is constructed:

[0286] Topological constraints of ship equipment: The physical connection relationship (topological structure) of ship equipment (such as engines, bearings, and sensors) constitutes a natural constraint. For example, "the engine is directly connected to the drive shaft, and the vibration data should be strongly correlated" and "there is a fixed spatial relationship between the navigation sensor and the hull position." These constraints must be reflected in data fusion to avoid alignment results that violate physical logic.

[0287] Binary representation of heterogeneous data activation state:

[0288] Whether the collected heterogeneous data is “effectively activated” (e.g., whether the sensor is working properly, whether the data is within a reasonable range) is represented as a binary variable (0 or 1):

[0289] The variable value is 1: it means that the data point is valid and can participate in the fusion;

[0290] The variable value is 0: It means that the data point is abnormal and needs to be excluded or corrected.

[0291] QUBO model construction: The quadratic unconstrained binary optimization model (QUBO) is a commonly used optimization model in quantum computing. Its objective function is a quadratic polynomial and its variables are binary numbers. Here:

[0292] Objective function: Comprehensively considers "minimization of spatiotemporal alignment error" (such as the time difference and spatial distance between different sensor data) and "topological constraint satisfaction" (such as compliance with the physical connection relationship of devices);

[0293] Constraint conversion: converting the topological constraints of ship equipment (e.g., "data of two devices must be synchronized") into penalty terms in the objective function (the objective function value increases when the constraint is violated);

[0294] The final model form is: ;

[0295] in, is a binary variable, 、 is the coefficient (reflecting data weight and constraint strength);

[0296] Furthermore, quantum annealing is used to search for the optimal space-time alignment path in parallel:

[0297] Quantum annealing technology: It is an optimization algorithm that uses the quantum tunneling effect to find the global optimal solution. It is good at handling complex combinatorial optimization problems (such as path planning and scheduling optimization). Compared with traditional optimization algorithms, its core advantage is the parallel search capability - it can explore multiple possible solution spaces simultaneously to avoid being trapped in the local optimum.

[0298] The "spatiotemporal alignment path" refers to a scheme for calibrating heterogeneous data from different sensors on the time axis (e.g., unified sampling time) and the space axis (e.g., the same position mapped to the ship coordinate system). The optimal path needs to meet the following conditions:

[0299] In time: The sampling time of different system data is as synchronized as possible (e.g., the time difference between engine temperature and vibration data is minimized);

[0300] In space: It conforms to the device topology relationship (e.g., the data correlation of devices close in distance is high);

[0301] Data validity: Preferentially retain data points with active state 1 (valid).

[0302] The solution process of quantum annealing:

[0303] Map the QUBO model to the physical system of the quantum annealer, control the evolution of the quantum bits, and gradually cool the system from a high-energy state to a low-energy state. The final low-energy state corresponds to the minimum value of the objective function - the optimal spatiotemporal alignment path. Due to quantum parallelism, this process can be completed in a very short time, and compared with traditional algorithms, the efficiency is significantly improved, especially when the data volume is large and the constraints are complex.

[0304] The quantum-enhanced fusion unit combines quantum computing technology with ship data characteristics through the process of "topology constraint QUBO modeling → binary data representation → quantum annealing optimization", solves the problem of spatiotemporal alignment of heterogeneous data, and provides high-quality data for the entire system, which is "spatiotemporally consistent, physically reliable, and real-time available". It is an important guarantee for the accuracy and efficiency of subsequent fault diagnosis and decision optimization. By incorporating the QUBO model into the topology constraints of ship equipment, it ensures that the data alignment results conform to the physical logic and avoid fusion errors caused by spatiotemporal misalignment (e.g., incorrectly associating vibration data from different positions). It improves the accuracy of data fusion. The parallel search capability of quantum annealing solves the "computational bottleneck" of traditional algorithms in the context of multiple sensors and large data volumes, quickly finds the optimal alignment path, and meets the real-time requirements (e.g., real-time data processing during ship navigation). It accelerates the solution of complex optimization problems. By representing the data activation state with binary variables, abnormal data is automatically filtered during the alignment process, improving the data quality input to subsequent modules (e.g., fault diagnosis, rule retrieval), and enhancing data validity filtering.

[0305] Example Two

[0306] Example in the context of VLCC turbocharger surge prevention and control

[0307] Taking the turbocharger surge prevention and control under the Indian Ocean 6-level wave (wave direction θ = 120°) as an example, the system operation process is as follows:

[0308] The data acquisition module obtains the data of the previous cycle: supercharger vibration frequency 850Hz, pressure fluctuation ±6%, wave direction θ=120°;

[0309] The rule retrieval module matches the "high wave direction + high frequency vibration + pressure fluctuation" rule to generate a surge diagnosis logic tree;

[0310] The fault diagnosis module integrates the vibration coupling features extracted by the spatiotemporal graph convolutional network with logic tree reasoning and outputs a surge probability of 0.88;

[0311] The control decision module generates parameters: the speed is reduced to 75%, and the anti-surge valve is opened to 30%;

[0312] The simulation calculation module predicts: the expected vibration frequency is 620Hz, and the pressure fluctuation is ±3%;

[0313] The feedback optimization module compared the actual data (vibration frequency 700Hz), found a deviation of 80Hz (out of tolerance range 50Hz), adjusted the speed to 72%, and re-simulated.

[0314] The final actual vibration frequency is 610 Hz (deviation 10 Hz), the parameters take effect, and the rule base is updated: "When θ = 120°, the surge probability is > 0.8, and the optimal speed is 72%."

[0315] The technical scope of the present invention is not limited to the contents of the above description. Those skilled in the art can make various deformations and modifications to the above embodiments without departing from the technical idea of ​​the present invention, and these deformations and modifications should all fall within the protection scope of the present invention.

Claims

1. Intelligent fault diagnosis and active control system based on multi-source heterogeneous data fusion of ships, characterized by: include: A data acquisition module is used to obtain heterogeneous data collected by the power system, navigation system, and monitoring system in the current cycle and the previous cycle, wherein the heterogeneous data includes engine data, position data, and vibration data; The rule retrieval module is used to retrieve the corresponding expert rules from the expert rule library based on the heterogeneous data of the previous cycle and form a logic tree; The fault diagnosis module is used to calculate the fault probability cloud using the spatiotemporal graph convolutional network and logic tree for the heterogeneous data of the previous cycle, and to parse out the probability of failure of each system; A control decision module is used to generate control parameters for each system in the current cycle based on the probability of failure of each system; Simulation calculation module: used to input control parameters into the pre-built twin digital simulation model to simulate and calculate the expected values ​​of the heterogeneous data of various ship systems in the current cycle; Feedback optimization module: Used to calculate the deviation between the expected value and the heterogeneous data of the current period and determine whether the deviation exceeds the tolerance range; If yes, the control parameters are updated based on the deviation value and fed back to the output simulation calculation module to calculate the expected value until the deviation value is within the tolerance range. The last updated control parameters are input into the expert rule library to realize real-time dynamic autonomous update of the logic tree.

2. The intelligent fault diagnosis and active control system for ship multi-source heterogeneous data fusion according to claim 1 is characterized by: The logic tree is a differentiable function that encodes the expert rule base: ； in, : sigmoid function; : The confidence level of proposition A / B; : rule weight; : Bias term.

3. The intelligent fault diagnosis and active control system for ship multi-source heterogeneous data fusion according to claim 2 is characterized by: The optimization process of the feedback optimization module is as follows: Encode the control parameters of the previous cycle and the control parameters of the current cycle into quantum states and construct quantum transformations that depend on rule weights; The quantum gradient value is calculated based on the quantum state of the control parameters of the previous cycle and the control parameters of the current cycle, and the rule weight is adjusted using quantum gradient descent.

4. The intelligent fault diagnosis and active control system for ship multi-source heterogeneous data fusion according to claim 3 is characterized by: The spatiotemporal graph convolutional network includes an edge weight model for host nodes and bearing nodes designed based on the coupling characteristics of physical distance and force transmission efficiency, and introduces a time decay factor. When the edge weight exceeds a preset threshold, a hidden fault path is predicted, and a new virtual node is added to represent the potential fault point. The edge weight is dynamically updated based on the hidden fault path by fusing the time decay effect and the fault probability.

5. The intelligent fault diagnosis and active control system for ship multi-source heterogeneous data fusion according to claim 1 is characterized by: The pre-built twin digital simulation model is a multi-physics field coupling model based on fluid-structure coupling and heat-force coupling, and is solved in real time through a ship-specific reduced-order model to obtain the expected values ​​of the heterogeneous data of various ship systems in the current period.

6. The intelligent fault diagnosis and active control system for ship multi-source heterogeneous data fusion according to claim 5 is characterized by: Also includes brain-computer collaboration module: Calculating the risk entropy of the control parameters of the current cycle based on the pre-built twin digital simulation model and in combination with the fault probability cloud, and determining whether the entropy value exceeds a threshold; If yes, the control parameters of the previous cycle are parsed through the security verification mechanism and overwriting instructions are generated.

7. The intelligent fault diagnosis and active control system for ship multi-source heterogeneous data fusion according to claim 6 is characterized by: The calculation method of the risk entropy is: Extract typical failure modes throughout the ship's life cycle from the heterogeneous database of the previous cycle and construct a failure mode probability distribution model: The risk entropy is calculated based on the failure mode probability distribution model.

8. The intelligent fault diagnosis and active control system for ship multi-source heterogeneous data fusion according to claim 6 is characterized by: The risk entropy threshold setting method is: The risk entropy value is 0: the probability of only one failure mode is 1; The risk entropy value increases: the probability distribution becomes more dispersed; The maximum value of risk entropy is ; Set risk entropy thresholds in combination with safety standards.

9. The intelligent fault diagnosis and active control system for ship multi-source heterogeneous data fusion according to claim 6 is characterized by: The brain-computer collaboration module includes a fuse mechanism: Used to monitor in real time whether the expert fatigue index is greater than the fuse threshold after the cover instruction is generated and before it is executed. If so, it will automatically switch to the AI ​​autonomous decision-making mode; The fuse threshold is dynamically adjusted according to the sea condition level.

10. The intelligent fault diagnosis and active control system for ship multi-source heterogeneous data fusion according to claim 1 is characterized by: The data acquisition module also includes a quantum enhanced fusion unit: A quadratic unconstrained binary optimization model for constructing ship equipment topology constraints is used to represent the activation state of the collected heterogeneous data as binary variables; The optimal space-time alignment path is quickly solved through quantum annealing parallel search capabilities.

Images