A ship main engine adaptive monitoring and early warning system

Abstract

This invention discloses an adaptive monitoring and early warning system for ship main engines, comprising a multi-source heterogeneous data acquisition and cleaning module, a dynamic operating condition self-identification module, an adaptive condition monitoring model library, an operating condition matching and model scheduling unit, a real-time monitoring and anomaly triggering unit, a multi-dimensional contribution tracing and decision support module, and a model adaptive update engine. The system performs anomaly detection by calling PCA-based sub-monitoring models to calculate and statistically analyze data through data cleaning, online operating condition identification and matching scheduling. When an anomaly is triggered, the system locates the suspected root cause by calculating the contribution of variables and generates maintenance suggestions based on a fault case library. If no anomaly is triggered, the system uses health data to continuously update model parameters and control limits through a sliding window mechanism. This invention achieves dynamic adaptive tracking of main engine operating conditions, self-learning evolution of monitoring models, and a closed loop from anomaly detection to intelligent decision support, improving the accuracy and reliability of ship main engine condition monitoring.

Description

Technical Field

[0001] This invention relates to the field of marine main engine technology, and more specifically to a marine main engine adaptive monitoring and early warning system. Background Technology

[0002] As the core power plant of a ship, the stability and reliability of its main engine directly affect the ship's navigation safety, energy efficiency, and operational economy. Traditional ship engine room monitoring systems mainly rely on setting independent, fixed alarm thresholds on the central control console. When a single parameter such as temperature or pressure detected by the sensor exceeds the preset threshold, the system will trigger an alarm.

[0003] However, this traditional monitoring method has the following significant drawbacks: First, alarm thresholds are usually set based on sea trial data or human experience, resulting in static values. This method ignores dynamic factors such as different sea states, different loading conditions, and the degradation of the main engine's performance over time, leading to poor threshold adaptability and a high likelihood of missed or false alarms. Second, traditional systems process multiple monitoring parameters in isolation, lacking systematic correlation analysis. A fault is often reflected by a combination of anomalies in multiple parameters, and threshold alarms for a single parameter are insufficient to capture this complex coupling relationship, failing to provide effective early warning before a fault occurs. Furthermore, when an alarm occurs, it usually means that a fault has already formed, and engine room personnel find it difficult to quickly and accurately analyze the root cause of the fault from numerous independent alarm points, especially when facing common alarms, where cause localization is particularly difficult. In addition, with the development of ship intelligence, there are numerous monitoring points in the engine room, generating massive amounts of data, but traditional systems lack effective data mining and utilization methods, resulting in a significant waste of data resources.

[0004] In recent years, some studies have attempted to apply data-driven methods to equipment fault diagnosis, such as building condition monitoring models based on principal component analysis (PCA) and... and Anomaly detection is performed using statistical methods, supplemented by contribution plots for source tracing analysis. However, most of these methods remain at the theoretical verification stage, facing the following challenges in engineering applications: First, the monitoring model is static and lacks adaptive updating capabilities. Existing technologies typically build PCA models offline based on a period of historical health data, and once deployed, they are no longer updated. However, in actual operation, ship main engines experience slow performance baseline drift due to factors such as component wear, marine organism attachment, and changes in fuel quality. Static models cannot adapt to this time-varying characteristic; over time, their monitoring accuracy gradually decreases, and the false alarm and false negative rates increase accordingly. Second, the operating condition classification is offline and cannot dynamically respond to changes in operating status. Existing methods typically collect historical data containing multiple operating conditions in advance, use clustering algorithms to offline classify several typical operating conditions, and build monitoring models for each condition. However, ships face complex and variable sea conditions during actual navigation, the main engine's operating conditions may change in real time, and new operating condition patterns may emerge as the ship ages. Offline, fixed-condition segmentation cannot handle such dynamic changes, leading to a mismatch between the model and the current operating conditions, significantly reducing monitoring effectiveness. Third, source tracing analysis remains at the chart level, lacking intelligent decision support. Existing technologies, after detecting anomalies, display the contribution rate of each variable to the statistical excess through contribution graphs, but only go as far as visualization. The contribution rate results require human experience for interpretation and analysis, failing to bridge the gap from display to decision-making. The lack of association and similarity matching mechanisms with fault case knowledge bases prevents the automatic generation of structured maintenance suggestions, limiting the ability to assist maintenance personnel in decision-making. Fourth, data preprocessing methods are simplistic and struggle to effectively handle multi-source heterogeneous data. Ship monitoring systems contain various sensors with varying sampling frequencies, resulting in noise and missing values ​​in the data. Existing preprocessing methods often employ single interpolation or filtering techniques, failing to combine density-based and time-series proximity-based dual detection mechanisms. This makes it difficult to effectively handle the identification and repair of abnormal data under complex operating conditions, affecting the accuracy of subsequent modeling and monitoring.

[0005] Therefore, there is an urgent need for a ship main engine monitoring and early warning system that can make full use of massive historical data of ships to achieve dynamic monitoring of main engine status, online adaptive identification of operating conditions, self-learning and updating of models, and intelligent traceability and decision-making capabilities. Summary of the Invention

[0006] To comprehensively improve the monitoring and early warning capabilities of ship main engines, this invention provides an adaptive monitoring and early warning system for ship main engines.

[0007] The first technical solution adopted in this invention is as follows: A ship main engine adaptive monitoring and early warning system, comprising: a multi-source heterogeneous data acquisition and cleaning module, used to acquire multi-source monitoring data of the ship main engine in real time at a sampling frequency matching the sensor type, and to perform real-time noise removal and missing value imputation on the multi-source monitoring data by combining a density-based local outlier detection algorithm and a time series proximity-based interpolation algorithm, generating a standardized real-time data stream; and a dynamic operating condition self-identification module, which is data-connected to the multi-source heterogeneous data acquisition and cleaning module, used to receive the real-time data stream, and to perform online clustering analysis on the main engine speed and fuel rack scale based on a Gaussian mixture model within a sliding time window, automatically identifying... It outputs the dynamic operating condition label of the host at the current moment; an adaptive state monitoring model library stores multiple sub-monitoring models based on principal component analysis, each sub-monitoring model corresponding to an operating condition defined by the dynamic operating condition self-identification module; an operating condition matching and model scheduling unit, connected to the dynamic operating condition self-identification module and the adaptive state monitoring model library, is used to retrieve the corresponding sub-monitoring model from the adaptive state monitoring model library as the current health status benchmark model based on the received current operating condition label; a real-time monitoring and anomaly triggering unit, connected to the operating condition matching and model scheduling unit, is used to substitute the current standardized real-time data stream into the current health status benchmark model and calculate its principal component space statistics. and residual space statistics and judge and Whether the current health status benchmark model exceeds the dynamic control limit; if it does, an anomaly trigger signal containing a timestamp, operating condition label, and out-of-limit statistics is generated; the multi-dimensional contribution tracing and decision support module, connected to the real-time monitoring and anomaly triggering unit, is used to calculate the impact of each original monitoring variable on the response to the anomaly trigger signal. and The contribution of statistical values ​​exceeding the limit is used to identify combinations of variables whose contribution exceeds a preset threshold as a set of suspected root causes, and to generate structured alarm and maintenance suggestions by combining them with a fault case knowledge base; and a model adaptive update engine is connected to the adaptive state monitoring model library and the real-time monitoring and anomaly triggering unit, which is used to feed back the current health data according to the working condition label to the corresponding sub-monitoring model in the adaptive state monitoring model library when no anomaly triggering signal is generated, and to perform rolling updates of model parameters and control limits based on a sliding window mechanism.

[0008] Preferably, the multi-source monitoring data includes key operating parameters required to construct the sub-monitoring model, including at least: main engine fuel inlet pressure, main engine lubricating oil inlet pressure, main engine lubricating oil inlet temperature, turbocharger speed, turbocharger lubricating oil outlet temperature, main engine speed, fuel rack scale, cylinder liner cooling water inlet pressure, cylinder liner cooling water outlet pressure, cylinder liner cooling water outlet temperature, average exhaust temperature of each cylinder, intercooler fresh water inlet pressure, intercooler fresh water outlet pressure, and intercooler low-temperature fresh water outlet temperature.

[0009] Preferably, the dynamic operating condition self-identification module further includes an online clustering validity evaluation unit, used to calculate the pseudo-value of the current clustering result in real time. Statistics and Bayesian information criteria; The pseudo The formula for calculating the statistic is:

[0010] in, The sum of squares of deviations between classes. For the sum of squared deviations within the class, This represents the current number of clusters. This represents the total number of samples within the sliding window; The formula for calculating the Bayesian information criterion is as follows:

[0011] in, This is the maximum likelihood estimate of the current Gaussian mixture model. The total number of model parameters; when the pseudo-parameters are used in multiple consecutive calculation cycles. The statistics show a downward trend and When the value does not converge to the minimum, the clustering of the Gaussian mixture model is triggered. Optimization and adjustments.

[0012] Preferably, the real-time monitoring and anomaly triggering unit is also used to calculate a comprehensive health index, the calculation formula of which is:

[0013] in, and Under the current operating conditions and Statistic in confidence level The control limit below; the generation condition of the abnormal trigger signal is: the... , and At least two of the three indicators exceed their respective control limits simultaneously.

[0014] Preferably, in the multidimensional contribution tracing and decision support module, the formula for calculating the contribution of variables includes: right Contribution of the statistic:

[0015] right Contribution of the statistic:

[0016] in, For the current sample vector, This is a diagonal matrix related to the principal components. For the first A unit vector of variables, , For the load matrix, It is an identity matrix.

[0017] Preferably, the multidimensional contribution tracing and decision support module performs similarity matching on identified suspected fault patterns by querying a fault case knowledge base to provide decision support. The similarity matching formula is as follows:

[0018] in, For the current root dependent variable set, the first The contribution of each standardized variable For the historical cases in the knowledge base The standardized contribution of each variable. The total number of variables; select the case with the highest similarity and output its processing solution.

[0019] Preferably, the model adaptive update engine uses a length of A sliding time window is used, within which data must belong to the same operating condition and be determined to be in a healthy state; model updates include: recalculating the covariance matrix, eigenvalues, and eigenvectors based on the window data to update the load matrix. And recalculate the control limits; Control limits of statistics The updated formula is:

[0020] in, Number of principal elements The total number of healthy samples used for modeling. For degrees of freedom of Distribution at confidence level The critical value below; Control of statistics The updated formula is:

[0021] in, The covariance matrix is ​​the first 1 eigenvalue, For the total number of variables, For a standard normal distribution at a confidence level The quantiles below.

[0022] Preferably, the length of the sliding time window Dynamically adjust based on host operational stability: calculate the variance of core state parameters within the window. If continuous Each window Less than the stability threshold Then increase If continuous Each window Greater than the fluctuation threshold If , then L decreases.

[0023] Preferably, it also includes a human-computer interaction and augmented reality display module, which is connected to the multi-dimensional contribution tracing and decision support module. It is used to highlight the physical device or sensor location corresponding to the suspected root cause variable set in the three-dimensional digital twin cabin model, and overlay the corresponding real-time data, historical trends and maintenance suggestion information in an augmented reality manner.

[0024] The second technical solution adopted in this invention is as follows: A monitoring and early warning method based on the aforementioned ship main engine adaptive monitoring and early warning system, comprising the following steps: Step S1. Through the multi-source heterogeneous data acquisition and cleaning module, multi-source monitoring data of the ship's main engine is acquired in real time, and noise and missing values ​​are processed to output a standardized real-time data stream; Step S2. Using the dynamic operating condition self-identification module, based on the speed and load parameters in the real-time data stream, the current operating condition of the host is identified through online clustering; Step S3. Through the working condition matching and model scheduling unit, schedule the matching benchmark monitoring model from the adaptive state monitoring model library according to the current working condition; Step S4. Through the real-time monitoring and anomaly triggering unit, the real-time data stream is calculated using the benchmark monitoring model to obtain a health index, which is then compared with the dynamic control limit to determine whether an anomaly has been triggered. Step S5. If an anomaly is triggered, the key variables causing the anomaly will be analyzed through the multidimensional contribution tracing and decision support module, and a suspected fault diagnosis report and maintenance suggestions will be generated. Step S6. If no anomaly is triggered, the current health data is used to update the parameters and control limits of the benchmark monitoring model through the model adaptive update engine. Step S7. Visualize the monitoring status, abnormal alarms, and maintenance guidance through the human-computer interaction and augmented reality display module.

[0025] The present invention has the following beneficial effects: 1. Accurate and reliable monitoring: Through a dual data cleaning mechanism based on density and time series proximity, noise is effectively removed and missing values ​​are repaired, providing a high-quality standardized data stream for subsequent modeling and significantly improving the input data quality of the monitoring model; 2. Dynamic operating condition adaptive identification: The Gaussian mixture model based on a sliding window is used to perform online clustering analysis on the host operating conditions. It can identify changes in operating status in real time and output operating condition labels, which solves the technical problem that traditional offline operating condition classification cannot adapt to dynamic changes and ensures that the monitoring model always matches the current operating conditions. 3. Multi-model adaptive monitoring system: PCA-based sub-monitoring models are constructed for different operating conditions to form a condition-model correspondence library. Accurate adaptive monitoring of operating conditions is achieved through condition matching and model scheduling units, which improves the accuracy and adaptability of condition monitoring under complex operating conditions. 4. Model self-learning update: An adaptive update engine is introduced. When no anomalies are detected, the current health data is fed back to the corresponding sub-monitoring model. The model parameters and control limits are updated on a rolling basis based on the sliding window mechanism, so that the monitoring model can adaptively adjust with the slow degradation of host performance, maintain high-precision monitoring in the long term, and avoid false alarms and missed alarms. 5. Multi-dimensional intelligent source tracing decision support: In response to anomaly trigger signals, the root cause variable set that causes the anomaly is accurately identified through the contribution calculation formula, and similarity matching is performed in combination with the fault case knowledge base to automatically generate structured alarm and maintenance suggestion information, realizing a closed loop from anomaly detection to intelligent decision-making, which greatly improves the efficiency of fault diagnosis. 6. Augmented Reality Visualization Interaction: Through human-computer interaction and augmented reality display modules, the physical equipment locations corresponding to abnormal variables are highlighted in the 3D digital twin cabin model, and real-time data, historical trends and maintenance suggestions are overlaid in an augmented reality manner to provide operation and maintenance personnel with intuitive and accurate on-site guidance; 7. Long-term operational stability: Through the adaptive update mechanism, the system's monitoring accuracy remains stable during long-term operation, with virtually no decline in the early stages and only a slight and steady decrease in the middle and later stages. This is significantly better than the continuous decline in accuracy of traditional static models, greatly extending the effective lifespan of the system. Attached Figure Description

[0026] Figure 1 This is the overall architecture diagram of the ship main engine adaptive monitoring and early warning system in this embodiment of the invention.

[0027] Figure 2 This is a flowchart of multi-source heterogeneous data acquisition and preprocessing in an embodiment of the present invention.

[0028] Figure 3 This is a schematic diagram of online identification of dynamic working conditions and model scheduling in an embodiment of the present invention.

[0029] Figure 4 This is a schematic diagram illustrating the construction principle of the PCA model in this embodiment of the invention.

[0030] Figure 5 This is a flowchart of real-time monitoring and statistical judgment in an embodiment of the present invention.

[0031] Figure 6 This is a schematic diagram of the sliding window mechanism in an embodiment of the present invention.

[0032] Figure 7 This is a flowchart of the model parameter update process in an embodiment of the present invention.

[0033] Figure 8 This is a flowchart of multidimensional contribution tracing and decision support in an embodiment of the present invention.

[0034] Figure 9 This is a comparison chart of the monitoring effects of the embodiments of the present invention and the prior art.

[0035] 100 - Multi-source heterogeneous data acquisition and cleaning module; 200-Dynamic Operating Condition Self-Identification Module; 300-Adaptive Condition Monitoring Model Library; 400-Working Condition Matching and Model Scheduling Unit; 500 - Real-time monitoring and anomaly triggering unit; 600-Multidimensional Contribution Tracing and Decision Support Module; 700-Model Adaptive Update Engine; 800 - Human-computer interaction and augmented reality display module. Detailed Implementation

[0036] The present invention will be further described below with reference to the embodiments and accompanying drawings.

[0037] like Figure 1As shown in the figure, an embodiment provides a ship main engine adaptive monitoring and early warning system. This system includes: a multi-source heterogeneous data acquisition and cleaning module 100, a dynamic operating condition self-identification module 200, an adaptive state monitoring model library 300, an operating condition matching and model scheduling unit 400, a real-time monitoring and anomaly triggering unit 500, a multi-dimensional contribution tracing and decision support module 600, a model adaptive update engine 700, and a human-computer interaction and augmented reality display module 800. These are described in detail below.

[0038] Figure 1 As shown, the multi-source heterogeneous data acquisition and cleaning module 100 is used to acquire multi-source monitoring data of the ship's main engine in real time at a sampling frequency matched to the sensor type. It employs a combination of a density-based local outlier detection algorithm and a time-series proximity-based interpolation algorithm to perform real-time noise removal and missing value imputation on the multi-source monitoring data, generating a standardized real-time data stream. This module, through its dual data cleaning mechanism, effectively solves the problems of inconsistent sampling frequencies, data noise, and missing values ​​in multi-source data, providing a high-quality data foundation for subsequent monitoring and significantly improving the input data quality of the monitoring model.

[0039] like Figure 1 As shown, the dynamic operating condition self-identification module 200 is data-connected to the multi-source heterogeneous data acquisition and cleaning module 100. It receives real-time data streams and performs online clustering analysis on the main engine's speed and fuel rack scale based on a Gaussian mixture model within a sliding time window. It automatically identifies and outputs the current dynamic operating condition label of the main engine. This module enables real-time perception of the main engine's operating status, dynamically responding to changes in sea state and operational adjustments, ensuring that the monitoring model always matches the current operating condition.

[0040] like Figure 1 As shown, the adaptive condition monitoring model library 300 stores multiple sub-monitoring models based on principal component analysis. Each sub-monitoring model corresponds to an operating condition defined by the dynamic operating condition self-identification module 200. This model library constructs an operating condition-model correspondence system, providing differentiated health benchmarks for different operating conditions.

[0041] like Figure 1 As shown, the operating condition matching and model scheduling unit 400 is connected to the dynamic operating condition self-identification module 200 and the adaptive state monitoring model library 300. It is used to retrieve the corresponding sub-monitoring model from the adaptive state monitoring model library 300 as the current health status benchmark model based on the received current operating condition label. This unit achieves accurate matching between the model and the operating condition, ensuring the targeted and accurate nature of the monitoring.

[0042] like Figure 1 and Figure 5As shown, the real-time monitoring and anomaly triggering unit 500 is connected to the operating condition matching and model scheduling unit 400, and is used to substitute the current standardized real-time data stream into the current health status benchmark model to calculate its principal component space statistics. and residual space statistics and judge and Whether the dynamic control limits corresponding to the current health status baseline model are exceeded; if exceeded, an anomaly trigger signal containing a timestamp, operating condition label, and over-limit statistics is generated. This unit performs anomaly detection based on PCA statistics, which can capture changes in multiple parameters and achieve early warning.

[0043] like Figure 1 and Figure 8 As shown, the multidimensional contribution tracing and decision support module 600 is connected to the real-time monitoring and anomaly triggering unit 500, and is used to calculate the contribution of each original monitoring variable in response to an anomaly triggering signal. and The module identifies variable combinations with contribution values ​​exceeding preset thresholds as potential root cause variables and generates structured alarm and maintenance suggestions based on a fault case knowledge base. This module achieves a leap from anomaly detection to intelligent decision-making, transforming data-driven analysis into actionable maintenance guidance and significantly improving fault diagnosis efficiency.

[0044] like Figure 1 , Figure 6 and Figure 7 As shown, the adaptive update engine 700 is connected to the adaptive condition monitoring model library 300 and the real-time monitoring and anomaly triggering unit 500. When no anomaly trigger signal is generated, it feeds back the current health data, categorized by operating condition label, to the corresponding sub-monitoring model in the adaptive condition monitoring model library 300, and performs rolling updates of model parameters and control limits based on a sliding window mechanism. This engine endows the system with self-learning capabilities, enabling the monitoring model to adaptively adjust as host performance degrades, maintaining high-precision monitoring over the long term and avoiding false alarms and missed alarms.

[0045] like Figure 2 As shown, the multi-source heterogeneous data acquisition and cleaning module 100 first collects raw data from data sources such as ship monitoring systems, AIS, depth sounders, and weather instruments at their respective sampling frequencies. Then, after data fusion and alignment, density-based local outlier detection, and time-series proximity-based interpolation algorithms, it outputs a standardized real-time data stream. This process ensures the effective fusion and quality improvement of multi-source heterogeneous data.

[0046] like Figure 3As shown, the dynamic operating condition self-identification module 200 performs online clustering of the main engine's speed and fuel rack scale based on a Gaussian mixture model within a sliding window, and outputs the current operating condition label in real time. The online clustering validity evaluation unit 210 calculates the pseudo-clustering value in real time. Statistics and Bayesian information criteria, when pseudo The statistics show a downward trend and If the model fails to converge, cluster number optimization is triggered. This mechanism ensures the accuracy of the workload classification and avoids model mismatch problems caused by an unreasonable number of clusters.

[0047] like Figure 4 As shown, for each operating condition, a PCA model is constructed based on historical health data: first, the data is standardized; then, the covariance matrix is ​​calculated, eigenvalue decomposition is performed to obtain eigenvalues ​​and the load matrix; finally, the number of principal components is determined based on a cumulative variance contribution rate ≥ 95%. This process establishes an accurate baseline health status model for each operating condition.

[0048] like Figure 5 As shown, the real-time monitoring and anomaly triggering unit 500 substitutes the real-time data stream into the current sub-monitoring model, projects it onto the principal component space and residual space, and calculates... and Statistics and comparisons with dynamic control limits. Comprehensive Health Index Fusion and Information, when , and The system triggers at least two abnormal signals when the time limit is exceeded, which improves the robustness of the judgment and effectively reduces the risk of false alarms or missed alarms for a single indicator.

[0049] like Figure 6 As shown, the model adaptive update engine 700 uses a length of The sliding time window selects only stable operating data within the window that is deemed healthy and belongs to the same operating condition for model updates. Window length It can be dynamically adjusted based on operational stability, balancing timely updates and data sufficiency.

[0050] like Figure 7 As shown, the covariance matrix, eigenvalues, and principal components are recalculated based on the health data within the window to obtain a new loading matrix. Update according to the formula and The control limits are set. The updated model parameters are stored in the adaptive state monitoring model library 300 to achieve adaptive evolution of the monitoring model.

[0051] like Figure 8As shown, when an anomaly is triggered, the multidimensional contribution tracing and decision support module 600 calculates the pairs of each variable. and The contribution of variables exceeding a threshold is used to group them into a set of suspected root cause variables. Then, by matching the fault case knowledge base with cosine similarity, the historical cases with the highest similarity are selected, and structured maintenance suggestions are output. This process transforms historical experience into decision support, achieving intelligent anomaly diagnosis.

[0052] like Figure 1 As shown, the human-computer interaction and augmented reality display module 800 is connected to the multi-dimensional contribution tracing and decision support module 600. This module highlights the locations of physical devices or sensors corresponding to suspected root cause variables in the 3D digital twin cabin model, and overlays corresponding real-time data, historical trends, and maintenance suggestions in an augmented reality manner. This module provides maintenance personnel with intuitive and accurate on-site guidance, significantly improving the human-computer interaction experience and emergency response efficiency.

[0053] like Figure 2 As shown, the multi-source monitoring data includes key operating parameters required to construct the sub-monitoring model, including at least: main engine fuel inlet pressure, main engine lubricating oil inlet pressure, main engine lubricating oil inlet temperature, turbocharger speed, turbocharger lubricating oil outlet temperature, main engine speed, fuel rack scale, cylinder liner coolant inlet pressure, cylinder liner coolant outlet pressure, cylinder liner coolant outlet temperature, average exhaust temperature of each cylinder, intercooler freshwater inlet pressure, intercooler freshwater outlet pressure, and intercooler low-temperature freshwater outlet temperature. By collecting at least 14 key operating parameters, a comprehensive monitoring indicator system reflecting the operating status of the main engine's major systems was constructed, providing rich feature information for the PCA model, ensuring that the monitoring model can capture multi-dimensional anomalies, and improving the comprehensiveness and accuracy of condition monitoring.

[0054] like Figure 1 , Figure 3 As shown, the dynamic operating condition self-identification module 200 performs online clustering of the main engine's speed and fuel rack scale based on a Gaussian mixture model within a sliding window, and outputs the current operating condition label in real time. It also includes an online clustering validity evaluation unit, used to calculate the pseudo-value of the current clustering result in real time. Statistics and Bayesian information criteria.

[0055] fake The formula for calculating the statistic is: ;in, The sum of squares of deviations between classes. For the sum of squared deviations within the class, This represents the current number of clusters. This represents the total number of samples within the sliding window.

[0056] The formula for calculating the Bayesian information criterion is: ;in, This is the maximum likelihood estimate of the current Gaussian mixture model. This represents the total number of model parameters.

[0057] When pseudo-multiple consecutive calculation cycles The statistics show a downward trend and When the value does not converge to the minimum, the clustering of the Gaussian mixture model is triggered. Optimization and adjustment. The pseudo-clustering validity is calculated in real time using an online clustering validity evaluation unit. The use of statistics and Bayesian information criteria automatically optimizes the number of clusters when cluster quality declines. This mechanism ensures the accuracy of the working condition division, avoids model mismatch problems caused by an unreasonable number of clusters, and improves the system's adaptability to new working conditions.

[0058] like Figure 1 , Figure 5 As shown, the real-time monitoring and anomaly triggering unit 500 is also used to calculate a comprehensive health index, the formula of which is: ;in, and Under the current operating conditions and Statistic in confidence level The control limits below.

[0059] The conditions for generating an abnormal trigger signal are: , and At least two of the three indicators simultaneously exceed their respective control limits. This is determined by calculating the comprehensive health index. It adopts a "choose two out of three" multi-indicator cross-judgment logic, which effectively reduces the risk of false alarms or missed alarms of a single indicator, enhances the robustness of anomaly detection, and is especially suitable for monitoring complex nonlinear systems such as ship main engines.

[0060] like Figure 1 , Figure 8 As shown, in the multidimensional contribution tracing and decision support module 600, the formula for calculating the contribution of variables includes: right Contribution of the statistic: ; right Contribution of the statistic: ; in, For the current sample vector, This is a diagonal matrix related to the principal components. For the first A unit vector of variables, , For the load matrix, It is an identity matrix.

[0061] Quantify each variable's contribution using a precise contribution calculation formula. and The extent of the impact of exceeding the statistical limit provides reliable data support for subsequent fault location, ensuring the accuracy and interpretability of anomaly tracing.

[0062] like Figure 1 , Figure 8 As shown, the multidimensional contribution tracing and decision support module 600 performs similarity matching on identified suspected fault patterns by querying the fault case knowledge base to provide decision support. The similarity matching formula is as follows: ;in, For the current root dependent variable set, the first The contribution of each standardized variable For the historical cases in the knowledge base The standardized contribution of each variable. The total number of variables is represented. The case with the highest similarity is selected and its processing solution is output. By matching the fault case knowledge base with cosine similarity, historical experience is transformed into intelligent decision support, realizing a closed loop from anomaly detection to structured maintenance suggestions, which greatly improves the efficiency of fault diagnosis and the level of intelligent decision-making.

[0063] like Figure 1 , Figures 6-7 As shown, the model adaptive update engine 700 uses a length of A sliding time window is used, within which data must belong to the same operating condition and be determined to be in a healthy state; model updates include: recalculating the covariance matrix, eigenvalues, and eigenvectors based on the window data to update the load matrix. And recalculate the control limits; Control limits of statistics The updated formula is: ;in, Number of principal elements The total number of healthy samples used for modeling. For degrees of freedom of Distribution at confidence level The critical value below.

[0064] Control of statistics The updated formula is:

[0065] in, The covariance matrix is ​​the first 1 eigenvalue, For the total number of variables, For a standard normal distribution at a confidence level The quantiles below.

[0066] By using a sliding window mechanism to continuously update model parameters and control limits, the monitoring model can adaptively adjust itself as the host performance degrades, maintaining high-precision monitoring over a long period of time, effectively solving the problem of decreased accuracy of static models over long-term operation.

[0067] like Figure 1 , Figure 6 As shown, the length of the sliding time window Dynamically adjust based on host operational stability: calculate the variance of core state parameters within the window. If continuous Each window Less than the stability threshold Then increase If continuous Each window Greater than the fluctuation threshold If the sliding window length is too large, then L is reduced. By dynamically adjusting the sliding window length, the window is increased under stable conditions to obtain more stable model updates, and decreased under fluctuating conditions to improve the response speed to state changes, thus balancing the stability and sensitivity of model updates.

[0068] like Figure 1 As shown, it also includes a human-computer interaction and augmented reality display module 800, connected to a multi-dimensional contribution tracing and decision support module 600. This module highlights the locations of physical devices or sensors corresponding to suspected root cause variables in the 3D digital twin cabin model, and overlays corresponding real-time data, historical trends, and maintenance suggestions in an augmented reality manner. By combining the 3D digital twin cabin model with augmented reality technology, the location of abnormal equipment is intuitively displayed, and maintenance suggestions are overlaid, providing precise on-site guidance for maintenance personnel and significantly improving the human-computer interaction experience and emergency response efficiency.

[0069] like Figures 1-8 As shown, the embodiment provides a monitoring and early warning method based on the above-mentioned ship main engine adaptive monitoring and early warning system, including the following steps: Step S1. The multi-source heterogeneous data acquisition and cleaning module 100 is used to acquire multi-source monitoring data of the ship's main engine in real time, and to process noise and missing values, and output a standardized real-time data stream. Step S2. The dynamic operating condition self-identification module 200 identifies the current operating condition of the host through online clustering based on the speed and load parameters in the real-time data stream; Step S3. The working condition matching and model scheduling unit 400 schedules the matching benchmark monitoring model from the adaptive state monitoring model library 300 according to the current working condition. Step S4. The real-time monitoring and anomaly triggering unit 500 uses the benchmark monitoring model to calculate the real-time data stream, obtains the health index, and compares it with the dynamic control limit to determine whether an anomaly is triggered. Step S5. If an anomaly is triggered, the key variables causing the anomaly are analyzed through the multidimensional contribution tracing and decision support module 600, and a suspected fault diagnosis report and maintenance suggestions are generated. Step S6. If no anomaly is triggered, the current health data is used to update the parameters and control limits of the benchmark monitoring model through the model adaptive update engine 700. Step S7. Visualize the monitoring status, abnormal alarms, and maintenance guidance through the human-computer interaction and augmented reality display module 800.

[0070] Each step corresponds one-to-one with the aforementioned system modules and works collaboratively. Specifically: Step S1 corresponds to the multi-source heterogeneous data acquisition and cleaning module 100, such as... Figure 2 As shown, standardized real-time data streams are generated through data cleaning; step S2 corresponds to the dynamic operating condition self-identification module 200, as follows: Figure 3 As shown, the current operating condition of the host is identified online; step S3 corresponds to the operating condition matching and model scheduling unit 400, as follows: Figure 3 As shown, according to the sub-monitoring model corresponding to the working condition scheduling; step S4 corresponds to the real-time monitoring and anomaly triggering unit 500, such as... Figure 5 As shown, calculate the health index and compare it with the control limits; step S5 corresponds to the multidimensional contribution tracing and decision support module 600, as follows. Figure 8 As shown, when an anomaly occurs, the source is traced and repair suggestions are generated; step S6 corresponds to the model adaptive update engine 700, such as... Figure 6 , Figure 7 As shown, the model parameters are updated when the system is healthy; step S7 corresponds to the human-computer interaction and augmented reality display module 800, such as... Figure 1 As shown, the monitoring status and maintenance instructions are displayed visually.

[0071] The above method achieves closed-loop management of the ship's main engine status throughout the entire process through seven steps: data cleaning, operating condition identification, model scheduling, real-time monitoring, anomaly tracing, model updating, and visualization, ensuring the accuracy and adaptability of monitoring and early warning.

[0072] like Figure 9The figure shows the accuracy comparison curves of the system in this embodiment and the static PCA model over 9 months. The accuracy of this embodiment slowly decreased from 97% initially to 94.5%, a decrease of 2.5 percentage points; while the accuracy of the simulated static PCA model (without enabling the model adaptive update engine 700) decreased from 96% to 73%, a decrease of 23 percentage points. The gap between the two widened from 1 percentage point initially to 21.5 percentage points, fully verifying the role of the adaptive update mechanism of this invention in ensuring long-term monitoring accuracy.

[0073] Specific running examples Taking a large ocean-going trawler as an example, the vessel has a full-load displacement of 3,800 tons, a MAN B&W 8L27 / 38 main engine with a rated power of 2,720 kW, and a design speed of 15.5 knots. The vessel mainly operates in the North Pacific Ocean, with a single fishing voyage lasting 45-60 days. In March 2024, after completing its five-year special maintenance, the vessel was equipped with the adaptive monitoring and early warning system for the main engine described in this embodiment. The following records the vessel's complete operational data from April 2024 to January 2025.

[0074] I. Initial System Deployment Phase (April 2024) The vessel's first voyage after leaving the factory was from Zhoushan to the high seas fishing grounds, with a sailing time of approximately 68 hours. The system entered a self-learning phase, with the multi-source heterogeneous data acquisition and cleaning module 100 collecting data from various main engine sensors at a frequency of 1Hz, including 24 monitoring points such as main engine fuel inlet pressure, lubricating oil pressure and temperature, exhaust temperature of each cylinder, turbocharger speed, and cylinder liner coolant temperature. The dynamic operating condition self-identification module 200 performed online cluster analysis on the main engine speed and fuel rack scale based on a sliding window (window length L = 200 samples).

[0075] Data shows that the main engine operated under three typical conditions during this voyage: Operating Condition 1 (Standby Status): Speed ​​0-20 r / min, fuel rack 0-8%, port entry / exit and maneuvering operations, accounting for 8%; Operating Condition 2 (Economic Navigation): Engine speed 550-620 r / min, fuel rack 42-55%, cruising speed 11-13 knots, accounting for 47%; Operating condition 3 (fishing operation): Rotation speed 380-450 r / min, fuel rack 60-72%, trawl load condition, accounting for 45%.

[0076] The online clustering validity evaluation unit 210 calculates pseudo-clustering validity. The statistic is 312.7. The value was -1382.4, indicating a stable clustering structure. The system automatically established PCA sub-monitoring models for each operating condition, with the number of principal components determined to be 6 based on a cumulative variance contribution rate ≥95%. Of particular note is that operating condition 3 (fishing operation) exhibited significant periodic load fluctuations, with each trawl cycle lasting approximately 4-6 hours. The system successfully identified this as an independent operating condition, laying the foundation for subsequent accurate monitoring.

[0077] II. Normal Operation Phase (May 2024 - August 2025) Between May and August, the vessel completed two full fishing voyages, accumulating 106 days of sailing and fishing operations, including 38 days of sailing and 68 days of fishing. The system operated continuously, with the real-time monitoring and anomaly triggering unit completing one operation per second at a rate of 500. and Statistical calculation. Statistical data display: Operating Condition 2 (Economic Navigation): Average Value 28.4 (control limit) ),average Value 21.6 (control limit) ), Comprehensive Health Index Average 0.63; Operating Condition 3 (Fishing Operation): Average Value 35.7 (control limit) ),average Value 29.3 (control limit) ), Comprehensive Health Index Average 0.71.

[0078] The adaptive model update engine 700 completed 41 model updates during this period, with each update based on the 400 most recent healthy samples within a sliding window (the window length was appropriately shortened due to frequent changes in fishing vessel operating conditions). Taking an update on July 8th as an example, before the model update for operating condition 3... The updated value is 52.8, a slight adjustment of 1.5%, reflecting the performance fine-tuning of the main engine after continuous high-load fishing operations. The Human-Machine Interaction and Augmented Reality Display Module 800 presents a real-time 3D engine room model on the bridge display screen, with all parameters displayed in green as normal status, allowing the engineer to intuitively monitor changes in the main engine load during fishing operations.

[0079] III. Early Warning Stage of Anomalies (September 2024) On September 15th, the vessel was conducting trawling operations in the North Pacific Ocean (165°E, 42°N). The sea state was 4, the ambient temperature was 18°C, and the seawater temperature was 15°C. At 08:47:22 that day, the real-time monitoring and anomaly triggering unit 500 detected... The statistical volume showed an abnormal increase:

[0080] at this time , and All three indicators have exceeded the limits, meeting the trigger condition of "at least two exceeding the limits". The system generates an abnormal trigger signal with the timestamp "20240915084724" and the working condition label "Working Condition 3 - Fishing Operation". The statistical value of the exceeding limit is recorded.

[0081] The multidimensional contribution tracing and decision support module 600 immediately starts contribution calculation. Figure 8 The process shown is executed automatically, calculating the results for each of the 24 monitored variables. Contribution and Contribution. The top five results are: ① Main engine lubricating oil inlet pressure (contribution 38.6%); ② Turbocharger speed (contribution 24.3%); ③ Cylinder 4 exhaust temperature (contribution 15.7%); ④ Cylinder liner coolant outlet temperature (contribution 12.2%); ⑤ Fuel rack scale fluctuation rate (contribution 9.2%).

[0082] Variables whose contribution exceeds the preset threshold of 15% are the main engine lubricating oil inlet pressure, turbocharger speed, and cylinder 4 exhaust temperature. The combination of these three variables is identified as a suspected root cause variable set. Module 600 performs cosine similarity matching between the current contribution vector and the fault case knowledge base, which contains 112 historical fault cases, including 38 fishing vessel cases. The matching results show that the highest similarity (0.91) is with the case "XXXX-2022-023", which is "poor lubrication of the turbocharger bearing leading to speed fluctuations, accompanied by abnormal single-cylinder exhaust temperature".

[0083] The system automatically generates a structured maintenance suggestion: "Suspected fault mode: Turbocharger lubrication system malfunction, which may affect the scavenging pressure of cylinder 4. Recommended measures: ① Check the turbocharger oil inlet filter; ② Measure the turbocharger bearing vibration value; ③ Check the clearance of the cylinder 4 injector and exhaust valve; ④ Monitor the oil pressure and temperature change trends." This suggestion is displayed in the 3D engine bay model through the human-computer interaction and augmented reality display module 800, highlighting the physical location of the turbocharger and cylinder 4 in red, and displaying the above maintenance suggestion text in a floating position.

[0084] IV. Maintenance Response and Effectiveness Verification Phase (September 2024) At 09:30 that day, the chief engineer organized maintenance according to system prompts. On-site inspection revealed: ① The pressure differential of the turbocharger lubricating oil inlet filter was 0.25 bar, higher than the normal value of 0.12 bar; ② Upon disassembly and inspection of the filter, approximately 0.3 grams of metal debris were found, initially judged to be early bearing wear; ③ The clearance of the exhaust valve in cylinder 4 was excessive (0.65 mm, standard 0.45 ± 0.05 mm); ④ Vibration measurement showed that the vibration value at the turbocharger bearing increased from 2.8 mm / s to 4.6 mm / s. Given that the fishing vessel was still in the operating area, the engine room personnel took emergency measures: replacing the turbocharger lubricating oil filter element, adjusting the clearance of the exhaust valve in cylinder 4 to 0.46 mm, and increasing the operation of the lubricating oil separator. The fault characteristics were recorded for in-depth maintenance upon returning to port.

[0085] Continuous monitoring from September 16th to 18th showed that: ① the main engine lubricating oil inlet pressure recovered from 2.1 bar before maintenance to 2.5 bar; ② the turbocharger speed fluctuation rate decreased from ±4.2% to ±1.8%; ③ the temperature difference between the exhaust temperature of cylinder 4 and other cylinders decreased from 35℃ to 12℃; ④ The statistic has fallen back to 31.5. The statistic has fallen back to 47.2. It fell back to 0.82.

[0086] The Model Adaptive Update Engine 700 incorporated health data from September 16th onwards into the sliding window, completing the model update on September 22nd. New control limits... , The performance has been slightly adjusted compared to before the anomaly, reflecting the performance improvement after the temporary repair, but it has not yet fully recovered to its initial state.

[0087] After the vessel returned to port on October 8th, the turbocharger was disassembled and inspected according to system recommendations. Slight wear was found on the bearing cage; the bearing was replaced and the system was readjusted. Monitoring data after the vessel set sail again on October 15th showed that all parameters had returned to normal range. The statistic has fallen back to 27.8. The statistic has fallen back to 44.3. The value dropped back to 0.67, and the system completed the model update again, restoring the control limits to normal levels.

[0088] V. Long-term operation effect tracking (October 2024 - January 2025) The system operated continuously until January 2025, accumulating a monitoring period of 9 months, including 3 complete fishing voyages. Statistical data shows: ① Anomaly warning accuracy: The system triggered 14 anomaly warnings, of which 13 were confirmed as genuine anomalies by manual verification, and 1 was due to parameter fluctuations caused by extreme sea conditions, resulting in a false alarm rate of 7.1%; ② Average warning lead time: Compared to traditional threshold alarms, this system detected abnormal trends on average 1-3 hours in advance, providing valuable time for crew emergency response; ③ Model update frequency: The model was updated on average every 2.8 days, with control limit fluctuations within ±6%, reflecting stable model convergence; ④ False alarm rate comparison: Compared to a sister vessel using traditional threshold alarms, 52 alarms occurred during the same period, of which 29 were false alarms, resulting in a false alarm rate of 54.7%.

[0089] VI. Supplementary Typical Fault Cases On December 18, 2024, while the vessel was operating in the North Pacific Ocean, the system triggered an anomaly warning again. Contribution analysis showed that the fuel rack scale fluctuation rate (44.7%) and main engine speed fluctuation rate (31.2%) were the main contributing variables, while the exhaust temperature of cylinder 2 contributed 12.8%. Matching the similar case "XXXX-2023-087" to the knowledge base, the diagnosis was "fuel system vapor lock or injector carbon buildup." The engine room personnel found that the fuel filter pressure differential was too high, the fuel contained air bubbles, and the injector of cylinder 2 had severe carbon buildup. After cleaning the fuel system, replacing the fuel filter, and inspecting the injector of cylinder 2 as recommended, the parameters returned to normal.

[0090] The above operational examples demonstrate that the system in this embodiment exhibits the following advantages in practical applications on ocean-going fishing vessels: multi-source data cleaning ensures input quality; dynamic operating condition identification successfully distinguishes complex operating conditions such as navigation and fishing operations; the adaptive update mechanism maintains long-term monitoring accuracy; multi-dimensional contribution tracing provides precise maintenance guidance; and augmented reality visualization improves the emergency response efficiency of engine room personnel in harsh sea conditions. The system has withstood long-term testing in the harsh marine environment of the North Pacific, proving its reliability, accuracy, and practicality, and providing effective technical support for the intelligent operation and maintenance of ocean-going fishing vessels.

[0091] Similarly, the system described in this embodiment can be further extended to ocean-going freight vessels. Ocean-going freighters are characterized by long routes, stable speeds, large load variations, and continuous long-term operation of main engines, placing extremely high demands on the reliability of the main engines. By deploying this system on ocean-going container ships, bulk carriers, and tankers, the multi-source heterogeneous data acquisition and cleaning module can effectively integrate ship monitoring systems, AIS, meteorological, and loading data, providing high-quality data sources for subsequent analysis; the dynamic operating condition self-identification module can accurately distinguish different operating modes such as port entry and exit, ballast navigation, full-load navigation, and economic speed; the adaptive condition monitoring model library and real-time monitoring unit can dynamically track and adaptively adjust thresholds for the slow degradation of main engine performance during long voyages, promptly detecting early potential problems such as turbocharger efficiency decline, cylinder liner wear, and fuel system anomalies; the multi-dimensional contribution tracing and decision support module, combined with a fault case knowledge base, can automatically generate targeted maintenance suggestions, significantly reducing the risk of unplanned downtime; and the augmented reality display module provides marine engineering managers with intuitive fault location and operation guidance, improving emergency response efficiency. The widespread application of this system in ocean-going cargo ships will effectively improve ship operational safety, reduce maintenance costs, and propel ship intelligent management to a new level.

[0092] Obviously, the above embodiments of the present invention are merely illustrative examples to illustrate the invention and are not intended to limit the implementation of the invention. Other obvious variations or modifications derived from the essential spirit of the invention still fall within the protection scope of the invention.

Claims

1. A ship main engine adaptive monitoring and early warning system, characterized in that, include: The multi-source heterogeneous data acquisition and cleaning module (100) is used to acquire multi-source monitoring data of the ship's main engine in real time at a sampling frequency that matches the sensor type, and to perform real-time noise removal and missing value filling on the multi-source monitoring data by combining a density-based local outlier detection algorithm with a time series proximity-based interpolation algorithm, thereby generating a standardized real-time data stream. The dynamic operating condition self-identification module (200) is connected to the multi-source heterogeneous data acquisition and cleaning module (100) to receive the real-time data stream, perform online clustering analysis on the main engine speed and fuel rack scale based on the Gaussian mixture model within the sliding time window, and automatically identify and output the dynamic operating condition label of the main engine at the current moment. An adaptive state monitoring model library (300) stores multiple sub-monitoring models based on principal component analysis, each sub-monitoring model corresponding to a running condition divided by the dynamic operating condition self-identification module (200); The working condition matching and model scheduling unit (400) is connected to the dynamic working condition self-identification module (200) and the adaptive state monitoring model library (300), and is used to retrieve the corresponding sub-monitoring model from the adaptive state monitoring model library (300) as the current health status benchmark model according to the received current working condition label. The real-time monitoring and anomaly triggering unit (500), connected to the working condition matching and model scheduling unit (400), is used to substitute the current standardized real-time data stream into the current health status benchmark model and calculate its principal component space statistics. and residual space statistics and judge and Whether it exceeds the dynamic control limit corresponding to the current health status benchmark model; If the limit is exceeded, an abnormal trigger signal containing a timestamp, operating condition label, and over-limit statistics will be generated. The multidimensional contribution tracing and decision support module (600), connected to the real-time monitoring and anomaly triggering unit (500), is used to calculate the pairs of each original monitoring variable in response to the anomaly triggering signal. and The contribution of statistical values ​​exceeding the limit is identified as a set of suspected root cause variables, and structured alarm and maintenance suggestion information is generated by combining the fault case knowledge base. as well as The model adaptive update engine (700) is connected to the adaptive state monitoring model library (300) and the real-time monitoring and anomaly triggering unit (500). When no anomaly triggering signal is generated, it feeds back the current health data according to the working condition label to the corresponding sub-monitoring model in the adaptive state monitoring model library (300) and performs rolling updates of model parameters and control limits based on the sliding window mechanism.

2. The ship main engine adaptive monitoring and early warning system according to claim 1, characterized in that, The multi-source monitoring data includes key operating parameters required to construct the sub-monitoring model, including at least: main engine fuel inlet pressure, main engine lubricating oil inlet pressure, main engine lubricating oil inlet temperature, turbocharger speed, turbocharger lubricating oil outlet temperature, main engine speed, fuel rack scale, cylinder liner cooling water inlet pressure, cylinder liner cooling water outlet pressure, cylinder liner cooling water outlet temperature, average exhaust temperature of each cylinder, intercooler fresh water inlet pressure, intercooler fresh water outlet pressure, and intercooler low-temperature fresh water outlet temperature.

3. The ship main engine adaptive monitoring and early warning system according to claim 1, characterized in that, The dynamic operating condition self-identification module (200) also includes an online clustering validity evaluation unit, used to calculate the pseudo-value of the current clustering result in real time. Statistics and Bayesian information criteria; The pseudo The formula for calculating the statistic is: in, The sum of squares of deviations between classes. For the sum of squared deviations within the class, This represents the current number of clusters. This represents the total number of samples within the sliding window; The formula for calculating the Bayesian information criterion is as follows: in, This is the maximum likelihood estimate of the current Gaussian mixture model. This represents the total number of model parameters. When the pseudo-calculation of multiple consecutive calculation cycles The statistics show a downward trend and When the value does not converge to the minimum, the clustering of the Gaussian mixture model is triggered. Optimization and adjustments.

4. The ship main engine adaptive monitoring and early warning system according to claim 1, characterized in that, The real-time monitoring and anomaly triggering unit (500) is also used to calculate a comprehensive health index, the calculation formula of which is: in, and Under the current operating conditions and Statistic in confidence level The lower control limit; The conditions for generating the abnormal trigger signal are: , and At least two of the three indicators exceed their respective control limits simultaneously.

5. The ship main engine adaptive monitoring and early warning system according to claim 1 or 4, characterized in that, In the multidimensional contribution tracing and decision support module (600), the formula for calculating the contribution of variables includes: right Contribution of the statistic: right Contribution of the statistic: in, For the current sample vector, This is a diagonal matrix related to the principal components. For the first A unit vector of variables, , For the load matrix, It is an identity matrix.

6. The ship main engine adaptive monitoring and early warning system according to claim 5, characterized in that, The multidimensional contribution tracing and decision support module (600) provides decision support by querying the fault case knowledge base to perform similarity matching on the identified suspected fault patterns. The similarity matching formula is as follows: in, For the current root dependent variable set, the first The contribution of each standardized variable For the historical cases in the knowledge base The standardized contribution of each variable. The total number of variables; Select the case with the highest similarity and output its processing solution.

7. The ship main engine adaptive monitoring and early warning system according to claim 1, characterized in that, The model adaptive update engine (700) uses a length of The sliding time window requires that the data within the window belong to the same working condition and be judged to be in a healthy state. Model updates include: recalculating the covariance matrix, eigenvalues, and eigenvectors based on window data to update the load matrix. And recalculate the control limits; Control limits of statistics The updated formula is: in, Number of principal elements The total number of healthy samples used for modeling. For degrees of freedom of Distribution at confidence level The critical value below; Control of statistics The updated formula is: in, The covariance matrix is ​​the first 1 eigenvalue, For the total number of variables, For a standard normal distribution at a confidence level The quantiles below.

8. The ship main engine adaptive monitoring and early warning system according to claim 7, characterized in that, The length of the sliding time window Dynamically adjust based on host operational stability: calculate the variance of core state parameters within the window. ; If continuous Each window Less than the stability threshold Then increase ; If continuous Each window Greater than the fluctuation threshold If , then L decreases.

9. The ship main engine adaptive monitoring and early warning system according to claim 1, characterized in that, It also includes a human-computer interaction and augmented reality display module (800), which is connected to the multidimensional contribution tracing and decision support module (600) and is used to highlight the physical equipment or sensor locations corresponding to the suspected root cause variable set in the three-dimensional digital twin cabin model, and overlay the corresponding real-time data, historical trends and maintenance suggestion information in an augmented reality manner.

10. A monitoring and early warning method based on the ship main engine adaptive monitoring and early warning system according to any one of claims 1 to 9, characterized in that, Includes the following steps: Step S1. The multi-source monitoring data of the ship's main engine is collected in real time through the multi-source heterogeneous data acquisition and cleaning module (100), and noise and missing value processing is performed to output a standardized real-time data stream; Step S2. The dynamic operating condition self-identification module (200) identifies the current operating condition of the host through online clustering based on the speed and load parameters in the real-time data stream; Step S3. Through the working condition matching and model scheduling unit (400), according to the current working condition, schedule the matching benchmark monitoring model from the adaptive state monitoring model library (300); Step S4. The real-time monitoring and anomaly triggering unit (500) uses the benchmark monitoring model to calculate the real-time data stream, obtains the health index, and compares it with the dynamic control limit to determine whether an anomaly is triggered. Step S5. If an anomaly is triggered, the key variables causing the anomaly are analyzed through the multidimensional contribution tracing and decision support module (600), and a suspected fault diagnosis report and maintenance suggestions are generated. Step S6. If no anomaly is triggered, the current health data is used to update the parameters and control limits of the benchmark monitoring model through the model adaptive update engine (700). Step S7. Visualize the monitoring status, abnormal alarms and maintenance guidance through the human-computer interaction and augmented reality display module (800).

Images