Method for dynamic analysis and early warning of ship equipment maintenance data

Abstract

The application provides a ship equipment maintenance data dynamic analysis and early warning method, and relates to the technical field of data processing.The method comprises the following steps: step 1, processing the multi-dimensional running state time series data of the ship equipment, constructing a unified time series data set, and extracting a health feature vector reflecting the mechanical performance and overall running state of the equipment; for the key rod structure in the equipment, based on the real-time monitored axial load data and the geometric parameters of the rod, the axial stress is calculated, and combined with the allowable stress of the material, the axial tension and compression strength safety factor is obtained by comparison, and the health feature vector is formed. The application realizes the early identification and intelligent early warning of faults and hidden dangers, improves the maintenance efficiency and guarantees the safety of ship operation.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method for dynamic analysis and early warning of ship equipment maintenance data. Background Technology

[0002] In the development of intelligent operation and maintenance management of ships, assessing and providing early warnings on the operating status of key mechanical equipment is an important foundation for ensuring navigation safety and reducing maintenance costs. Currently, health management of ship equipment mostly relies on preset single threshold alarms or regular offline inspection mechanisms.

[0003] However, in practical applications, existing monitoring and analysis methods have some limitations. For example, in scenarios where key components of a ship's propulsion system, such as the intermediate shaft and thrust shaft, are the monitoring objects, most existing methods acquire axial load data through strain sensors arranged on the shaft system and calculate the axial stress accordingly. Then, a simple comparison is made with the allowable stress of the material to determine whether the strength safety factor is within the allowable range. Although this approach can reflect the static stress of the component at a specific moment, it is mostly difficult to effectively cope with the drastic fluctuations of dynamic loads under complex sea conditions. Due to the coupling effect of wave impact, propeller excitation force, and hull deformation, axial load data often contains a large amount of non-stationary noise and instantaneous impact signals. If the safety factor is directly calculated based on these single-point data that have not been thoroughly cleaned, and then mapped to the health feature vector of the entire equipment, it may lead to bias in the health status assessment, making it difficult for the assessment results to accurately reflect the continuous degradation trend of the equipment's mechanical performance. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a method for dynamic analysis and early warning of ship equipment maintenance data, so as to realize the early identification and intelligent early warning of potential faults, improve maintenance efficiency and ensure the safety of ship operation.

[0005] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:

[0006] Firstly, a method for dynamic analysis and early warning of ship equipment maintenance data, the method comprising:

[0007] Step 1: Process the multi-dimensional operating status time-series data of the ship equipment, construct a unified time-series dataset, and extract health feature vectors that reflect the mechanical performance and overall operating status of the equipment; for the key rod structures in the equipment, calculate the axial stress based on the real-time monitored axial load data and the geometric parameters of the rods, and combine it with the allowable stress of the material to obtain the axial tensile and compressive strength safety factor by comparison, and form a health feature vector.

[0008] Step 2: Input the health feature vector into the pre-trained deep learning ship equipment status prediction model to obtain the health assessment results and anomaly probability prediction values.

[0009] Step 3: Based on the real-time data of the three core monitoring indicators in the unified time series dataset, generate dynamic correction coefficients through multi-parameter correlation analysis and correct the health assessment results and abnormal probability prediction values ​​to obtain the corrected health assessment results and abnormal probability prediction values.

[0010] Step 4: Based on the corrected health assessment results and the predicted abnormality probability, a real-time early warning judgment is made in conjunction with a dynamic adaptive threshold judgment strategy; when the predicted value exceeds the dynamic early warning threshold, an alarm message is generated.

[0011] Step 5: Initiate the maintenance coordination and management response process based on the alarm information, integrate the alarm information into the ship maintenance management platform, generate maintenance work orders, recommend maintenance time windows and personnel configuration schemes, and form maintenance execution records; at the same time, combine existing early warning records and maintenance feedback data to dynamically optimize the maintenance plan and form a closed-loop maintenance business management process.

[0012] The above-described solution of the present invention has at least the following beneficial effects:

[0013] This system unifies the processing of multi-dimensional time-series data of ship equipment and extracts health feature vectors to achieve a comprehensive quantitative characterization of equipment mechanical performance and overall operating status. Simultaneously, it calculates axial stress and strength safety factors for key structural members, improving the relevance and accuracy of equipment health status descriptions. A pre-trained deep learning model is used for health assessment and anomaly probability prediction, enabling intelligent prediction of equipment operating trends. Multi-parameter correlation analysis is introduced to generate dynamic correction coefficients, enhancing the reliability of status judgments. Real-time early warnings are provided based on a dynamic adaptive threshold strategy, reducing false alarms and missed alarms. Early warning information is deeply integrated with the ship maintenance management platform, automatically generating maintenance work orders, recommending maintenance plans, and creating execution records. Simultaneously, maintenance plans are continuously optimized, constructing a closed-loop management system to improve ship equipment maintenance efficiency and reduce operating costs. Attached Figure Description

[0014] Figure 1 This is a flowchart illustrating the dynamic analysis and early warning method for ship equipment maintenance data provided in an embodiment of the present invention.

[0015] Figure 2 This is a schematic diagram illustrating how a health feature vector is input into a pre-trained deep learning ship equipment status prediction model to obtain health assessment results and anomaly probability prediction values, as provided in an embodiment of the present invention. Detailed Implementation

[0016] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0017] like Figure 1 As shown, an embodiment of the present invention proposes a method for dynamic analysis and early warning of ship equipment maintenance data, the method comprising the following steps:

[0018] Step 1: Process the multi-dimensional operating status time-series data of the ship equipment, construct a unified time-series dataset, and extract health feature vectors that reflect the mechanical performance and overall operating status of the equipment; for the key rod structures in the equipment, calculate the axial stress based on the real-time monitored axial load data and the geometric parameters of the rods, and combine it with the allowable stress of the material to obtain the axial tensile and compressive strength safety factor by comparison, and form a health feature vector.

[0019] Step 2: Input the health feature vector into the pre-trained deep learning ship equipment status prediction model to obtain the health assessment results and anomaly probability prediction values.

[0020] Step 3: Based on the real-time data of the three core monitoring indicators in the unified time series dataset, generate dynamic correction coefficients through multi-parameter correlation analysis and correct the health assessment results and abnormal probability prediction values ​​to obtain the corrected health assessment results and abnormal probability prediction values.

[0021] Step 4: Based on the corrected health assessment results and the predicted abnormality probability, a real-time early warning judgment is made in conjunction with a dynamic adaptive threshold judgment strategy; when the predicted value exceeds the dynamic early warning threshold, an alarm message is generated.

[0022] Step 5: Initiate the maintenance coordination and management response process based on the alarm information, integrate the alarm information into the ship maintenance management platform, generate maintenance work orders, recommend maintenance time windows and personnel configuration schemes, and form maintenance execution records; at the same time, combine existing early warning records and maintenance feedback data to dynamically optimize the maintenance plan and form a closed-loop maintenance business management process.

[0023] In this embodiment of the invention, multi-dimensional time-series data of ship equipment are processed uniformly and health feature vectors are extracted to achieve a comprehensive quantitative characterization of the equipment's mechanical performance and overall operating status. Simultaneously, axial stress and strength safety factors are calculated for key structural members to improve the relevance and accuracy of equipment health status descriptions. A pre-trained deep learning model is used for health assessment and anomaly probability prediction to achieve intelligent prediction of equipment operating trends. Multi-parameter correlation analysis is introduced to generate dynamic correction coefficients, improving the reliability of status judgment. Real-time early warnings are provided based on a dynamic adaptive threshold strategy to reduce false alarms and missed alarms. Early warning information is deeply integrated with the ship maintenance management platform to automatically generate maintenance work orders, recommend maintenance plans, and form execution records. Simultaneously, maintenance plans are continuously optimized to build a closed-loop management system, improving ship equipment maintenance efficiency and reducing operation and maintenance costs.

[0024] In a preferred embodiment of the present invention, step 1 above processes the multi-dimensional operational status time-series data of the ship equipment to construct a unified time-series dataset and extracts a health feature vector reflecting the mechanical performance and overall operational status of the equipment; for key structural members in the equipment, based on real-time monitored axial load data and member geometric parameters, axial stress is calculated, and combined with the allowable stress of the material, the axial tensile and compressive strength safety factor is obtained by comparison to form a health feature vector, which may include:

[0025] In this embodiment of the invention, step 110 involves cleaning the multidimensional operating status time-series data collected from different sensors, removing outliers caused by noise or transmission errors, and filling in missing data using linear interpolation to obtain cleaned time-series data. Specifically, this includes: sequentially reading the raw time-series data collected by vibration sensors, temperature sensors, pressure sensors, load sensors, and strain sensors; arranging the values ​​collected by each sensor at consecutive time points to form the raw time-series data sequence for the corresponding sensor channel; for each raw time-series data sequence, comparing the current collected value with the collected values ​​at multiple adjacent times before and after that time; when the difference between the collected value at a certain time and the collected values ​​at multiple adjacent times before and after exceeds a preset normal fluctuation range, the collected value at that time is determined to be due to… Abnormal values ​​caused by noise interference or data transmission errors are directly removed from the original time series data sequence and do not participate in subsequent calculations. After removing abnormal values, time points with missing values ​​in the original time series data sequence are identified to determine the time position of the missing data. For the identified missing data positions, linear interpolation is used to fill the data. Specifically, the valid acquisition values ​​at the time before and after the missing data position are found. The valid acquisition values ​​at the time before and after the missing data position are added together and then divided by two. The intermediate value is used as the filling value at the missing position. This filling value is filled into the corresponding time point to keep the time series data continuous and without gaps in time. Finally, the cleaned time series data after outlier removal and missing value filling is obtained.

[0026] Step 111: Based on the cleaned time-series data, resample according to a unified time frequency to achieve time alignment of multi-source data and form a structured unified time-series dataset. Specifically, this includes: pre-setting a unified time sampling frequency for the entire system, using this unified time sampling frequency as the benchmark for aligning all sensor data; re-dividing the cleaned time-series data of each sensor channel according to the unified time sampling frequency to generate a time axis with the same starting point, time interval, and duration for each sensor channel; for each sensor channel where the original collected values ​​are missing at the newly divided time points, supplementing the corresponding values ​​using the same linear interpolation method as in Step 110, i.e., adding the effective values ​​of the two most recent moments before and after that time point and dividing by two to obtain the resampled value at that time point; horizontally stitching the resampled data from all sensor channels according to the same time points, so that the monitoring values ​​of all sensors at the same moment correspond to the same timestamp, completing the time alignment of multi-source heterogeneous data; and arranging all the time-aligned data sequentially according to time order to form a unified time-series dataset with a regular structure, unified time, and complete channels.

[0027] Step 112: Based on the unified time-series dataset, extract feature parameters that characterize the mechanical performance and overall operating status of the equipment to obtain an initial feature set. Specifically, this includes extracting feature parameters for the time-series data corresponding to each sensor in the unified time-series dataset. The extracted feature parameters are mainly divided into three categories: statistical features, frequency domain features, and time-frequency features. These three types of features complement each other and can comprehensively characterize the mechanical performance and overall operating status of the equipment. Among them, statistical features are mainly used to reflect the overall distribution and numerical fluctuation characteristics of the data from each sensor. The specific extraction method is as follows: select a continuous time window, usually one hour, which can be adjusted according to the actual operating conditions. Perform a series of simple numerical calculations on all the collected values ​​within this time window to obtain the corresponding statistical features. Add all the collected values ​​within the time window together, and then divide the sum by the time window. The total number of data points collected within a window yields the average value of the data within that time window, reflecting the overall level of the data. For each data point collected within the time window, the difference between the collected value and the calculated average is calculated. Then, the square of each deviation is calculated. The sum of all squared deviations is then divided by the total number of data points collected, yielding the variance of the data within that time window, reflecting the degree of data dispersion. Taking the square root of the calculated variance yields the root mean square (RMS) value, which more directly reflects the data's fluctuation range. Simultaneously, the maximum and minimum values ​​among all collected values ​​within the time window are identified. Subtracting the minimum value from the maximum value yields the range of data fluctuations within that time window, reflecting extreme data changes.

[0028] For the time-series data corresponding to vibration sensors, additional frequency domain features reflecting the signal frequency composition are extracted. Specifically, a continuous time window is selected, and the vibration time-series data within this window is divided according to certain frequency intervals. The entire vibration signal is decomposed into signal components of different frequency intervals. The energy magnitude of the vibration signal within each frequency interval is then statistically analyzed to obtain the energy value corresponding to each frequency point. Among these energy values, the frequency point with the highest energy is identified; the energy value corresponding to this frequency point is the dominant frequency energy. The dominant frequency energy reflects the main frequency characteristics of the vibration signal, thus reflecting the mechanical vibration state of the equipment. When the equipment experiences mechanical wear, loosening, or other faults, the dominant frequency energy will show significant changes. Furthermore, for the time-series data corresponding to vibration sensors, temperature sensors, and pressure sensors, extraction... The time-frequency local features that can reflect the changes of a signal with both time and frequency are extracted by selecting multiple consecutive, relatively short time sub-windows (usually 5 minutes as a time sub-window). For the time series data within each time sub-window, the main frequency components are analyzed, and the changes in the main frequency components within different time sub-windows are recorded. At the same time, the changing trends of the average value and fluctuation range of the data within different time sub-windows are also recorded. Integrating these changes gives us the time-frequency local features. Time-frequency local features can reflect the dynamic changes in the operating status of equipment and capture subtle changes in early equipment faults. All the statistical features, frequency domain features, and time-frequency features extracted above are summarized together, and duplicate feature parameters are removed to form a set containing multi-dimensional and multi-type features. This set is the initial feature set.

[0029] Step 113: Based on the initial feature set, key features are selected using a random forest-based feature importance assessment to form a health feature vector. Specifically, this includes: determining the core purpose of feature selection, namely, selecting feature parameters from the initial feature set that are most sensitive to changes in equipment operating status and can most accurately distinguish between normal and abnormal equipment states, eliminating redundant features with low correlation to equipment operating status and weak distinguishing effect, reducing the complexity of subsequent data processing, and improving the accuracy of equipment health status assessment; next, a random forest-based feature importance assessment method is used for selection. Specifically, each feature parameter in the initial feature set is treated as an independent feature item and included in the feature assessment process one by one; simultaneously, past operating data of the ship's equipment is retrieved. This past data includes feature data of the equipment under normal operating conditions, as well as feature data when the equipment malfunctions or is abnormal, determining the equipment state (normal or abnormal) corresponding to each set of past feature data; using these past feature data and corresponding equipment states as references, the effectiveness of each feature in the initial feature set in distinguishing between normal and abnormal equipment states is determined sequentially.

[0030] The specific judgment logic is as follows: For each feature, compare the numerical distribution of that feature under normal equipment conditions with that under abnormal conditions. If the difference in the numerical distribution of the feature under normal and abnormal conditions is very significant, and it can quickly and accurately distinguish between the two states, it indicates that the feature has a strong discriminative effect and is highly sensitive to changes in equipment state. If the difference in the numerical distribution of the feature under the two states is small, and it is difficult to distinguish whether the equipment is operating normally, it indicates that the feature has a weak discriminative effect and is less sensitive to changes in equipment state. The discriminative effect of each feature is quantified and scored. The scoring criteria are set according to the discriminative sensitivity of the feature. The more obvious the discriminative effect and the higher the sensitivity, the higher the importance score is assigned; the weaker the discriminative effect and the lower the sensitivity, the lower the importance score is assigned. After scoring, the features are ranked in descending order of importance score. All features in the initial feature set are sorted. Then, a feature screening threshold is set, and features with importance scores higher than the threshold and that can effectively reflect the equipment's operating status are selected as key features. At the same time, feature parameters with importance scores lower than the threshold, those that contribute little to the equipment status assessment, and those with high redundancy are removed to ensure that the selected key features are both highly sensitive and free of redundant information. All the selected key features are combined and arranged in a preset fixed order. The arrangement order can be based on the feature importance scores from high to low, or it can be classified and arranged according to the monitoring dimensions (vibration, temperature, pressure, etc.) corresponding to the features, forming an ordered feature set. This feature set is the health feature vector used to characterize the equipment's health status. Each feature component in this vector can sensitively reflect a certain aspect of the equipment's operating status or mechanical performance.

[0031] Step 114: Real-time acquisition of axial load data borne by key members using strain sensors or load sensors deployed on them. Simultaneously, obtaining the geometric parameters and material identification of the members from the equipment design parameter database. Geometric parameters include at least cross-sectional area, length, and moment of inertia. Specifically, this involves identifying the key structural members in the ship's equipment, primarily including core load-bearing components such as the intermediate shaft, thrust shaft, and thrust blocks in the ship's propulsion system. These key members directly bear the axial tensile and compressive forces during ship operation, and their structural strength directly affects the safe operation of the ship's equipment. Therefore, specialized monitoring and feature extraction are necessary. Next, strain sensors or load sensors are deployed at appropriate locations on or inside these key members. The placement locations need to be carefully selected to ensure that the sensors can directly and accurately sense the axial loads on the key members during operation. During navigation, sensors are subjected to axial tensile or compressive forces. To avoid inaccurate monitoring data due to improper placement, sensors are typically placed at the points where stress is most concentrated and deformation is most pronounced, while avoiding connection points and areas of severe wear. This ensures stable operation and monitoring accuracy. After placement, the sensors are activated for real-time data acquisition. The sensors continuously sense the load response signals generated by key components under current navigation conditions, such as different navigation speeds, sea states, and loads. These analog response signals are converted into digital axial load values ​​that can be directly used for calculation and processing. These axial load values ​​are recorded sequentially according to the acquisition time, forming continuous axial load time-series data. This time-series data can reflect the changes in axial load borne by key components in real time.

[0032] While collecting axial load data, a pre-established equipment design parameter database is retrieved. This database stores detailed design information for all ship equipment and its components, including geometric parameters, material information, and design standards for key members. Each piece of information in the database corresponds one-to-one with a unique identifier for the equipment and component, ensuring accurate retrieval of relevant parameters for the corresponding member. Based on the unique identifier of the currently monitored key member, a search and match is performed in the equipment design parameter database to find the design information entry that perfectly corresponds to that key member. From this entry, the geometric parameters of the key member are read and obtained, including at least three core geometric parameters: first, the area of ​​the member's cross-section, i.e., the cross-sectional area of ​​the member perpendicular to the axial force direction; this value is used for calculation... The key parameters for calculating axial stress include: 1) the overall length of the member, i.e., the effective length between its two ends; 2) the moment of inertia of the member's cross-section, which reflects the member's bending and torsional resistance and indirectly reflects its structural strength; and 3) the material identification information uniquely identifying the material used in the key member, typically a unique code that accurately distinguishes the type of material used. The collected axial load time series data is then associated and stored with the acquired key member geometric parameters and material identification information to ensure that each set of axial load data corresponds to a specific key member and its related parameters, providing complete data support for the calculation of axial stress and safety factor.

[0033] Step 115: Based on the axial load data and cross-sectional area, calculate the axial stress value of the member under the current working condition. According to the material identifier, retrieve the allowable stress value of the corresponding material from the preset material property database. Specifically, this includes: extracting the axial load value at the current moment from the axial load time series data collected in Step 114. This value represents the actual axial load magnitude borne by the key member under the current navigation conditions and load. Accuracy of the value must be ensured during extraction to avoid extracting incorrect historical or abnormal values. If the axial load value at the current moment is an abnormal value, it has already been removed through the cleaning operation in Step 110. To ensure calculation continuity, the most recent effective axial load value before and after the current moment is extracted as a replacement. Next, the axial stress value of the key member under the current working condition is calculated. The specific calculation logic is as follows: axial stress refers to the axial force per unit area on the cross-section of the member. Therefore, the axial load value extracted at the current moment is divided by the area value of the cross-section of the key member obtained in step 114. The result obtained through this division operation is the axial force per unit area on the cross-section of the key member at the current moment. This force value is the axial stress value of the key member under the current working condition. This value can directly... This reflects the current stress and strength of the key members. A higher value indicates a more concentrated stress on the member and greater pressure on the structure's strength. Simultaneously with calculating the axial stress, a pre-established material property database is retrieved. This database stores detailed property information for various commonly used marine equipment materials, including the maximum allowable safe stress (i.e., allowable stress value) that different materials can withstand under different operating environments and temperatures. The material information in the database corresponds one-to-one with material identifiers, allowing for quick and accurate retrieval of the corresponding material's allowable stress value based on the material identifier. Based on the material of the key members obtained in step 114... The identification information is matched and searched in the material property database to find the material property entry that completely corresponds to the material identification. From the entry, the maximum safe stress value that the material can withstand under the current ship operating environment and working temperature is read and retrieved. This value is the allowable stress value of the material used in the current key member. This value is the core benchmark value for judging whether the structural strength of the member is safe. The calculated axial stress value at the current moment is associated with the retrieved material allowable stress value and stored to ensure that the two values ​​correspond to the same key member and the same working condition at the same time, thus preparing for the calculation of the axial tensile and compressive strength safety factor.

[0034] Step 116: Compare the allowable stress value with the axial stress value, and calculate the axial tensile and compressive strength safety factor based on the ratio. Specifically, this involves determining the core function of the axial tensile and compressive strength safety factor. This safety factor is an important indicator for measuring the safety reserve of the current structural strength of key members relative to the material's ultimate strength. A larger safety factor indicates that the current stress of the member is far lower than the material's ultimate bearing capacity, and the structure is safer. A smaller safety factor indicates that the current stress of the member is closer to the material's ultimate bearing capacity, and the structure faces a higher risk of failure. When the safety factor is less than 1, it indicates that the current stress of the member has exceeded the material's bearing capacity, and there is a risk of structural damage. Next, obtain the two core values ​​obtained in Step 115: the allowable stress value of the material used in the key member and the current axial stress value of the key member. Ensure that the units of the two values ​​are consistent (both are the same pressure unit) to avoid calculation errors caused by inconsistent units. Calculate the axial tensile and compressive strength safety factor using a ratio calculation method. The specific calculation logic is as follows: The allowable stress value of the material... As the numerator in the division operation, the axial stress value calculated at the same moment is used as the denominator. The allowable stress value of the material is divided by the axial stress value. The ratio obtained through this division operation is the axial tensile and compressive strength safety factor of the current critical member under the current load. For example, if the allowable stress value of a critical member is a fixed value, and the axial stress value at the current moment is another fixed value, the ratio obtained by dividing the allowable stress value by the axial stress value is the safety factor at that moment. If the ratio is 2.5, it means that the current stress strength of the member is only 40% of the maximum allowable strength of the material, and there is sufficient safety reserve. If the ratio is 1.2, it means that the current stress strength of the member is close to the maximum allowable strength of the material, and it needs to be closely monitored and the monitoring frequency increased. After the calculation is completed, the axial tensile and compressive strength safety factor at the current moment is recorded and associated with the corresponding axial stress value, allowable stress value, and timestamp for storage, forming continuous time-series data of safety factors, which can reflect the changes in the safety reserve of the structural strength of critical members in real time.

[0035] Step 117: The axial tensile and compressive strength safety factor is used as a feature component in the health feature vector, and combined with geometric parameters and related monitoring data to form a health feature vector that can comprehensively characterize the structural strength state of key members. Specifically, this includes: determining that the health feature vector obtained in step 113 mainly reflects the overall operating state and mechanical performance of the ship's equipment, but does not specifically characterize the structural strength of key members. Therefore, the axial tensile and compressive strength safety factor calculated in step 116 needs to be included to supplement the strength characteristics of key members. The axial tensile and compressive strength safety factor calculated in step 116 is added as a new independent feature component to the health feature vector obtained in step 113. The addition position can be set according to the importance of the feature, usually added adjacent to core features such as vibration and pressure to ensure the rationality and orderliness of the feature vector. After adding this feature component, the health feature vector will have the ability to characterize the structural strength safety reserve of key members. Simultaneously, the key members obtained in step 114... The geometric parameters of the members, including cross-sectional area, length, and moment of inertia, are also added as feature components to the health feature vector. These geometric parameters reflect the structural characteristics of the key members themselves. Members with different geometric parameters have different load-bearing capacities and structural strengths. Including these parameters in the feature vector can further improve the accuracy of the feature vector in representing the structural strength state of the key members. In addition, other monitoring data related to the key members, including temperature monitoring data and vibration monitoring data around the key members, are also added to the health feature vector. These data can reflect the environmental conditions and overall state of the key members during operation. In conjunction with the axial tensile and compressive strength safety factor and geometric parameters, they can more comprehensively reflect the structural strength state of the key members. All feature components are systematically integrated and arranged in the order of overall equipment operation characteristics—key member geometric characteristics—key member strength characteristics—key member environmental characteristics to ensure that the feature vector is structurally regular and logically clear, forming a health feature vector with more complete dimensions and more comprehensive information.

[0036] By unifying frequency resampling and time alignment, evaluation bias caused by data distortion is avoided. By comprehensively extracting statistical features, frequency domain features, and time-frequency features, an initial feature set representing the equipment's operating status from multiple perspectives is formed. Then, through feature importance assessment, the key features that are most sensitive to changes in equipment status and have the strongest distinguishing effect are selected, thereby improving the health assessment and anomaly early warning.

[0037] like Figure 2 As shown, in a preferred embodiment of the present invention, step 2 above, which involves inputting the health feature vector into a pre-trained deep learning ship equipment status prediction model to obtain health assessment results and anomaly probability prediction values, may include:

[0038] In this embodiment of the invention, step 220 involves collecting operational data of the ship's equipment under past operating conditions as sample operational data. The sample operational data is preprocessed and multi-dimensional health features are extracted to form a sample health feature vector. Simultaneously, equipment status labels corresponding to the sample health feature vectors are obtained to construct a training sample set. The scope and standards for collecting the sample operational data are determined to ensure the comprehensiveness, representativeness, and effectiveness of the sample data. This ensures that the data covers various typical operating conditions of the ship's equipment, including normal navigation, low-speed navigation, berthing, navigation under complex sea conditions, and operating conditions under various states such as minor, moderate, and severe equipment anomalies. This avoids insufficient subsequent model training and prediction accuracy due to a single type of sample data. The problem of low accuracy is addressed here. The data collection source for the sample operation is consistent with that for the real-time operation data in step 110, both originating from various sensors deployed on the ship's equipment, including vibration sensors, temperature sensors, pressure sensors, load sensors, strain sensors, etc. The collected data includes monitoring values ​​of various sensors under different operating conditions and at different time points in the past, while also recording the specific operating condition information and timestamp corresponding to each set of monitoring values ​​to ensure that the sample data can be mapped to a specific operating scenario. After collection, all past monitoring data are organized and summarized in chronological order of timestamps to form the original sample operation data set. Subsequently, this sample operation data set is processed in a manner completely consistent with steps 110, 111, 112, and 113. The preprocessing operation is thorough and leaves no step out. The specific steps are as follows: First, data cleaning is performed, checking the sample data from each sensor one by one, removing abnormal values ​​caused by noise interference and transmission errors. The removal criteria are the same as in step 110: when the monitored value at a certain moment deviates significantly from the values ​​at multiple adjacent moments, it is considered an outlier and removed. Second, for sample data with missing values ​​after cleaning, linear interpolation is used to fill them in. The filling method is the same as in step 110: finding the two closest valid values ​​before and after the missing position, adding the two valid values, dividing by 2, and using the intermediate value as the filling value to ensure continuous sample data without gaps. Third... The first step involves resampling the data at a uniform time frequency, consistent with the preset uniform sampling frequency in step 111. The time axis of the sample data from each sensor is redefined, and linear interpolation is used to fill in the blank time points, achieving time alignment of the multi-source sample data. The second step involves feature extraction from the time-aligned sample data, using the same extraction method as steps 112 and 113. This involves first extracting statistical features (mean, variance, root mean square value, fluctuation range), frequency domain features (main frequency energy of the vibration signal), and time-frequency features (changes in the main frequency components over different time periods) to form an initial sample feature set. Then, a feature importance evaluation method based on random forest is used to screen out key features sensitive to changes in equipment status and eliminate redundant features.

[0039] After feature extraction, the selected key features are arranged in a fixed order as set in steps 113 and 117. Simultaneously, feature components such as the safety factor of the axial tensile and compressive strength of key members and the geometric parameters of the members, calculated in steps 114 to 116, are added to form a sample health feature vector corresponding to each time point. Each sample health feature vector contains multi-dimensional feature information on the overall operating status of the equipment and the structural strength status of key members at that time point, and is completely consistent with the dimension, feature order, and feature type of the current-moment health feature vector generated in step 117. Next, a corresponding equipment status label is assigned to each sample health feature vector. The equipment status label is used to clarify the actual operating status of the ship's equipment corresponding to the sample health feature vector. The label assignment is based on the ship's past actual operation records, maintenance records, and fault records to ensure the accuracy and authenticity of the labels. Specifically, the labels are divided into three categories: Normal status label, abnormal status label, and fault status label are assigned as follows: When the device is operating normally without any abnormal signs at the time point corresponding to the health feature vector of the sample, and no problems are found in the subsequent maintenance records during that time period, the normal status label is assigned; when the device experiences a minor abnormality at the time point corresponding to the health feature vector of the sample, but it does not affect the normal operation of the device, and a minor potential problem is confirmed in the subsequent maintenance, the abnormal status label is assigned; when the device experiences a serious abnormality at the time point corresponding to the health feature vector of the sample, affecting the normal operation of the device, and subsequent fault shutdown or emergency maintenance occurs, the fault status label is assigned. All sample health feature vectors are associated with their corresponding device status labels one by one, and divided into training set, validation set, and test set according to a preset ratio (usually 7:2:1). The three sets together constitute a complete training sample set. After the training sample set is constructed, it is organized and archived according to the order of feature vectors and labels.

[0040] Step 221: Based on the training sample set, construct a deep learning ship equipment status prediction model. The model is trained using sample health feature vectors as input and equipment health status labels at corresponding times as output. The parameters of the deep learning ship equipment status prediction model are optimized by minimizing the prediction error to obtain the trained model. Specifically, this includes: determining the construction principles of the deep learning ship equipment status prediction model; the model construction must adapt to the temporal characteristics and multi-dimensional features of ship equipment operation data; it must be able to capture the correlation between sample health feature vectors and equipment status labels; it must also have strong generalization ability; and it must be able to adapt to the dynamic changes in equipment operation status under complex sea conditions, avoiding model errors. Overfitting (can only accurately predict training samples, unable to predict new real-time data) or underfitting (cannot accurately capture the correlation in samples, resulting in large prediction errors); Next, a deep learning ship equipment status prediction model is constructed in detail. This model adopts a multi-layer neural network structure and does not involve a complex model framework. The specific construction process is as follows: First, build the input layer. The number of neurons in the input layer is exactly the same as the dimension of the sample health feature vector. Each neuron corresponds to a feature component. The role of the input layer is to receive each feature value in the sample health feature vector and pass it completely to the next layer of the network to ensure that feature information is not lost; Second, build the hidden layers. The hidden layers are set with multiple layers ( Typically, there are 3-5 layers (adjustable based on the number of samples and feature dimensions). Each hidden layer has a certain number of neurons (the number of neurons is reasonably set according to the number of neurons in the input layer, usually 2-3 times the number of neurons in the input layer). Adjacent hidden layers are interconnected. Each neuron in a hidden layer receives information from the neurons in the previous layer and integrates and processes the received information. The processing uses simple numerical operations: first, the value transmitted by each neuron in the previous layer is multiplied by a preset initial weight value; then, all the weighted values ​​are added together, and a preset initial bias value is added to obtain the preliminary calculation result of that neuron. The preliminary calculation result is then processed... First, a simple nonlinear transformation is performed (to ensure the model can capture the nonlinear relationship between features and states). The transformation method is to retain positive values ​​in the initial calculation results and convert negative values ​​to 0 to ensure the rationality of the calculation results. Second, an output layer is built. The output layer has two neurons. The first neuron is used to output the device health score, and the second neuron is used to output the device abnormality probability value. Each neuron in the output layer is connected to all neurons in the previous hidden layer, receiving the calculation results passed from the previous layer. Then, through simple numerical calculations, the calculation results are converted into output values ​​that meet the actual needs (health score range is 0-100 points, and abnormality probability value range is 0-1).

[0041] After the model structure is built, the model parameters are initialized, including the weight values ​​between neurons in each hidden layer and neurons in the previous layer, the bias values ​​of each hidden layer, the weight values ​​between neurons in the output layer and neurons in the previous hidden layer, and the bias values ​​of the output layer. Initial weight values ​​are all set to random decimals between 0 and 1, and initial bias values ​​are all set to random decimals between 0 and 0.5 to ensure the rationality of the initial parameters and avoid model training failure due to excessively large or small initial parameters. Subsequently, the model training process is started. The training process uses the health feature vectors of the training set samples as input and the corresponding device status labels as output, inputting the training set samples into the built model batch by batch. The specific training logic is as follows. As follows, each feature component of the health feature vector of each sample is input into the corresponding neuron in the input layer. The input layer neuron directly transmits the feature values ​​to the first hidden layer. Each neuron in the first hidden layer receives the feature values ​​transmitted from the input layer, multiplies each input value by its corresponding weight value, adds all the products, and adds the bias value of that layer to obtain the preliminary calculation result of each neuron. The preliminary calculation result is then subjected to a non-linear transformation, and the transformed result is transmitted to the second hidden layer. The second hidden layer performs the same calculation and transformation on the received values ​​in the same way as the first hidden layer, and then transmits them to the next hidden layer, and so on, until the calculation result is transmitted to the next hidden layer. The results are passed to the output layer; the two neurons in the output layer receive the computation results from the previous hidden layer, and perform the same product-addition-plus-bias operation to obtain the preliminary output result. This preliminary output result is then converted into a health score between 0-100 and an anomaly probability value between 0-1, which are the model's predicted output values. During training, for each batch of training samples, the prediction error between the model's predicted value and the actual device status label corresponding to the training sample is calculated. The prediction error is calculated as follows: for the health score, the model's predicted health score is compared with the sample's actual health score (set according to the actual device status; normal status corresponds to 80-100 points, minor anomalies to...). The health prediction error for a single sample is calculated by squaring the difference between the model's predicted abnormality score (60-80 points for moderate abnormality, 40-60 points for moderate abnormality, and 0-40 points for severe abnormality). For the abnormality probability value, the difference between the model's predicted abnormality probability value and the actual abnormality probability value for that sample (0-0.2 for normal state, 0.2-0.4 for slight abnormality, 0.4-0.6 for moderate abnormality, and 0.6-1.0 for severe abnormality) is calculated, and then squared to obtain the abnormality probability prediction error for a single sample. The health prediction error and abnormality probability prediction error for all samples in a batch are summed, and then the two sums are added together to obtain the total prediction error for that batch of samples.

[0042] After obtaining the total prediction error, the model's parameters (weights and biases) are adjusted to reduce it. The adjustment logic is as follows: if the total prediction error is large, it indicates a significant deviation between the model's prediction and the actual state. In this case, the principle of adjusting parameters according to the larger the error is applied is followed. Specifically, the weights and biases of each layer are fine-tuned in the opposite direction to the prediction error. That is, when the predicted value is greater than the actual value, the corresponding weights and biases are appropriately decreased; when the predicted value is less than the actual value, the corresponding weights and biases are appropriately increased. If the total prediction error is small, it indicates that the model's prediction is close to the actual state. In this case, only minor adjustments are made to the parameters to avoid excessive adjustment that could lead to model instability. The process of inputting samples—calculating predicted values—calculating prediction errors—adjusting model parameters is repeated to train all training set samples in batches. After training, the health feature vectors of the validation set samples are input into the currently trained model, and the model's predicted values ​​are compared with the actual values ​​of the validation set. The validation error between the actual state labels is considered. If the validation error is small and the model's prediction accuracy on the validation set reaches the preset standard (usually above 90%), it indicates that the model training effect is good and can accurately capture the correlation between sample features and equipment status. If the validation error is large and the prediction accuracy does not reach the preset standard, the model parameters are adjusted, training batches are increased, and training is repeated until the validation error meets the requirements and the prediction accuracy reaches the standard. After the model training is qualified, the health feature vectors of the test set samples are input into the model to perform a final test on the model's prediction accuracy and generalization ability. If the test accuracy is close to the validation accuracy and both reach the preset standard, it indicates that the model training is complete, has good prediction performance, and can be used for subsequent real-time status prediction. If the test accuracy is significantly lower than the validation accuracy, it indicates that the model has an overfitting problem, and the model structure and parameters need to be readjusted and trained again until the test effect meets the requirements, finally obtaining a well-trained deep learning ship equipment status prediction model.

[0043] Step 222: Real-time acquisition of current operational data from ship equipment; preprocessing and feature extraction of the current operational data to generate a health feature vector for the current moment; specifically, this includes: activating all sensors on the ship equipment and, according to the acquisition standards of steps 110 and 114, acquiring multi-dimensional operational data of the ship equipment in real time. The sensor types and acquisition content are completely consistent with steps 110 and 114, including various monitoring values ​​such as vibration, temperature, pressure, axial load of key components, and strain. Simultaneously, the current operating condition information (sailing speed, sea state level) and timestamp are recorded to ensure the current... To ensure the real-time performance and accuracy of the running data, after collection, the current running data undergoes preprocessing and feature extraction operations identical to steps 110, 111, 112, 113, and 117. This ensures that the generated current-time health feature vector is completely consistent with the sample health feature vector generated in step 220 and the health feature vector generated in step 117 in terms of dimension, feature order, and feature type. This avoids inconsistencies in feature vectors that could prevent input into the trained model or lead to inaccurate prediction results. The specific operation process is as follows: First, the current real-time collected multidimensional running data... The data is cleaned by checking the current monitoring values ​​of each sensor one by one, removing abnormal values ​​caused by real-time noise and transmission interference, with the removal criteria consistent with steps 110 and 220. Secondly, if there are missing values ​​in the cleaned current operating data (e.g., a temporary malfunction of a sensor causing missing data at the current moment), linear interpolation is used to quickly fill them in, with the filling method consistent with the previous steps, ensuring the integrity of the current data. Thirdly, the current operating data is resampled according to a preset unified time frequency to ensure that the time frequency of the current data is consistent with the time frequency of the sample data during model training. Fourthly, feature extraction is performed on the time-aligned current data, extracting statistical features, frequency domain features, and time-frequency features, and key features are selected using a feature importance assessment based on random forest. Fifthly, the axial tensile and compressive strength safety factor of the key members at the current moment is calculated. This safety factor, the geometric parameters of the key members, and the current relevant monitoring data are integrated with the selected key features, arranged in a fixed order, to finally generate a health feature vector for the current moment. This feature vector contains all the key information about the overall operating status of the equipment and the structural strength status of the key members at the current moment, and can be directly input into the trained model for prediction.

[0044] Step 223 involves inputting the current health feature vector into the trained ship equipment status prediction model. Through forward propagation, the current equipment health score and anomaly probability value are obtained. Specifically, this includes confirming that the trained ship equipment status prediction model is operating normally, and that all model parameters (weights and biases at each layer) are optimal after training, without any parameter errors or loss. This ensures the model can normally receive input, perform calculations, and output prediction results. Then, the current health feature vector generated in step 222 is completely input into the input layer of the trained model. During input, it is ensured that each feature component corresponds to its corresponding neuron in the input layer, without feature misalignment or omission, ensuring the integrity and accuracy of the input data and avoiding prediction deviations due to input errors. After input, the model performs step-by-step calculations on the current health feature vector using the same forward propagation calculation method as in step 221. The specific calculation process is as follows. Each neuron in the input layer directly transmits the received feature components to the first hidden layer without any additional computation, ensuring complete transmission of feature information. Each neuron in the first hidden layer receives the feature values ​​transmitted from the input layer, multiplies each feature value by the weight value between that neuron and the corresponding neuron in the input layer, adds all the multiplied values ​​together, and adds the bias value of the first hidden layer to calculate the preliminary computation result for each neuron. Then, a non-linear transformation is performed on the preliminary computation result, retaining positive values ​​and converting negative values ​​to 0. After the transformation, the computation results of all neurons in this layer are transmitted to the second hidden layer. Each neuron in the second hidden layer receives the computation results transmitted from the previous layer according to the exact same computation logic as the first hidden layer, performs multiplication, addition, and bias addition operations and non-linear transformation, and then transmits the computation result to the next hidden layer. This process continues until the last hidden layer completes the computation and transmits the result to the output layer.

[0045] The two neurons in the output layer receive all the computation results from the last hidden layer. They calculate two preliminary output results using the same product-addition-plus-bias operation. These preliminary output results are then transformed to meet the actual prediction requirements. The first preliminary output result is converted into a value between 0 and 100, representing the current device health score. A higher score indicates better current operating status, more stable mechanical performance, and higher health. The second preliminary output result is converted into a value between 0 and 1, representing the current device anomaly probability value. A value closer to 1 indicates a higher probability of an anomaly or malfunction, while a value closer to 0 indicates normal operation and a lower probability of malfunction. After the computation, the model outputs the current device health score and anomaly probability value, and stores the two prediction results in association with the current timestamp, operating condition information, and current health feature vector.

[0046] Collect past operating data covering various working conditions, and adopt the same preprocessing process as real-time data to ensure that the sample features are consistent with the real-time features. The preprocessing and feature extraction of current operating data are completely consistent with the sample data processing process, which meets the needs of real-time monitoring and real-time prediction of ship equipment.

[0047] In a preferred embodiment of the present invention, step 3 above, based on real-time data of three core monitoring indicators in a unified time-series dataset, generates dynamic correction coefficients through multi-parameter correlation analysis and corrects the health assessment results and anomaly probability prediction values ​​to obtain corrected health assessment results and anomaly probability prediction values, may include:

[0048] In this embodiment of the invention, step 330 involves extracting the current-time data and continuous data within a preset time window of three core monitoring indicators—vibration amplitude, temperature, and pressure—from a unified time-series dataset to obtain a real-time monitoring data set. Specifically, this includes: defining the specific definitions of the three core monitoring indicators to ensure consistency with the terminology used in the preceding steps. Vibration amplitude refers to the maximum fluctuation value of the vibration signal collected by the vibration sensor, directly reflecting the intensity of vibration in the ship's equipment; temperature refers to the real-time temperature value of key parts of the equipment collected by the temperature sensor, reflecting the temperature state during equipment operation; and pressure refers to the real-time pressure value of the equipment's hydraulic and pneumatic systems collected by the pressure sensor, reflecting the operating state of the equipment's fluid circuits. These three core indicators are interrelated and mutually influential, comprehensively reflecting the current actual operating conditions of the equipment. Next, a fixed time window is preset, taking into account both data representativeness and real-time performance, typically set to 10 minutes. This time window provides a fixed time range benchmark for subsequent calculations of correlation coefficients and comparisons of deviations from operating conditions. Subsequently, data of the three core monitoring indicators are extracted in real-time from the unified time-series dataset constructed in step 111, ensuring the extraction process is accurate and complete. Without omitting any key data, the specific extraction process is as follows: First, extract the three core indicator data for the current moment: the vibration amplitude, temperature, and pressure values ​​corresponding to the current timestamp. Verify the accuracy of each data point to ensure that the extracted values ​​are completely consistent with the values ​​in the unified time-series dataset at the current moment, avoiding extraction errors. Second, extract continuous data within a preset time window. That is, trace back from the current moment and extract continuous data for the vibration amplitude, temperature, and pressure values ​​corresponding to each timestamp within the entire preset time window. Arrange these continuous data points sequentially according to their timestamps to ensure data continuity. To ensure data integrity, if any data remains missing after cleaning in step 110 within the time window, the missing data is extracted and filled to avoid data gaps. After extraction, all extracted data are integrated to form a real-time monitoring data set. This data set contains three independent continuous data sequences: a continuous vibration amplitude data sequence within a preset time window, a continuous temperature value data sequence within a preset time window, and a continuous pressure value data sequence within a preset time window. Each data sequence contains multiple timestamps and their corresponding monitoring values, and the timestamps of the three data sequences are completely consistent to ensure that pairwise correlation analysis can be performed.

[0049] Step 331: Based on the real-time monitoring data set, calculate the real-time correlation coefficients between each pair of vibration amplitude, temperature, and pressure values ​​at the current moment, generating a real-time correlation matrix. Simultaneously, retrieve the baseline correlation coefficients within the corresponding time window under standard operating conditions or fault-free operation from the pre-stored baseline operating condition feature library, generating a baseline correlation matrix. Specifically, this includes: determining the core role of the correlation coefficient, which is a numerical value used to quantify the degree of correlation between two monitoring indicators. Its value ranges from -1 to 1. The closer the value is to 1, the stronger the positive correlation between the two indicators; that is, when the value of one indicator increases, the value of the other indicator also increases. The closer the value is to -1, the stronger the positive correlation between the two indicators. The stronger the negative correlation between indicators, the more likely it is that when one indicator increases, the other decreases. The closer the value is to 0, the less correlation there is between the two indicators, indicating they are independent. By calculating the correlation coefficients between the three core indicators, the correlation between them under the current working condition can be accurately captured. Next, based on the real-time monitoring data set obtained in step 330, the real-time correlation coefficients between the three core indicators at the current moment are calculated. The specific calculation process is as follows: First, calculate the real-time correlation coefficient between vibration amplitude and temperature value. First, extract all vibration amplitude data and corresponding temperature value data within the preset time window from the real-time monitoring data set to ensure that the two sets of data are consistent. The quantities are completely identical (each timestamp corresponds to a set of vibration amplitude and temperature values); calculate the average value of the vibration amplitude data set, that is, add up all vibration amplitude values ​​within the preset time window, and then divide the sum by the total number of vibration amplitude data to obtain the average vibration amplitude; simultaneously, calculate the average value of the temperature data set, that is, add up all temperature values ​​within the preset time window, and then divide the sum by the total number of temperature value data to obtain the average temperature; then, calculate the difference between the vibration amplitude value corresponding to each timestamp and the average vibration amplitude, and then calculate the difference between the temperature value corresponding to that timestamp and the average temperature, multiply these two differences to obtain... The product of differences corresponding to the timestamps is calculated; all the product of differences corresponding to the timestamps within the preset time window are summed to obtain the sum of the product of differences; then, the square of the vibration amplitude difference corresponding to each timestamp is calculated, and all the squared values ​​are summed to obtain the sum of the squared vibration amplitude differences; the square of the temperature difference corresponding to each timestamp is calculated, and all the squared values ​​are summed to obtain the sum of the squared temperature differences; the sum of the squared vibration amplitude differences and the sum of the squared temperature differences are multiplied to obtain the product of squares, and then the square root of the product of squares is taken; the sum of the product of differences is divided by the square root of the product of squares to obtain the real-time correlation coefficient between the vibration amplitude and the temperature value.

[0050] The second step involves calculating the real-time correlation coefficient between vibration amplitude and pressure value. This calculation logic is identical to that used for the correlation coefficient between vibration amplitude and temperature. Vibration amplitude and pressure value data are extracted within a preset time window, and the average vibration amplitude and pressure value are calculated. The difference in vibration amplitude and pressure value corresponding to each timestamp is calculated, and these differences are multiplied to obtain a product. All difference products are then summed. The sum of squares of the vibration amplitude and pressure value differences is calculated, multiplied, and the square root of the product is obtained. Finally, the sum of the difference products is divided by this square root to obtain the vibration amplitude and pressure value difference. The third step is to calculate the real-time correlation coefficient between the amplitude and pressure values. The calculation logic is completely consistent with the previous two steps: extract the temperature and pressure data within the preset time window, calculate the average temperature and average pressure values; calculate the temperature and pressure differences corresponding to each timestamp, multiply them, and sum the products of the differences; calculate the sum of squares of the temperature and pressure differences, multiply them, and take the square root of the product of the sums of squares; divide the sum of the product of the differences by the square root to obtain the real-time correlation coefficient between the temperature and pressure values; these three real-time correlation coefficients... After calculation, a real-time correlation matrix is ​​generated. This matrix is ​​a 3x3 matrix, with each row and column corresponding to one of the three core monitoring indicators: vibration amplitude, temperature, and pressure. The value at each position in the matrix represents the real-time correlation coefficient between the corresponding two indicators. The value in the first row and first column is the correlation coefficient between vibration amplitude and itself (fixed to 1); the value in the first row and second column is the real-time correlation coefficient between vibration amplitude and temperature; the value in the first row and third column is the real-time correlation coefficient between vibration amplitude and pressure; and the value in the second row and first column is the real-time correlation coefficient between temperature and vibration amplitude (and...). The values ​​in the first row and second column are the same. The values ​​in the second row and second column are the correlation coefficients between the temperature value and itself (fixed to 1). The values ​​in the second row and third column are the real-time correlation coefficients between the temperature value and the pressure value. The values ​​in the third row and first column are the real-time correlation coefficients between the pressure value and the vibration amplitude (the same as the values ​​in the first row and third column). The values ​​in the third row and second column are the real-time correlation coefficients between the pressure value and the temperature value (the same as the values ​​in the second row and third column). The values ​​in the third row and third column are the correlation coefficients between the pressure value and itself (fixed to 1). This forms a complete real-time correlation matrix, which intuitively presents the correlation between the three core indicators under the current working conditions.

[0051] Simultaneously, the pre-stored benchmark operating condition feature library is retrieved. This library is pre-established and stored, containing benchmark correlation coefficients for different time windows corresponding to the ship's equipment under standard operating conditions, i.e., the normal operating conditions set during equipment design, such as rated sailing speed, operating conditions under stable sea states, or fault-free operating conditions. These benchmark correlation coefficients are obtained by collecting historical data under long-term normal operating conditions of the equipment and using the same calculation method as described above. They can reflect the standard correlation relationship of the three core indicators when the equipment is operating normally. Furthermore, the time windows in the benchmark operating condition feature library correspond to the time windows preset in step 330. Completely consistent; retrieve the benchmark correlation coefficients corresponding to the preset time window in step 330 from the benchmark operating condition feature library. When retrieving, ensure that the type of benchmark operating condition (standard operating condition or fault-free operation state) is consistent with the current design standard of the equipment. The three benchmark correlation coefficients retrieved are the benchmark correlation coefficients of vibration amplitude and temperature value under standard operating condition, the benchmark correlation coefficients of vibration amplitude and pressure value under standard operating condition, and the benchmark correlation coefficients of temperature and pressure value under standard operating condition. Then, according to the matrix structure that is exactly the same as the real-time correlation matrix, fill the three benchmark correlation coefficients into the corresponding positions to generate the benchmark correlation matrix.

[0052] Step 332 involves comparing the real-time correlation matrix with the benchmark correlation matrix element-by-element, calculating the deviation values ​​of the corresponding correlation coefficients, and combining these deviation values ​​to obtain a correlation deviation quantity that can quantify the degree of deviation of the current operating condition from the standard operating condition. Specifically, this includes determining the principle of element-by-element comparison: the value at each position in the real-time correlation matrix is ​​compared one-to-one with the value at the corresponding position in the benchmark correlation matrix to ensure the accuracy of the comparison and prevent misalignment. Because the two matrices have completely identical structures (3 rows and 3 columns) and correspond to the same core indicators, each element has a unique corresponding benchmark element. This allows for accurate reflection of the deviation between the current correlation and the standard correlation. Next, an element-by-element comparison is performed, and the deviation value is calculated. The specific steps are as follows: First, compare the elements in the first row and first column of the real-time correlation matrix with those in the benchmark correlation matrix. Both elements are 1 (correlation coefficient between themselves). Calculate the deviation value, which is the real-time element value minus the benchmark element value; the resulting deviation value is 0. Second, compare the elements in the first row and second column (correlation coefficient between vibration amplitude and temperature). Subtract the benchmark correlation coefficient in the benchmark correlation matrix from the real-time correlation coefficient at that position in the real-time correlation matrix to obtain the vibration amplitude. The first step is to compare the correlation coefficient between the vibration amplitude and the pressure value; the second step is to compare the correlation coefficient between the vibration amplitude and the pressure value with the elements in the first row and third column (correlation coefficient between vibration amplitude and pressure value), subtracting the baseline correlation coefficient from the real-time correlation coefficient to obtain the deviation value; the third step is to compare the correlation coefficient between the vibration amplitude and the pressure value with the elements in the second row and first column (correlation coefficient between temperature value and vibration amplitude), subtracting the baseline correlation coefficient from the real-time correlation coefficient, and obtaining the same deviation value as the deviation value in the second row and second column; the fifth step is to compare the elements in the second row and second column, both of which are 1, so the deviation value is 0; the sixth step is to compare the correlation coefficient between the temperature value and the pressure value with the elements in the second row and third column (correlation coefficient between temperature value and pressure value). The first step is to compare the correlation coefficients of the pressure and vibration amplitudes in the third row with the second column. The deviation value obtained by subtracting the reference correlation coefficient from the real-time correlation coefficient is the same as the deviation value in the third column of the first row. The second step is to compare the correlation coefficients of the pressure and temperature in the third row with the reference correlation coefficient. The deviation value obtained by subtracting the reference correlation coefficient from the real-time correlation coefficient is the same as the deviation value in the third column of the second row. The third step is to compare the elements in the third row and third column. Both elements are 1, so the deviation value is 0.

[0053] After calculating the deviation values, nine deviation values ​​are obtained. Three of these deviation values ​​are fixed at 0 (the deviation from their own correlation coefficients), while the remaining six are the deviations from the correlation coefficients between the three core indicators (each pair corresponds to one effective deviation value). These deviation values ​​are then processed to obtain the correlated deviation, which quantifies the degree of deviation between the current operating condition and the standard operating condition. The greater the deviation, the larger the correlated deviation; the smaller the deviation, the smaller the correlated deviation. The specific processing procedure is as follows: all nine calculated deviation values ​​are squared to obtain the square of each deviation value. The purpose of squaring is to eliminate the deviation values. Positive and negative effects (whether the deviation is positive or negative, it indicates a deviation from the standard value, and the squared value is positive) are considered. Larger deviations are amplified to highlight obvious deviations from the operating condition. The squares of all deviation values ​​are summed to obtain the sum of squares of deviations. Then, the average of the sum of squares of deviations is calculated by dividing the sum of squares of deviations by the total number of deviation values ​​(9). The square root of the average of squares of deviations is then taken to obtain the correlation deviation. This correlation deviation is a non-negative value. The larger the value, the further the correlation between the three core indicators deviates from the correlation under the standard operating condition, and the more unstable the current operating condition is. The smaller the value, the closer the current correlation is to the standard operating condition, and the more stable the current operating condition is.

[0054] Step 333: Based on the associated deviation, dynamically generate correction coefficients using preset mapping rules. Specifically, when the associated deviation exceeds a preset silent threshold range, a change in operating condition is determined, and a nonlinear correction coefficient is obtained. This includes: presetting a silent threshold range, which is set based on the normal operating characteristics of the ship's equipment, historical operating condition data, and actual monitoring experience. This range is used to determine whether the current operating condition has changed significantly. The silent threshold range is typically set to 0-0.2, indicating that when the associated deviation is within this range, the current operating condition is considered to have not changed significantly and remains in a stable state close to the standard operating condition. When the associated deviation exceeds this range (i.e., greater than 0.2), a significant change in the current operating condition is determined, and a nonlinear correction coefficient needs to be generated to specifically correct the initial predicted value. Simultaneously, a complete set of mapping rules is preset and stored. These rules are used to dynamically map and generate corresponding correction coefficients based on the specific value of the associated deviation. The correction coefficients are divided into linear and nonlinear correction coefficients, corresponding to scenarios where the operating condition has not changed significantly and scenarios where the operating condition has changed significantly, respectively. The core logic is that the smaller the correlation deviation, the closer the correction coefficient is to 1 (the smaller the correction magnitude, the closer the initial predicted value is to the actual state, and no significant correction is needed); the larger the correlation deviation, the greater the deviation of the correction coefficient from 1 (the greater the correction magnitude, the greater the deviation between the initial predicted value and the actual state, and a significant correction is needed). The specific mapping rules and correction coefficient generation process are as follows: First, determine whether the correlation deviation is within the preset silent threshold range. If the correlation deviation is greater than or equal to 0 and less than or equal to 0.2 (within the silent threshold range), then it is determined that no significant change has occurred in the current operating condition. The correlation between the three core indicators is close to that of standard operating conditions, and the initial predicted values ​​deviate little from the actual state of the equipment. At this point, a linear correction coefficient is generated. The logic for generating the linear correction coefficient is to subtract the correlation deviation from 1. For example, if the correlation deviation is 0.1, the linear correction coefficient is 1 minus 0.1, which equals 0.9; if the correlation deviation is 0, the linear correction coefficient is 1 minus 0, which equals 1. The value range of the linear correction coefficient is between 0.8 and 1. The smaller the correlation deviation, the closer the correction coefficient is to 1, and the smaller the correction magnitude.

[0055] The second step involves determining that if the correlation deviation exceeds 0.2 (exceeding the silent threshold range), the current operating condition has significantly changed, and the correlation between the three core indicators deviates considerably from the standard operating condition. The initial predicted value may not accurately reflect the actual state of the equipment. In this case, a non-linear correction coefficient needs to be generated. The logic for generating the non-linear correction coefficient is as follows: first, subtract the maximum silent threshold value (0.2) from the correlation deviation to obtain the deviation excess; then multiply the deviation excess by 2 to obtain the amplified excess; finally, subtract the amplified excess from 1 to obtain the non-linear correction coefficient. For example, if the correlation deviation is 0.3, the deviation excess is 0.3 minus 0.2, equaling 0.1; the amplified excess is 0.1 multiplied by 2, equaling 0.2; and the non-linear correction coefficient is 1. Subtracting 0.2 equals 0.8; if the correlation deviation is 0.4, then the deviation excess is 0.4 minus 0.2, equaling 0.2; the amplified excess is 0.2 multiplied by 2, equaling 0.4; the nonlinear correction coefficient is 1 minus 0.4, equaling 0.6. The value range of the nonlinear correction coefficient is between 0 and 0.8. The larger the correlation deviation, the smaller the correction coefficient and the larger the correction magnitude, ensuring that the prediction deviation caused by changes in working conditions can be corrected in a targeted manner; after the correction coefficient is generated, it is clear that the correction coefficient is used for subsequent correction of the health assessment results and the anomaly probability prediction values. That is, the same correction coefficient is applied to the correction of the two initial prediction values ​​respectively, ensuring the consistency of the correction logic. At the same time, the correction coefficient is associated with and stored one by one with the correlation deviation and the working condition judgment results at the current time.

[0056] Step 334 involves applying correction coefficients to the initial health assessment results and anomaly probability predictions, using weighted summation to correct the original predictions, ultimately obtaining corrected health assessment results and anomaly probability predictions that better reflect the current actual physical state of the equipment. Specifically, this includes: determining the initial data requiring correction, namely the current equipment health score and current equipment anomaly probability value obtained in step 223; verifying the accuracy of these two initial predictions to ensure they are the original values ​​output by the model in step 223, without any modification; and confirming the validity of the correction coefficients generated in step 333, ensuring that the values ​​of the correction coefficients conform to the mapping rules and match the current associated deviation and operating condition. Next... A weighted summation method is used, applying correction coefficients to the two initial predicted values ​​for correction. The core logic of the weighted summation is to calculate the corrected predicted value based on the initial predicted value and the weights of the correction coefficients, ensuring that the corrected predicted value better reflects the actual physical state of the equipment under the current operating conditions. The specific correction process is as follows: the health score and the anomaly probability value are corrected in two steps. First, the initial health assessment result is corrected to obtain the corrected health assessment result. Two weights for the weighted summation are determined: the first weight is the correction coefficient generated in step 333, and the second weight is 1 minus the correction coefficient. The sum of the two weights is 1, ensuring the rationality of the weighted summation. The weighted summation operation involves multiplying the initial health score by the first weight (correction coefficient) to obtain a weighted value for the initial health score; then multiplying the preset health baseline value by the second weight (1 minus the correction coefficient) to obtain a weighted value for the health baseline value. The health baseline value is a preset standard value for the equipment's health during normal operation, typically set at 90 points (corresponding to a fault-free and stable operating state). This baseline value is set based on equipment design standards and long-term normal operation data, and is used to provide a correction benchmark for the health score when operating conditions change. The weighted value of the initial health score is added to the weighted value of the health baseline value, and the sum obtained is the corrected health assessment result. For example... If the initial health score is 85, the correction coefficient is 0.9, the second weight is 1 minus 0.9 equals 0.1, and the baseline health score is 90, then the initial weighted health score is 85 multiplied by 0.9 equals 76.5, the baseline health score is 90 multiplied by 0.1 equals 9, and the corrected health assessment result is 76.5 plus 9 equals 85.5. If the correction coefficient is 0.8 (due to changes in working conditions), the initial health score is 70, the second weight is 0.2, then the initial weighted health score is 70 multiplied by 0.8 equals 56, the baseline health score is 90 multiplied by 0.2 equals 18, and the corrected health assessment result is 56 plus 18 equals 74.

[0057] The second step is to correct the initial anomaly probability prediction value to obtain the corrected anomaly probability prediction value. The correction logic is completely consistent with the correction logic of the health score, and it also adopts a weighted summation method. The specific calculation process is as follows: Determine two weights. The first weight is the correction coefficient generated in step 333, and the second weight is 1 minus the correction coefficient. The sum of the two weights is 1. Multiply the initial anomaly probability value by the first weight (correction coefficient) to obtain the weighted value of the initial anomaly probability value. Then multiply the preset anomaly probability benchmark value by the second weight (1 minus the correction coefficient) to obtain the weighted value of the anomaly probability benchmark value. The anomaly probability benchmark value is a preset standard value of anomaly probability when the equipment is running normally, usually set to 0.1. This benchmark value is set based on historical data of the equipment running without faults and is used to provide a correction benchmark for the anomaly probability value when the operating conditions change. Add the weighted value of the initial anomaly probability value to the weighted value of the anomaly probability benchmark value. The sum obtained is the corrected anomaly probability prediction. For example, if the initial anomaly probability is 0.3, the correction coefficient is 0.9, the second weight is 0.1, and the baseline anomaly probability is 0.1, then the weighted initial anomaly probability is 0.3 multiplied by 0.9, which equals 0.27; the weighted baseline anomaly probability is 0.1 multiplied by 0.1, which equals 0.01; and the corrected anomaly probability prediction is 0.27 plus 0.01, which equals 0.28. If the correction coefficient is 0.6, the initial anomaly probability is 0.5, and the second weight is 0.4, then the weighted initial anomaly probability is 0.5 multiplied by 0.6, which equals 0.3; the weighted baseline anomaly probability is 0.1 multiplied by 0.4, which equals 0.04; and the corrected anomaly probability prediction is 0.3 plus 0.04, which equals 0.34. After correction, the corrected health assessment result and the corrected anomaly probability prediction are obtained. These two corrected values ​​are then associated and stored with the current timestamp, the associated deviation, the correction coefficient, and the initial prediction value to ensure data traceability.

[0058] Focusing on three core monitoring indicators—vibration amplitude, temperature, and pressure—continuous data within a preset time window is extracted to avoid correction deviations caused by incomplete data. By calculating the correlation coefficients between the three core indicators in detail, a real-time correlation matrix is ​​generated and compared with a benchmark correlation matrix to reduce the probability of false alarms or missed alarms.

[0059] In a preferred embodiment of the present invention, step 4 above, based on the corrected health assessment result and the predicted abnormality probability value, combines a dynamic adaptive threshold judgment strategy to perform real-time early warning judgment; when the predicted value exceeds the dynamic early warning threshold, an alarm message is generated, which may include:

[0060] In this embodiment of the invention, step 440 involves, based on the corrected health assessment result and the anomaly probability prediction value, merging the health assessment result as the latest data point with the corrected health assessment result stored in the database from previous moments to form a health assessment time series including the current and past states. Specifically, this includes: obtaining the current health assessment result and anomaly probability prediction value of the equipment after correction in step 3, wherein the health assessment result is a specific value that quantifies the current performance degradation of the equipment, and the anomaly probability prediction value is a specific value that characterizes the probability of the equipment currently experiencing a failure; determining the current time node, using the corrected health assessment result corresponding to this time node as the latest data point, and clarifying the time information and health value corresponding to this data point; and calling the database of the ship equipment maintenance management platform to extract data from the database. All previously stored health assessment results (i.e., every monitoring moment before the current moment) that have been corrected in step 3 are stored to ensure that each extracted health assessment result from a previous moment corresponds to a specific time node and uses the same quantitative standards and terminology definitions as the health assessment result at the current moment, thus ensuring data consistency. All extracted corrected health assessment results from previous moments are arranged in order from earliest to latest time nodes, and the health assessment result at the current moment is added to the end of this sequence to form a complete health assessment time series that includes health data from the current moment and all previous monitoring moments. Each data point in this time series corresponds to a unique time node and a corresponding corrected health assessment value, which can fully reflect the change process of the device's health status from the past to the present.

[0061] Step 441: Select the most recent N corrected health assessment results from the health assessment time series, calculate the arithmetic mean of these N results as the current health baseline value, and calculate the standard deviation of these N results as the current health fluctuation range. Specifically, this includes: determining the selected time window parameter N, where N is a preset positive integer. Its value can be set according to the type of ship equipment, operating conditions, and the accumulation of historical monitoring data to ensure that the selected N health data accurately reflect the recent health status changes of the equipment, avoiding data lacking representativeness due to N being too small, or failing to capture recent health fluctuations of the equipment in a timely manner due to N being too large; from the health assessment time series formed in step 440, select the N corrected health assessment results whose time nodes are closest to the current time. The selection process strictly follows the order from latest to earliest time to ensure that the latest N health data are selected, and past health data exceeding the N time range are removed; calculate the corrected health assessment results of these N times. The arithmetic mean of the estimated results is used as the baseline health value at the current moment. The specific calculation process involves summing the values ​​of all N selected health assessment results, dividing this sum by the number of selected moments N, and obtaining the quotient, which is the baseline health value at the current moment. This baseline value reflects the average level of the equipment's recent health status. The standard deviation of these N corrected health assessment results is calculated as the range of health fluctuations at the current moment. The specific calculation process involves first calculating the difference between each selected health assessment result and the baseline health value calculated above, then squaring each difference, summing all the squared differences, dividing this sum by the number of selected moments N, and obtaining the quotient, which is the variance. The square root of this variance is then taken to obtain the standard deviation. This standard deviation reflects the fluctuation range of the equipment's recent health status; a larger standard deviation indicates more drastic fluctuations in the equipment's recent health status, while a smaller standard deviation indicates a more stable recent health status.

[0062] Step 442: Based on the health baseline value and health fluctuation range, and according to the preset conversion rules, dynamically generate the current abnormal probability warning threshold, so that the warning threshold adaptively follows the normal decline trend of the equipment and changes in operating conditions. Specifically, this includes: determining the preset conversion rules, which are pre-set based on historical operating data, fault case data, and maintenance experience of the ship's equipment. These rules establish a correspondence between the health baseline value and health fluctuation range and the abnormal probability warning threshold. Furthermore, the conversion rules can be dynamically adjusted according to changes in the equipment's operating status and maintenance feedback data to ensure the rationality and accuracy of the conversion; obtaining the current health baseline value and health fluctuation range (standard deviation) calculated in Step 441, ensuring that the calculation of these two values ​​is accurate and that the units and quantification standards are consistent with the requirements of the conversion rules; and substituting the health baseline value and health fluctuation range into the conversion process according to the preset conversion rules. Specifically, the conversion process is based on the normal decline characteristics of the equipment. When the health baseline value decreases (i.e., the equipment performance deteriorates), the warning threshold is adjusted accordingly. When the health status declines, the corresponding anomaly probability warning threshold will be appropriately lowered to improve the sensitivity of the warning and avoid missed alarms. When the health status fluctuation range (standard deviation) increases (i.e., the equipment health status fluctuates more), the corresponding anomaly probability warning threshold will be appropriately lowered to capture potential anomaly risks in a timely manner. When the health status benchmark value increases and the fluctuation range decreases, the corresponding anomaly probability warning threshold will be appropriately increased to reduce false alarms. Specifically, the conversion can be performed by multiplying the health status benchmark value by a preset benchmark coefficient and the health status fluctuation range by a preset fluctuation coefficient, and then adding or subtracting the two products to obtain the anomaly probability warning threshold at the current moment. The generated anomaly probability warning threshold is stored in the database as the standard for warning judgment at the current moment. This warning threshold is not a fixed value, but is dynamically updated with the changes in the health status benchmark value and fluctuation range at each monitoring moment. It can adaptively follow the normal decline trend of the equipment and the changes in complex operating conditions to ensure the rationality and pertinence of the warning threshold.

[0063] Step 443: Compare the corrected anomaly probability prediction value with the current warning threshold to obtain a comparison result; and based on the comparison result, determine whether the preset alarm triggering conditions are met, including the corrected anomaly probability prediction value exceeding the dynamic warning threshold multiple times consecutively, or exceeding the dynamic warning threshold once with an exceedance exceeding a preset range; specifically, this includes: obtaining the corrected anomaly probability prediction value at the current time in Step 3, and the anomaly probability warning threshold at the current time generated in Step 442, ensuring that the quantification standards and units of the two values ​​are consistent to avoid errors in the comparison result due to inconsistent standards; directly comparing the corrected anomaly probability prediction value with the current anomaly probability warning threshold to clarify the relationship between the two and obtain a comparison result, specifically divided into three cases: the corrected anomaly probability prediction value is less than the warning threshold, the corrected anomaly probability prediction value is equal to the warning threshold, and the corrected anomaly probability prediction value is greater than the warning threshold; retrieving the preset alarm triggering conditions, which are pre-set based on the importance of the ship's equipment, the scope of the fault's impact, and the maintenance requirements, clarifying two core triggering scenarios. The first scenario involves the corrected anomaly probability prediction value exceeding the dynamic warning threshold multiple times consecutively. The number of consecutive times is a preset value. Specifically, the corrected anomaly probability prediction value and the corresponding warning threshold for the most recent consecutive monitoring times are retrieved, and each prediction value is compared with the warning threshold. If the prediction value for multiple consecutive times is greater than the warning threshold for the corresponding time, this trigger scenario is met. The second scenario involves the corrected anomaly probability prediction value exceeding the dynamic warning threshold once, and the exceedance exceeds a preset range. The exceedance is calculated as (corrected anomaly probability prediction value - current warning threshold) ÷ current warning threshold × 100%, where the preset range is a pre-defined percentage. Specifically, the exceedance is calculated first, and then compared with the preset range. If the exceedance is greater than the preset range, this trigger scenario is met. Based on the comparison results and the alarm triggering conditions, a comprehensive judgment is made as to whether the alarm triggering conditions are met. If either of the two scenarios is met, the alarm triggering conditions are met; if neither scenario is met, the alarm triggering conditions are not met.

[0064] Step 444: If the alarm triggering conditions are met, generate alarm information, including equipment identifier, anomaly occurrence time, anomaly probability value, relevant monitoring indicator data, and recommended maintenance level. Specifically, this includes: confirming that the judgment result of step 443 meets the alarm triggering conditions, immediately initiating the alarm information generation process to ensure timely generation and push of alarm information; collecting all specific data required for the alarm information, where the equipment identifier is the unique identification number of the ship's equipment, which can be retrieved from the equipment information database of the ship maintenance management platform to ensure accurate location of the specific equipment experiencing the anomaly; the anomaly occurrence time is the current specific time, consistent with the time nodes in the health assessment time series to ensure accurate recording of the specific time of the anomaly; the anomaly probability value is the predicted anomaly probability value for the current time after correction in step 3, directly retrieving this value and filling it into the alarm information to ensure numerical accuracy. Correct; the relevant monitoring index data are real-time data of core monitoring indicators related to the equipment anomaly, specifically including vibration amplitude, temperature value, pressure value used to generate dynamic correction coefficients in step 3, as well as axial load data of key components, axial tensile and compressive strength safety factors, etc. The current time data of these indicators are retrieved in real time from a unified time-series dataset to ensure data integrity and real-time performance; the recommended maintenance level is pre-set based on the magnitude of the anomaly probability value, the degree of anomaly of the relevant monitoring indicators, and the importance of the equipment, and is divided into different levels. The higher the anomaly probability value and the more serious the anomaly of the monitoring indicators, the higher the recommended maintenance level. In specific judgment, the current anomaly probability value is compared with the preset maintenance level classification standard, and the corresponding recommended maintenance level is determined in combination with the anomaly of the relevant monitoring indicators; the above-collected data are organized and integrated according to the preset format to form complete alarm information.

[0065] By constructing a health assessment time series, integrating current and past health data, the system reflects the continuous changing trend of equipment health status, avoiding bias caused by relying solely on single-point data for early warning judgment. Based on the health baseline value and fluctuation range, the system dynamically generates anomaly probability early warning thresholds, enabling the early warning thresholds to adaptively follow the normal decline trend of equipment and complex operating conditions, thereby improving the accuracy and relevance of early warnings.

[0066] In a preferred embodiment of the present invention, step 5 above, which initiates a maintenance coordination and management response process based on alarm information, integrates the alarm information into the ship maintenance management platform, generates a maintenance work order, and recommends maintenance time windows and personnel configuration schemes, forming a maintenance execution record; simultaneously, it dynamically optimizes the maintenance plan by combining existing early warning records and maintenance feedback data, forming a closed-loop maintenance business management process, may include:

[0067] In this embodiment of the invention, step 550 involves automatically pushing alarm information to the ship maintenance management platform, extracting the equipment type, fault mode, and anomaly occurrence time, and generating a maintenance work order including a detailed task description based on the ship's current navigation status data. Specifically, after determining in step 4 that the alarm triggering conditions are met and generating alarm information, the system automatically initiates the alarm information push process. Through the ship's internal communication link, the complete alarm information is pushed to the ship maintenance management platform in real time, ensuring no delay or data loss during the push process, and guaranteeing that maintenance personnel can promptly obtain equipment anomaly information. Upon receiving the alarm information, the ship maintenance management platform automatically parses the alarm information, extracting three core key pieces of information: equipment type, fault mode, and anomaly occurrence time. The equipment type specifically refers to the category of the ship's equipment experiencing the anomaly; the fault mode specifically refers to the specific manifestation of the equipment anomaly; and the anomaly occurrence time specifically refers to the... The current time of alarm information generation must be completely consistent with the time of anomaly occurrence in step 4. The system automatically retrieves the ship's current navigation status data, including the ship's current speed, heading, location in the sea area, sea state level, cargo load, and current navigation task, ensuring that the acquired navigation status data is updated in real time and accurately reflects the ship's current actual operating status. The extracted equipment type, fault mode, and anomaly occurrence time are integrated with the retrieved ship's current navigation status data and combined with the preset work order generation specifications to generate a complete maintenance work order. This maintenance work order must include a detailed task description, which must clearly state the specific circumstances of the equipment anomaly, the maintenance operations to be performed, and the basic requirements of the maintenance work. The work order generation time and work order number must also be marked to ensure that the content of the maintenance work order is complete, clear, and specific, avoiding maintenance errors due to ambiguous work order information.

[0068] Step 551: Based on the maintenance work order, obtain the ship's navigation plan, the current availability status of the equipment, and the stored past maintenance data. Calculate and recommend the final maintenance time window using a preset optimization algorithm. Specifically, this includes: using the maintenance work order generated in step 550 as the core basis, determining the equipment name, equipment type, and urgency of the maintenance task corresponding to the work order. The system automatically retrieves the ship's complete navigation plan for the current and future period from the ship navigation management system. This navigation plan includes the ship's future route, expected speed, expected arrival time at each port, port stay duration, anchoring plan, and navigation task arrangements, ensuring that the obtained navigation plan is complete and accurate. This reflects the ship's future operational arrangements; simultaneously, it retrieves the current availability status data of the malfunctioning equipment. The current availability status specifically includes whether the equipment is currently operating, its operating load, whether it can be temporarily shut down, and whether shutdown would affect the ship's normal navigation, determining whether the equipment currently meets the basic conditions for maintenance. Furthermore, it retrieves the equipment's stored past maintenance data from the ship's maintenance management platform database. This past maintenance data includes the equipment's historical maintenance time, maintenance content, maintenance duration, problems encountered during maintenance, post-maintenance operational effects, and historical maintenance work orders, ensuring the completeness and relevance of past maintenance data. The retrieved ship navigation plan and current equipment availability data will then be used to determine whether the equipment meets the basic conditions for maintenance. The system integrates three types of data: equipment availability status, past maintenance data, and current availability. These are then fed into a pre-defined optimization algorithm to determine the recommended maintenance time window. The specific calculation process is as follows: Based on the ship's navigation plan, time periods when the ship is in a non-emergency navigation state and has the conditions for maintenance are selected, excluding time periods when the ship is in an emergency navigation state or in severe sea conditions, making maintenance impossible. Considering the current availability status of the equipment, it is determined whether the equipment can be shut down for maintenance within the selected time periods. If the equipment cannot be shut down, further selection is made for interval time periods with low equipment load that do not affect the ship's normal navigation. Finally, combined with past maintenance data, the average duration of each maintenance operation historically for this equipment is calculated and used as a filter. The total duration of the selected available time slots is subtracted from the average maintenance duration to obtain the effective time available for maintenance operations. This ensures that the effective time is sufficient to complete the maintenance task. Simultaneously, the selected available time slots are prioritized based on the urgency of the maintenance work order; higher urgency results result in higher priority time slots. Combining all the above filtering and calculation results, one or more final maintenance time slots are determined as recommended final maintenance time windows. These recommended maintenance time windows must clearly indicate the specific start and end times, along with an explanation of the reasons for recommending the time slot. This ensures that the recommended time windows are scientific and reasonable, neither affecting the ship's normal navigation nor hindering the smooth completion of maintenance operations.

[0069] Step 552: Based on the technical requirements of the maintenance task in the maintenance work order and the recommended maintenance time window, the system calls the maintenance personnel skill level database and personnel availability information to automatically match personnel available within the maintenance time window who possess the corresponding skill level, generating a recommended personnel configuration plan. Specifically, this includes: carefully analyzing the maintenance work order generated in step 550, extracting the technical requirements of the maintenance task, which specifically include the professional skills, skill proficiency, required equipment knowledge, and the number of personnel required to complete the maintenance task, thus determining the specific personnel requirements for this maintenance operation; simultaneously, determining the maintenance time window recommended in step 551, specifying the start and end times of the maintenance operation, and using this as the time basis for matching personnel availability; the system automatically calls the maintenance personnel skill level database in the ship maintenance management platform, which stores basic information, skill levels, areas of expertise, skill assessment scores, and past maintenance performance of all maintenance personnel, ensuring that the information in the database is updated in real time and is accurate; and simultaneously, calling the maintenance personnel availability information, which includes the current work of all maintenance personnel. The availability of each maintenance worker within the recommended maintenance time window is determined by considering their current status, future work schedule, and available time. Personnel matching then begins, following this process: First, based on the technical requirements of the maintenance task, personnel with the corresponding professional skills and meeting the required skill level are selected from the maintenance personnel skill level database, excluding those who do not meet the skill standards or are not proficient in maintaining this type of equipment. From the selected qualified personnel, those with available time within the recommended maintenance time window and able to participate fully in the maintenance work are further selected, excluding those who cannot be present within the time window or whose work schedules conflict. Next, based on the number of personnel specified in the technical requirements of the maintenance task, the personnel selected in the above two rounds are prioritized based on their past maintenance performance and skill proficiency, prioritizing those with high skill levels, good past performance, and expertise in maintaining this type of equipment. Based on the ranking results, a specific list of personnel participating in this maintenance operation is determined, clarifying the responsibilities of each person, and noting each person's skill level, area of ​​expertise, and arrival time, thus forming a recommended personnel configuration plan.

[0070] Step 553: After performing the maintenance work according to the recommended personnel configuration plan, collect the actual operation content, time consumption, and spare parts replacement information of this maintenance, and enter it into the database of the ship maintenance management platform to form a complete maintenance execution record associated with the maintenance work order. Specifically, this includes: maintenance personnel performing the maintenance work strictly according to the task description and technical requirements in the maintenance work order within the recommended maintenance time window, following the personnel configuration plan recommended in Step 552. During the maintenance work, a dedicated recorder is assigned to record all specific information of the maintenance work in real time, ensuring that the recorded information is true, accurate, and complete; collecting the actual operation content of this maintenance work, specifically including each operation step actually performed by the maintenance personnel, any additional problems discovered during the operation, and the measures taken to deal with these additional problems, ensuring that the operation content is recorded in detail and can completely reconstruct the entire process of this maintenance work; recording the actual time consumption of this maintenance work. The actual maintenance time includes the start and end times of the maintenance operation. The total maintenance time is calculated by subtracting the start time from the end time. The time for each specific operation step is also recorded to ensure accuracy and provide a basis for subsequent statistical analysis of maintenance efficiency. Furthermore, information on spare parts replaced during the maintenance operation is collected, including the spare part name, model, quantity, specifications, replacement location, and reason for replacement. If no spare parts are replaced during the maintenance operation, this must also be clearly recorded to ensure complete spare part information for subsequent spare part management and cost accounting. The collected actual operation content, actual time, spare part replacement information, as well as the names of maintenance personnel and the acceptance results of the maintenance operation, are all entered into the database of the ship maintenance management platform. During entry, the corresponding maintenance work order number must be associated to ensure a one-to-one correspondence between the maintenance execution record and the maintenance work order, forming a complete maintenance execution record.

[0071] Step 554: Periodically summarize existing alarm records and maintenance execution records stored in the ship maintenance management platform. Through statistical analysis, identify maintenance patterns and optimizable aspects. Based on the statistical analysis results, dynamically adjust the maintenance cycle, early warning threshold settings, and resource allocation strategies in the maintenance plan to achieve closed-loop maintenance business management optimization. Specifically, this includes: pre-setting the summary analysis cycle, which can be set according to the ship's sailing frequency, equipment operating conditions, and operation and maintenance management needs, ensuring a reasonable summary cycle that can promptly identify problems during maintenance without excessively increasing the workload; within each pre-set summary cycle, the system automatically summarizes all stored alarm records and maintenance execution records from the ship maintenance management platform's database. The alarm records include information such as the device identifier, occurrence time, probability, fault mode, and alarm level of all historical alarms. Maintenance execution records include information such as the actual operation content, time consumed, replacement parts, maintenance personnel, and maintenance results of all historical maintenance work orders. This ensures that the compiled records are complete and without omissions, and that the two types of records correspond one-to-one (i.e., the maintenance execution record corresponding to each alarm record can be accurately found). Statistical analysis is then performed on the compiled alarm records and maintenance execution records. The specific analysis process is as follows: First, the alarm frequency of different equipment types is calculated by dividing the total number of alarms for a certain type of equipment by the total operating time of that type of equipment to obtain the alarm frequency per unit time for that type of equipment, thus identifying the equipment types with higher alarm frequencies. The analysis focuses on four key aspects: First, monitoring the operational status of this type of equipment; second, statistically analyzing the occurrence frequency and handling effectiveness of different fault modes. The probability of a fault mode is calculated by dividing the total occurrence frequency of a particular fault mode by the total occurrence frequency of all fault modes. The runtime after each maintenance for that fault mode is also analyzed to assess its handling effectiveness and identify fault modes with poor handling effects and a high recurrence rate. Third, the average maintenance time is calculated by dividing the total maintenance time by the total number of maintenance operations. The differences in maintenance time among different maintenance personnel and equipment types are analyzed to identify areas with low maintenance efficiency. Fourth, the frequency and quantity of spare parts replacements are calculated by dividing the total replacement quantity of a particular type of spare parts by the total replacement quantity of that type of spare parts. Using the total maintenance time, the replacement frequency of this type of spare parts is obtained, identifying easily worn and frequently replaced spare parts. Through the above statistical analysis, the patterns and optimizable aspects of the maintenance process are clarified. Based on the results of the above statistical analysis, the existing maintenance plan is dynamically adjusted. The specific adjustments include: first, adjusting the maintenance cycle. For equipment with high alarm frequency and prone to failure, the maintenance cycle is shortened, while for equipment with stable operation and low alarm frequency, the maintenance cycle is appropriately extended to reduce unnecessary maintenance work; second, adjusting the warning threshold setting. Based on the failure occurrence patterns and equipment operating characteristics obtained from the statistical analysis, the preset conversion rules of the dynamic adaptive threshold in step 4 are adjusted to make the warning threshold more closely match the actual operating conditions of the equipment, reducing false alarms and missed alarms.Third, adjust resource allocation strategies. For equipment with high alarm frequency and maintenance needs, increase the number of maintenance personnel and spare parts. For areas with low maintenance efficiency, strengthen skills training for maintenance personnel and optimize personnel allocation to ensure the rational use of maintenance resources. Through these dynamic adjustments, achieve closed-loop maintenance business management of early warning, maintenance, feedback, and optimization, improve the efficiency and quality of ship equipment maintenance management, and reduce operation and maintenance costs.

[0072] By automatically pushing alarm information and generating detailed maintenance work orders, seamless integration of alarm information and maintenance management is achieved, avoiding blind maintenance operations. The system recommends maintenance time windows based on ship navigation plans, equipment status, and past maintenance data, reducing navigation impact and increased maintenance costs caused by inappropriate maintenance timing.

[0073] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for dynamic analysis and early warning of ship equipment maintenance data, characterized in that, The method includes: Step 1: Process the multi-dimensional operational status time-series data of the ship's equipment to construct a unified time-series dataset and extract health feature vectors reflecting the mechanical performance and overall operational status of the equipment. For key structural members in the equipment, calculate the axial stress based on real-time monitored axial load data and member geometric parameters, and combine this with the allowable stress of the material to obtain the axial tensile and compressive strength safety factor, forming a health feature vector. The health feature vector includes the root mean square value of the vibration signal, the dominant frequency energy of the spectrum, and the changing trends of temperature and pressure. Specifically, this includes: real-time acquisition of axial load data borne by the members through strain sensors or load sensors arranged on the key members, and simultaneously... The geometric parameters and material identifiers of the members are obtained from the equipment design parameter database. The geometric parameters include at least the cross-sectional area, length, and moment of inertia. Based on the axial load data and cross-sectional area, the axial stress value of the member under the current working condition is calculated. According to the material identifier, the allowable stress value of the corresponding material is retrieved from the preset material property database. The allowable stress value is compared with the axial stress value, and the axial tensile and compressive strength safety factor is calculated according to the ratio. The axial tensile and compressive strength safety factor is used as a feature component in the health feature vector and combined with the geometric parameters and related monitoring data to form a health feature vector that can comprehensively characterize the structural strength state of key members. Step 2: Input the health feature vector into the pre-trained deep learning ship equipment status prediction model to obtain the health assessment results and anomaly probability prediction values. Step 3: Based on the real-time data of the three core monitoring indicators in the unified time series dataset, generate dynamic correction coefficients through multi-parameter correlation analysis and correct the health assessment results and abnormal probability prediction values ​​to obtain the corrected health assessment results and abnormal probability prediction values. Step 4: Based on the corrected health assessment results and the predicted abnormality probability, a real-time early warning judgment is made in conjunction with a dynamic adaptive threshold judgment strategy; when the predicted value exceeds the dynamic early warning threshold, an alarm message is generated. Step 5: Initiate the maintenance coordination and management response process based on the alarm information, integrate the alarm information into the ship maintenance management platform, generate maintenance work orders, recommend maintenance time windows and personnel configuration schemes, and form maintenance execution records; at the same time, combine existing early warning records and maintenance feedback data to dynamically optimize the maintenance plan and form a closed-loop maintenance business management process.

2. The method for dynamic analysis and early warning of ship equipment maintenance data according to claim 1, characterized in that, The multidimensional operational status time-series data of ship equipment is processed to construct a unified time-series dataset, and health feature vectors reflecting the mechanical performance and overall operational status of the equipment are extracted, including: Data cleaning is performed on the multidimensional operating status time series data collected from different sensors to remove outliers caused by noise or transmission errors, and linear interpolation is used to fill in the missing data to obtain the cleaned time series data. Based on the cleaned time-series data, resampling is performed according to a unified time frequency to achieve time alignment of multi-source data and form a structured unified time-series dataset. Based on the unified time-series dataset, feature parameters that can characterize the mechanical performance and overall operating status of the equipment are extracted to obtain an initial feature set; Based on the initial feature set, key features are selected using a random forest-based feature importance assessment to form a healthy feature vector.

3. The method for dynamic analysis and early warning of ship equipment maintenance data according to claim 2, characterized in that, The characteristic parameters include statistical characteristics reflecting the data distribution characteristics, frequency domain characteristics reflecting the signal frequency composition, and time-frequency characteristics reflecting the local time-frequency characteristics of the signal.

4. The method for dynamic analysis and early warning of ship equipment maintenance data according to claim 3, characterized in that, The health feature vector is input into a pre-trained deep learning ship equipment condition prediction model to obtain health assessment results and anomaly probability prediction values, including: The ship's equipment operation data under past working conditions is collected as sample operation data. The sample operation data is preprocessed and multi-dimensional health features are extracted to form a sample health feature vector. At the same time, the equipment status label corresponding to the sample health feature vector is obtained to construct a training sample set. Based on the training sample set, a deep learning ship equipment status prediction model is constructed. The model is trained by taking the health feature vector of the sample as input and the equipment health status label at the corresponding time as output. The parameters of the deep learning ship equipment status prediction model are optimized by minimizing the prediction error to obtain the trained ship equipment status prediction model. Real-time acquisition of current operating data of ship equipment; preprocessing and feature extraction of current operating data to generate a health feature vector for the current moment; The health feature vector at the current moment is input into the trained ship equipment status prediction model, and the equipment health score and anomaly probability value at the current moment are obtained through forward propagation calculation.

5. The method for dynamic analysis and early warning of ship equipment maintenance data according to claim 4, characterized in that, The health score is used to quantify the degree of equipment performance degradation, and the anomaly probability value is used to characterize the likelihood of equipment failure.

6. The method for dynamic analysis and early warning of ship equipment maintenance data according to claim 5, characterized in that, Based on real-time data from three core monitoring indicators in a unified time-series dataset, dynamic correction coefficients are generated through multi-parameter correlation analysis to correct the health assessment results and anomaly probability predictions, resulting in corrected health assessment results and anomaly probability predictions, including: Real-time monitoring data sets are obtained by extracting the current moment data of three core monitoring indicators—vibration amplitude, temperature value, and pressure value—as well as continuous data within a preset time window from a unified time series dataset. Based on the real-time monitoring data set, the real-time correlation coefficients between each pair of vibration amplitude, temperature value and pressure value at the current moment are calculated to generate a real-time correlation matrix; at the same time, the benchmark correlation coefficients of the equipment in the corresponding time window under standard working conditions or fault-free operation are retrieved from the pre-stored benchmark working condition feature library to generate a benchmark correlation matrix. The real-time correlation matrix is ​​compared element by element with the benchmark correlation matrix, the deviation value of each corresponding correlation coefficient is calculated, and these deviation values ​​are combined to obtain a correlation deviation quantity that can quantitatively reflect the degree of deviation of the current working condition from the standard working condition. Based on the correlation deviation, a correction coefficient is dynamically generated according to a preset mapping rule; when the correlation deviation exceeds the preset silent threshold range, it is determined that the working condition has changed and a nonlinear correction coefficient is obtained. The correction coefficients are applied to the initial health assessment results and the anomaly probability predictions. The original predictions are corrected by weighted summation, and finally, the corrected health assessment results and anomaly probability predictions are obtained that are more consistent with the current actual physical state of the equipment.

7. The method for dynamic analysis and early warning of ship equipment maintenance data according to claim 6, characterized in that, Based on the corrected health assessment results and the predicted abnormality probability, a real-time early warning judgment is made in combination with a dynamic adaptive threshold judgment strategy. When the predicted value exceeds the dynamic early warning threshold, an alarm message is generated, including: Based on the corrected health assessment results and the predicted abnormality probability, the health assessment results are used as the latest data points and merged with the corrected health assessment results from previous moments stored in the database to form a health assessment time series that includes the current and past states. Select the corrected health assessment results from the most recent N time points in the health assessment time series, calculate the arithmetic mean of these N results as the health baseline value at the current time, and calculate the standard deviation of these N results as the health fluctuation range at the current time. Based on the health baseline value and the health fluctuation range, and in accordance with the preset conversion rules, the abnormal probability warning threshold for the current moment is dynamically generated, so that the warning threshold adaptively follows the normal decline trend of the equipment and changes in operating conditions. The corrected anomaly probability prediction value is compared with the current warning threshold to obtain the comparison result; and based on the comparison result, it is determined whether the preset alarm triggering conditions are met, including the corrected anomaly probability prediction value exceeding the dynamic warning threshold multiple times consecutively, or exceeding the dynamic warning threshold once and the exceedance exceeds the preset range. If the alarm triggering conditions are met, an alarm message is generated, including the device identifier, the time of the anomaly, the probability value of the anomaly, relevant monitoring index data, and the recommended maintenance level.

8. The method for dynamic analysis and early warning of ship equipment maintenance data according to claim 7, characterized in that, Based on the alarm information, the maintenance coordination and management response process is initiated, integrating the alarm information into the ship maintenance management platform, generating maintenance work orders, and recommending maintenance time windows and personnel allocation plans, thus creating a maintenance execution record. Simultaneously, by combining existing early warning records and maintenance feedback data, the maintenance plan is dynamically optimized, forming a closed-loop maintenance business management process, including: The alarm information is automatically pushed to the ship maintenance management platform, which extracts the equipment type, fault mode and time of anomaly occurrence, and generates a maintenance work order including a detailed task description by combining the ship's current navigation status data. Based on maintenance work orders, the system obtains the ship's navigation plan, the current availability status of the equipment, and the stored past maintenance data. It then calculates and recommends the final maintenance time window using a preset optimization algorithm. Based on the technical requirements of the maintenance tasks in the maintenance work order and the recommended maintenance time window, the system calls the maintenance personnel skill level database and personnel availability information to automatically match personnel who are available within the maintenance time window and have the corresponding skill level, and generates a recommended personnel configuration plan. After the maintenance work is actually carried out according to the recommended personnel configuration plan, the actual operation content, time consumption, and spare parts replacement information of this maintenance are collected and entered into the database of the ship maintenance management platform to form a complete maintenance execution record associated with the maintenance work order; Regularly summarize existing alarm records and maintenance execution records stored in the ship maintenance management platform, identify maintenance patterns and optimizable aspects through statistical analysis, and dynamically adjust maintenance cycles, early warning threshold settings, and resource allocation strategies in the maintenance plan based on the statistical analysis results to achieve closed-loop optimization of maintenance business management.

Images